The international journal of science / 10 December 2020

Eyes of the world

50 national regulatory bodies from 31 European countries.
The European Medicines Agency (EMA), created in 1995,
sits at the centre of this network. All countries have their

are on medicines own medical regulators, but the EMA provides manufac-
turers with a single place for scientific evaluation of drug

regulators applications, if they want Europe-wide approval.

In other regions — Africa, for example — countries are
also inching towards an approach that allows for the
pooling of regulatory expertise. This is the purpose of the
The race to produce COVID-19 vaccines African Vaccine Regulatory Forum, set up in 2006.
is a chance to create a more harmonized And at the global level, in 2012, the member states of the
approvals process. World Health Organization (WHO) agreed to establish the
International Coalition of Medicines Regulatory Authori-

T
he roll out of COVID-19 vaccines is under way,
but without, it seems, much global coordina-
tion where timing is concerned. China, Russia
and the United Arab Emirates began adminis-
tering vaccines before the conclusion of clinical
trials. Last week, the United Kingdom issued emergency
approval for a vaccine developed by the US biopharmaceu-
tical company Pfizer and BioNTech of Germany following
positive results from phase III testing (see page 205). The US
Food and Drug Administration (FDA) has needed longer to
make its decision on the same vaccine. And the regulatory
agencies of Australia, the European Union and Switzerland
are taking longer still.
This patchwork of different approvals processes, despite
COVID-19 being the one common enemy, has revived a
long-standing question of how to enhance harmonization
in vaccine regulation. Researchers reviewing the regulatory
landscape found at least 50 pathways to various types of
accelerated vaccine approval in a group of 24 countries
(S. Simpson et al. npj Vaccines 5, 101; 2020).
Greater harmonization would bring many benefits. Drug
companies could look forward to agreed definitions for
different types of approval, and would also benefit from
agreed guidelines for criteria that their vaccine candidates
would need to meet. If countries’ regulators were to ask
for broadly the same things, companies could cut the time
needed to prepare their drug applications. Companies, for
their part, would need to allow — or help to create — a secure
way for regulators to share data, which they are often not
permitted to do at present.
By assessing the same data, regulators could more easily
compare their findings and analyses with those of others,
and their decisions would not only be more robust, but
also be seen to be so. That, in turn, would shore up public
confidence in a world in which vaccine hesitancy is rising
and in which many citizens already have the means to com-
pare regulatory verdicts. This would be an evolutionary
shift, not a revolutionary one, because in recent years — and
particularly after the Ebola crisis — regulators have made
unprecedented efforts to discuss, coordinate and begin
to harmonize some of their processes.
The FDA, which was set up in 1906, is the world’s old-
est national medicines regulator. But the world has been
moving towards greater regulatory coordination for some
time. Europe’s regulatory system comprises a network of
ties (ICMRA) to allow regulators to share information and
agree on approaches. The ICMRA has 29 members, includ-
ing regulators from China, Europe and the United States.
Through it, members have been able to reach a consensus
on the best animal models for testing COVID-19 vaccines,
the ideal clinical-trial end points and the complicated issue
of continuing placebo-controlled trials after vaccine roll
out begins. The coalition’s COVID-19 working group is
Regulators now trying to harmonize the monitoring of vaccines once
want to be they have been deployed, because faint signals of adverse
effects might be too weak to spot in any one country.
able to talk And then there is the WHO itself. Low- and middle-in-
to each other come countries can now benefit from the work that goes
in the same into its Emergency Use Listing (EUL) process. On 13 Novem-
ber, the agency issued its first ever such vaccine listing for
units and a new polio vaccine. Around the end of October, the WHO
about the requested that both the FDA and the EMA assess the suit-
same end ability of COVID-19 vaccines for low- and middle-income
points.” countries as they consider whether to issue emergency
authorizations. It is not clear whether the regulators will
agree — but if either does, the WHO can draw on that anal-
ysis and issue its own EUL within days of the decision. That
would be collaboration indeed.
These are all important and necessary efforts. The need
now is to go a step further and find a path through the many
different types of vaccine approval. Before the pandemic,
the Coalition for Epidemic Preparedness, a global group of
funding agencies, companies and non-governmental organ-
izations, set up a working group to map out obstacles to bet-
ter regulatory alignment, in anticipation of a new infectious
disease. This process confirmed how regulatory agencies
differ on issues such as the use of genetic modification in
vaccine development, trials in pregnant women, and even
vial labelling. But it also meant that inconsistencies were
already mapped out and under discussion when the pan-
demic struck. COVID-19 has intensified these discussions.
The next step will not be easy. Regulators want to be able
to exchange data. Their experiences during the pandemic
have convinced many that they are moving towards a point
at which this will be possible. They want to be able to talk to
each other in the same units and about the same end points;
and to make decisions based on the same data.
Ultimately, each country must make its own decisions
about what’s best. But the goal of a harmonized regulatory
dossier for vaccines, conforming to an agreed set of interna-
tional regulatory requirements, would be transformative.

Nature | Vol 588 | 10 December 2020 | 195

©
2
0
2
0
S
p
r
i
n
g
e
r
N
a
t
u
r
e
L
i
m
i
t
e
d
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
 Editorials

Accounting for sex

because women had not been properly represented among
trial participants when the drugs were first evaluated.
Although sex and gender analysis is improving in drugs

and gender makes trials, it remains a work in progress in many fields, Londa
Schiebinger, a science historian at Stanford University in

science better California, told Nature (see page 209). Researchers have
been highlighting the harms caused by failing to account
for sex and gender for decades, but it wasn’t until after the
turn of the millennium that funding bodies really started
The European Commission is set to insist on to address the problem. The Canadian Institutes of Health
steps that will make research design more Research began to request that analyses of sex and gen-
inclusive. der be included in grant applications in 2010, and the US
National Institutes of Health followed suit in 2016.

A
t the end of last month, the European Com-
mission announced that its grant recipients
will be required to incorporate sex and gender
analyses into the design of research studies.
The policy will affect researchers applying
for grants that are part of the commission’s seven-year,
�85-billion (US$100-billion) Horizon Europe programme,
which is due to begin next year.
The funding is still awaiting sign-off from the European
Union’s 27 member states. But if all goes to plan, the com-
mission will be the largest funder to require sex and gender
analyses — along with analyses of other aspects of inclu-
sion, also known as intersectionality — in research design.
Such analyses could include disaggregating data by sex
when examining cells, or considering how a technology
might perpetuate gender stereotypes.
It’s a significant achievement. Science will be strength-
ened by researchers incorporating analyses of sex and
gender into their work at every stage — from study design
to gathering data, analysing those data and drawing
conclusions.
The European Commission is not the first funding
agency to make such changes. And this isn’t the first time
it has requested that studies account for sex and gender.
But in Horizon Europe, the requirement becomes a man-
date, and is expected to extend, by default, to most grant
recipients. Exceptions will be made only for those working
on topics for which the commission thinks such studies
would not be relevant, such as in pure mathematics.
Science and scientists have a troubled history of failing
to account for sex and gender when designing research. For
decades, crash test dummies were based on male bodies.
Even though smaller models are now used to represent
women, they fail to account for some other typical differ-
ences, such as neck strength1. The inclusion of sex and gen-
der analyses can also be revelatory. Sea turtles in Australia’s
Great Barrier Reef are being born mostly female because
of warming temperatures — a discovery that was made
when researchers were able to analyse male and female
populations2.
In some cases, the results of not accounting for sex and
gender have been catastrophic. Between 1997 and 2001, ten
prescription drugs were withdrawn from use in the United
States, eight of which had been found to be more danger-
ous for women than men. This had been missed, in part,
The European Commission began asking grant recipi-
ents to include sex and gender analysis in their research
design in 2013, a request which, by 2020, covered around
one-third of research fields. But according to later eval-
uation reports, fewer researchers than expected imple-
mented this request.
An analysis of researchers funded by the Canadian Insti-
tutes of Health Research, published in 2014, revealed that
Researchers some had pushed back when asked to consider sex and
have been gender3. And both this analysis and the European Com-
mission’s evaluation highlighted that some grant recipi-
highlighting ents used sex (which refers to biological characteristics)
the harms interchangeably with gender (which is a social construct
caused by and is not necessarily aligned with a person’s sex). To help
researchers to better appreciate the value of sex and gen-
failing to der analysis, the commission’s expert advisory group —
acount for which Schiebinger chairs — has published 15 case studies
sex and as examples of good practice (go.nature.com/33vxcxz).
gender for Another positive action could be for research teams to
include appropriate specialists to advise on, participate in,
decades.” or lead the design of more-inclusive research. Groups could
include researchers from the social or health sciences — the
Canadian Institutes of Health Research analysis revealed
that health- and social-science researchers are more likely
to include sex and gender analyses in project design than
are researchers in the biomedical sciences.
Ultimately, inclusive research design cannot be the
sole responsibility of funders. Some journals — includ-
ing Nature — are requesting that authors include sex and
gender analyses, when appropriate. Universities and
research supervisors also need to incorporate inclusive
design into the research methodology training they pro-
vide to students.
The European Commission is rightly adding its consid-
erable voice to the effort to ensure that science is designed
and carried out in a more inclusive way. But to change prac-
tices that have existed for centuries, more researchers —
especially research leaders — need to accept where they
have been going wrong, and how research and individuals
have suffered as a result. The foundations are being laid for
better science, and the more hands join in this important
effort, the better.
1. Linder, A. & Svedberg, W. Accid. Anal. Prev. 127, 156–162 (2019).
2. Jensen, M. P. et al. Curr. Biol. 28, 154–159 (2018).
3. Johnson, J. , Sharman, Z., Vissandjée, B. & Stewart, D. E. PLoS ONE 9,
e99900 (2014).

196 | Nature | Vol 588 | 10 December 2020

©
2
0
2
0
S
p
r
i
n
g
e
r
N
a
t
u
r
e
L
i
m
i
t
e
d
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
 A personal take on science and society

World view
By Ulrich Dirnagl

Institutions can retool for

more-rigorous research
Big moves to rebuild the scientific engagement. A QUEST good-evaluation-practice officer
infrastructure are possible. has sat as an independent assessor on hiring commissions
for 10 of the past 29 hiring calls.
We made

F
ive years ago, I was part of a small group of ‘activ-
ists’ who convinced the Berlin Institute of Health
(BIH), where I work, to try out a set of reforms
intended to improve the trustworthiness, use-
fulness and ethics of research. Things grew from
there: three years ago, with the help of government grants
and some nudging by a retired local politician, we secured
€2.5 million (US$2.9 million) per year for efforts to build up
incentives and technologies that increase rigour.
We were inspired by initiatives at other universities, such
as the reforms that Frank Miedema introduced during his
deanship at the University Medical Center Utrecht in the
Netherlands. But when the QUEST Center (QUEST stands
for Quality, Ethics, Open Science and Translation) launched
at the BIH, there was no precedent or blueprint for a pro-
gramme of this scale.
From the beginning, we presumed that researchers and
clinician–scientists are skilled professionals who want to
‘do the right thing’ but are also under pressure to accrue
publications to advance their careers. Doing quality
research takes time and humility, so unless we changed
the system, researchers who pursued quality-enhancing
practices could have found themselves at a disadvantage.
What was the solution? We made sure that we were viewed
as a resource, not a policing unit. We selected interventions
that we thought we could implement. Alongside introduc-
ing courses on experimental design and methods aimed at
reducing bias, we focused on practices to increase the trans-
parency of research. One push was for the use of electronic
laboratory notebooks (ELNs), which improve research doc-
umentation and make collaboration easier. We made sure
that QUEST, and not individual labs, covered the licence
fees and provided plenty of support. So far, nearly 2,000
of our 7,000 researchers, PhD students and technicians are
registered ELN users; my guess is that about half of these
have an ELN as their primary lab notebook. For many, ELNs
are a necessary first step towards systematically managing
their research data, which QUEST also supports.
We simultaneously adjusted the incentive and reward
system. When hiring professors and awarding institutional
funds, we now consider how thoroughly and quickly people
share their results. Those who make original data availa-
ble in publications are rewarded with a financial bonus
that can be spent on research. QUEST works with the BIH
THOMAS RAFALZYK
We tried to craft a system designed for its own improve-
sure that we ment. For example, we have developed an anonymous
were viewed online tool through which researchers have reported
as a resource, hundreds of errors and worrying incidents (U. Dirnagl
et al. PLoS Biol. 14, e2000705; 2016). This has allowed us
not a policing to learn from errors — for example, a technician realized
unit.” that ambiguous labelling of cell-culture media by a man-
ufacturer had spoiled her experiment. Her swift report-
ing prevented others from making the same mistake. The
company changed the labels on its flasks and alerted other
customers. After we saw many errors stemming from the
use of pipettes outside the calibrated range, we set up
‘pipetting exercises’ and saw the rate of these errors fall.
Three years in, we’re seeing more papers published
open access and with open data. We’re also seeing greater
participation in educational activities and in intramural
programmes using responsible selection criteria, such as
engagement with patient communities, reuse of data or
preregistration. Of course, funders and journals are also
pulling in the same direction, so it is impossible to know to
which changes are due to the efforts of QUEST.
However, we still have a long way to go. Our benchmark-
ing study found that, within 2 years of completion, only 40%
of studies sponsored by the Charité had reported results
(S. Wieschowski et al. J. Clin. Epidemiol. 115, 37–45; 2019).
Furthermore, 5 years after completion, more than 30% of
results remained unavailable. But we hope to correct this.
We use counselling and web tools to offer guidance on how
to publish null, inconclusive, negative and other ‘nonstand-
ard’ results, and award monetary research bonuses for
the publication of negative results or replication studies.
Most faculty members welcome our activities, and we are
working to expand student and researcher engagement.
For example, using funding from the biomedical
research charity Wellcome in London, we have established
fellowships for mid-career researchers who collaborate
to develop and track initiatives for improving science in
their own research groups. Our experience shows that
structured programmes can be rolled out by any academic
institution that is willing and able to improve its research in
a systematic fashion. The budget of QUEST is less than 1%
of our institution’s state funding for research and teaching,
not including monies from third-party funders.
Ulrich Dirnagl directs QUEST started from scratch. But many institutions
the QUEST Center at already promote activities such as open science, data

and the leadership of the Charité, Berlin’s university med- the Berlin Institute management and responsible research. If they align their
ical centre, to ensure that evaluation criteria encompass of Health. efforts, they can expand them and incorporate scientific
responsible research practices, including publication e-mail: ulrich. ideals into incentive structures. The quality of science and
of null results, provision of open data and community dirnagl@charite.de the culture of the workplace will be better off.

Nature | Vol 588 | 10 December 2020 | 197

©
2
0
2
0
S
p
r
i
n
g
e
r
N
a
t
u
r
e
L
i
m
i
t
e
d
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
 The world this week

News in brief
COVID-19 VACCINES
ARE NOT BEING
SHARED EQUALLY
Vaccine developers who have
already reported promising
phase III trial results against
COVID-19 estimate that,
between them, they can make
sufficient doses for more
than one-third of the world’s
MILKY WAY MAP
population by the end of 2021.
But many people in low- REVEALS ONE BILLION
income countries might have STARS IN MOTION
to wait until 2023 or 2024 for
vaccination. Five rich countries A huge data update from the
and the European Union have Gaia space observatory — which
pre-ordered about half of is tracking more than one billion
expected capacity for 2021, stars in the Galaxy — offers a
according to data from Airfinity, picture of what Earth’s night sky
a life-sciences market analytics will look like for 1.6 million years
firm in London. Canada leads to come.
on vaccine deals per capita,
with nearly nine doses per
The European Space Agency
probe lifted off in late 2013, and Arecibo telescope collapses
person. Most low- and middle-
income countries will rely on
began observing stars in July
2014 from a perch 1.5 million
in gut-wrenching display
contributions from COVAX, a kilometres from Earth. Gaia The iconic radio telescope at the Arecibo Observatory
joint fund for equitable vaccine continuously scans the sky as in Puerto Rico has collapsed, leaving astronomers and
distribution. it slowly spins, and it has now
the Puerto Rican scientific community to mourn its
measured the positions of the
BEST AND WORST SUPPLIED same stars multiple times. This demise.
Canada has pre-ordered almost 9 doses enables scientists to track stars’ Engineers had warned that the 900-tonne platform
of COVID-19 vaccines per person.
nearly imperceptible motions suspended above the telescope’s 305-metre-wide dish
Pre-ordered across the Galaxy year after
Potential for expansion in deal could fall at any moment, given that one of the main
year, and to triangulate their
positions using a technique
cables supporting it had snapped in early November.
Canada
called parallax. Last month, the US National Science Foundation,
United States
The latest update is based which owns the observatory, announced that it would
PHOTOGRAPHS L TO R: ESA/GAIA/DPAC (CC BY-SA 3.0 IGO); RICARDO ARDUENGO/AFP VIA GETTY

United Kingdom on around three years of data, shut down the telescope permanently, citing safety
Australia and includes a complete census
concerns over its instability, and damage too extensive
European Union of the Sun’s neighbourhood:
all but the faintest stars to repair.
SOURCE: DATA FROM AIRFINITY, UP TO 19 NOVEMBER/NATURE TABULATIONS;

Japan
within 100 parsecs (326 light The platform plummeted into the dish after
Vietnam years), totalling more than some cables failed just before 8 a.m. local time on
India 300,000 objects. The mission 1 December. No one was injured.
Israel has expanded its catalogue
Once the world’s largest single-dish radio telescope,
of stars by 15%, and its
Switzerland
measurements have become the Arecibo facility has been the site of many key
Indonesia more precise astronomical discoveries over the years, including
Brazil The data will underpin studies observations of the spinning stars known as pulsars
Latin America that range from the origins
(excl. Brazil) that led to the 1993 Nobel Prize in Physics.
and evolution of the Galaxy to
Egypt “Our hearts are heavy about this,” said Thomas
locating its dark matter.
Mexico Zurbuchen, NASA’s associate administrator for
China science, at a 1 December NASA advisory meeting.
COVAX It is unclear whether the dish will be demolished,
0 2 4 6 8 10 rebuilt or left in ruins.
Doses per person

Nature | Vol 588 | 10 December 2020 | 201

©
2
0
2
0
S
p
r
i
n
g
e
r
N
a
t
u
r
e
L
i
m
i
t
e
d
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
 The world this week

News in focus
EDWARD KINSMAN/SPL

A protein’s function is determined by its 3D shape.

‘IT WILL CHANGE EVERYTHING’:

AI MAKES GIGANTIC LEAP IN
SOLVING PROTEIN STRUCTURES
DeepMind’s program for determining the 3D shapes
of proteins stands to transform biology, say scientists.
By Ewen Callaway

A
n artificial intelligence (AI) network
developed by Google AI offshoot
DeepMind has made a gargantuan
leap in solving one of biology’s grand-
est challenges — determining a pro-
tein’s 3D shape from its amino-acid sequence.
DeepMind’s program, called AlphaFold,
outperformed around 100 other teams
in a biennial protein-structure prediction
challenge called CASP, short for Critical
Assessment of Structure Prediction. The
results were announced on 30 November, at
the start of the conference — held virtually this
year — that takes stock of the exercise.
“This is a big deal,” says John Moult, a compu-
tational biologist at the University of Maryland
in College Park, who co-founded CASP in 1994
to improve computational methods for accu-
rately predicting protein structures. “In some
sense the problem is solved.”
The ability to accurately predict proteins’
structures from their amino-acid sequences
would be a huge boon to life sciences and
medicine. It would vastly accelerate efforts
to understand the building blocks of cells and
aid more advanced drug discovery.
AlphaFold came top of the table at the
last CASP — in 2018, the first year that
London-based DeepMind participated. But,
this year, the outfit’s deep-learning net-
work was head-and-shoulders above other
teams and, say scientists, performed so
mind-bogglingly well that it could herald a
revolution in biology.
“It’s a game changer,” says Andrei Lupas, an
evolutionary biologist at the Max Planck Insti-
tute for Developmental Biology in Tübingen,

Nature | Vol 588 | 10 December 2020 | 203

©
2
0
2
0
S
p
r
i
n
g
e
r
N
a
t
u
r
e
L
i
m
i
t
e
d
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
 News in focus
Germany, who assessed the performance of
different teams in CASP. AlphaFold has helped
him to find the structure of a protein that has
vexed his laboratory for a decade. “This will
change medicine. It will change research. It
will change bioengineering. It will change
everything,” Lupas adds.
In some cases, AlphaFold’s structure pre-
dictions were indistinguishable from those
determined using ‘gold standard’ experimen-
tal methods such as X-ray crystallography and,
in recent years, cryo-electron microscopy
(cryo-EM). AlphaFold might not obviate the
need for these laborious and expensive meth-
ods — yet — say scientists, but the AI will make
it possible to study living things in new ways.
The structure problem
Proteins are the building blocks of life, respon-
sible for most of what happens inside cells.
How a protein works and what it does is deter-
mined by its 3D shape. Proteins tend to adopt
their shape without help, guided only by the
laws of physics.
For decades, laboratory experiments have
been the main way to obtain good protein struc-
tures. The first complete structures of proteins
were determined, starting in the 1950s, using
a technique in which X-ray beams are fired at
crystallized proteins and the diffracted light
translated into a protein’s atomic coordinates.
X-ray crystallography has produced the lion’s
share of protein structures. But, over the past
decade, cryo-EM has become the favoured tool
of many structural-biology labs.
Scientists have long wondered how a pro-
tein’s constituent amino acids map out the
twists and folds of its eventual shape. Early
attempts to use computers to predict protein
structures in the 1980s and 1990s performed
poorly. Lofty claims for methods in published
papers tended to disintegrate when other
scientists applied them to other proteins.
Moult started CASP to bring rigour to these
STRUCTURE SOLVER
DeepMind’s AlphaFold 2 algorithm significantly outperformed other teams at the CASP14
protein-folding contest — and its previous version’s performance at the last CASP.
100
90
A score above 90
Global distance test (GDT_TS; average)
efforts. The event challenges teams to pre-
dict the structures of proteins that have been
solved using experimental methods, but for
which the structures are not public.
DeepMind’s 2018 performance at CASP13
startled many scientists in the field, which
has long been the bastion of small academic
groups. But its approach was broadly similar
to those of other teams that were applying AI,
says Jinbo Xu, a computational biologist at the
University of Chicago, Illinois.
The first iteration of AlphaFold applied the
AI method known as deep learning to struc-
tural and genetic data to predict the distance
between pairs of amino acids in a protein.
In a second step, which does not invoke AI,
AlphaFold uses this information to come up
“This is going to empower
a new generation of
molecular biologists to ask
more advanced questions.”
with a ‘consensus’ model of what the pro-
tein should look like, says John Jumper at
DeepMind, who is leading the project.
The team tried to build on that approach
but eventually hit the wall. So it changed tack,
says Jumper, and developed an AI network that
incorporated additional information about
the physical and geometric constraints that
determine how a protein folds. The team also
set it a more difficult task: instead of predict-
ing relationships between amino acids, the
network predicts the final structure of a target
protein sequence. “It’s a more complex system
by quite a bit,” Jumper says.
Startling accuracy
CASP takes place over several months. Tar-
get proteins or portions of proteins called
domains — about 100 in total — are released
AlphaFold 2
on a regular basis, and teams have several
weeks to submit their structure predictions.
A team of independent scientists assesses the
predictions using metrics that gauge how simi-
lar a predicted protein is to the experimentally
determined structure. The assessors don’t
know who is making a prediction.
AlphaFold’s predictions arrived under the
name ‘group 427’, but the startling accuracy
of many of its entries made them stand out,
says Lupas. “I had guessed it was AlphaFold.
Most people had,” he says.
Some predictions were better than others,
but nearly two-thirds were comparable in qual-
ity to experimental structures. In some cases,
says Moult, it was not clear whether the dis-
crepancy between AlphaFold’s predictions and
the experimental results was a prediction error
or an artefact of the experiment. AlphaFold
also struggled to model individual structures
in protein complexes.
Faster structures
An AlphaFold prediction helped to determine
the structure of a bacterial protein that Lupas’s
lab has been trying to crack for years. Lupas’s
team had previously collected raw X-ray diffrac-
tion data, but transforming these patterns into
a structure requires some information about
the shape of the protein. Tricks for getting this
information, as well as other prediction tools,
had failed. “The model from group 427 gave us
our structure in half an hour,” Lupas says.
Demis Hassabis, DeepMind’s co-founder and
chief executive, says that the company plans to
make AlphaFold useful to other scientists. (It
previously published enough details about the
first version of AlphaFold for other researchers
to replicate the approach.) It can take AlphaFold
days to come up with a predicted structure,
which includes estimates on the reliability of
different regions of the protein. “We’re just
starting to understand what biologists would
want,” adds Hassabis.
In early 2020, the company released pre-
dictions for a handful of SARS-CoV-2 protein
structures that hadn’t been determined experi-
mentally. DeepMind’s predictions for a protein
called Orf3a ended up being similar to one later
determined through cryo-EM, says Stephen
Brohawn, a molecular neurobiologist at the

80 is considered roughly University of California, Berkeley, whose team

equivalent to the
70 released the structure in June. “What they have
experimentally
determined structure AlphaFold been able to do is very impressive,” he adds.
60
AlphaFold is unlikely to remove the need
50 for labs, such as Brohawn’s, that use experi-
mental methods to solve protein structures.
40
But it could mean that lower-quality experi-
30 mental data would be all that’s needed to get
a good structure. Some applications, such as
20
the evolutionary analysis of proteins, are set
SOURCE: DEEPMIND

10 to flourish. “This is going to empower a new

0
generation of molecular biologists to ask more
2006 2008 2010 2012 2014 2016 2018 2020 advanced questions,” says Lupas. “It’s going
Contest year to require more thinking and less pipetting.”

204 | Nature | Vol 588 | 10 December 2020

©
2
0
2
0
S
p
r
i
n
g
e
r
N
a
t
u
r
e
L
i
m
i
t
e
d
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
 remain susceptible to asymptomatic infection
— and could transmit that infection to others
who remain vulnerable. “In the worst-case
scenario, you have people walking around
feeling fine, but shedding virus everywhere,”
says virologist Stephen Griffin at the Univer-
sity of Leeds, UK.
Pfizer has said that its scientists are looking
at ways to assess virus transmission in future
studies. For now, AstraZeneca and the Univer-
sity of Oxford might be able to provide the
first hints as to whether a vaccine can protect
against such transmission. Although they have
yet to publish complete results, their trial did
routinely test participants for SARS-CoV-2,
allowing investigators to track whether people
JEROME DELAY/AP/SHUTTERSTOCK

became infected without developing symp-

toms. Early indications are that the vaccine
might have reduced the frequency of such
infections, which would suggest that trans-
mission might also be reduced.

How long will vaccine-induced

Large-scale trials of COVID-19 vaccines have been conducted at an unprecedented speed. immunity last?
There is no quick way to determine how long

COVID VACCINES:
immunity to the SARS-CoV-2 virus will last, and
researchers will need to monitor this closely

WHAT SCIENTISTS
in the coming months and years.
There have been some reports that peo-

NOW WANT TO KNOW

UK approves Pfizer–BioNTech vaccine. Researchers
are watching how it and others will perform.
By Heidi Ledford, David Cyranoski &
Richard Van Noorden
a request for emergency approval to the US
ple who have had one bout of COVID-19 and
developed antibodies against it can experi-
ence falling antibody levels and even reinfec-
tion months later, but it is still unclear how
prevalent reinfection is. There are signs that
the immune system preserves a memory of
coronavirus infection in the form of special-
ized memory cells that could kick into action
rapidly if the virus is encountered again.

W
ith striking speed, the United
Kingdom has become the first
country to approve a COVID-19
vaccine that has been tested in a
large clinical trial. On 2 Decem-
ber, UK regulators granted emergency-use
authorization to a vaccine from drug firms
Pfizer and BioNTech, just seven months after
the start of clinical trials. Hospitals have
already administered the first doses; front-
line health-care workers, care-home staff and
residents are at the head of the queue.
China and Russia have approved vaccines
already, but without waiting for the immuni-
zations to complete the final round of tests in
people. Regulators in the United States and the
European Union are expected to issue their
decisions on the Pfizer vaccine in the coming
weeks.
Tests on more than 43,000 people have
shown that it is 95% effective at preventing
disease when measured a week after partic-
ipants are given their second dose, the New
York City-based firm said in November when it
and BioNTech, in Mainz, Germany, submitted
Food and Drug Administration. The trial has
so far gathered data from only 170 cases of
COVID-19 across its control and intervention
arms, and real-world efficacy might be lower
than in a trial, but it is still an extraordinarily
promising result, says immunologist Danny
Altmann at Imperial College London: “This is
brilliant news.”
The approval is a historic moment. But
scientists still have many questions about how
this and other vaccines will perform as they’re
rolled out to millions of people.
Do the vaccines prevent
transmission of SARS-CoV-2?
In addition to the Pfizer vaccine, regulators
are poring over data from a similar vaccine
made by Moderna of Cambridge, Massachu-
setts, and a third produced by AstraZeneca of
Cambridge, UK, and the University of Oxford,
UK. All three have been tested in large clinical
trials, and have shown promise in preventing
disease symptoms.
But none has demonstrated that it prevents
infection altogether, or reduces the spread of
the virus in a population. This leaves open the
chance that those who are vaccinated could
And vaccines, Altmann says, are deliberately
designed to provoke strong responses from
the immune system.
Still, it will be important for public-health
officials to monitor immunity — and to know
when it begins to wane. One way to do that,
in addition to keeping track of infections
among people who have received the shots,
is to assess their levels of antibodies and
immune cells periodically. Tracking how
these immune responses change could give
an early indication of when they are waning to
worrisome levels, says Altmann. But the wide
variation in people’s immune responses could
make it a challenge to understand the circum-
stances in which a vaccine doesn’t work, and
such studies will need to track many people.
“You need to have a good stab at some high-
level population analysis to work out whether
you’re winning or losing,” says Altmann. “Oth-
erwise, you might be a government kidding
yourself in years’ time.”
How well do the vaccines work in
older people and other groups?
The major vaccine trials so far have enrolled
tens of thousands of people, but for each one,

Nature | Vol 588 | 10 December 2020 | 205

©
2
0
2
0
S
p
r
i
n
g
e
r
N
a
t
u
r
e
L
i
m
i
t
e
d
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
News in focus
conclusions about effectiveness are drawn
from fewer than 200 people who have devel-
oped disease. As a result, it can be difficult to
break up the data to look at efficacy in differ-
ent groups — such as people who are obese or
elderly — without losing statistical power. More
data are needed across demographics, says
Michael Head, an infectious-disease researcher
at the University of Southampton, UK.
There are early indications that the three
leading vaccines protect people over 65. But
researchers will probably need real-world
data from large numbers of vaccinated peo-
ple before they can get the demographic
granularity necessary to ensure that parts of
the population aren’t left unprotected.
There are no data yet on how the vaccines
fare in children and pregnant women. On
2 December, Moderna unveiled plans to test
its vaccine in adolescents.
How do the vaccines stack up
against each other?
All three leading vaccines have probably
beaten the goal of achieving 50% efficacy,
and all seem to be safe, on the basis of the
clinical-trial data so far. But there might be
differences in how well they work.
The vaccines from Pfizer and Moderna rely
on RNA encased in a lipid particle that ferries
it into cells, where it helps to generate a viral
genome, however, seems to be fairly stable
so far. Most of the vaccines being developed,
including the three that lead the pack, target
a molecule called the spike protein, which
the virus needs to infect cells. And immune
responses elicited by those vaccines will prob-
ably target multiple sites on that protein.
This gives researchers some reassurance
that the virus might not evolve ways to evade
immunity. But mass vaccination campaigns
will, for the first time, put enormous pressure
on SARS-CoV-2 to adapt, and will select for any
strain of the virus that might be able to escape
immune defences. “We’ve never seen a virus
like this under selective pressure,” says Griffin.
“So we don’t know how it’s going to respond.”
As a result, researchers will need to monitor
samples of SARS-CoV-2 for signs of change,
says Charlie Weller, head of vaccines at the bio-
medical research charity Wellcome in London.
“Robust surveillance with ongoing sampling
and sequencing will be key,” she says.
How will scientists monitor for
long-term safety concerns?
The Pfizer vaccine has completed only a few
months of the two-year clinical-trial period
needed before it is approved to be sold freely
on the market. As a result, people will be
watching closely for as-yet unobserved signs
of danger.
Clinical trials vet vaccines rigorously for
potential side effects with a combination of
self-reporting from participants and data col-
lection by clinicians. Pfizer’s trials revealed
that some recipients experienced pain at the
injection site, along with fever, fatigue, sore
muscles and headaches — although these
symptoms are generally not serious.
But after a vaccine is approved, whether
fully or only for emergency use, clinicians are
expected to continue reporting any adverse
reactions. Many countries have some kind of
programme, such as the US Vaccine Adverse
Event Reporting System, that collects reports
of serious symptoms after people receive
a vaccine. US doctors are legally bound to
report such symptoms. For COVID-19 drugs
and vaccines, the United Kingdom has set
up a specialized Coronavirus Yellow Card
reporting site.
Such systems work, says Jerome Kim,
director-general of the International Vaccine
Institute in Seoul. “You still need strong sur-
veillance. These rare events can be important,”
he says.

CAN JOE BIDEN MAKE

protein that stimulates the immune system.
AstraZeneca’s vaccine uses DNA that is shut-

GOOD ON HIS AMBITIOUS

tled into cells inside a harmless virus.
Early data suggest that the RNA approach

CLIMATE AGENDA?
might be more effective for preventing disease
symptoms developing. But there are subtle dif-
ferences in the immune responses provoked
by each approach, notes Griffin. Research-
ers might eventually find that one approach The US president-elect faces an uphill battle, but
works better than another in certain groups
of people, or that one is the best at limiting there are levers he can pull to curb global warming.
transmission.
By Jeff Tollefson

Differences in costs and logistics will also
shape which vaccine is best for which region.
Shortly after the UK government announced
the authorization of the Pfizer vaccine, offi-
cials acknowledged that getting the vaccine
to residents in individual care homes would
be a challenge, because it needs to be stored
at extremely low temperatures (−70 °C). The
other two vaccines do not need to be kept at
such low temperatures, and the AstraZeneca
immunization is likely to be the easiest and
cheapest to store, says Head.
Comparisons between the effectiveness
of the different vaccines are important and
should be done, but until then, the path for-
ward is clear, says Altmann. “Grab any vaccine
that your government can buy,” he says.
Could the virus evolve to evade
immunity given by vaccines?
Some viruses, such as the influenza virus,
are notorious for mutating. The SARS-CoV-2
W
hen Joe Biden won the US presi-
dency last month, it seemed like
a huge opportunity to restore the
country’s position as a leader in
the fight against climate change.
But whether he’ll be able to deliver on his
aggressive climate agenda remains to be
seen, especially because he will face a powerful
Republican opposition in Congress.
Still, climate-policy experts say that there is
a lot the former senator and vice-president to
Barack Obama can do, including exerting his
authority over federal agencies and leveraging
his experience working with both parties in the
Senate to push legislation in Congress.
“This is really the first time that a US presi-
dent is leading with climate,” says Vicki Arroyo,
executive director of Georgetown University’s
Climate Center in Washington DC. That’s excit-
ing, she says, but suggests cautious optimism:
global warming is still a partisan issue on
Capitol Hill, and “that is going to limit what
Biden can accomplish”.
Biden’s election comes at a crucial juncture.
President Donald Trump pulled the United
States out of the Paris climate agreement last
month, but other players on the world stage,
from China to the European Union, are prepar-
ing to present a new round of commitments
at the United Nations climate conference in
Glasgow, UK, next year.
Having the United States back on board
will give an important boost to these negoti-
ations, says Jean-Pascal van Ypersele, a clima-
tologist at the Catholic University of Louvain
in Louvain-la-Neuve, Belgium, and former
vice-chair of the Intergovernmental Panel on
Climate Change. “The stars are much better
aligned for a successful outcome in Glasgow
than they would have been if Trump had been
re-elected.”
Biden’s first opportunity to advance his

206 | Nature | Vol 588 | 10 December 2020

©
2
0
2
0
S
p
r
i
n
g
e
r
N
a
t
u
r
e
L
i
m
i
t
e
d
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
 greenhouse-gas emissions directly through the
Environmental Protection Agency (EPA). Over
the past four years, Trump’s EPA has reversed
or weakened dozens of environmental regu-
lations, including a trio of Obama-era climate
policies targeting emissions from vehicles,
power plants and oil and gas facilities. Biden
is expected to move immediately to restore —
and strengthen — those efforts, but this means
starting again and crafting new rules.
“That is the killer, in terms of workload,” says
Betsy Southerland, who spent more than three
decades at the EPA before resigning in 2017
in protest against Trump. “The Biden admin-
istration is going to have to make a decision:
do they laboriously reverse each one of those
rules, or is there something more effective and
efficient they can do?”
In the case of the policy on fuel-efficiency
ALEX WONG/GETTY

standards for vehicles, the administration

Biden’s climate platform was the most aggressive put forth by a leading presidential candidate.
agenda through Congress could come, as it
did for Obama, in the form of an economic
stimulus bill. With the US economy reeling
from the pandemic, many analysts expect
this to be at the top of Biden’s agenda when
he enters office. His team has made climate a
central feature of its economic plan and could
use a stimulus package to increase federal
investments in low-carbon energy and green
infrastructure.
Unlike Obama, Biden will probably next look
for ways to advance smaller climate meas-
ures through Congress rather than pushing
sweeping legislation, because of Republican
opposition. One possibility would be bipar-
tisan legislation that creates a carbon tax to
reduce US greenhouse-gas emissions — an
idea that has backing among many conserva-
tives and business leaders who are concerned
about the climate. One proposal developed
by the Climate Leadership Council, a non-
profit organization based in Washington DC,
would levy a tax on carbon dioxide emissions,
starting with a modest US$40 per tonne and
increasing over time, with the goal of cutting
US emissions in half by 2035. The proceeds
would be refunded to taxpayers.
Getting such legislation through the Senate
won’t be easy, but it’s not impossible, says Bob
Inglis, who heads the Energy and Enterprise
Initiative, a think tank advocating politically
conservative environmental solutions at
George Mason University in Fairfax, Virginia.
“This is a major opportunity,” says Inglis.
Regardless of what happens in Congress,
many scientists and environmentalists expect
that Biden will immediately use his executive
authority to advance his climate agenda across
the full suite of federal agencies. (It wasn’t
until Obama’s second term, in 2012, that he
got serious about sidestepping Congress with
his authority to battle climate change.)
Under Biden, the interior department,
for instance, could hasten the processing
of federal permits to build offshore wind
farms and other renewable-energy projects.
And the Department of Energy could raise
energy-efficiency standards for appliances.
“There’s no need for Biden to wait,” says Tim
Profeta, who leads Duke University’s Nicholas
Institute for Environmental Policy Solutions
in Durham, North Carolina. “There’s a lot the
president can do using his own authority, start-
ing from day one.”
Profeta co-chairs the Climate 21 Project,
an independent group of academics, policy
specialists and former government officials
“There’s a lot the
president can do using
his own authority,
starting from day one.”
that has crafted a blueprint for executive
action across 11 federal offices and agencies
to address global warming. The top-line rec-
ommendation from the group is that the new
administration should establish a National
Climate Council led by an official who reports
directly to the president. This person would
help to advance Biden’s climate agenda by
coordinating with various US agencies. “You
need somebody in the West Wing who has the
president’s ear and who is focused on mak-
ing climate action happen across the federal
govern­ment,” says Profeta.
The president’s most powerful tool when
it comes to climate change is regulating
might move forward with an entirely new rule.
The Trump administration rolled back stand-
ards put in place under Obama so that the car
industry has to boost average fuel efficiency
by only around 1.5% per year between 2022 and
2025, instead of Obama’s 5% per year. Rather
than reworking the rule, the Biden admin-
istration will probably move to develop an
entirely new set of regulations that look for-
ward another 10–15 years for longer-lasting
impact, says David Doniger, strategic director
of the Climate and Clean Energy Program at
the Natural Resources Defense Council, an
environmental group based in New York City.
The road to Glasgow
Getting an early start on implementing
his climate agenda will be crucial as Biden
re­integrates the country into the Paris climate
agreement.
The president-elect will need to develop
a climate pledge and present it to the world
at next year’s conference in Glasgow, where
countries are expected to update their com-
mitments for the first time since the agree-
ment was signed in 2015. Under Obama,
the United States initially committed to cut
greenhouse-gas emissions by at least 26%
below 2005 levels by 2025. The challenge is
to make sure the new US pledge is both strong
and credible, says Joseph Aldy, an economist
at Harvard University in Cambridge, Massa-
chusetts, who served as a White House climate
adviser under Obama.
“We have lost credibility on many fronts as
a result of Donald Trump,” says Aldy. If Biden
wants to take a leadership role in the Paris pro-
cess and push other countries to do more, Aldy
says, the president-elect will need to convince
the global community that any regulations
or legislation he puts in place are going to be
effective and won’t be easily reversed in four
or eight years, when a new president is elected.
“Our counterparts around the world will be
looking very closely at what we are doing.”

Nature | Vol 588 | 10 December 2020 | 207

©
2
0
2
0
S
p
r
i
n
g
e
r
N
a
t
u
r
e
L
i
m
i
t
e
d
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
News in focus
is on the table,” says Koopmans.
Another member, Hung Nguyen, an environ­
ment and food-safety researcher at the
International Livestock Research Institute in
Nairobi, will contribute his knowledge on how
pathogens spread in wet markets, similar to
the Huanan seafood market in Wuhan, which
many of the first people reported to have
COVID-19 had visited. Nguyen has investigated
how salmonella and other bacteria spread
through smallholder farms, slaughterhouses
and live-animal markets in his home country
of Vietnam and across southeast Asia.

MARKO KONIG/IMAGEBROKER/SHUTTERSTOCK
Also on the team is Peter Daszak, president of
the non-profit research organization Ecohealth
Alliance in New York City, who has spent more
than a decade studying coronaviruses. He has
worked closely with the Wuhan Institute of
Virology (WIV) to test bats for coronaviruses
with the potential to spill over into people.
“It is an honour to be part of this team,” says
Daszak. “There hasn’t been a pandemic on this
scale since the 1918 flu, and we’re still close
SARS-CoV-2 probably originated in bats, but how it passed to people is being investigated. enough to the origin to really find out more
details about where it has come from.”

THE SCIENTISTS
Another team member, Fabian Leendertz,
a veterinary researcher at the Robert Koch

INVESTIGATING THE
Institute in Berlin, will bring his expertise in
spillover events. In April 2014, Leendertz vis-

PANDEMIC’S ORIGINS
ited Meliandou village in Guinea, months after
a two-year-old died of Ebola — the first person
reported to be infected in West Africa.
Work by Leendertz, including interviews
The World Health Organization will draw on a diverse with locals and environmental sampling,
suggests that the outbreak started in bats that
team to examine a major mystery about SARS-CoV-2. lived in a hollow tree where the children used
to play. The tree was burnt down days before
By Smriti Mallapaty

A
n epidemiologist who helped to tie
the 2012 outbreak of Middle East res-
piratory syndrome (MERS) to camels;
a food-safety officer who studies how
pathogens spread in markets; and a
veterinarian who found evidence linking
the 2014 West Africa Ebola outbreak to bats
roosting in a hollow tree. These researchers
are among the team that the World Health
Organization (WHO) has assembled to explore
the origins of the coronavirus pandemic.
The investigation aims to find out how
and when the virus SARS-CoV-2 first infected
people. Strong evidence suggests that the
corona­virus originated in bats, but its jour-
ney to people remains a mystery. Scientists
say the team is highly qualified, but its task will
be challenging.
“This is an excellent team with a lot of expe-
rience,” says Martin Beer, a virologist at the
Federal Research Institute for Animal Health
in Greifswald, Germany.
The group will be working with research-
ers in China and professionals from several
other international agencies, and will start
the search in Wuhan — the Chinese city where
SARS-CoV-2 was first identified — and expand
across China and beyond.
The international group comes with a
breadth of knowledge. Marion Koopmans is a
virologist specializing in molecular epidemiol-
ogy at the Erasmus University Medical Centre
in Rotterdam, the Netherlands. She was on
the team that found, in 2013, that dromedary
camels were an intermediate host for the virus
that causes MERS, which has killed more than
850 people. She has since worked with another
team member — Elmoubasher Farag, an epi-
demiologist at the Ministry of Public Health
in Doha — to test camels for MERS antibodies.
During the COVID-19 pandemic, Koopmans
has tracked the rapid spread of SARS-CoV-2
in mink farms in Europe. Studies on the pan-
demic’s origin will need to explore the role of
animals kept for fur and food, she says.
Koopmans says that the group is keeping an
open mind about how the pandemic started
and will not exclude any scenarios, including
the unlikely one that SARS-CoV-2 accidentally
escaped from a laboratory. Scientists have pre-
viously told Nature that the virus is likely to
have passed from bats to humans, probably
through an intermediate animal — but ruling
out the lab scenario will be difficult. “Anything
his arrival and no Ebola virus was detected in
nearby bats, which he says highlights the dif-
ficulties of finding an outbreak’s beginnings.
Considerable time has passed since the
emergence of COVID-19, and many people have
only mild or no symptoms, which will make it
challenging to identify the first infected per-
son, says Leendertz.
Other team members include researchers
from Denmark, the United Kingdom, Australia,
Russia and Japan.
Although the team members are highly
qualified, eight out of ten are men and inves-
tigators from Europe dominate the group;
none is from Africa or South America, says
Angela Rasmussen, a virologist at Georgetown
University, who is based in Seattle, Washing-
ton. “It could be more representative of the
larger global scientific community,” she says.
She also says that Daszak’s ties to the WIV
could raise a conflict of interest, given the
unsubstantiated claims that the virus acci-
dentally leaked from the lab.
Daszak says that he has been transparent
about his work in China. The trust he has built
with researchers there will help the team to
gain a deeper understanding of the pandemic’s
early days, he says.

208 | Nature | Vol 588 | 10 December 2020

©
2
0
2
0
S
p
r
i
n
g
e
r
N
a
t
u
r
e
L
i
m
i
t
e
d
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
REVERSAL OF BIOLOGICAL Q&A
CLOCK RESTORES
VISION IN OLD MICE How sex and
gender analysis
‘Reprogramming’ approach seems
to make old cells young again. improves science
The European Commission has said
By Heidi Ledford that it aims to make sex and gender

R
esearchers have restored vision in old
mice and in mice with damaged retinal
nerves by resetting some of the thou-
sands of chemical marks that accu-
mulate on DNA as cells age. The work,
published on 2 December in Nature, suggests a
new approach to reversing age-related decline,
by reprogramming some cells to a ‘younger’
state in which they are better able to repair or
replace damaged tissue.
“It is a major landmark,” says Juan Carlos
Izpisua Belmonte, a developmental biologist
at the Salk Institute for Biological Studies in La
Jolla, California, who was not involved in the
study. “These results clearly show that tissue
regeneration in mammals can be enhanced.”
Visionary approach
Ageing affects the body in myriad ways —
among them, adding, removing or altering
chemical groups such as methyls on DNA.
These ‘epigenetic’ changes accumulate as a
person ages, and some researchers have pro-
posed tracking the changes as a way of calibrat-
ing a molecular clock to measure biological
age, an assessment that takes into account
biological wear-and-tear and can differ from
chronological age.
“We set out with a question: if epigenetic
changes are a driver of ageing, can you reset
the epigenome?” says David Sinclair, a genet-
icist at Harvard Medical School in Boston,
Massachusetts, and a co-author of the Nature
study (Y. Lu et al. Nature 588, 124–129; 2020).
“Can you reverse the clock?”
There were suggestions that the approach
could work: in 2016, Belmonte and his col-
leagues reported the effects of expressing
four genes in mice genetically engineered
to age more rapidly than normal (A. Ocampo
et al. Cell 167, 1719–1733; 2016). It was already
known that triggering these genes could cause
cells to lose their developmental identity —
the features that make, for example, a skin cell
look and behave like a skin cell. But rather than
turn the genes on and leave them that way,
Belmonte’s team turned them on for only a few
days, then switched them off again. The result
was mice that aged more slowly, and had a pat-
tern of epigenetic marks indicative of younger
animals. But the technique had disadvantages:
previous work had shown that if the genes are
present in extra copies or expressed for too
long, some mice will develop tumours.
In Sinclair’s lab, geneticist Yuancheng Lu
looked for a safer approach. He dropped one of
the four genes used by Belmonte’s team — one
that is linked to cancer — and put the remaining
three into a virus that could shuttle them into
cells. He included a switch that would allow
him to turn the genes on by giving mice water
spiked with a drug. Withholding the drug
would switch the genes off again.
Because mammals lose the ability to
regenerate components of the central nerv-
ous system early in development, Lu and his
colleagues tested their approach there — in
the eye’s retinal nerves. They first injected the
virus into the eye to see whether expression of
the three genes would allow mice to regenerate
injured nerves — something that no treatment
had yet been shown to do.
Lu remembers the first time that he saw a
nerve regenerating from injured eye cells. “It
was breathtaking,” he says.
“If epigenetic changes are
a driver of ageing, can you
reset the epigenome? Can
you reverse the clock?”
The team went on to show that its system
improved visual acuity in mice with age-related
vision loss, or with increased pressure inside
the eye — a hallmark of the disease glaucoma.
The approach also reset epigenetic patterns
to a more youthful state in mice and in human
cells grown in the laboratory. It is still unclear
how cells preserve a memory of a more youth-
ful epigenetic state, says Sinclair, but he and
his colleagues are trying to find out.
In the meantime, Harvard has licensed
the technology to Boston company Life
Biosciences, which, Sinclair says, is carrying
out preclinical safety assessments with a view
to developing it for use in people. It would be
an innovative approach to treating vision loss,
says Botond Roska, director of the Institute
of Molecular and Clinical Ophthalmology in
Basel, Switzerland, but will probably need con-
siderable refinement before it can be deployed
safely in humans.
analysis mandatory in the research it
funds through its €85-billion (US$100-
billion) Horizon Europe programme.
The strengthened policy is a result of
recommendations made in a report (see
go.nature.com/3mryv1a), produced last
month by an expert group chaired by
Londa Schiebinger, who studies gender
and science at Stanford University in
California. Nature spoke to Schiebinger
about the group’s work.
How do you convince people of the need
for sex and gender analysis in research?
Our iconic example of failure when you
don’t do this analysis is that between 1997
and 2001, ten prescription drugs were
withdrawn from the US market, eight of
which were more dangerous for women
than for men. When drugs fail, you’re losing
money and people are suffering and dying.
From preclinical studies to human clinical
trials, you have to collect data on males
and females and analyse them separately.
What mistakes do researchers make in
these analyses?
The biggest mistake is simply ignoring sex,
gender and intersectionality. Another is to
not distinguish between biological sex and
sociocultural gender. Gender is specific
to ethnicity, age and culture. Researchers
need to get the right variables, collect their
data correctly and do the analysis well.
Are there research areas where people
might be surprised that sex and gender
analysis is essential?
For some marine organisms, sex is
determined by temperature. Our report
includes a fascinating study from Australia,
where they found that the turtles in the
north of the Great Barrier Reef were 99%
female, whereas in the cooler south, it was
about 67% female. It’s important that we
understand how global warming is skewing
these ratios, so that we can efficiently
manage ecosystems.
Interview by Elizabeth Gibney.
Edited for length and clarity.

Nature | Vol 588 | 10 December 2020 | 209

©
2
0
2
0
S
p
r
i
n
g
e
r
N
a
t
u
r
e
L
i
m
i
t
e
d
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
Feature

PLANET OBSERVER/UNIVERSAL IMAGES GROUP/GETTY

Life might have begun in bodies of water on land, perhaps in craters similar to Canada’s Lake Manicouagan, formed by an ancient impact.

THE WATER PARADOX

AND THE ORIGINS OF LIFE
Water is essential for life, but it breaks down DNA and other key molecules. So how did the first
cells deal with such a necessary and dangerous substance? By Michael Marshall

O
n 18 February next year, a NASA
spacecraft will plummet through
the Martian atmosphere, fire its
retro-rockets to break its fall and
then lower a six-wheeled rover
named Perseverance to the sur-
face. If all goes according to plan,
the mission will land in Jezero Cra-
ter, a 45-kilometre-wide gash near the planet’s
equator that might once have held a lake of
liquid water.
Among the throngs of earthlings cheering
on Perseverance, John Sutherland will be pay-
ing particularly close attention. Sutherland, a
biochemist at the MRC Laboratory of Molecu-
lar Biology in Cambridge, UK, was one of the
scientists who lobbied NASA to visit Jezero
Crater, because it fits his ideas about where life
might have originated — on Mars and on Earth.
The choice of landing site reflects a shift in
thinking about the chemical steps that trans-
formed a few molecules into the first biologi-
cal cells. Although many scientists have long
speculated that those pioneering cells arose
in the ocean, recent research suggests that the
key molecules of life, and its core processes,
can form only in places such as Jezero — a rel-
atively shallow body of water fed by streams.
That’s because several studies suggest
that the basic chemicals of life require ultra-
violet radiation from sunlight to form, and
that the watery environment had to become
highly concentrated or even dry out com-
pletely at times. In laboratory experiments,
Sutherland and other scientists have produced
DNA, proteins and other core components of

210 | Nature | Vol 588 | 10 December 2020

©
2
0
2
0
S
p
r
i
n
g
e
r
N
a
t
u
r
e
L
i
m
i
t
e
d
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
cells by gently heating simple carbon-based
chemicals, subjecting them to UV radiation
and intermittently drying them out. Chem-
ists have not yet been able to synthesize such
a wide range of biological molecules in condi-
tions that mimic seawater.
The emerging evidence has caused many
researchers to abandon the idea that life
emerged in the oceans and instead focus on
land environments, in places that were alter-
nately wet and dry. The shift is hardly unani-
mous, but scientists who support the idea of
a terrestrial beginning say it offers a solution
to a long-recognized paradox: that although
water is essential for life, it is also destructive
to life’s core components.
Surface lakes and puddles are highly prom-
ising, says David Catling, a planetary scientist
at the University of Washington in Seattle.
“There’s a lot of work that’s been done in the
last 15 years which would support that direc-
tion.”
Primordial soup
Although there is no standardized definition
of life, most researchers agree that it needs
several components. One is information-car-
rying molecules — DNA, RNA or something
else. There must have been a way to copy
these molecular instructions, although the
process would have been imperfect to allow
for mistakes, the seeds of evolutionary change.
Furthermore, the first organisms must have
had a way to feed and maintain themselves,
perhaps using protein-based enzymes.
Finally, something held these disparate parts
together, keeping them separate from their
environment.
When laboratory research into life’s origins
started in earnest in the 1950s, many research-
ers assumed that life began in the sea, with a
rich mix of carbon-based chemicals dubbed
the primordial soup.
This idea was independently proposed in
the 1920s by biochemist Alexander Oparin, in
what was then the Soviet Union, and geneticist
J. B. S. Haldane in the United Kingdom. Each
imagined the young Earth as a huge chemi-
cal factory, with multitudes of carbon-based
chemicals dissolved in the waters of the early
oceans. Oparin reasoned that increasingly
complicated particles were formed, culmi-
nating in carbohydrates and proteins: what
he called “the foundation of life”.
In 1953, a young researcher named Stanley
Miller at the University of Chicago in Illinois
described a now-famous experiment that was
seen as confirming these ideas1. He used a glass
flask holding water to mimic the ocean, and
another flask containing methane, ammonia
and hydrogen to simulate the early atmos-
phere. Tubes connected the flasks, and an
electrode simulated lightning. A few days of
heating and electric shocks were enough to
make glycine, the simplest amino acid and
an essential component of proteins. This
suggested to many researchers that life arose
near the surface of the ocean.
But many scientists today say there’s a fun-
damental problem with that idea: life’s corner-
stone molecules break down in water. This is
because proteins, and nucleic acids such as
DNA and RNA, are vulnerable at their joints.
Proteins are made of chains of amino acids,
and nucleic acids are chains of nucleotides.
If the chains are placed in water, it attacks the
links and eventually breaks them. In carbon
chemistry, “water is an enemy to be excluded
as rigorously as possible”, wrote the late bio-
chemist Robert Shapiro in his totemic 1986
book Origins, which critiqued the primordial
ocean hypothesis2.
This is the water paradox. Today, cells solve
it by limiting the free movement of water in
their interiors, says synthetic biologist Kate
Adamala at the University of Minnesota in
Minneapolis. For this reason, popular images
of the cytoplasm — the substance inside the
cell — are often wrong. “We are taught that
cytoplasm is just a bag that holds everything,
and everything is swimming around,” she
adds. “That’s not true, everything is incredi-
bly scaffolded in cells, and it’s scaffolded in
“Wet–dry cycles are
everywhere. It’s as simple as
rainwater evaporating on
wet rocks.”
a gel, not a water bag.”
If living things keep water controlled, then
the implication, say many researchers, is obvi-
ous. Life probably formed on land, where water
was only intermittently present.
Land start
Some of the key evidence in favour of this idea
emerged in 2009, when Sutherland announced
that he and his team had successfully made two
of the four nucleotides that comprise RNA3.
They started with phosphate and four simple
carbon-based chemicals, including a cyanide
salt called cyanamide. The chemicals were
dissolved in water throughout, but they were
highly concentrated, and crucial steps required
UV radiation. Such reactions could not take
place deep in an ocean — only in a small pool or
stream exposed to sunlight, where chemicals
could be concentrated, he says.
Sutherland’s team has since shown that the
same starter chemicals, if they are treated sub-
tly differently, can also produce precursors to
proteins and lipids4. The researchers suggest
that these reactions might have taken place if
water containing cyanide salts was dried out by
the Sun, leaving a layer of dry, cyanide-related
chemicals that was then heated by, say, geo-
thermal activity. In the past year, his team
has produced the building blocks of DNA
— something previously thought implausi-
ble — using energy from sunlight and some of
the same chemicals at high concentrations5.
This approach has been extended by
biochemist Moran Frenkel-Pinter at the
NSF–NASA Center for Chemical Evolution in
Atlanta, Georgia, and her colleagues. Last year,
they showed that amino acids spontaneously
linked up to form protein-like chains if they
were dried out6. And those kinds of reaction
were more likely to occur with the 20 amino
acids found in proteins today, compared with
other amino acids. That means intermittent
drying could help to explain why life uses only
those amino acids, out of hundreds of possi-
bilities. “We saw selection for today’s amino
acids,” says Frenkel-Pinter.
Wet and dry
Intermittent drying out can also help to drive
these molecular building blocks to assemble
into more-complex, life-like structures.
A classic experiment along these lines
was published in 1982 by researchers David
Deamer and Gail Barchfeld, then at the Univer-
sity of California, Davis7. Their aim was to study
how lipids, another class of long-chain mol-
ecule, self-organize to form the membranes
that surround cells. They first made vesicles:
spherical blobs with a watery core surrounded
by two lipid layers. Then the researchers dried
the vesicles, and the lipids reorganized into
a multi-layered structure like a stack of pan-
cakes. Strands of DNA, previously floating in
the water, became trapped between the layers.
When the researchers added water again, the
vesicles reformed — with DNA inside them.
This was a step towards a simple cell.
“These wet–dry cycles are everywhere,” says
Deamer, who is now at the University of Cali-
fornia, Santa Cruz. “It’s as simple as rainwater
evaporating on wet rocks.” But when they are
applied to biological chemicals such as lipids,
he says, remarkable things happen.
In a 2008 study, Deamer and his team mixed
nucleotides and lipids with water, then put
them through wet–dry cycles. When the lipids
formed layers, the nucleotides linked up into
RNA-like chains — a reaction that would not
happen in water unaided8.
Other studies are pointing to a different
factor that seems to be a key part of life’s
origins: light. That’s one of the conclusions
coming from the team of synthetic biologist
Jack Szostak at Massachusetts General Hos-
pital in Boston, which works with ‘protocells’
— simple versions of cells that contain a hand-
ful of chemicals, but can grow, compete and
replicate themselves. The protocells display
more-lifelike behaviours if they are exposed to
conditions similar to those on land. One study,
on which Adamala was a co-author, found that
the protocells could use energy from light to
divide, in a simple form of reproduction9.

Nature | Vol 588 | 10 December 2020 | 211

©
2
0
2
0
S
p
r
i
n
g
e
r
N
a
t
u
r
e
L
i
m
i
t
e
d
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
Feature
Similarly, Claudia Bonfio, now also at the
MRC Laboratory of Molecular Biology, and her
colleagues showed in 2017 that UV radiation
drives the synthesis of iron-sulfur clusters10,
which are crucial to many proteins. These
include those in the electron transport chain,
which helps to power all living cells by driving
the synthesis of the energy-storage molecule
ATP. The iron–sulfur clusters would break
apart if they were exposed to water, but Bon-
fio’s team found they were more stable if the
clusters were surrounded by simple peptides
3–12 amino acids long.
Water, but not too much
Such studies have given momentum to the
idea that life began on a well-lit surface with a
limited amount of water. However, there is still
debate over how much water was involved, and
what part it played in starting life.
Like Deamer, Frenkel-Pinter argues that
wet–dry cycles were crucial. Dry conditions,
she says, provided an opportunity for chain
molecules such as proteins and RNA to form.
But simply making RNA and other mole-
cules is not life. A self-sustaining, dynamic
system has to form. Frenkel-Pinter suggests
that water’s destructiveness could have helped
to drive that. Just as prey animals evolved to
run faster or secrete toxins to survive preda-
tors, the first biological molecules might have
evolved to cope with water’s chemical attacks
— and even to harness its reactivity for good.
This year, Frenkel-Pinter’s team followed
up on its previous study6 showing that drying
caused amino acids to link up spontaneously.
The team found that their proto-proteins
could interact with RNA, and that both became
more stable in water as a result11. In effect,
water acted as a selection pressure: only those
combinations of molecules that could survive
in water would continue, because the others
would be destroyed.
The idea is that, with each cycle of wetting,
the weaker molecules, or those that could
not protect themselves by binding to others,
were destroyed. Bonfio and her team demon-
strated this in a study this year12, in which they
attempted to convert simple fatty acids into
more-complex lipids resembling those found
in modern cell membranes. The researchers
created mixtures of lipids, and found that the
simple ones were destroyed by water, while the
larger, more complex ones accumulated. “At
some point, you would have enough of these
lipids for them to form membranes,” she says.
In other words, there might be a Goldilocks
amount of water: not so much that biological
molecules are destroyed too quickly, but not
so little that nothing changes.
Warm little ponds
Where might all this have happened? On this
point, there is a generational divide in the
field. Many senior researchers are committed
to one scenario or another, whereas younger
researchers often argue that the question is
wide open.
The open ocean is unviable, says
Frenkel-Pinter, because there is no way for
chemicals to become concentrated. “That’s
really a problem,” agrees Bonfio.
An alternative marine idea has been cham-
pioned since the 1980s by geologist Michael
Russell, an independent researcher formerly at
the Jet Propulsion Laboratory in Pasadena, Cal-
ifornia. Russell argues that life began in vents
on the seabed, where warm alkaline water
seeps up from geological formations below.
Interactions between warm water and rocks
would provide chemical energy that would
first drive simple metabolic cycles, which
would later start making and using chemicals
such as RNA.
Russell is critical of Sutherland’s approach.
“He’s doing all these fantastic bits of chem-
istry,” he says, but for Russell, none of it is
relevant. That’s because modern organisms
use completely different chemical processes
to make substances such as RNA. He argues
that these processes must have arisen first, not
the substances themselves. “Life, it picks very
particular molecules. But you can’t pick them
from the bench. You’ve got to make them from
scratch and that’s what life does.”
Sutherland counters that once RNA, pro-
teins and so forth had formed, evolution would
have taken over and enabled proto-organisms
to find new ways to make these molecules and
thus sustain themselves.
Meanwhile, many researchers have
expressed scepticism about Russell’s alka-
line-vent hypothesis, arguing that it lacks
experimental support.
By contrast, chemical experiments that
simulate surface conditions have made the
building blocks of nucleic acids, proteins and
lipids. “None of that synthesis exists in that
deep-sea hydrothermal vent hypothesis. It just
simply hasn’t been done, and possibly because
it can’t be done,” says Catling.
Frenkel-Pinter is also critical of the vent
idea, because the molecules she works with
wouldn’t survive long in those conditions.
“The formation of these protopeptides is not
very compatible with hydrothermal vents,”
says Frenkel-Pinter.
BETTMANN/GETTY

A possible solution was proposed in May

by geochemist Martina Preiner, a postdoc
at the University of Düsseldorf in Germany,
and her colleagues. She argues that in the
In experiments in the 1950s, Stanley Miller created amino acids from simple building blocks. rocks beneath hydrothermal vents, heat and
chemical reactions bind up water molecules

212 | Nature | Vol 588 | 10 December 2020

©
2
0
2
0
S
p
r
i
n
g
e
r
N
a
t
u
r
e
L
i
m
i
t
e
d
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
 or break them apart — creating dry spaces13.
“There are rock–water interactions getting
rid of the water to a certain extent,” she says.
Intermittently, more seawater would trickle in,
giving “something like a wet–dry cycling”. This
ought to make the deep-sea rocks much more
suitable for the formation of key molecules,
argues Preiner, although she acknowledges
this is still a hypothesis. “Of course, you still
have to do the according experiments to prove
that this could do certain reactions.”
At present, however, that evidence doesn’t
exist. Meanwhile, experimental support is
growing for the idea that life started in small
bodies of water on land.
Sutherland favours a meteorite impact
crater, heated by the Sun and by the residual
ESA/FU-BERLIN

energy of the impact, with multiple streams

of water running down the sloping sides,
and finally meeting in a pool at the bottom.
This would have been a complex, 3D environ-
ment with mineral surfaces to act as catalysts,
where carbon-based chemicals could have
been alternately dissolved in water and dried
out in the Sun. “You can say with some degree
of confidence we need to be on the surface,
we can’t be deep in the ocean or 10 kilometres
down in the crust,” says Sutherland. “Then we
need phosphate, we need iron. A lot of those
things are very easily delivered by iron–nickel
meteorites.” The impact scenario has a fur-
ther advantage: meteorite impacts shock
the atmosphere, producing cyanide, says
Sutherland.
Deamer has long championed a different
suggestion: volcanic hot springs. In a study this
year, he and his colleague Bruce Damer argued
that lipids would have formed protocells in the
hot waters14, as his earlier experiments indi-
cated. The wet–dry cycles on the edges of the
pools would have driven the formation and
copying of nucleic acids such as RNA.
Deamer has conducted several experiments
in modern volcanic hot springs to test his ideas.
In 2018, his team showed that vesicles could
form in hot spring water15, and even enclose
nucleic acids — but they would not form in sea-
water. A follow-up study last year found that
when the resulting vesicles were dried, nucle-
otides linked up to form RNA-like strands16.
Narrowing down the location where life
started will require understanding of the
broader picture of prebiotic chemistry: how
the many reactions fit together, and the ranges
of conditions under which they occur. That
mammoth task has been attempted by a group
led by chemist Sara Szymkuć, president of the
start-up firm Allchemy in Highland, Indiana.
The team published a comprehensive study in
September that used a computer algorithm to
explore how a vast network of known prebiotic
reactions could have produced many of the
biological molecules used in life today17.
The network was highly redundant, so key
biological compounds could still form even
NASA’s Perseverance rover will search for signs of life in Jezero Crater on Mars.
if multiple reactions were blocked. For this
reason, Szymkuć argues that it is too early
to rule out any of the scenarios for where life
originated. That will require systematically
testing a range of different environments, to
see which reactions occur where.
Beyond Earth
If experiments such as Sutherland’s do point
the way to how life began on Earth, they can
also help to explore where life might have
started elsewhere in the cosmos.
Mars has attracted the most attention,
because there is clear evidence it once had
liquid water on its surface. The landing site for
NASA’s Perseverance rover, the Jezero Crater,
was chosen in part because it seems to have
once been a lake — and could have hosted the
“You can say with some
degree of confidence we
need to be on the surface, we
can’t be deep in the ocean.”
chemistry Sutherland has studied. He helped
to write a 2018 presentation to NASA led by
Catling, which summarized the prebiotic
chemistry findings and advised on where
Perseverance should look. “We presented this
chemistry and said this Jezero Crater, which is
the one they eventually chose, is the one where
there was the highest likelihood of this chem-
istry playing out,” says Sutherland.
It will be two months before Perseverance
reaches Mars — and years before the samples it
collects are returned to Earth by an as-yet-un-
named future mission. So, there is still a long
wait before we find out whether Mars harbours
life, or if it did so billions of years ago. But even
if it did not, it might reveal traces of prebiotic
chemistry.
The best case, says Catling, is that
Perseverance finds complicated carbon-based
molecules in the layers of Martian sediment,
such as lipids or proteins, or their degraded
remains. He also hopes for evidence of
wet–dry cycles. This might come in the form
of carbonate layers that formed when a lake
dried and refilled many times. He suspects
that “life didn’t get particularly far on Mars”,
because we haven’t seen any obvious signs of
it, such as clear fossils or carbon-rich black
shales. “What we’re looking for is pretty sim-
ple, maybe even to the point of being prebiotic
rather than the actual cells themselves.”
It could be that Mars took only the first few
chemical steps towards life, and did not go all
the way. In that case, we might find fossils — not
of life, but of pre-life.
Michael Marshall is a science writer based
in Devon, UK, and the author of The Genesis
Quest.
1. Miller, S. L. Science 117, 528–529 (1953).
2. Shapiro, R. Origins: A Skeptic’s Guide to the Creation of
Life on Earth (Summit, 1986).
3. Powner, M. W., Gerland, B. & Sutherland, J. D. Nature 459,
239–242 (2009).
4. Patel, B. H., Percivalle, C., Ritson, D. J., Duffy, C. D. &
Sutherland, J. D. Nature Chem. 7, 301–307 (2015).
5. Xu, J. et al. Nature 582, 60–66 (2020).
6. Frenkel-Pinter, M. et al. Proc. Natl Acad. Sci. USA 116,
16338–16346 (2019).
7. Deamer, D. W. & Barchfeld, G. L. J. Mol. Evol. 18, 203–206
(1982).
8. Rajamani, S. et al. Orig. Life Evol. Biosph. 38, 57–74 (2008).
9. Zhu, T. F., Adamala, K., Zhang, N. & Szostak, J. W. Proc.
Natl Acad. Sci. USA 109, 9828–9832 (2012).
10. Bonfio, C. et al. Nature Chem. 9, 1229–1234 (2017).
11. Frenkel-Pinter, M. et al. Nature Commun. 11, 3137 (2020).
12. Bonfio, C., Russell, D. A., Green, N. J., Mariani, A. &
Sutherland, J. D. Chem. Sci. 11, 10688–10697 (2020).
13. do Nascimento Vieira, A., Kleinermanns, K., Martin, W. F. &
Preiner, M. FEBS Lett. 594, 2717–2733 (2020).
14. Damer, B. & Deamer, D. Astrobiology 20, 429–452 (2020).
15. Milshteyn, D., Damer, B. Havig, J. & Deamer, D. Life 8, 11
(2018).
16. Deamer, D., Damer, B. & Kompanichenko, V. Astrobiology
19, 1523–1537 (2019).
17. Wołos, A. et al. Science 369, eaaw1955 (2020).

Nature | Vol 588 | 10 December 2020 | 213

©
2
0
2
0
S
p
r
i
n
g
e
r
N
a
t
u
r
e
L
i
m
i
t
e
d
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
 Science in culture

Books & arts

In Silico
Director: Noah Hutton
Sandbox Films (2020)

Kasparov in 1997 — simulate an entire rodent

brain within a decade. He planned to build it
from information about the brain’s tens of
millions of individual neurons.
Entranced, Hutton sought permission
to film with the project annually over those
ten years. He had no idea that he would end
up tracking one of the twenty-first century’s
most explosive scientific ventures. Nor that
the ten-year horizon would never get closer.

Rise and fall

In 2010, the first year of filming, Hutton cap-
tures Markram’s boastful mood: “I believe we
will understand the brain before we even finish
building it.” In 2011, Blue Brain ran a simulation
that for the first time generated something
the team hadn’t programmed — a wave that
seemed to mimic the spontaneous, synchro-
nized electrical activity in real brains. “This is
it,” gasps Markram.
SANDBOX FILMS

But that year, Hutton also started to encoun-

ter critics in the neuroscience community. They
claimed that the simulation project was pre-
In Silico focuses on Henry Markram’s attempts to model rodent and human brains. mature because too little was known about the
different types of neuron in the brain and how

Implosion of a billion-euro
they were wired. Anyone can repair a broken
watch by putting its known components in the
right places, neuroscientist Zachary Mainen at

brain model: the movie

the Champalimaud Centre for the Unknown
in Lisbon, Portugal, tells the camera. Try this
with the incompletely understood components
of the brain, he says, “and you’ll end up with a
bunch of parts that doesn’t tell the time”.
Annual footage offers tantalizing glimpses inside a As for that real-looking wave? Sebastian
Seung, now at Princeton University in New Jer-
troubled European flagship project. By Alison Abbott sey, says: “How would you know if that activity

I
n October 2013, I attended the launch of the
Human Brain Project in Lausanne, Switzer-
land, as correspondent for Nature. I hoped
to leave with a better understanding of the
exact mission of the baffling billion-euro
enterprise, but I was frustrated. Things became
clear the following year, when the project fell
spectacularly, and very publicly, apart.
Noah Hutton’s documentary In Silico cap-
tures a sense of what it was like behind the
scenes of the project, which was supported
with great fanfare by the European Commis-
sion. It had been hyped as a quantum leap in
understanding how the human brain works.
Instead, it left a trail of angry neuroscientists
across Europe. Yet aspects of what went so
expensively wrong still remain elusive.
In Silico is more about the back story of
the Human Brain Project (HBP). Hutton was
22 years old when he watched a 2009 talk
by Henry Markram, the controversial figure
who later became the first director of the HBP.
Markham was speaking about the Blue Brain
Project, a major initiative he had launched a
few years before at one of Europe’s top univer-
sities, the Swiss Federal Institute of Technology
in Lausanne, with generous funding from the
Swiss government. He claimed that he would
— with the help of a supercomputer related to
the one that beat world chess champion Garry
pattern was right or wrong?”
When Hutton visited the next year, Markram
ticked him off for contacting critics without
informing him. The commission was decid-
ing which two projects would become its bil-
lion-euro Future and Emerging Technologies
Flagships and Markram didn’t want any con-
troversy to upset his chances.
The film suggests (as other commentators
have) that Markram saw the flagship pro-
gramme as a means to expand Blue Brain. But
to win the money, it had to be more than that.
He had to team up with top scientists in other
European Union countries to present an inter-
disciplinary collaboration. He persuaded some
initially sceptical cognitive neuroscientists
to join. Their job, it was understood, would

Nature | Vol 588 | 10 December 2020 | 215

©
2
0
2
0
S
p
r
i
n
g
e
r
N
a
t
u
r
e
L
i
m
i
t
e
d
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
Books & arts
Books in brief
Wall Disease
Jessica Wapner The Experiment (2020)
Since the Berlin Wall fell in 1989, border walls have multiplied, notes
science journalist Jessica Wapner in her compelling, dispiriting, global
survey. In the decade after the September 2001 terrorist attacks,
47 appeared worldwide; Wapner investigates their geography and
psychological effects. “Wall disease” — a translation of Mauerkrankheit,
coined in 1973 by a former Berlin psychiatrist who had abandoned East
Germany for the West — consists of fear, isolation, a sense of immobility,
financial insecurity and suspicion of “the other” on the far side.
The Brutish Museums
Dan Hicks Pluto (2020)
This timely book echoes the British Museum’s decision this year to
redisplay a bust of its founder with labels about his links to the slave
trade. Dan Hicks is a curator at the Pitt Rivers Museum in Oxford,
UK, which, like the British Museum, holds many prized objects
murderously looted by colonial forces in 1897 from Benin, in what is
now Nigeria. Rejecting the view of Oxford colleague John Boardman
that “the rape proved to be a rescue”, Hicks vehemently advocates
that “brutish” museums urgently begin restitution of stolen objects.
be to ensure that brain simulations would be
linked to behavioural outcomes, so they would
always know whether any simulated activity
was ‘right or wrong’. The film only scratches
the surface of this thorny issue, although it is
central to the scientific controversy.
What comes across more strongly is how
Markram’s frequent overblown claims for the
simulation projects — that they would obviate
the need for animal experiments, for example
— irritated many in the community. “Henry
has two personalities,” says Christof Koch,
“A fascinating window
into the trouble grandiose
projects and grandiose
personalities can generate.”
president of the Allen Institute for Brain Science
in Seattle, Washington. “One is a fantastic, sober
scientist … the other is a PR-minded messiah.”
Markram’s answers to these charges on
camera are often evasive; his critics, he says,
simply don’t accept an unconventional way
of doing science.

What Is a Complex System?
James Ladyman & Karoline Wiesner Yale Univ. Press (2020)
The Santa Fe Institute in New Mexico inaugurated the study of complex
systems, but its founding workshops in 1984 did not define the topic.
Even today there is no agreement on a definition, nor whether one
is possible, remark philosopher of science James Ladyman and
mathematician Karoline Wiesner. After a clear analysis of systems
ranging from radiation to human brains, they conclude: there is no
“single natural phenomenon of complexity”, but ‘complexity science’
does exist, rather than being “merely branches of different sciences”.
A Manual of the Mammalia
Douglas A. Kelt & James L. Patton Univ. Chicago Press (2020)
The subtitle of this comprehensive, lavishly illustrated reference book
terms it “an homage” to Timothy Lawlor’s acclaimed Handbook to the
Orders and Families of Living Mammals, which was published in 1979,
revised, but out of date following Lawlor’s death in 2011. As wildlife
ecologist Douglas Kelt and mammal curator James Patton note,
Lawlor’s final edition featured about 4,170 species of mammal; today’s
figure is 6,495. “Do not be overwhelmed”, they advise students,
“simply revel in the diversity that is the Mammalia.” Andrew Robinson
Yellowstone Wolves
Eds Douglas W. Smith et al. Univ. Chicago Press (2020)
Twenty-five years ago, the authors reintroduced wolves to Yellowstone
National Park in Wyoming — the first deliberate return of an apex
carnivore to a big ecosystem. Here, they relate what they’ve learnt of
the animals’ predation, mating, play, genetics, disease and more, and
their impact on other species and the landscape. Also detailed are the
fraught history, politics and implications of rewilding. Glorious pictures
bear witness to fragile gains. US President Donald Trump’s silver-
anniversary gift? Rolling back protections on the wolves. Sara Abdulla
Internal tensions
Early optimism is quickly strained, as project
members are sidelined. Hutton returns to find
that just nine months after the launch, Mainen
and some colleagues had written a public let-
ter calling on the commission to rethink the
project, claiming that autocratic management
was distorting its mission. The letter attracted
around 800 signatories from neuroscientists
globally. (Two years later, they set out an alter-
native approach in this journal: Z. F. Mainen
et al. Nature 539, 159–161; 2016).
By 2016, Markram had been removed from
the leadership (see Nature https://doi.org/
fkgx; 2015). The final two years of filming fol-
low him back on Blue Brain. The simulation
progresses, the 3D visualizations get more
impressive, research papers emerge — but the
project’s pep seems to drain away. Markram’s
insistence that a complete brain simulation is
still just ten years away sounds hollow. Mean-
while, the HBP continues with a more distrib-
uted, democratic structure.
In Silico is a fascinating window into the
trouble grandiose research projects and
grandiose personalities can generate, even if
it fails to get to the heart of what specifically
went wrong with the HBP. Hutton hints that the
disputes were driven by money. I disagree; my
sense is that it came down to leadership style
and irresolvable differences in scientific opin-
ion. There is a bolder, even more interesting,
story waiting to be told.
Alison Abbott writes from Munich, Germany.
e-mail: alison.abbott.consultant@
springernature.com

216 | Nature | Vol 588 | 10 December 2020

©
2
0
2
0
S
p
r
i
n
g
e
r
N
a
t
u
r
e
L
i
m
i
t
e
d
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
 Setting the agenda in research

Comment
BUDA MENDES/GETTY

Smoke rises from a fire in Brazil’s Pantanal, the world’s largest tropical wetland, in September.

Rescue Brazil’s burning Pantanal

wetlands
Renata Libonati, Carlos C. DaCamara, Leonardo F. Peres, Lino A. Sander de Carvalho & Letícia C. Garcia

B
Climate extremes, poor razil has changed. As well as the
COVID-19 pandemic killing more than
management and lax laws 170,000 of its citizens so far, 2020 has
are making this World seen almost one-third of the Pantanal,
the largest tropical wetland in the
Heritage Site prone to world, on fire. Four million hectares of for-
fierce fires. Researchers and est, savannah and shrub-land (an area bigger
governments must develop than the US state of Maryland) have gone up
in flames since January (see go.nature.com/
a plan to manage these risks 2jtw6va). Almost all the Indigenous territories
together. and conservation facilities were burnt, as was
much private land. Conservation areas such
as Encontro das Águas State Park have been
devastated — it contained one of the largest
populations of jaguars in the world.
Fires’ impacts have been felt nationwide.
Smoke has spread thousands of kilometres,
reducing air quality in São Paulo, Rio de
Janeiro and Curitiba. Southern states have
experienced showers of black rain. The fires
are decimating Brazil’s economy, curbing
inward investment as well as sectors such as

Nature | Vol 588 | 10 December 2020 | 217

©
2
0
2
0
S
p
r
i
n
g
e
r
N
a
t
u
r
e
L
i
m
i
t
e
d
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
Comment
air travel and tourism that are already hit hard
by the pandemic.
The public is worried. The fires have
made the headlines for months. Thousands
of Brazilians have volunteered to fight the
flames, rescue wildlife or donate money. Yet,
Brazil’s government is doing little. It is ignor-
ing the causes of the fires: a combination
of inadequate fire management, climate
extremes, human behaviour and weak envi-
ronmental regulations. Worse, it has slashed
funding for fire prevention and has been
slow to contract firefighters. It has even
cast doubts on the reliability of satellite fire
detections.
On the scientific front, fire risks and impacts
in the region are under-studied. Deeper
research is needed on the weather condi-
tions that fan fires, as well as the influences
of ecology and management. Scientists need
to know how the many factors behind large
fires interact, including vegetation stress,
extreme weather and human activities. And
more studies are needed to inform fire man-
agement strategies in the region.
This year’s fire season in the Pantanal is
exceptional. But the conditions that led to
these blazes are becoming increasingly com-
mon as the area warms. In response, political,
socio-economic and scientific approaches
need to change. Researchers and governments
need to come together to develop a compre-
hensive strategy for preventing and managing
fires. Otherwise this great tropical wilderness
will not bounce back.
Devastating impacts
With more than 84% of its territory conserved,
the Pantanal is the largest remaining wetland
area of natural vegetation in the world. It’s a
UNESCO World Heritage Site. Indigenous, riv-
erine and quilombo communities live there.
Traditional farmers practice unique forms
of sustainable agriculture, including grazing
to push back exceptionally fierce flames (see
‘Pantanal fire crisis’).
The total loss will take months to calculate.
But the impacts are long lasting. Charcoal and
ash contaminate rivers and promote harm-
ful bacteria that poison drinking-water sup-
plies and kill fish. Eroded soils are flushed
downstream. Fire-sensitive plants struggle
to produce seeds. Vast tracts of land will need
to be assessed to understand whether they
can be restored. Communities will have to be
rebuilt.
Growing risk
What lies behind these fires? The Pantanal
is no stranger to burning, even though it is a
wetland2. For half the year it is dry and prone
to catching alight, especially during drought.
Sometimes lightning causes the spark. More
often it is human-related — flashes from elec-
trical cables, burning garbage and wood from
BOLIVIA
PANTANAL
FIRE CRISIS
Almost one-third of this
wetland World Heritage Site
burnt down in 2020. Drought,
climate change, inadequate
government response and the
pandemic all played a part.
Paraguay
River
Brazil
cattle fences, use of fire to ward off bee attacks
when collecting honey, and even car crashes
and damaged agricultural machinery. Cattle-
men burn the landscape to remove shrubs and
stimulate the growth of native grasses, which
are adapted to fire and sprout after pruning
or scorching. Such fires regularly get out of
control, especially in areas where there’s no
system for managing them3.
The frequency and severity of fire out-
breaks are worsening, as the climate warms
and human impacts increase. Since 1980, aver-
age temperatures there have risen by 2 °C and
humidity has fallen by 25%, according to the
European Centre for Medium-range Weather
Forecasts (ECMWF). This year saw the worst
drought recorded in the Pantanal in 60 years
(see go.nature.com/2jpdubc), induced by unu-
sually warm waters in the North Atlantic4. The
wet season saw 57% less rain than normal. By
June, the Paraguay River was at half its usual
Encontro das
Águas State Park
contains the
most jaguars
BRAZIL
in the Pantanal.
Pantanal
Amolar
region
100% of
Matogrossense
National Park was
burnt. It’s one of
South America’s
premier wetlands
for waterbirds.

SOURCE: LABORATORY FOR ENVIRONMENTAL SATELLITE APPLICATIONS, FED. UNIV. RIO DE JANEIRO
cattle on native pastures and moving animals
to higher land when lowlands flood. Tourists
flock to the region for its spectacular scenery,
safaris and sport-fishing.
Each rainy season, from October to April, PARAGUAY
pulses of floods swell the Paraguay River to sup-
port ecosystems found nowhere else on Earth. Fire danger ratings are rising
Endangered jaguar, giant otter, marsh deer as the region warms. 2020
saw the worst conditions in
and hyacinth macaws roam wild. Thousands three decades.
of birds pass through on their migrations1. Indigenous Kadiwéu
Difficulty of 9.9 people are trained
It’s a haven for caimans, capybaras, monkeys,
controlling fires to fight fires in their
deer, coatis, tapirs, snakes and the jabiru stork (DSR index*) territory.
( Jabiru mycteria) — the region’s symbol.
The fires have affected all aspects of life.
COVID-19 has made things worse. PREVFOGO, Conservation Indigenous
areas territories
the national centre for forest fire prevention
4 Fires
and fighting, has struggled to hire and train
2019 2020 Both years
firefighters. Many fires broke out in remote
regions, even underground, that were hard 1980 1990 2000 2010 2020
to reach. Local firefighters in the Kadiwéu ter- *Averaged daily severity rating (DSR) from January to
August of each year for the Pantanal biome.
ritory, for example, struggled almost alone

218 | Nature | Vol 588 | 10 December 2020

©
2
0
2
0
S
p
r
i
n
g
e
r
N
a
t
u
r
e
L
i
m
i
t
e
d
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
level. This combination of hot, dry conditions
pushed flammability thresholds to their
highest since 1980. Such thresholds indicate
the difficulty of controlling fires, a scale that
is quantified using the averaged daily sever-
ity rating (DSR) index, which is derived from
ECMWF data. Deforestation in Amazonia has
also been linked to reduced rainfall in the
Pantanal, although this is debated.
Environmental regulations are failing to
keep up5. In July, Brazil’s government issued
a 120-day ban on the use of fire in the Amazon
and the Pantanal. It seems to have been
widely ignored. The government has denied
responsibility, blaming Indigenous peoples
and traditional communities for starting fires,
and criticized campaigns by the media and
non-governmental organizations highlighting
the exceptionality of the fire season.
Resources for environmental protection
and climate actions have been slashed, espe-
cially in the past two years. The Ministry of
the Environment’s US$630-million budget
was cut by around 20% in 2020 and looks set
to fall by a further 35% in 2021. Brazil is also
failing to meet its commitment to reduce
greenhouse-gas emissions under the Paris
climate agreement6. Licensing requirements
for dams, roads and mines have been weak-
ened (Nature 572, 161–162; 2019). Last year,
to promote agricultural and biofuel produc-
tion, the government revoked the law that has
prohibited new sugar-cane plantations in the
Amazon and the Pantanal that has been in
place since 2009 (ref. 7). The decree was pro-
visionally suspended by the Brazilian federal
court in April, and is awaiting a final decision.
Researchers need to bolster evidence to
back a new approach. Until now, most studies
in the Pantanal have focused on a single dis-
cipline, plant ecology for example. Research
on other topics, such as climate, isn’t granu-
lar enough. There are few studies of human
causes and responses to fires in the Pantanal,
to inform fire-management strategies. A full
understanding of cycles of burning and long-
term trends is missing.
Fire science is multidisciplinary, spanning
fields from climate to chemistry, ecology to
economics, as well as risk analysis and com-
putational modelling. A task force is needed to
bring together researchers from all these areas,
along with technicians working in the field.
Neglecting the connections between
climate, land use and fire management will
make it impossible to restore the Pantanal to
its former state, let alone protect the region in
the future. Any change to the natural pattern of
burning disrupts ecosystems and food chains,
sometimes completely. For instance, jaguars
will struggle to find herbivores to eat, if the
latter are killed by flames or are unable to find
fruits and leaves in a scorched landscape. Gen-
erations of fire-sensitive trees could be lost,
including Genipa americana3, fruits of which
are a staple for fauna and used by Indigenous
people to make black ink for body painting.
The impacts cascade quickly. Repeated
wildfires lower the resilience of communities
and vegetation; forests are replaced by open
landscapes with fewer resources.
Economic fallout
Brazil must act on deforestation and forest
fires to protect its economy. After earlier
fires in 2019, Norway and Germany froze
their donations to the Brazilian govern-
ment’s Amazon Fund, after having contrib-
uted more than $1.2 billion and $68 million,
respectively. Around 250 investors, including
the California Public Employees’ Retirement
System (CalPERS), representing approxi-
mately $17.7 trillion in assets, endorsed an
open letter pointing out the financial impacts
that deforestation may have on investee com-
panies (see go.nature.com/36gzirt).
“Resources for
environmental protection
and climate actions have
been slashed.”
In June, 7 European investment firms,
managing $2 trillion in assets ($5 billion linked
to Brazil), announced they might divest from
beef producers, grain traders and govern-
ment bonds in Brazil if there was no progress
in stopping deforestation and fires. Soon
after, 34 companies (including the Church
of England and KPL, Norway’s pension fund,
managing around $4 trillion) wrote to Brazilian
embassies in their countries (including Norway,
Sweden, France, Denmark, the Netherlands,
the United States and the United Kingdom)
expressing concerns over the dismantling of
environmental policies in Brazil.
European countries (France, Austria and
the Netherlands) threaten not to ratify the
provisional trade deal between the European
Union and the Mercosur bloc (comprising
Brazil, Argentina, Uruguay and Paraguay),
unless Brazil achieves its Paris climate com-
mitments. The EU–Mercosur agreement was
negotiated for 20 years and is considered
the largest free-trade agreement in history.
It accounts for $20 trillion of global gross
domestic product (GDP), about one-quarter
of the world’s economy, and the consumer
market in the 32 countries reaches 780 million
people. Currently, Brazilian companies export
almost $20 billion to the EU; the deal would
lead to an increase of $100 billion for Brazil’s
GDP by 2035.
Steps forward
Brazil’s government must develop a long-
term strategy to mitigate damage from wild-
fires in the Pantanal that takes all factors into
account, including effective fire management
and environmental protection policies.
Researchers need to shore up knowledge about
the fire regime there to inform this strategy.
First, gather satellite and other data about
the time, location and intensity of fires, burnt
area and vegetation conditions before and
after. This information can then be used to
assess factors behind the onset and spreading
of fires. Scientists should model the impacts of
current and future land use and climate change
on fire events, as well as feedbacks such as
between biomass burning and global warming.
Second, model fire management and
response strategies, including the impacts
on biota, pasture, communities, economies,
ecology, weather and fire risk. Fire managers
need to decide which areas to protect and
which activities to prohibit, taking into account
scientific, Indigenous and local knowledge.
Some areas could be kept fire free, or have
carefully managed blazes outside the dry sea-
son to protect biodiversity. Other areas might
accommodate agriculture, cattle or tourism, as
long as fire-management principles, as well as
state and federal legislation on environmental
protection are followed (such as the 2012 Bra-
zilian Forest Code). Near-real-time informa-
tion about the location, intensity and spread
of wildfires in the Pantanal should be dissemi-
nated, along with daily forecasts of fire danger.
Funding should be directed towards fire
management and environmental protection,
as well as to law enforcement and fine
collecting by environmental inspectors.
Education and information programmes
in schools or by the media would make the
population more aware of the consequences
of irresponsible behaviour.
A warming and fast-changing world demands
a new proactive approach to fighting wildfires.
The authors
Renata Libonati and Lino A. Sander
de Carvalho are adjunct professors of
climatology and remote sensing, and
Leonardo F. Peres is an associate professor of
atmospheric sciences and remote sensing at
the Federal University of Rio de Janeiro, Brazil.
Carlos C. DaCamara is an associate professor
of climate science at the University of Lisbon,
Portugal. Letícia C. Garcia is an adjunct
professor of restoration ecology at the Federal
University of Mato Grosso do Sul, Brazil.
e-mail: renata.libonati@igeo.ufrj.br
1. De Pinho, J. B., Aragona, M., Hakamada, K. Y. P. & Marini,
M. Â. Bird Conserv. Int. 27, 371–387 (2017).
2. Pivello, V. R. Fire Ecol. 7, 24–39 (2011).
3. Pott, A. & Pott, V. J. Wetl. Ecol. Manag. 12, 547–552 (2004).
4. Thielen, D. et al. PLoS ONE 15, e0227437 (2020).
5. Abessa, D., Famá, A. & Buruaem. L. Nature Ecol. Evol. 3,
510–511 (2019).
6. da Silva Junior, C. A. et al. Sci. Rep. 10, 16246 (2020).
7. Ferrante, L. & Fearnside, P. M. Science 359, 1476 (2018).

Nature | Vol 588 | 10 December 2020 | 219

©
2
0
2
0
S
p
r
i
n
g
e
r
N
a
t
u
r
e
L
i
m
i
t
e
d
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
Readers respond

Correspondence
Another diversity
problem — scientists’
politics
According to your poll before
the US presidential election
(see Nature 586, 654; 2020), the
political leaning of scientists
was 86% in favour of Democrat
Joe Biden, now president-
elect, with just 8% supporting
Republican Donald Trump, the
outgoing president. However,
this finding is glaringly out
of step with the voting of the
population from which the US
scientists were drawn (about 51%
versus 47%, respectively).
This misalignment could
be attributed to differences in
education, understanding and
awareness of the issues at stake.
But such a gulf risks isolating
science further from society at a
time when we should be building
bridges beyond this election.
As academics become more
aware of the importance of
diversity of thought, we must be
careful not to recreate different
forms of the old elitist patterns
of collective behaviour recently
challenged by anti-racism. Any
association of science with
political archetypes could turn
some against it by enhancing
the view that it is an exclusive
pursuit.
Andrew Isaac Meso King’s
College London, UK.
andrew.meso@kcl.ac.uk
Land use predicts What counts as
pandemic disparities climate finance?
Define urgently
COVID-19 morbidity is
linked to social, economic
and environmental factors,
including residential
location, air pollution and
median household income
(H. A. Washington Nature 581,
241; 2020). These have an
overlapping determinant that
could prove to be an important
predictor of COVID‑19
disparities: land use.
The United States has a
strained history of land use and
land governance, including
ethnic constraints on land
ownership and unfair mortgage-
lending practices. Decisions
on land-use classification
have led to hazardous and
polluting facilities being
sited next to minority and
other vulnerable residential
communities. Despite policies
enacted in 1968 to protect
against housing discrimination
(go.nature.com/39v1bt3), the
United States is witnessing
a correlation of historical
‘redlining’ — the systematic
denial of services to residents
of certain areas, on the basis
of race or ethnicity — with
COVID‑19 incidence today.
It is crucial that land-use
practices are considered
when making public-health
management decisions. This
could help to mitigate the multi-
generational, compounding
impacts of isolated or confined
residential spaces. Those who
live in such areas will continue
to take a disproportionate hit
unless land-use equity is made a
priority in governance.
Cesunica Ivey University of
California, Riverside, USA.
cesunica@ucr.edu
To resolve arguments over
what funding actually flows
from developed to developing
nations, the United Nations
Framework Convention on
Climate Change needs to
draw up a definition of what
constitutes climate finance.
At the 2009 UN climate
summit, developed countries
pledged to mobilize US$100
billion annually by 2020 to
help developing countries
mitigate and adapt to climate
change. Has the promise
been met? The answer to this
question will be available only
in “the first quarter of 2022 at
the earliest”, according to a
report published last month
(go.nature.com/2kdeklu) by
the Organisation for Economic
Co-operation and Development
(OECD), a club of wealthy
countries.
Letting the OECD decide what
counts as climate finance on the
world’s behalf risks introducing
questionable accounting
practices (see R. Weikmans and
J. T. Roberts Clim. Dev. 11, 97–111;
2019). The OECD, for example,
continues to account loans
at face value, which equates a
$10‑million loan (which has to
be paid back) to a $10‑million
grant. It is therefore no surprise
that developing countries
have found OECD reports
unacceptable before (see Nature
573, 328–331; 2019).
Romain Weikmans Free University
of Brussels, Belgium.
romain.weikmans@ulb.be
J. Timmons Roberts Brown
University, Providence, Rhode
Island, USA.
Stacy-ann Robinson Colby
College, Waterville, Maine, USA.
Combine resilience
and efficiency in
post-COVID societies
As countries prepare to remodel
themselves after the COVID-19
pandemic, they must tackle
growth and development
expectations by using resources
more sustainably, and by
ensuring that their societies are
better placed to weather future
disruptions.
The COVID-19 experience
indicates that society could
become more vulnerable to
systemic shocks and cascading
disruption if the practices on
which it depends excessively
prioritize system efficiency
over resilience. Efficiency
emphasizes performance
at maximum capacity with
minimal use of scarce resources.
To meet the rising demands
of society, efficiency-based
approaches often rely on
increasingly complex and
interconnected systems. But
when a tightly interdependent
society encounters acute or
chronic stressors beyond its
expectations or operating
capabilities, such highly
efficient systems are prone to
catastrophic failure that can
delay or prevent recovery.
More-resilient systems
might be less efficient, but they
recover better from systemic
disruptions. Building resilience
does not mean abandoning
efficiency, but rather maximizing
socio-economic systems’ long-
term sustainability in the face
of future disruptions. Marrying
resilience with efficiency would
allow society to preserve or
even improve living standards in
current and future crises.
Benjamin D. Trump, Igor Linkov
US Army Corps of Engineers,
Boston, Massachusetts, USA.
igor.linkov@usace.army.mil
William Hynes OECD, Paris,
France.

220 | Nature | Vol 588 | 10 December 2020

©
2
0
2
0
S
p
r
i
n
g
e
r
N
a
t
u
r
e
L
i
m
i
t
e
d
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
Expert insight into current research

News & views

predatory dinosaurs6. Despite the present-day
catalogue of approximately 200 Meso­zoic bird
Palaeontology
species from around the world7, ranging in age

The changing face of birds

from about 150 million to 66 million years old,
none has a skull resembling anything like that
of Falcatakely. Its discovery reveals a skull

from the age of dinosaurs
Daniel J. Field
The fossil record traces the origin of the modern bird skull
as birds evolved from their dinosaurian ancestors. Now the
discovery of a bizarre fossil reveals a surprising diversion
during this process of facial transformation. See p.272
As living dinosaurs, birds are the product of
a long and complex evolutionary history that
has given rise to more than 11,000 living spe-
cies1. The past decade has witnessed a surge
of interest in the evolution of the avian skull
— a structure that is hugely variable across the
diversity of living birds2. However, our abil-
ity to test hypotheses of how and when key
transformations of the bird skull took place
is limited if we can’t incorporate fossils into
evolutionary models. On page 272, O’Connor
et al.3 report a stunning fossil-bird discovery
from the age of the dinosaurs that reminds us
of the crucial value of fossils for casting light
on unexpected complexities in avian evolu-
tionary history.
This striking addition to the aviary of
the Mesozoic era is between 72 million and
66 million years old (corresponding to the
latest stage of the Cretaceous period). It
comes from Madagascar, and is named
Falcatakely forsterae, which roughly translates
as Forster’s small scythe beak. The name refer-
ences the distinctive shape of the fossil’s bill
and honours Catherine Forster’s numerous
contributions to vertebrate palaeontology in
Madagascar. The specimen is small (less than
9 centimetres long) and delicate (paper thin
in places), yet the stunning bone preservation
provides a spectacular look at this ancient
creature’s anatomy.
Although the fossil consists of only the
front half of a skull, it’s clear that Falcatakely
is more than just a pretty face. The skull is
utterly bizarre, characterized by a deep and
elongated snout (Fig. 1) unlike those seen in any
other Mesozoic birds. The skull’s architecture
becomes even weirder. The very tip of its snout
has one small preserved tooth (the tip possibly
had more teeth that were not preserved); how-
ever, there are clearly no teeth anywhere else
shape previously unknown for any bird from
the age of the dinosaurs.
The exceptional degree of preservation of
Falcatakely enabled the authors to make other
astonishing findings. Imaging using a method
called high-resolution micro­c omputed
tomography enabled them to digitally ‘extract’
the fragile skull bones from the surrounding
rock. O’Connor and colleagues could then
reassemble the delicate components of the
bill, including elements such as the paper-
thin palate bones, which are rarely found
along its jaws. By contrast, the closest relatives preserved, into a compelling 3D model (see
of modern birds from the time of the dinosaurs Supplementary Videos 1–8 of ref. 3).
show the opposite pattern, with teeth found Studying the palate, the authors spotted a
throughout the jaws, but none at the tip of surprising bone called the ectopterygoid. This
the beak (Fig. 1)4. These features give the skull is absent in living birds, but is a component of
of Falcatakely an almost comical profile — the palate of non-avian dinosaurs and early
imagine a creature resembling a tiny, buck- bird-like forms, such as the iconic early birds
toothed toucan flitting from branch to branch, Archaeopteryx and Sapeornis8. However, on the
occasionally glancing down at Madagascar’s basis of detailed analyses, O’Connor et al. infer
formidable Late Cretaceous inhabitants, which that Falcatakely belongs to a group of Meso-
included equally bizarre mammals5 and giant zoic ‘pre-modern’ birds called Enantiornithes
Falcatakely Ichthyornis Asteriornis
Lacrimal Nasal
Maxilla Premaxilla
(upper jaw) (upper beak)
Towards
living birds
Figure 1 | The evolution of ancient bird skulls. Discoveries of bird skulls from the Mesozoic era (the age of
the dinosaurs) have revealed both how the skull of modern birds arose and the surprising variability of these
ancient skulls (as illustrated by these fossils, reported between 2018 and 2020). O’Connor et al.3 present
their discovery of the skull of a bird specimen they name Falcatakely forsterae, which shows an unusually
deep and elongated snout, with teeth (at least one tooth and possibly more) positioned only at the very tip
of the upper jaw in a skull region called the premaxilla. Like other distant relatives of modern birds, such
as non-avian dinosaurs, the upper jaw of Falcatakely consists mainly of a region called the maxilla. Closer
relatives of modern birds, such as Ichthyornis4, had teeth throughout the jaws, except at the tip, and retained
the ancestrally large maxilla. Early modern birds, including Asteriornis (an ancient relative of chickens and
ducks)13, lost their teeth completely, and had upper jaws dominated by the premaxilla. Nasal bones are
shown in grey and lacrimal bones (inferred for Asteriornis) are in beige. (Figure adapted from Fig. 2 of ref. 3,
Fig. 3 of ref. 4 and Fig. 1 of ref. 13.)

Nature | Vol 588 | 10 December 2020 | 221

©
2
0
2
0
S
p
r
i
n
g
e
r
N
a
t
u
r
e
L
i
m
i
t
e
d
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
News & views
(a name that means ‘opposite birds’, in refer-
ence to their atypical shoulder-joint articula-
tions), which occupy a branch of the dinosaur
family tree that is much closer to that of mod-
ern birds than the branches occupied by either
Archaeopteryx or Sapeornis. The presence of
an ectopterygoid in Enantiornithes has been
suggested previously9, but this identification
has been questioned10. Thus, the detection of
an ecto­pterygoid in Falcatakely either shows
that this ancestral component of the palate was
indeed retained in Enantiornithes (at a rela-
tively late stage in avian evolutionary history),
or challenges the identification of Falcatakely
as a member of Enantiornithes, suggesting
instead that it belongs on a deeper branch of
Daniel J. Field is in the Department of
Earth Sciences, University of Cambridge,
Cambridge CB2 3EQ, UK.
e-mail: djf70@cam.ac.uk
1. del Hoyo, J. All the Birds of the World (Lynx, 2020).
2. Felice, R. N. & Goswami, A. Proc. Natl Acad. Sci. USA 115,
555–560 (2018).
3. O’Connor, P. M. et al. Nature 588, 272–276 (2020).
4. Field, D. J. et al. Nature 557, 96–100 (2018).
5. Krause, D. W. et al. Nature 515, 512–517 (2014).
6. Lavocat, R. Bull. Mus. Natl Hist. Nat. 27, 256–259 (1955).
7. Pittman, M. et al. in Pennaraptoran Theropod Dinosaurs:
Past Progress and New Frontiers (eds Pittman, M. &
Particle physics
Xing, X.) 37–95 (Bull. Am. Mus. Natl Hist. No. 440;
Am. Mus. Natl Hist., 2020).
8. Hu, H. et al. Proc. Natl Acad. Sci. USA 116, 19571–19578
(2019).
9. Elzanowski, A. Cour. Forschungsinst. Senckenb. 181,
37–53 (1995).
10. Mayr, G. Avian Evolution: The Fossil Record of Birds and its
Paleobiological Significance (Wiley, 2016).
11. O’Connor, P. M. & Forster, C. A. J. Vert. Paleontol. 30,
1178–1201 (2010).
12. Li, Y., Ruta, M. & Wills, M. A. Syst. Biol. 69, 638–659 (2020).
13. Field, D. J., Benito, J., Chen, A., Jagt, J. W. M. &
Ksepka, D. T. Nature 579, 397–401 (2020).
14. Longrich, N. R., Tokaryk, T. & Field, D. J. Proc. Natl Acad.
Sci. USA 108, 15253–15257 (2011).
This article was published online on 25 November 2020.

How protons interact

the family tree of Mesozoic birds.
Although it is impossible to decide defin-
itively between these two options without
access to further fossil material, O’Connor
et al. grapple with this uncertainty to an
impressively thorough degree, showing that
Falcatakely nests with Enantiornithes in evo-
lutionary trees constructed under a range of
alternative analytical approaches. Moreover,
the identification of Falcatakely as a member
of Enantiornithes makes sense in light of the
previous identification of fragmentary bones
assigned to Enantiornithes from the same
Madagascan fossil locality11. Nonetheless,
some research has indicated that family-tree
reconstructions of dinosaurs can return con-
flicting results when skulls, instead of com-
plete skeletons, are analysed12. This lack of
certainty is all the more reason for the team
to continue its productive fieldwork in the
hope of discovering more-complete material.
Modern birds originated in the Late Creta-
ceous13, and it has become increasingly appar-
ent that the final 20 million years of the age of
the dinosaurs (86 million to 66 million years
ago) was a pivotal time in avian evolutionary
history. The discovery of Falcatakely shows
us that the importance of this window in time
for bird evolution extends well beyond the ori-
gin of modern birds. Apparently, ‘pre-modern’
bird lineages such as Enantiornithes were
still experimenting with bold new forms —
and possibly previously unfilled ecological
niches — well into the terminal stages of the
Cretaceous.
The pre-modern birds were wiped out in the
end-Cretaceous mass extinction event, along
with all other dinosaurs, apart from modern
birds14. Considering the impressive diversity
and global distribution of Enantiornithes in
the Late Cretaceous, determining why they
disappeared in that mass extinction, whereas
the earliest modern-bird lineages survived,
remains one of the greatest mysteries in
avian evolutionary history. The answers to
such questions, much like the unexpected
anatomy of creatures such as Falcatakely, can
be revealed only by evidence from the fossil
record. So, let’s keep digging.
with their exotic siblings
Manuel Lorenz
The nuclear forces that act on short-lived subatomic particles
have been hard to study. This problem has now been solved by
smashing high-energy protons together and measuring the
momenta of the unstable particles produced. See p.232
On page 232, the ALICE Collaboration 1
reports that data from high-energy collisions
between protons can be used to investigate the
little-understood nuclear forces between pro-
tons and subatomic particles called hyperons.
The measurements have comparable precision
to state-of-the-art numerical calculations of
the forces, thereby allowing conclusive quanti-
tative comparisons of experimental data with
theory. Accurate knowledge of these forces
is needed for various aspects of physics
research, for example in efforts to understand
the stability of neutron stars.
The nuclear force between neutrons and
protons (which are known collectively as
nucleons) is a residual effect of the strong
interaction that acts between their ele-
mentary constituents (quarks and gluons).
First-principles calculations of the nuclear
force have been challenging because of the
peculiarities of the strong interaction. Our
knowledge of this force is, therefore, based
largely on simplified models and theories2,
guided by experimental data3. The strong
interaction between hadrons (subatomic par-
ticles, such as nucleons, that consist of two
or more quarks bound together by the strong
interaction) at low energies is therefore often
referred to as the final frontier of the standard
model of particle physics.
The interaction between nucleons has been
measured with high accuracy3, but the inter-
action of nucleons with their heavier siblings,
the hyperons, is less well assessed. Hyperons
consist of three quarks, at least one of which
must be a type (flavour) known as a strange
quark; the other quarks can be up or down,
the two lightest quark flavours. Hyperons are
not present in the everyday matter that sur-
rounds us on Earth, but — depending on their
interactions with nucleons — might affect the
compressibility of nuclear matter at high den-
sities. This means they could be relevant to the
stability of neutron stars4. Precise knowledge
of hyperon–nucleon interactions is therefore
of great importance not only for nuclear phys-
ics, but also for astrophysics. However, meas-
urements of these interactions are difficult
to make in conventional experiments involv-
ing direct particle collisions in accelerators,
because hyperons are short-lived (their life-
times are about 10−10 s; ref. 5) and fly only a few
centi­metres, on average, before they decay.
The ALICE Collaboration now reports that
proton–hyperon interactions can be investi-
gated using high-energy collisions between
protons carried out at the Large Hadron Col-
lider (LHC) at CERN, Europe’s particle phys-
ics laboratory near Geneva, Switzerland. The
technique depends on measurements of corre-
lations between the momenta of protons and
hyperons produced in the collisions.
The process studied in the experiments
involves three steps (Fig. 1). First, protons are
collided at extremely high energies, taking
advantage of the fact that the LHC produces
higher collision energies than any other
accelerator. Second, hadrons are emitted

222 | Nature | Vol 588 | 10 December 2020

©
2
0
2
0
S
p
r
i
n
g
e
r
N
a
t
u
r
e
L
i
m
i
t
e
d
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
 LHC are expected to go into full operation in
a b c the coming years, including NICA in Russia,
J-PARC in Japan and FAIR in Germany. Although
fewer proton–hyperon pairs are generated per
Particle Interactions Detector collision in lower-energy collisions, a greater
source
Proton proportion of those pairs will be emitted at low
momenta — which might turn out to be advan-
Hyperon
tageous, because more data are needed to
reduce the statistical errors in measurements
Figure 1 | Investigating the proton–hyperon interaction. a, The ALICE Collaboration1 smashed together of low-momentum systems. Increases in com-
high-energy protons in CERN’s Large Hadron Collider. b, The collisions generate a ‘particle source’ — a puting power should also substantially reduce
volume of space in which components of the colliding protons interact and become confined within new the uncertainties of first-principles calcula-
particles. These new particles are emitted from the source, and include protons that pair up with heavier tions of nuclear forces. Taken together, these
particles known as hyperons. c, The paired-up protons and hyperons interact with each other in a way that developments bode well for future research
alters the relative momentum of the system, which is then measured by a detector. These measurements are
into the final frontier of the standard model
then used to determine the nuclear force between the proton–hyperon pair.
of particle physics.

by a ‘source’ produced by the collision — a
volume of space in which quarks and gluons
that originally came from the protons interact
and become confined within new hadrons. The
source emits various types of hadron, includ-
ing protons and hyperons, some of which form
proton–hyperon pairs. Finally, the proton and
hyperon in each of these pairs interact with
each other in ways that alter the momentum
of the paired system. This momentum is meas-
ured by a detector and used to determine the
momentum correlations.
The momentum correlations reflect the size
of the hadron source and the properties of the
interaction between the produced proton–
hyperon pairs. Such correlation analyses were
originally used to determine the source size in
collisions of heavy ions6, but in the new work,
be calculated from first principles so that the
results can be compared with experimental
findings. The precision with which nucleon–
nucleon interactions can be determined from
experimental data is still superior to that
obtained from these calculations, but the
ALICE Collaboration’s measurements of the
proton–hyperon interactions almost exactly
match those obtained from theory.
A wealth of high-precision measurements of
proton–hyperon interactions is expected from
the LHC in the next decade, following on from
its recent upgrade. Moreover, various other
facilities that will study particle collisions at
lower energies than those produced at the
Virology
Manuel Lorenz is at the Institute for Nuclear
Physics, Goethe University, Frankfurt 60438,
Germany.
e-mail: m.lorenz@gsi.de
1. ALICE Collaboration. Nature 588, 232–238 (2020).
2. Epelbaum, E., Hammer, H.-W. & Meissner, U.-G. Rev. Mod.
Phys. 81, 1773–1825 (2009).
3. Stoks, V. & de Swart, J. Phys. Rev. C 47, 761–767 (1993).
4. Weissenborn, S., Chatterjee, D. & Schaffner-Bielich, J.
Phys. Rev. C 85, 065802 (2012).
5. Tanabashi, M. et al. Phys. Rev. D 98, 030001 (2018).
6. Lisa, M. A., Pratt, S., Soltz, R. & Wiedemann, U. Annu. Rev.
Nucl. Part. Sci. 55, 357–402 (2005).
7. Adamczewski-Musch, J. et al. Phys. Rev. C 94, 025201 (2016).
8. Acharya, S. et al. Phys. Rev. C 99, 024001 (2019).
9. Sasaki, K. et al. Nucl. Phys. A 998, 121737 (2020).
10. Iritani, T. et al. Phys. Lett. B 792, 284–289 (2019).

Cracking the cell access

they are instead used to investigate the inter-
action between the particles of interest. This
approach to studying particle interactions was
pioneered by the HADES Collaboration7 at the
GSI Helmholtz Centre for Heavy Ion Research
in Darmstadt, Germany, and was further devel-
oped by the ALICE collaboration8 at the LHC.
The current work depends on the fact that the
extremely high-energy proton–proton colli-
sions carried out at the LHC produce a high
abundance of hyperons from small-volume
hadron sources. The authors used this method
to measure the strong force between pro-
tons and Ω– hyperons (which consist of three
strange quarks) and between protons and Ξ
hyperons (which consist of two strange quarks
and one up or down quark).
The ALICE Collaboration’s findings open
up a new ‘laboratory’ for investigating other
nucleon–hyperon interactions, including the
little-explored interactions with hyperons that
contain two or three strange quarks. This will
aid our understanding of metastable states
of hyperon pairs or of the compressibility of
nuclear matter at high densities. The latter is
relevant not only for the stability of neutron
stars, but also for neutron-star mergers and
heavy-ion collisions.
In a lucky coincidence, recent developments
in theoretical physics9,10 allow nuclear forces to
code for a deadly virus
James Zengel & Jan E. Carette
The discovery that the receptor protein LDLRAD3 is essential
for infection of human cells by Venezuelan equine encephalitis
virus could inform strategies to combat this potentially
lethal infection. See p.308
When viruses jump from animals to humans, membrane and initiate its replication cycle. On
disease outbreaks can follow. A striking exam- page 308, Ma et al.3 describe the long-sought
ple is Venezuelan equine encephalitis virus receptor for VEEV, and show that it is essential
(VEEV). This virus causes sporadic disease for viral replication in both human cells and
outbreaks in horses in Latin America that fre- mouse models.
quently spill over into humans, resulting in Interactions between a virus and its host
often-deadly neurological disease1. Because receptor protein can control which tissue
of its pathogenicity in livestock and humans, types in the body support viral growth, thus
VEEV has been studied as a biological weapon influencing the type of disease that results.
by several countries, including the United Furthermore, these interactions can deter-
States2. Treatments for the disease are there- mine how well the virus spreads through a host
fore highly desirable. It has been unknown how population. During the continuing SARS-CoV-2
VEEV co-opts cellular pathways to establish pandemic, for instance, viral strains that had a
infection in people — in particular, which host specific mutation in the virus’s spike protein
receptor protein allows VEEV to cross the cell became predominant soon after the virus

Nature | Vol 588 | 10 December 2020 | 223

©
2
0
2
0
S
p
r
i
n
g
e
r
N
a
t
u
r
e
L
i
m
i
t
e
d
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
 LHC are expected to go into full operation in
a b c the coming years, including NICA in Russia,
J-PARC in Japan and FAIR in Germany. Although
fewer proton–hyperon pairs are generated per
Particle Interactions Detector collision in lower-energy collisions, a greater
source
Proton proportion of those pairs will be emitted at low
momenta — which might turn out to be advan-
Hyperon
tageous, because more data are needed to
reduce the statistical errors in measurements
Figure 1 | Investigating the proton–hyperon interaction. a, The ALICE Collaboration1 smashed together of low-momentum systems. Increases in com-
high-energy protons in CERN’s Large Hadron Collider. b, The collisions generate a ‘particle source’ — a puting power should also substantially reduce
volume of space in which components of the colliding protons interact and become confined within new the uncertainties of first-principles calcula-
particles. These new particles are emitted from the source, and include protons that pair up with heavier tions of nuclear forces. Taken together, these
particles known as hyperons. c, The paired-up protons and hyperons interact with each other in a way that developments bode well for future research
alters the relative momentum of the system, which is then measured by a detector. These measurements are
into the final frontier of the standard model
then used to determine the nuclear force between the proton–hyperon pair.
of particle physics.

by a ‘source’ produced by the collision — a
volume of space in which quarks and gluons
that originally came from the protons interact
and become confined within new hadrons. The
source emits various types of hadron, includ-
ing protons and hyperons, some of which form
proton–hyperon pairs. Finally, the proton and
hyperon in each of these pairs interact with
each other in ways that alter the momentum
of the paired system. This momentum is meas-
ured by a detector and used to determine the
momentum correlations.
The momentum correlations reflect the size
of the hadron source and the properties of the
interaction between the produced proton–
hyperon pairs. Such correlation analyses were
originally used to determine the source size in
collisions of heavy ions6, but in the new work,
be calculated from first principles so that the
results can be compared with experimental
findings. The precision with which nucleon–
nucleon interactions can be determined from
experimental data is still superior to that
obtained from these calculations, but the
ALICE Collaboration’s measurements of the
proton–hyperon interactions almost exactly
match those obtained from theory.
A wealth of high-precision measurements of
proton–hyperon interactions is expected from
the LHC in the next decade, following on from
its recent upgrade. Moreover, various other
facilities that will study particle collisions at
lower energies than those produced at the
Virology
Manuel Lorenz is at the Institute for Nuclear
Physics, Goethe University, Frankfurt 60438,
Germany.
e-mail: m.lorenz@gsi.de
1. ALICE Collaboration. Nature 588, 232–238 (2020).
2. Epelbaum, E., Hammer, H.-W. & Meissner, U.-G. Rev. Mod.
Phys. 81, 1773–1825 (2009).
3. Stoks, V. & de Swart, J. Phys. Rev. C 47, 761–767 (1993).
4. Weissenborn, S., Chatterjee, D. & Schaffner-Bielich, J.
Phys. Rev. C 85, 065802 (2012).
5. Tanabashi, M. et al. Phys. Rev. D 98, 030001 (2018).
6. Lisa, M. A., Pratt, S., Soltz, R. & Wiedemann, U. Annu. Rev.
Nucl. Part. Sci. 55, 357–402 (2005).
7. Adamczewski-Musch, J. et al. Phys. Rev. C 94, 025201 (2016).
8. Acharya, S. et al. Phys. Rev. C 99, 024001 (2019).
9. Sasaki, K. et al. Nucl. Phys. A 998, 121737 (2020).
10. Iritani, T. et al. Phys. Lett. B 792, 284–289 (2019).

Cracking the cell access

they are instead used to investigate the inter-
action between the particles of interest. This
approach to studying particle interactions was
pioneered by the HADES Collaboration7 at the
GSI Helmholtz Centre for Heavy Ion Research
in Darmstadt, Germany, and was further devel-
oped by the ALICE collaboration8 at the LHC.
The current work depends on the fact that the
extremely high-energy proton–proton colli-
sions carried out at the LHC produce a high
abundance of hyperons from small-volume
hadron sources. The authors used this method
to measure the strong force between pro-
tons and Ω– hyperons (which consist of three
strange quarks) and between protons and Ξ
hyperons (which consist of two strange quarks
and one up or down quark).
The ALICE Collaboration’s findings open
up a new ‘laboratory’ for investigating other
nucleon–hyperon interactions, including the
little-explored interactions with hyperons that
contain two or three strange quarks. This will
aid our understanding of metastable states
of hyperon pairs or of the compressibility of
nuclear matter at high densities. The latter is
relevant not only for the stability of neutron
stars, but also for neutron-star mergers and
heavy-ion collisions.
In a lucky coincidence, recent developments
in theoretical physics9,10 allow nuclear forces to
code for a deadly virus
James Zengel & Jan E. Carette
The discovery that the receptor protein LDLRAD3 is essential
for infection of human cells by Venezuelan equine encephalitis
virus could inform strategies to combat this potentially
lethal infection. See p.308
When viruses jump from animals to humans, membrane and initiate its replication cycle. On
disease outbreaks can follow. A striking exam- page 308, Ma et al.3 describe the long-sought
ple is Venezuelan equine encephalitis virus receptor for VEEV, and show that it is essential
(VEEV). This virus causes sporadic disease for viral replication in both human cells and
outbreaks in horses in Latin America that fre- mouse models.
quently spill over into humans, resulting in Interactions between a virus and its host
often-deadly neurological disease1. Because receptor protein can control which tissue
of its pathogenicity in livestock and humans, types in the body support viral growth, thus
VEEV has been studied as a biological weapon influencing the type of disease that results.
by several countries, including the United Furthermore, these interactions can deter-
States2. Treatments for the disease are there- mine how well the virus spreads through a host
fore highly desirable. It has been unknown how population. During the continuing SARS-CoV-2
VEEV co-opts cellular pathways to establish pandemic, for instance, viral strains that had a
infection in people — in particular, which host specific mutation in the virus’s spike protein
receptor protein allows VEEV to cross the cell became predominant soon after the virus

Nature | Vol 588 | 10 December 2020 | 223

©
2
0
2
0
S
p
r
i
n
g
e
r
N
a
t
u
r
e
L
i
m
i
t
e
d
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
News & views
a VEEV
D1
LDLRAD3
Cell membrane
Cytoplasm
Severe disease
Death
Figure 1 | Preventing infection by Venezuelan equine encephalitis virus (VEEV) in mice. a, Ma et al.3
report that LDLRAD3 is the mammalian receptor protein for VEEV. Entry of VEEV into cells is mediated
by binding to LDLRAD3’s domain 1 (D1). VEEV infection in mice causes severe disease and death in all cases.
b, The authors fused D1 to part of an antibody. This construct acts as an antiviral decoy, binding to VEEV and
so preventing it from interacting with LDLRAD3. The decoy treatment protected mice from VEEV infection —
the animals showed few signs of disease and had a much higher survival rate than did untreated animals.
jumped to humans — this mutation enhances
binding between the spike protein and its
receptor on human cells, ACE2 (ref. 4).
Despite the importance of host receptors
for understanding infection, their identities
for VEEV and other alphaviruses (a category
of mainly mosquito-borne RNA viruses) have
mostly been elusive . Alphaviruses that infect
humans can cause either severe arthritis or
— as VEEV does — inflammation of the brain
(encephalitis). In 2018, previous work5 from
some of the authors of the current study
uncovered Mxra8 as a mammalian receptor
protein for multiple arthritogenic alpha-
viruses, but not for encephalitis-causing
alphaviruses.
Ma et al. therefore went in search of the
mammalian receptor for VEEV. The authors
made use of a gene-editing tool called
CRISPR–Cas9 to introduce mutations into
more than 20,000 genes in mouse neuronal
cells. They then screened the cells to deter-
mine which mutations prevented infection
by a modified form of VEEV (the version used
was less pathogenic than normal, to enable
safe experimentation in the laboratory). The
screen revealed that Ldlrad3 was the gene most
commonly mutated in infection-resistant cells.
Subsequent experiments in a broad range of
human and mouse cell types demonstrated
that the LDLRAD3 protein is essential for VEEV
entry into host cells.
LDLRAD3 is a poorly characterized mem-
ber of a large group of membrane-bound
receptors called the LDL scavenger receptor
family. This family is mainly known for its role
in bringing lipoprotein particles into the cell in
vesicles (a process called endocytosis). Other
members of the family have been shown to be
co-opted by viruses unrelated to alphaviruses
to gain entry into the cell6.
Ma and colleagues identified a specific
region called domain 1 (D1) in the extracellular
portion of LDLRAD3 through which VEEV-like
particles directly bind to the receptor (Fig. 1a).
b infection in the blood after a mosquito bite?
D1
Antiviral
decoy
Few signs of disease
High survival rate
But, intriguingly, deletion of the intracellular
domain of LDLRAD3 — which typically medi-
ates endocytosis in this receptor family — did
not prevent VEEV entry. This could mean that
binding of VEEV to LDLRAD3 triggers fusion
of the viral and cell membranes, resulting
in direct release of viral RNA into the cell.
Alternatively, LDLRAD3 might mainly medi-
ate virus binding, with another, unknown
factor controlling endocytosis. Future stud-
ies are needed to distinguish between these
possibilities.
Finally, the authors investigated whether
modulating LDLRAD3 could protect mice from
VEEV infection. Strikingly, deletion of Ldlrad3
completely protected the animals from other­
wise-lethal infection with highly pathogenic
VEEV strains (Fig. 1b). The authors gave wild-
type mice a soluble form of LDLRAD3 in which
D1 of the receptor was fused to part of an anti-
body. The construct binds to VEEV, preventing
interactions with LDLRAD3 on cells. Admin-
istration of the soluble construct before or
after infection with VEEV led to near-complete
protection in wild-type mice.
A key question is whether LDLRAD3 mainly
mediates infection in the brain, where it causes
encephalitis, or whether it also takes part in
VEEV infection in the different cell types
involved in spreading the virus through the
body after an initial mosquito bite. The broad
expression pattern of LDLRAD3 in many tis-
sues suggests that the protein has roles in
viral spread throughout the body. Indeed, the
authors show that soluble LDLRAD3 almost
totally blocked virus replication in several tis-
sues that are involved in such spread, including
the blood serum, spleen and brain.
In the future, deletion of Ldlrad3 in specific
mouse tissues could help to reveal more about
how VEEV spreads and causes disease. By abol-
ishing infection in chosen tissues in this way,
one could rigorously test various unknown
aspects of disease progression. Are blood
cells (specifically, a type called myeloid cells)
key in initiating viral spread and fuelling viral
Does VEEV reach the central nervous system
through the peripheral nervous system, or
more directly by crossing the blood–brain
barrier7? And do these steps require LDLRAD3?
LDLRAD3 mediates cell entry of VEEV, but
Ma and colleagues found that it does not con-
trol entry of related encephalitic alphaviruses
such as Western and Eastern equine enceph-
alitis viruses. This is somewhat unexpected,
given the strong similarities in structure8 and
pathogenesis between the three. An intriguing
possibility is that other members of the LDL
scavenger receptor family act as receptors
for distinct encephalitic alphaviruses. Further
structural studies defining the LDLRAD3–
VEEV interface will provide clues to why this
interaction is seemingly so specific.
What are the therapeutic implications of Ma
and colleagues’ work? Precise characterization
of the LDLRAD3-binding site on VEEV could
aid the development of highly neutralizing
antibodies that block the VEEV–LDLRAD3
inter­action. A similar antibody therapy that
prevents interactions between the Ebola virus
and its human receptor protein NPC1 has
shown success, reducing the mortality from
Ebola9,10. Another strategy is to use soluble
LDLRAD3 as an antiviral decoy. The authors
have provided strong proof of principle in
mice that this might work, although further
optimization to enhance the binding affin-
ity and half-life of soluble LDLRAD3 in vivo
might be required, equivalent to developing
engineered ACE2 that has a greatly enhanced
potency in blocking SARS-CoV-2 infection11.
The discovery of LDLRAD3 has therefore
revealed a range of ways in which we might,
in the future, combat severe VEEV disease.
James Zengel and Jan E. Carette are in the
Department of Microbiology and Immunology,
Stanford University School of Medicine,
Stanford, California 94305, USA.
e-mail: carette@stanford.edu
1. Weaver, S. C., Ferro, C., Barrera, R., Boshell, J. &
Navarro, J.-C. Annu. Rev. Entomol. 49, 141–174 (2004).
2. Bronze, M. S., Huycke, M. M., Machado, L. J.,
Voskuhl, G. W. & Greenfield, R. A. Am. J. Med. Sci. 323,
316–325 (2002).
3. Ma, H. et al. Nature 588, 308–314 (2020).
4. Yurkovetskiy, L. et al. Cell 183, 739–751 (2020).
5. Zhang, R. et al. Nature 557, 570–574 (2018).
6. Hofer, F. et al. Proc. Natl Acad. Sci. USA 91, 1839–1842
(1994).
7. Charles, P. C., Walters, E., Margolis, F. & Johnston, R. E.
Virology 208, 662–671 (1995).
8. Hasan, S. S. et al. Cell Rep. 25, 3136–3147 (2018).
9. Mulangu, S. et al. N. Engl. J. Med. 381, 2293–2303 (2019).
10. Misasi, J. et al. Science 351, 1343–1346 (2016).
11. Chan, K. K. et al. Science 369, 1261–1265 (2020).
This article was published online on 18 November 2020.

224 | Nature | Vol 588 | 10 December 2020

©
2
0
2
0
S
p
r
i
n
g
e
r
N
a
t
u
r
e
L
i
m
i
t
e
d
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
 widely variable carbon prices across regions
to ensure fair effort-sharing. This avoids trans-
Climate science
fers, but is less economically efficient globally.

Trade-offs for equitable

The authors then explored a series of hybrid
scenarios in between these two extremes,
changing the degree of variation in regional

climate policy assessed carbon prices and calculating the international

Wei Peng
Computational models show that regionally varied prices
for carbon emissions can greatly reduce the need for poor
countries to receive financial assistance to tackle climate
change, while still stabilizing global warming. See p.261
International agreements for tackling climate
change face many challenges. One of the
knottiest is how to allocate mitigation efforts
fairly across nations, without increasing the
overall global cost, and without asking poor
countries to accept large amounts of financial
assistance that raise concerns about infringe-
ments of national sovereignty. On page 261,
Bauer et al.1 report an analysis of the trade-
off between cost and sovereignty for various
international climate policies. They conclude
that sovereignty concerns can be allayed sub-
stantially with only slightly higher global costs
by using a strategy in which the carbon price
— the charge per tonne of carbon dioxide emis-
sions — is varied modestly to account for each
country’s ability to pay.
After decades of gridlock in climate diplo-
macy, the 2015 Paris agreement received
support from nearly 200 countries for the
collective goal of limiting global warming by
the end of this century to well below 2 °C above
pre-industrial levels. The key to its success
was that it provided much-needed flexibility
for countries to make their own nationally
determined contributions to reducing green-
house-gas emissions, instead of setting tar-
gets through a centralized treaty, as was the
case for the 1997 Kyoto Protocol (go.nature.
com/3oa6uvl).
By allowing countries to tailor commit-
ments to what can realistically be deliv-
ered, the Paris approach reflects the United
Nations’ equity principle of ‘common but
differ­entiated responsibilities and respec-
tive capabilities’. It also reduces the need for
financial transfer — the provision of money
GCF/ANGELI MENDOZA
countries pay a uniform carbon price, because
emissions might not be reduced in the places
in which it is cheapest to do this. The benefits
of the Paris approach for feasibility, equity
and sovereignty therefore come at the cost
of economic inefficiencies.
Bauer et al. quantify the trade-offs between
the global cost of climate mitigation (low cost
equates to high economic efficiency) and the
amount of international financial transfer
needed (a proxy for concerns about sover-
eignty), to find a balance that would enable
the 2 °C goal to be achieved with equitable
effort-sharing — ensuring the same ratio of
mitigation cost to income for all countries.
The authors started by analysing two extreme
policies. The first involves setting a globally
uniform carbon price. This is the cheapest
mitigation strategy overall, but requires
large international transfer to ensure fair
effort-sharing. The second policy is to have
transfer required to achieve fair effort-sharing
in each case. By combining and comparing
these scenarios, the authors plot a curve that
depicts the trade-off between global cost and
financial transfers (see Fig. 3a of the paper1).
Bauer et al. show that small deviations from
a globally uniform carbon price can achieve
the 2 °C goal with slightly higher mitigation
costs, but much lower transfers. For instance,
a modest regional variation in carbon prices
with a standard deviation of US$14 per tonne
in 2030 leads to a negligible increase in
global mitigation costs, but an 18% reduction
in required transfers. This implies that it is
highly possible to deliver reasonably good
outcomes for equity, economic efficiency
and sovereignty, if policymakers are willing to
deviate modestly from the economically most
efficient strategy (a uniform carbon price).
The authors’ findings highlight an advantage
of the Paris approach: an equitable outcome
can still be achieved when international trans-
fers are reduced. At the 2009 United Nations
Climate Change Conference, wealthy countries
pledged $100 billion a year by 2020 to help
developing countries tackle climate change.
But the total amount collected from public and
private sources in 2018 was less than $80 billion
(go.nature.com/39gnts7). US President Donald
Trump’s decision to withdraw $2 billion that
had been promised to the Green Climate Fund
— the largest international fund for financing

for poor countries to help them tackle cli-

mate change and to compensate for the neg-
ative economic consequences of mitigating
global warming. Reducing this need is polit-
ically beneficial for poor countries because
they sometimes perceive a heavy reliance Figure 1 | Panel inspection at the Sumber Solar Plant in Mongolia. The construction of this solar plant
on international financial transfer as under- was supported by the Green Climate Fund, the largest international fund for financing efforts to tackle
mining their national sovereignty. However, global warming. However, a heavy reliance on such support is sometimes seen by poor countries as an
the Paris approach could be more expensive infringement of their national sovereignty. Bauer et al.1 report an analysis of climate policies that suggests a
overall than the idealized strategy in which all way to reduce the need for financial support.

Nature | Vol 588 | 10 December 2020 | 225

©
2
0
2
0
S
p
r
i
n
g
e
r
N
a
t
u
r
e
L
i
m
i
t
e
d
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
News & views
efforts to mitigate and adapt to climate change
(Fig. 1) — contributed to the shortfall. These
realities demonstrate the uncertainties in
mobilizing large-scale international transfer.
By reducing the need for transfer, the Paris
approach creates a more collaborative space
for engaging both developed and developing
countries in climate diplomacy.
Bauer and colleagues also point out plau-
sible unintended consequences of the Paris
approach for environmental sustainability. If
the variation in carbon price across countries
is large and the 2 °C target remains the global
objective, then developing countries might
make only limited efforts to tackle climate
change. This could push developed coun-
tries to adopt costly options, such as using
a technology called bioenergy with carbon
capture and storage (BECCS). The authors
find that, in this scenario, countries that do
not belong to the Organisation for Economic
Co-operation and Development (OECD) will
export bioenergy to OECD countries for use
in BECCS — exacerbating deforestation and
land-use intensification in the global south.
The negative effects of BECCS on sustainabil-
ity have received wide attention2, but Bauer
and co-workers’ study emphasizes that these
effects could worsen under the Paris approach,
if mitigation efforts shift across countries.
The authors’ assessment is based on the
REMIND-MAgPIE computational model, which
is one of the ‘integrated assessment models’
(IAMs) used by the Intergovernmental Panel
on Climate Change to explore how different
policy and technology pathways might affect
global emissions and future climate3,4. Such
models include representations of economic,
energy and land systems, as well as the inter-
actions between them. IAMs have been impor-
tant analytical tools for designing climate
policies5, but Bauer and colleagues’ study is
exemplary in that an IAM is used to evaluate
competing considerations faced by climate
policymakers.
The new findings are a useful contribution
to the high-level debate about the modes
and objectives of climate policy, but crucial
aspects of the modelling framework lack
the granularity needed to inform real-world
decision-making. For instance, the model
considers just 12 world regions, whereas
climate decisions are often made at national
and subnational levels. Bauer et al. also use
carbon prices as a proxy for climate policy,
whereas policymakers need to choose from a
range of low-carbon policies that vary in cost
and feasibility.
Furthermore, countries decide their own
carbon prices in the real world, whereas
regional carbon prices are not allowed to
vary independently in Bauer and colleagues’
model. Instead, the authors use a rescaling
method to adjust the whole set of regional
carbon prices, to make computation easier.
Finally, inter­national transfer is modelled
as the total financial flow from rich to poor
countries, without taking into account specific
financing mechanisms. However, a wide range
of financing mechanisms exist, such as grants,
subsidized loans and private finance, which
involve different actors and terms. These limi-
tations do not overshadow the value of the new
work, but future efforts should engage model
developers and users to bring IAMs closer to
real-world problems.
Wei Peng is in the School of International
Affairs and the Department of Civil and
Environmental Engineering, Pennsylvania
State University, University Park, Pennsylvania
16802, USA.
e-mail: weipeng@psu.edu
1. Bauer, N. et al. Nature 588, 261–266 (2020).
2. Heck, V., Gerten, D., Lucht, W. & Popp, A. Nature Clim.
Change 8, 151–155 (2018).
3. Clarke, L. et al. in Climate Change 2014: Mitigation of
Climate Change. Contribution of Working Group III to
the Fifth Assessment Report of the Intergovernmental
Panel on Climate Change (eds Edenhofer, O. et al.) Ch. 6,
413–510 (Cambridge Univ. Press, 2014).
4. Hoegh-Guldberg, O. et al. in Global Warming of 1.5°C
(eds Masson-Delmotte, V. et al.) Ch. 3, 175–311 (IPCC, 2018).
5. van Beek, L., Hajer, M., Pelzer, P., van Vuuren, D. &
Cassen, C. Glob. Environ. Change 65, 102191 (2020).

nature
communications

Publishing high-quality Submit your research and benefit from a fast

decision process, full open access, CC BY
open access research from licensing as standard, the #3 most-cited*
multidisciplinary open access journal in the
across all areas of the world and 3 million monthly page views on
journal site.
natural sciences. *Clarivate Analytics, 2019

nature.com/ncomms @NatureComms @NatureCommunications A81661

©
2
0
2
0
S
p
r
i
n
g
e
r
N
a
t
u
r
e
L
i
m
i
t
e
d
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
Article

Detection of large-scale X-ray bubbles in the

Milky Way halo

https://doi.org/10.1038/s41586-020-2979-0 P. Predehl1 ✉, R. A. Sunyaev2,3 ✉, W. Becker1,4, H. Brunner1, R. Burenin2, A. Bykov5,

A. Cherepashchuk6, N. Chugai7, E. Churazov2,3 ✉, V. Doroshenko8, N. Eismont2, M. Freyberg1,
Received: 10 July 2020
M. Gilfanov2,3 ✉, F. Haberl1, I. Khabibullin2,3, R. Krivonos2, C. Maitra1, P. Medvedev2, A. Merloni1 ✉,
Accepted: 25 September 2020 K. Nandra1 ✉, V. Nazarov2, M. Pavlinsky2,11, G. Ponti1,9, J. S. Sanders1, M. Sasaki10, S. Sazonov2,
A. W. Strong1 & J. Wilms10
Published online: 9 December 2020

Check for updates

The halo of the Milky Way provides a laboratory to study the properties of the shocked
hot gas that is predicted by models of galaxy formation. There is observational
evidence of energy injection into the halo from past activity in the nucleus of the Milky
Way1–4; however, the origin of this energy (star formation or supermassive-black-hole
activity) is uncertain, and the causal connection between nuclear structures and
large-scale features has not been established unequivocally. Here we report
soft-X-ray-emitting bubbles that extend approximately 14 kiloparsecs above and
below the Galactic centre and include a structure in the southern sky analogous to the
North Polar Spur. The sharp boundaries of these bubbles trace collisionless and
non-radiative shocks, and corroborate the idea that the bubbles are not a remnant of a
local supernova5 but part of a vast Galaxy-scale structure closely related to features
seen in γ-rays6. Large energy injections from the Galactic centre7 are the most likely
cause of both the γ-ray and X-ray bubbles. The latter have an estimated energy of
around 1056 erg, which is sufficient to perturb the structure, energy content and
chemical enrichment of the circumgalactic medium of the Milky Way.

eROSITA8 is a large-collecting-area and wide-field-of-view X-ray tel- Although less evident at first glance, close inspection of the
escope, launched into space onboard the Spektr-RG mission on 13 medium-energy-band (0.6–1.0 keV) image in the hemisphere below the
July 2019. Over the course of six months (December 2019–June 2020), plane of the Milky Way (‘south’) reveals an astonishing new feature—a
Spektr-RG and eROSITA have completed a survey of the whole sky at huge circular annulus of similar shape and scale to the structure seen
energies of 0.2–8 keV—much deeper than the only other all-sky survey in the north (Fig. 2). Together, they seem to form a pair of ‘bubbles’ that
with an X-ray-imaging telescope, which was performed by ROSAT in emerge from the Galactic centre. They are traceable at various levels
1990 at energies of 0.1–2.4 keV. of intensity throughout most of the sky, and should represent a very
The sky map from the first eROSITA all-sky survey is shown in Fig. 1. large object (several kiloparsecs), akin to the Fermi bubbles, because
This image has been created from calibrated events in the energy range local features are unlikely to exhibit the fourfold symmetry around the
0.3–2.3 keV (Methods). A preliminary analysis indicates that more than direction towards the centre of the Galaxy.
one million X-ray point sources and about 20,000 extended ones are The Fermi bubbles were discovered in 20101 with the Fermi-LAT
detected in the survey. This is comparable to, and may exceed, the total (Fermi large-area telescope) γ-ray instrument. They have a hard,
number of X-ray sources known before eROSITA launched. Multiwave- non-thermal spectrum, which shows up clearly in maps at energies of
length identifications using the WISE and Gaia catalogues9,10 suggest more than 1 GeV. Their emission is probably due to inverse Compton
that about 80% of the point sources are distant active galactic nuclei scattering of cosmic-ray electrons on the cosmic microwave back-
(AGN; comprising about 80% of all known blazars) and that around ground and other radiation fields. This kiloparsec-scale structure was
20% are coronally active stars in the Milky Way, including about 150 quickly interpreted as a possible manifestation of past activity of the
planet-hosting stars (roughly 10% of all known outside of the Kepler now dormant supermassive black hole in the centre of the Milky Way,
field). thus linking it with AGN observed outside the Galaxy12–15. Alternatively,
Various very large and diffuse extended structures are visible in a burst of star formation could power the bubbles16–18. In either case,
the all-sky survey map. The most obvious is a quasi-circular feature, the energy needed to power their formation must have been very large,
which is part of the North Polar Spur and Loop I (northwest quadrant) at roughly 1055 erg1,19.
discovered in the early days of X-ray and radio astronomy, respec- X-ray emission from the North Polar Spur had already been found
tively5,11. by ROSAT5. Although considered in most early models to be a nearby
1
Max-Planck-Institut für Extraterrestrische Physik, Garching, Germany. 2Space Research Institute of the Russian Academy of Sciences, Moscow, Russia. 3Max-Planck-Institut für Astrophysik,
Garching, Germany. 4Max-Planck-Institut für Radioastronomie, Bonn, Germany. 5Ioffe Institute, St Petersburg, Russia. 6M. V. Lomonosov Moscow State University, P. K. Sternberg Astronomical
Institute, Moscow, Russia. 7Institute of Astronomy, Russian Academy of Sciences, Moscow, Russia. 8Institut für Astronomie und Astrophysik, Tübingen, Germany. 9INAF-Osservatorio
Astronomico di Brera, Merate, Italy. 10Dr. Karl-Remeis-Sternwarte Bamberg and Erlangen Centre for Astroparticle Physics, Universität Erlangen-Nürnberg, Bamberg, Germany.
11
Deceased: M. Pavlinsky. ✉e-mail: predehl@mpe.mpg.de; sunyaev@iki.rssi.ru; churazov@iki.rssi.ru; gilfanov@iki.rssi.ru; am@mpe.mpg.de; knandra@mpe.mpg.de

Nature | Vol 588 | 10 December 2020 | 227

Article
Fig. 1 | The Spektr-RG–eROSITA all-sky map. An RGB map of the first
Spektr-RG–eROSITA all-sky survey (red for 0.3–0.6 keV, green for 0.6–1.0 keV,
blue for 1.0–2.3 keV) is shown in Galactic coordinates, using a Hammer–Aitoff
projection. The original image, with a resolution of about 12″, was smoothed
supernova remnant nearly surrounding us, the possibility that the
North Polar Spur is of Galactic scale has been proposed6,20, and is sup-
ported by several observational arguments7,21. In particular, study of
absorption in X-ray and radio bands places a lower limit of 300 pc on
the distance to the structure21, which rules out a nearby supernova
remnant. In addition, evidence for a large-scale bipolar wind has been
presented, based purely on X-ray and mid-infrared data, even before
the discovery of the Fermi bubbles22.
With the eROSITA data, the full scope and morphology of these gigan-
tic X-ray structures has become evident. ROSAT, owing to a combination
of its lower sensitivity and softer energy response, could reveal only
the brightest part of the southern loop closest to the Galactic plane1,22,
not the whole structure. More recently, the 0.7–1-keV all-sky map from
the solid-state slit camera (SSC) of MAXI also provided evidence of a
southern enhancement on these large scales, and a close north–south
symmetry23.
The Fermi bubbles and large-scale X-ray emission revealed by eROS-
ITA show remarkable morphological similarity. We therefore suggest
that the Fermi bubbles and the eROSITA structure are physically related,
and refer to the latter as ‘eROSITA bubbles’. Our discovery confirms the
previously suggested common origin of the two objects6,7. The motiva-
tion for a separate name is that, despite the probably common origin,
the two structures differ in some important respects.
First, we compare their morphologies on the sky (Fig. 3). The Fermi
bubbles are roughly elliptical, about 55° × 45° (north–south, east–west)
in diameter, symmetric about the Galactic centre, with vertical axis per-
pendicular to the Galactic plane, and roughly uniform in γ-ray intensity.
The eROSITA bubbles appear as extended as 80° in longitude, roughly
80°–85° in latitude and concentrated in annuli or shells. This suggests
that they are, to first-order, close to spherical, with a radius of about
6–7 kpc along the plane, extending radially on the Milky Way close to
the Sun, so that their northern and southern edges are imprinted by
the closer rim of the bubble. The full vertical extent of the eROSITA
bubbles is more difficult to determine; assuming a spherical geometry,
228 | Nature | Vol 588 | 10 December 2020
(with a Gaussian with a full-width at half-maximum (FWHM) of 10′) to generate
this one. Image adapted from ref. 34. Credit: Jeremy Sanders, Hermann Brunner,
Andrea Merloni and the eSASS team (MPE); Eugene Churazov, Marat Gilfanov
(on behalf of IKI).
they would extend roughly 14 kpc above and below the Galactic plane
(Extended Data Fig. 1).
Second, from a preliminary spectral analysis of eROSITA data, the
absorbing column density of the diffuse emission in the southwestern
bright rim of the eROSITA bubbles (white rectangle in Fig. 2) can be
constrained to NH = (1.0–3.5) × 1021 cm−2, consistent with what has been
measured previously21 for the northern structure. One-dimensional
cross-sections of the observed surface brightness at various latitudes
(Fig. 2) are qualitatively consistent with the projection of (quasi-)spher-
ical thick shells with an outer diameter of 14 kpc. Regardless of the
uncertainties on these numbers, it is clear that the eROSITA bubbles
are comparable in size to the Galactic disk24.
We note that the extended X-ray emission revealed by eROSITA coin-
cides spatially with the soft component of the GeV emission reported
to surround the Fermi bubbles2,7,25. A possible connection with polar-
ized radio-continuum emission at 2.3 GHz and 23 GHz26 has yet to be
explored.
An episodic or continuous energy release in the region of the Galactic
centre is expected to generate a series of distinct structures: shocks and
contact discontinuities. We see two prominent structures in our maps:
one is the outer boundary of the eROSITA bubbles; the other separates
the eROSITA bubbles and the Fermi bubbles. The sharp boundary of
the eROSITA bubbles—which appears bright in X-rays, indicative of
hotter gas at the boundary than outside it—clearly traces the presence
of a non-radiative (or adiabatic) shock (see Methods for an estimate of
the gas cooling time). We associate the boundary with a forward shock
linked to the onset of large energy release at the Galactic centre. The
nature of the boundary between the eROSITA and Fermi bubbles is less
clear. It could be another forward shock (in the case of a sequence of
energy releases), a reverse shock, a wind-termination shock or a contact
discontinuity. The reverse or termination shock models for the Fermi
bubbles would imply an additional contact discontinuity somewhere
between the Fermi and eROSITA bubbles, which is not apparent in the
data. Instead, we consider the simplest scenario in which the eROSITA
 85° 85
a 80° 80°
70° 70°
60° 60°
50° 50°
40° 40°

30° 30°

20° 20°

10° 10°

23 h

22 h

21 h

19 h

17 h
20 h

18 h

16 h

15 h
0h
3h

2h

1h

14 h
5h
10 h

7h

6h

4h
9h

8h

13 h
11 h

12 h
12 h

–20° –20°

–30° –30°

–40° –40°

–50° –50°

–60° –60°
–70° –70°
–80° –80°
–90°

b
Surface brightness (counts s–1 deg–2)

20
+60° +50° +40°

10

20
Surface brightness (counts s–1 deg–2)

–60° –50° –40°

10

100 50 0 −50 −100 100 50 0 −50 −100 100 50 0 −50 −100

Galactic longitude (°) Galactic longitude (°) Galactic longitude (°)

Fig. 2 | The soft-X-ray eROSITA bubbles. a, False-colour map of extended
emission detected by eROSITA in the 0.6–1.0-keV range. The contribution of
the point sources has been removed and the scaling adjusted to enhance
large-scale structures in the Galaxy. b, One-dimensional surface-brightness
profiles in the same energy band (red lines with pink shading representing
statistical uncertainties), cut at various galactic latitudes (as labelled). For
comparison, we also show the predictions of four possible geometric models
(not normalized to the data): a full sphere (yellow), a very thick shell (thickness,
4 kpc; brown), a thick shell (thickness, 2 kpc; cyan) and a thin shell (thickness,
0.2 kpc; green). The thick shell (cyan) is the most consistent with the data (see
Extended Data Fig. 2 for a two-dimensional projection of this model). The
region indicated by the white rectangle is where a preliminary spectral analysis
was performed to constrain the line-of-sight absorption column density
towards the southern eROSITA bubbles.

and Fermi bubbles are causally connected, with the Fermi bubbles driv- outflow, and the boundary of the eROSITA bubbles is the shock that
ing the expansion of the eROSITA bubbles and both structures being propagates through the halo gas. The pressure is thus continuous
associated with the same (gradual or instantaneous) energy release in across the interface between the eROSITA and Fermi bubbles and the
the nuclear region of the Milky Way. In this scenario, the outer bound- total thermal energies of the two features simply reflect their volumes
ary of the Fermi bubbles plausibly represents a contact discontinuity (ignoring the effects of stratification, which may be non-negligible).
that separates the shock-heated interstellar medium from the shocked Given that their characteristic sizes differ by a factor of about 2, the

Nature | Vol 588 | 10 December 2020 | 229

Article
Fig. 3 | Comparison of the morphology of the γ-ray and X-ray bubbles.
A composite Fermi–eROSITA image is shown. The X-ray extended emission
revealed by eROSITA (0.6–1-keV band; cyan) encloses the hard component of
total thermal energy of the eROSITA bubbles is almost 10 times larger
than that of the Fermi bubbles.
The obser ved average X-ray surface brightness of
(2–4) × 10−15 erg cm−2 s−1 arcmin−2 in the eROSITA bubbles (Methods),
which decreases with Galactic latitude, is in broad agreement with the
above scenario. The observed surface brightness, integrated over the
full extent of the eROSITA bubbles, implies a total luminosity of hot
X-ray-emitting plasma of L ≈ 1 × 1039 erg s−1.
To inflate the eROSITA bubbles, an average luminosity of the order of
1041 erg s−1 during the past tens of millions of years would be required,
and could arise from either star-forming or AGN activity in the Galactic
centre. As discussed above, the arguments in favour of each interpreta-
tion in the context of the Fermi bubbles have been debated extensively.
In the case of the eROSITA bubbles, the energetics are such that they are
at the limit of what the past starburst activity at the centre of the Milky
Way could provide. Alternatively, the eROSITA bubbles could be inflated
by a period (about 1–2 Myr) of Seyfert-like activity (L ≈ 1043 erg s−1) of
the central supermassive black hole (Sgr A*). The long cooling time of
the hot plasma is consistent with such a hypothesis.
The structures seen here are reminiscent of similar effects seen in
AGN that host rapidly accreting supermassive black holes1. These can
inject a vast amount of mechanical energy into the ambient gas, as
revealed by radio-bright bubbles embedded in the X-ray cocoons27. This
process, known as AGN feedback, is seen in objects ranging from indi-
vidual early-type galaxies, such as Centaurus A28, to massive clusters,
such as A426 (Perseus)29,30, and is thought to have potentially marked
effects on the evolution of galaxies. On the other hand, explosions of
supernova associated with star formation yield kinetic energy of the
order of 1051 erg per supernova in the ejecta (also known as stellar feed-
back), which may drive an outflow from the central region of a galaxy31.
M82 provides a good example of the latter mechanism32. The energet-
ics and the most salient features of the observed eROSITA bubbles are
such that neither of the two mechanisms could be excluded a priori.
Irrespective of the specific source of energy, our results cor-
roborate the notion that inactive disk galaxies, such as the Milky
230 | Nature | Vol 588 | 10 December 2020
the extended gigaelectronvolt emission traditionally referred to as Fermi
bubbles (red; Fermi map adapted from ref. 35), unequivocally establishing their
close relation.
Way, have hot plasma in their haloes that is highly perturbed by
activity in their disks, demonstrating the presence of a feedback
mechanism in apparently quiescent galaxies. Galaxies are thought
to grow via the slow recondensation of the hot halo plasma, which
was shock-heated during the collapse of the dark-matter halo33.
The cooling time of the hot plasma in the halo is comparable to
the Hubble time, so the process of growing a galaxy is assumed
to be steady (apart from mergers) and slow. Here we have direct
evidence of the re-heating of such plasma, to considerable heights
above the Galactic disk.
The detection of these X-ray bubbles was enabled by the combined
capabilities of the eROSITA instrument and the Spektr-RG mission
profile. More detailed analysis following accurate calibration of the
instrument, substantial increases in data quality from the ongoing sky
survey and follow-up observations in other parts of the electromagnetic
spectrum will reveal further details of the properties of the eROSITA
bubbles and the implications for the structure and evolution of galax-
ies, including the Milky Way.
Online content
Any methods, additional references, Nature Research reporting sum-
maries, source data, extended data, supplementary information,
acknowledgements, peer review information; details of author con-
tributions and competing interests; and statements of data and code
availability are available at https://doi.org/10.1038/s41586-020-2979-0.
1. Su, M., Slatyer, T. R. & Finkbeiner, D. P. Giant gamma-ray bubbles from Fermi-LAT: active
galactic nucleus activity or bipolar Galactic wind? Astrophys. J. 724, 1044–1082 (2010).
2. Ackermann, M. et al. The spectrum and morphology of the Fermi bubbles. Astrophys. J.
793, 64 (2014).
3. Heywood, I. et al. Inflation of 430-parsec bipolar radio bubbles in the Galactic centre by
an energetic event. Nature 573, 235–237 (2019).
4. Ponti, G. et al. An X-ray chimney extending hundreds of parsecs above and below the
Galactic centre. Nature 567, 347–350 (2019).
5. Egger, R. & Aschenbach, B. Interaction of the Loop I supershell with the local hot bubble.
Astron. Astrophys. 294, L25–L28 (1995).
6. Sofue, Y. Bipolar hypershell Galactic center starburst model: further evidence from 22.
ROSAT data and new radio and X-ray simulations. Astrophys. J. 540, 224–235 (2000).
7. Kataoka, J. et al. X-ray and gamma-ray observations of the Fermi bubbles and NPS/Loop I 23.
structures. Galaxies 6, 27 (2018).
8. Merloni, A. et al. eROSITA science book: mapping the structure of the energetic Universe. 24.
Preprint at https://arxiv.org/abs/1209.3114 (2012).
9. Gaia Collaboration. Gaia data release 2. Summary of the contents and survey properties. 25.
Astron. Astrophys. 616, A1 (2018).
10. Eisenhardt, P. R. M. et al. The CatWISE preliminary catalog: motions from WISE and 26.
NEOWISE data. Astrophys. J. Suppl. Ser. 247, 69 (2020).
11. Berkhuijsen, E. M. A survey of the continuum radiation at 820 MHz between declinations 27.
-7° and +85°. A study of the Galactic radiation and the degree of polarization with special
reference to the loops and spurs. Astron. Astrophys. 14, 359–386 (1971). 28.
12. Zubovas, K., King, A. R. & Nayakshin, S. The Milky Way’s Fermi bubbles: echoes of the last
quasar outburst? Mon. Not. R. Astron. Soc. 415, L21–L25 (2011). 29.
13. Guo, F. & Mathews, W. G. The Fermi bubbles. I. Possible evidence for recent AGN jet
activity in the galaxy. Astrophys. J. 756, 181 (2012). 30.
14. Mou, G. et al. Fermi bubbles inflated by winds launched from the hot accretion flow in Sgr
A*. Astrophys. J. 790, 109 (2014). 31.
15. Zhang, R. & Guo, F. Simulating the Fermi bubbles as forward shocks driven by AGN jets.
Astrophys. J. 894, 117 (2020). 32.
16. Crocker, R. M. & Aharonian, F. Fermi bubbles: giant, multibillion-year-old reservoirs of
Galactic center cosmic rays. Phys. Rev. Lett. 106, 101102 (2011). 33.
17. Lacki, B. C. The Fermi bubbles as starburst wind termination shocks. Mon. Not. R. Astron.
Soc. 444, L39–L43 (2014). 34.
18. Crocker, R. M., Bicknell, G. V., Taylor, A. M. & Carretti, E. A unified model of the Fermi
bubbles, microwave haze, and polarized radio lobes: reverse shocks in the Galactic 35.
center’s giant outflows. Astrophys. J. 808, 107 (2015).
19. Miller, M. J. & Bregman, J. N. The Interaction of the Fermi Bubbles with the Milky Way’s Hot
Gas Halo. Astrophys. J. 829, 9 (2016).
20. Sofue, Y. Propagation of magnetohydrodynamic waves from the Galactic center. Origin Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in
of the 3-kpc arm and the North Polar Spur. Astron. Astrophys. 60, 327–336 (1977). published maps and institutional affiliations.
21. Lallement, R. et al. On the distance to the North Polar Spur and the local CO-H2 factor.
Astron. Astrophys. 595, A131 (2016). © The Author(s), under exclusive licence to Springer Nature Limited 2020
Bland-Hawthorn, J. & Cohen, M. The large-scale bipolar wind in the Galactic center.
Astrophys. J. 582, 246–256 (2003).
Nakahira, S. et al. MAXI/SSC all-sky maps from 0.7 keV to 4 keV. Publ. Astron. Soc. Japan
72, 17 (2020).
Bland-Hawthorn, J. & Gerhard, O. The galaxy in context: structural, kinematic, and
integrated properties. Annu. Rev. Astron. Astrophys. 54, 529–596 (2016).
Casandjian, J.-M. The Fermi-LAT model of interstellar emission for standard point source
analysis. Preprint at https://arxiv.org/abs/1502.07210 (2015).
Carretti, E. et al. Giant magnetized outflows from the centre of the Milky Way. Nature 493,
66–69 (2013).
Böhringer, H. et al. A ROSAT HRI study of the interaction of the X-ray emitting gas and
radio lobes of NGC 1275. Mon. Not. R. Astron. Soc. 264, L25–L28 (1993).
Kraft, R. et al. X-ray emission from the hot interstellar medium and southwest radio lobe
of the nearby radio galaxy Centaurus A. Astrophys. J. 592, 129–146 (2003).
Churazov, E. et al. Asymmetric, arc minute scale structures around NGC 1275. Astron.
Astrophys. 356, 788–794 (2000).
Fabian, A. C. et al. Chandra imaging of the complex X-ray core of the Perseus cluster.
Mon. Not. R. Astron. Soc. 318, L65–L68 (2000).
Strickland, D. K. & Stevens, I. R. Starburst-driven galactic winds – I. Energetics and
intrinsic X-ray emission. Mon. Not. R. Astron. Soc. 314, 511–545 (2000).
Rieke, G. H. et al. The nature of the nuclear sources in M82 and NGC 253. Astrophys. J.
238, 24–40 (1980).
Tumlinson, J., Peeples, M. S. & Werk, J. K. The circumgalactic medium. Annu. Rev. Astron.
Astrophys. 55, 389–432 (2017).
Sanders, J. et al. Annotated version of the eROSITA first all-sky image. http://www.mpe.
mpg.de/7461950/erass1-presskit (2020).
Selig, M., Vacca, V., Oppermann, N. & Enßlin, T. A. The denoised, deconvolved, and
decomposed Fermi γ-ray sky. An application of the D3PO algorithm. Astron. Astrophys.
581, 126 (2015).
Nature | Vol 588 | 10 December 2020 | 231
Article
Methods
eROSITA aboard Spektr-RG
On 13 July 2019, the X-ray observatory Spektrum-Roentgen-Gamma
(Spektr-RG or SRG; a joint Russian–German mission; R. Sunyaev et al.,
manuscript in preparation) was launched into space from the Rus-
sian cosmodrome Baikonur in Kazakhstan. The Navigator spacecraft
platform for SRG (NPO Lavochkin) and launcher (Proton M + Block
DM03 upper stage) were both provided by the Russian space agency
Roscosmos. The observatory carries two X-ray telescopes: ART-XC
(astronomical roentgen telescope X-ray concentrator; M. Pavlinsky
et al., manuscript in preparation), a Russian hard-X-ray instrument,
and the German-led eROSITA (extended roentgen survey imaging tel-
escope array)36. eROSITA was designed to address fundamental ques-
tions of astrophysical cosmology such as the nature of dark energy
and dark matter, the origin and growth of supermassive black holes,
and the expansion of discovery space for rare objects including those
of unknown nature8.
eROSITA consists of seven identical and coaligned X-ray telescopes
housed in a common optical bench that connects seven mirror mod-
ules with their associated cameras. The dimensions of the telescope
structure are approximately 1.9 m diameter and 3.2 m height; the total
mass is 810 kg.
The mirror modules comprise 54 paraboloid or hyperboloid mirror
shells in a Wolter-I geometry, with an outer diameter of 360 mm and
a common focal length of 1,600 mm. The on-axis spatial resolution of
all mirror modules is around 16 arcsec (half energy width) at 1.5 keV.
X-ray baffles in front of the mirrors effectively suppress stray X-ray
light from sources outside the field of view, while magnetic electron
deflectors behind the mirrors help to reduce background that arises
from low-energy cosmic-ray electrons.
Each mirror module has a p–n charge-coupled device (pnCCD) cam-
era at its focus, with 384 × 384 pixels on a single-chip image area of
28.8 mm × 28.8 mm, which provides a circular field of view of 1.03°
diameter. The nominal integration time for all eROSITA cameras is
50 ms. For calibration purposes, each camera has its own filter wheel
with a radioactive 55Fe source and an aluminium or titanium target that
provides three spectral lines, at 5.9 keV (Mn Kα), 4.5 keV (Ti Kα) and
1.5 keV (Al Kα). During science operations, the CCDs are passively cooled
down to −85 °C via heat pipes and radiators. Noteworthy features of
the detectors include the low and very stable in-orbit background and
excellent spectral response at low energies.
Two star-trackers are mounted on eROSITA to ensure accurate bore-
sight and attitude reconstruction. Thanks to these and the excellent
stability of the spacecraft, point-source positional accuracy of about
4″ (1σ) is achieved. The effective area at 1 keV of all seven telescopes
together is similar to that of all three XMM-Newton EPIC cameras; the
grasp, defined as the product of field of view and average effective area,
is about four times larger. Together with the ability of the spacecraft to
continuously scan large areas, this makes eROSITA an extremely fast
survey machine, capable of imaging uniformly large swaths of sky.
On its three-month cruise to its halo orbit around Lagrangian point
L2 of the Sun–Earth system, all SRG systems were put into operation
and checked. The instruments were calibrated and their performance
verified with a series of scientific observations.
SRG–eROSITA observations
The first SRG all-sky survey was conducted between 13 December 2019
and 11 June 2020. Eight all-sky surveys are planned in total, each deliver-
ing an average exposure with eROSITA of about 200 s/cos(lat), where
lat is the ecliptic latitude; the 1-square-degree area around the ecliptic
poles is revisited every four hours, accumulating an exposure of about
30 ks per survey.
During the all-sky survey, the spacecraft rotated continuously with
a scan rate of 90° h−1, giving a four-hour period per revolution. The
rotation axis is oriented to the neighbourhood of the Sun, with an aver-
age progression in ecliptic longitude of about 1° day−1; it thus completes
one all-sky survey in half a year. The scan speed guarantees that the
angular resolution is not degraded by smearing of the photons during
the 50-ms CCD read-out cycle, and provides sufficient overlap between
individual scans to enable source variability analysis and homogeneous
survey exposure. Since the start of the survey, the eROSITA cameras
have been operating with high efficiency, with more than 96% of the
observing time resulting in scientifically useful data. The only sub-
stantial loss of efficiency is due to single event upsets in parts of the
firmware that control the seven cameras, most probably caused by
heavy particles (cosmic rays). Mitigation strategies and procedures
have since been put in place by the operations teams. This, combined
with the very stable particle background along the large halo L2 orbit
of SRG and the flexible mission planning, has resulted in a gap-free
coverage for the first all-sky survey.
As SRG scans the sky, for each X-ray photon detected the eROSITA
instrument records its position on the detector, energy and time tag.
These data, along with detector housekeeping and star-tracker infor-
mation, are telemetered to ground stations in Russia daily. They are
immediately sent from these ground stations, via NPOL Lavochkin and
IKI (the Space Research Institute of the Russian Academy of Science) to
the Max Planck Institute for Extraterrestrial Physics (MPE), where they
are checked and converted to FITS (flexible image transport system)
format. These files are then cut into 4-h intervals, the duration of a
great circle scan. A pipeline based on the eROSITA standard analysis
software system (eSASS) determines good time intervals, dead times
and corrupted events and frames, masks bad pixels, and applies pattern
recognition and energy calibration. Finally, star-tracker and gyro data
are used to assign celestial coordinates to each reconstructed X-ray
photon. These (time-ordered) photons are then projected into sky
tiles of size 3.6° × 3.6°, where images and exposure maps are created
for downstream analysis.
Individual sources are detected in the sky tiles using a three-step
procedure. First, a local sliding-window detection algorithm identifies
enhancements, which are excised from the images. Second, adaptively
filtered background maps of the source-free regions are created. Third,
the sliding-window detection is repeated, and the identified excesses
with respect to the background maps provide an input list of source
candidates from which a maximum-likelihood point-spread-function
fitting algorithm determines the best-fitting source parameters and
detection likelihoods. Data from all seven telescope modules are
merged in this approach.
Generation of all-sky maps
To produce the sky map shown in Fig. 1, we projected the detected X-ray
photons onto the sky, applying smoothing using a two-dimensional
Gaussian with FWHM = 10 arcmin, significantly larger than either the
native resolution of the map (pixel size of 4″) or the point-spread func-
tion of the instrument (FWHM ≈ 12″).
Three sky maps were made in three energy bands (0.3–0.6 keV,
0.6–1.0 keV and 1.0–2.3 keV), which split the X-ray events according
to the calibrated and recombined amplitude values from the CCDs.
Lower-energy events (recorded down to 0.12 keV) were discarded,
owing to the presence of artefacts generated by yet-uncalibrated detec-
tor noise; higher energy events (2.3–8 keV) are not shown, owing to the
lower sensitivity of eROSITA and the effects of the particle background,
which greatly reduce the signal-to-noise ratio in this harder band.
The three images were reprojected into Galactic coordinates using a
Hammer–Aitoff projection. This projection maintains equal area on the
sky. We similarly computed the total exposure time that each pixel on
the sky was observed for during the survey, applying the same Gauss-
ian smoothing. This exposure time takes into account various factors
that reduce observed photon rates, such as the number of cameras
active at any one time, periods removed by screening software, the
effect of vignetting as a source moves away from the optical axis of
the telescopes, and bad pixels in the detectors.
Images of the three broadband maps were made using a logarith-
mic scaling between the upper and lower bounds, chosen to show the
majority of the large-scale structure. These bounds give dynamical
ranges of 13, 35 and 25. The three maps were then combined as red (0.3–
0.6 keV), green (0.6–1.0 keV) and blue (1.0–2.3 keV) channels to make an
RGB image. Brightness and contrast in each channel was adjusted
manually.
The 0.3–0.6-keV band (red), reveals large, soft structures such as the
20°-wide Monogem supernova remnant and the Eridanus superbubble
located southwest of the Orion nebula, which cover a large fraction
of the X-ray sky. Further towards the direction of the Galactic centre,
the Vela supernova remnant, together with the overlapping Puppis A
and Vela–Junior remnants, appear as prominent bright and extended
X-ray sources. Almost symmetric to Vela with respect to the Galactic
centre, the Cygnus superbubble and Cygnus loop are most prominent,
with emission connecting Cygnus and distant Draco regions. Between
these, just north of the Galactic centre, the brightest persistent X-ray
source in the sky (Sco X-1) can be seen; owing to its exceptional bright-
ness, a stray-light halo around the source gives it the appearance of an
extended object34.
Although a hint of extended emission associated with the eROSITA
bubbles is directly evident in the all-sky map (Fig. 1), its appearance
can be enhanced (Fig. 2). First, it appears most clearly in the 0.6–
1.0 keV energy band, where both the effective area of eROSITA and
the intensity of the roughly 0.3-keV hot bubbles peak, so we use only
this band in the analysis. Further improvement may be achieved by
removing the contribution of the point sources from the map. Because
point-source analysis in the survey is still ongoing, whereas the eROSITA
point-spread-function ground calibration is verified against in-flight
data, this task was accomplished using the full sky map obtained as
described above, using the Montage, scikit-image and astropy pack-
ages. In particular, point sources were identified and masked on
small areas of the map reprojected to a set of non-overlapping tiles
(to avoid elongation of point sources). This was accomplished using
a difference-of-Gaussians filter with a radius corresponding to the
mean size of a point source in the image. The filtered image was then
thresholded using the levels calculated with the threshold_local func-
tion in scikit-image (which allows us to leave all large-scale structures
unmasked) and the resulting mask was applied to the observed image.
The masked image was convolved using a Gaussian kernel with size
of five pixels (astropy.convolve) and reprojected to the initial Ham-
mer–Aitoff projection. Finally, the logarithmic intensity scale and
cubehelix colour scale were then applied to the image to allow a large
dynamic range of observed intensities to be displayed in print. All local
manipulations described above affect only the scales comparable with
the point-source size and do not affect the large-scale structures dis-
cussed in the paper.
To compare the morphology of the eROSITA and Fermi bubbles
(Fig. 3), we used the image of a bubble-like component35 reprojected
into a Hammer–Aitoff projection. The final composite was obtained
by applying sequential cyan and red colour scales for the eROSITA and
Fermi images, respectively.
Estimating the energetics of the eROSITA bubbles
In the 0.6–1.0-keV band, average count rates within the northern
and southern eROSITA bubbles of 0.0038 photons s−1 arcmin−2 and
0.0026 photons s−1 arcmin−2, respectively, were observed. These
count rates translate to fluxes of 3.6 × 10−15 erg cm−2 s−1 arcmin−2 and
2.5 × 10−15 erg cm−2 s−1 arcmin−2, assuming a plasma with temperature
kT = 0.3 keV (here k is the Boltzmann constant) and abundances 0.2
times solar19,21,37,38. Assuming a projected area of the eROSITA bubbles
of 35° × 35° × π for each bubble, the total flux is 5 × 10−8 erg cm−2 s−1
and 3.4 × 10−8 erg cm−2 s−1 for the northern and southern bubbles,
respectively. Assuming a distance of 10.6 kpc, we derive luminosities
of 6.0 × 1038 erg s−1 and 4.3 × 1038 erg s−1.
The total energy content of the eROSITA bubbles may be estimated
from their sizes and estimates of the density and temperature of the
X-ray-emitting gas. We considered two approaches. First, we assumed
a Mach number of the shock of M ≈ 1.5, as follows from the Rankine–
Hugoniot condition for the temperature increase from about 0.2 keV
outside the bubbles to around 0.3 keV inside7,19, and adopted an electron
density of the upstream halo gas at a distance of 10 kpc from the centre
of roughly 4 × 10−4 cm−3 (ref. 19), on the basis of the spectral analysis of
numerous XMM-Newton and Suzaku lines of sight. For such a Mach
number, the downstream pressure is a factor of about 2.6 larger than the
upstream pressure, yielding a total energy of two 7-kpc-radius spheres
of roughly 8 × 1055 erg. The contribution of the kinetic energy of the
expanding gas to the total energy is modest. Second, the density of
the gas in the eROSITA bubbles may be estimated from the observed
X-ray surface brightness (or X-ray luminosity), assuming a plasma with
kT ≈ 0.3 keV, a metallicity roughly 0.2 times solar37,38 and size along the
line of sight of about 5 kpc (or assuming a shell-like geometry of the
bubbles; Fig. 2, Extended Data Fig. 2). Assuming that the X-ray emission
from the eROSITA bubbles comes from a shell with inner radius of 3 kpc
and outer radius of 7 kpc, we estimate the electron density of the hot
emitting plasma to be 0.002 cm−3 and the thermal energy within each
bubble to be Eth ≈ 1.3 × 1056 erg.
The velocity of the M ≈ 1.5 shock in a 0.2-keV gas is approximately
340 km s−1. This velocity implies a characteristic expansion time to the
present size of around 20 Myr, which translates to an energy-release
rate of roughly (1–3) × 1041 erg s−1. The cooling time39 of such tenuous
hot gas (density n = 0.002 cm−3, logT = 6.5) is approximately 1.9 × 108 yr,
much longer than the estimated age of the bubbles. Therefore, once
heated, the interior of the bubbles will be visible in X-rays for a long
time, even after the energy release has ceased.
Data availability
The datasets analysed during this study are not yet publicly available.
Their proprietary scientific exploitation rights were granted by the
project funding agencies (Roscosmos and DLR) to two consortia led by
MPE (Germany) and IKI (Russia), respectively. The SRG–eROSITA all-sky
survey data will be released publicly after a minimum period of 2 years.
36. Predehl, M. et al. The eROSITA X-ray telescope on SRG. Astron. Astrophys. https://doi.
org/10.1051/0004-6361/202039313 (2020).
37. Kataoka, J. et al. Suzaku observations of the diffuse X-ray emission across the Fermi
bubbles’ edges. Astrophys. J. 779, 57 (2013).
38. Ursino, E., Galeazzi, M. & Liu, W. Studying the Interstellar medium and the inner region of
NPS/LOOP 1 with shadow observations toward MBM36. Astrophys. J. 816, 33 (2016).
39. Sutherland, M. S. & Dopita, M. A. Cooling functions for low-density astrophysical plasmas.
Astrophys. J. Suppl. Ser. 88, 253–327 (1993).
Acknowledgements This work is based on data from eROSITA, the primary instrument aboard
SRG, a joint Russian–German science mission supported by the Russian Space Agency
(Roskosmos), in the interests of the Russian Academy of Sciences, represented by its Space
Research Institute (IKI), and the Deutsches Zentrum für Luft- und Raumfahrt (DLR). The SRG
spacecraft was built by Lavochkin Association (NPOL) and its subcontractors, and is operated
by NPOL with support from IKI and the Max Planck Institute for Extraterrestrial Physics (MPE).
The development and construction of the eROSITA X-ray instrument was led by MPE, with
contributions from the Dr. Karl Remeis Observatory Bamberg and ECAP (FAU
Erlangen-Nuernberg), the University of Hamburg Observatory, the Leibniz Institute for
Astrophysics Potsdam (AIP), and the Institute for Astronomy and Astrophysics of the University
of Tübingen, with the support of DLR and the Max Planck Society. The Argelander Institute for
Astronomy of the University of Bonn and the Ludwig Maximilians Universität Munich also
participated in the science preparation for eROSITA. The eROSITA data shown here were
processed using the eSASS/NRTA software system developed by the German eROSITA
consortium. We thank the entire eROSITA collaboration team, in Germany and Russia, who,
over many years, have given fundamental contributions to the development of the mission, the
instrument and the science exploitation of the eROSITA data. SRG–eROSITA data processing
and calibration and data analyses were performed by a large number of collaboration
members in both the German and Russian teams, who also discussed and approved the
scientific results presented here. This research made use of Montage. It is funded by the
National Science Foundation under grant number ACI-1440620, and was previously funded by
NASA’s Earth Science Technology Office, Computation Technologies Project, under
Article
cooperative agreement number NCC5-626 between NASA and the California Institute of
Technology. G.P. acknowledges funding from the European Research Council (ERC) under the
European Union’s Horizon 2020 research and innovation programme (grant agreement
number 865637).
Author contributions H.B., M.F., C.M. and J.S.S. developed software to process the eROSITA
data and processed the German proprietary data that resulted in the all-sky maps. E.C., M.G.,
I.K., and P.M. processed the Russian proprietary data. H.B., E.C., M.G., C.M. and J.S.S. performed
the analysis that resulted in Fig. 1. V.D., I.K. and J.S.S. performed the image processing that
resulted in Figs. 2, 3 and Extended Data Figs. 1, 2. The majority of the text was written by P.P.,
W.B., M.F., M.G., E.C., G.P., A.W.S., M.S., H.B. and V.D. V.D. and E.C. worked on Fig. 2; Fig. 3 was
created by I.K. and V.D. Extended Data Fig. 1 was prepared by A.M. with the support of an MPE
graphic design expert. K.N., A.M. and R.A.S. contributed to writing and editing the manuscript.
The above-named authors all contributed to the discussion and interpretation of the results
and their implications. The remaining co-authors made important contributions to SRG
mission planning and operations, eROSITA data acquisition and analysis, and software
development for SRG–eROSITA.
Competing interests The authors declare no competing interests.
Additional information
Supplementary information is available for this paper at https://doi.org/10.1038/s41586-020-
2979-0.
Correspondence and requests for materials should be addressed to P.P., R.A.S., E.C., M.G.,
A.M. or K.N.
Peer review information Nature thanks Jun Kataoka and Roland Crocker for their contribution
to the peer review of this work. Peer reviewer reports are available.
Reprints and permissions information is available at http://www.nature.com/reprints.
Extended Data Fig. 1 | Schematic of the eROSITA and Fermi bubbles. approximate sizes of these structures, as derived from our analysis, are also
Schematic of the geometry of the eROSITA bubbles (EBs; yellow) and Fermi marked (green and purple arrows).
bubbles (FBs; purple) with respect to the Galaxy and the Solar System. The
Article
Extended Data Fig. 2 | Soft-X-ray data compared to a thick-shell model for
the eROSITA bubbles. Comparison between the thick-shell model (cyan line in
Fig. 2) and eROSITA data (0.6–1.0-keV band) in a Lambert zenithal equal-area
projection. The model is in red; the data are in cyan. The northern bubble is
shown on the left (N); the southern bubble is shown on the right (S). The
northern bubble is spherical, with an outer radius of 7 kpc and an inner radius of
5 kpc. It is slightly offset from the vertical above the Galactic centre. The
southern shell is instead an ellipse, slightly elongated in the north–south
direction (semi-major axis is 7 kpc; semi-minor axis 4.9 kpc).
Article

Unveiling the strong interaction among

hadrons at the LHC

https://doi.org/10.1038/s41586-020-3001-6 ALICE Collaboration*

Received: 3 June 2020

Accepted: 20 October 2020 One of the key challenges for nuclear physics today is to understand from first
Published online: 9 December 2020 principles the effective interaction between hadrons with different quark content.
First successes have been achieved using techniques that solve the dynamics of
Open access
quarks and gluons on discrete space-time lattices1,2. Experimentally, the dynamics of
Check for updates the strong interaction have been studied by scattering hadrons off each other. Such
scattering experiments are difficult or impossible for unstable hadrons3–6 and so
high-quality measurements exist only for hadrons containing up and down quarks7.
Here we demonstrate that measuring correlations in the momentum space between
hadron pairs8–12 produced in ultrarelativistic proton–proton collisions at the CERN
Large Hadron Collider (LHC) provides a precise method with which to obtain the
missing information on the interaction dynamics between any pair of unstable
hadrons. Specifically, we discuss the case of the interaction of baryons containing
strange quarks (hyperons). We demonstrate how, using precision measurements of
proton–omega baryon correlations, the effect of the strong interaction for this
hadron–hadron pair can be studied with precision similar to, and compared with,
predictions from lattice calculations13,14. The large number of hyperons identified in
proton–proton collisions at the LHC, together with accurate modelling15 of the small
(approximately one femtometre) inter-particle distance and exact predictions for the
correlation functions, enables a detailed determination of the short-range part of the
nucleon-hyperon interaction.

Baryons are composite objects formed by three valence quarks
bound together by means of the strong interaction mediated
through the emission and absorption of gluons. Between baryons,
the strong interaction leads to a residual force and the most common
example is the effective strong force among nucleons (N)—baryons
composed of up (u) and down (d) quarks: proton (p) = uud and
neutron (n) = ddu. This force is responsible for the existence of a
neutron–proton bound state, the deuteron, and manifests itself in
scattering experiments7 and through the existence of atomic nuclei.
So far, our understanding of the nucleon–nucleon strong interaction
relies heavily on effective theories16, where the degrees of freedom
are nucleons. These effective theories are constrained by scattering
measurements and are successfully used in the description of
nuclear properties17,18.
The fundamental theory of the strong interaction is quantum chromo-
dynamics (QCD), in which quarks and gluons are the degrees of free-
dom. One of the current challenges in nuclear physics is to calculate
the strong interaction among hadrons starting from first principles.
Perturbative techniques are used to calculate strong-interaction
phenomena in high-energy collisions with a level of precision of a
few per cent19. For baryon–baryon interactions at low energy such
techniques cannot be employed; however, numerical solutions on
a finite space-time lattice have been used to calculate scattering
parameters among nucleons and the properties of light nuclei1,2. Such
approaches are still limited: they do not yet reproduce the properties
of the deuteron20 and do not predict physical values for the masses
of light hadrons21.
Baryons containing strange (s) quarks, exclusively or combined with
u and d quarks, are called hyperons (Y) and are denoted by uppercase
Greek letters: Λ = uds, Σ0 = uds, Ξ− = dss, Ω− = sss. Experimentally, little
is known about Y–N and Y–Y interactions, but recently, major steps
forward in their understanding have been made using lattice QCD
approaches13,14,22. The predictions available for hyperons are character-
ized by smaller uncertainties because the lattice calculation becomes
more stable for quarks with larger mass, such as the s quark. In particu-
lar, robust results are obtained for interactions involving the heaviest
hyperons, such as Ξ and Ω, and precise measurements of the p–Ξ− and
p–Ω− interactions are instrumental in validating these calculations.
From an experimental point of view, the existence of nuclei in which
a nucleon is replaced by a hyperon (hypernuclei) demonstrates the
presence of an attractive strong Λ–N interaction23 and indicates the
possibility of binding a Ξ− to a nucleus24,25. A direct and more precise
measurement of the Y–N interaction requires scattering experiments,
which are particularly challenging to perform because hyperons are
short-lived and travel only a few centimetres before decaying. Previ-
ous experiments with Λ and Σ hyperons on proton targets3–5 delivered
results that were two orders of magnitude less precise than those for
nucleons, and such experiments with Ξ (ref. 6) and Ω beams are even
more challenging. The measurement of the Y–N and Y–Y interactions
has further important implications for the possible formation of a

*A list of members and their affiliations appears at the end of the paper.

232 | Nature | Vol 588 | 10 December 2020

 a b Interaction
V(r*) (MeV)
0 Repulsive
p2
r*
0 0.5
p1
Emission source S(r*) Schrödinger equation
Two-particle wavefunction
\ (k*, r*)
c
C(k*, r*) = ∫ S(r*) \ (k*, r*) 2 d3r* = [(k*)
Fig. 1 | Schematic representation of the correlation method. a, A collision of
two protons generates a particle source S(r*) from which a hadron–hadron pair
with momenta p1 and p2 emerges at a relative distance r* and can undergo a
final-state interaction before being detected. Consequently, the relative
momentum k* is either reduced or increased via an attractive or a repulsive
interaction, respectively. b, Example of attractive (green) and repulsive
(dotted red) interaction potentials, V(r*), between two hadrons, as a function
of their relative distance. Given a certain potential, a non-relativistic
Schrödinger equation is used to obtain the corresponding two-particle
Y–N or Y–Y bound state. Although numerous theoretical predictions
exist13,26–30, so far no clear evidence for any such bound states has been
found, despite many experimental searches31–35.
Additionally, a precise knowledge of the Y–N and Y–Y interactions
has important consequences for the physics of neutron stars. Indeed,
the structure of the innermost core of neutron stars is still completely
unknown and hyperons could appear in such environments depending
on the Y–N and Y–Y interactions36. Real progress in this area calls for
new experimental methods.
Studies of the Y–N interaction via correlations have been pioneered
by the HADES collaboration37. Recently, the ALICE Collaboration has
demonstrated that p–p and p–Pb collisions at the LHC are best suited
to study the N–N and several Y–N, Y–Y interactions precisely8–12. Indeed,
the collision energy and rate available at the LHC opens the phase
space for an abundant production of any strange hadron38, and the
capabilities of the ALICE detector for particle identification and the
momentum resolution—with values below 1% for transverse momentum
pT < 1 GeV/c—facilitate the investigation of correlations in momen-
tum space. These correlations reflect the properties of the interaction
and hence can be used to test theoretical predictions by solving the
Schrödinger equation for proton–hyperon collisions39. A fundamen-
tal advantage of p–p and p–Pb collisions at LHC energies is the fact
that all hadrons originate from very small space-time volumes, with
typical inter-hadron distances of about 1 fm. These small distances
are linked through the uncertainty principle to a large range of the
relative momentum (up to 200 MeV/c) for the baryon pair and enable
us to test short-range interactions. Additionally, detailed modelling
of a common source for all produced baryons15 allow us to determine
accurately the source parameters.
Similar studies were carried out in ultrarelativistic Au–Au colli-
sions at a centre-of-mass energy of 200 GeV per nucleon pair by the
STAR collaboration for Λ–Λ40,41 and p–Ω−42 interactions. This collision
system leads to comparatively large particle emitting sources of
3–5 fm. The resulting relative momentum range is below 40 MeV/c,
implying reduced sensitivity to interactions at distances shorter
than 1 fm.
d
Repulsive
Attractive
Attractive
C(k*)
1.0 1.5 2.0 1
r* (fm)
50 100 150 200
k* (MeV/c)
Correlation function
Nsame(k*)
Nmixed(k*)
wavefunction, ψ(k*, r*). c, The equation of the calculated (second term) and
measured (third term) correlation function C(k*), where Nsame(k*) and Nmixed(k*)
represent the k* distributions of hadron–hadron pairs produced in the same
and in different collisions, respectively, and ξ(k*) denotes the corrections for
experimental effects. d, Sketch of the resulting shape of C(k*). The value of the
correlation function is proportional to the interaction strength. It is above
unity for an attractive (green) potential, and between zero and unity for a
repulsive (dotted red) potential.
In this work, we present a precision study of the most exotic among
the proton–hyperon interactions, obtained via the p–Ω− correlation
function in p–p collisions at a centre-of-mass energy s = 13 TeV at the
LHC. The comparison of the measured correlation function with
first-principle calculations13 and with a new precision measurement
of the p–Ξ− correlation in the same collision system provides the first
observation of the effect of the strong interaction for the p–Ω− pair.
The implications of the measured correlations for a possible p–Ω−
bound state are also discussed. These experimental results challenge
the interpretation of the data in terms of lattice QCD as the precision
of the data improves.
Our measurement opens a new chapter for experimental methods
in hadron physics with the potential to pin down the strong interaction
for all known proton–hyperon pairs.
Analysis of the correlation function
Figure 1 shows a schematic representation of the correlation method
used in this analysis. The correlation function can be expressed theo-
retically43,44 as C(k*) = ∫d3r*S(r*) × |ψ(k*, r*)|2, where k* and r* are the
relative momentum and relative distance of the pair of interest. S(r*)
is the distribution of the distance r* = |r*| at which particles are emitted
(defining the source size), ψ(k*, r*) represents the wavefunction of the
relative motion for the pair of interest and k* = |k*| is the reduced rela-
⁎
tive momentum of the pair (k = |p⁎2 − p⁎1 |/2). Given an interaction poten-
tial between two hadrons as a function of their relative distance, a
non-relativistic Schrödinger equation can be used39 to obtain the
corresponding wavefunction and hence also predict the expected
correlation function. The choice of a non-relativistic Schrödinger
equation is motivated by the fact that the typical relative momenta
relevant for the strong final-state interaction have a maximal value of
200 MeV/c. Experimentally, this correlation function is computed as
C(k*) = ξ(k*)[Nsame(k*)/Nmixed(k*)], where ξ(k*) denotes the corrections
for experimental effects, Nsame(k*) is the number of pairs with a given
k* obtained by combining particles produced in the same collision
(event), which constitute a sample of correlated pairs, and Nmixed(k*) is
Nature | Vol 588 | 10 December 2020 | 233
Article
30,000 a
3.5
ALICE data p−Ξ −
25,000 ALICE data
Signal + background fit
3.0 Coulomb
S– mΩ = 1.6725 GeV/c2
p
20,000 V = 1.82 MeV/c2 Coulomb + p − Ξ − HAL QCD
dN/dm (GeV/c2)–1

Background fit 2.5 Coulomb + p − 1 − HAL QCD elastic

Κ–

C(k*)
15,000 Coulomb + p − 1 − HAL QCD elastic + inelastic
Λ
2.0
10,000

Ω– 1.5
5,000

1.0
0
1.65 1.66 1.67 1.68 1.69 1.70
b
mΛΚ (GeV/c2) 7
p−Ω −
+
Fig. 2 | Reconstruction of the Ω and Ω̄ signals. Sketch of the weak decay
−
6
of Ω− into a Λ and a Κ−, and measured invariant mass distribution (blue points)
¯ + combinations. The dotted red line represents the fit to the data
of ΛΚ− and ΛK 1.2
5
including signal and background, and the black dotted line the background

C(k*)
alone. The contamination from misidentification is ≤5%.

C( k *)
1.0
4

the number of uncorrelated pairs with the same k*, obtained by com-
bining particles produced in different collisions (the so-called
mixed-event technique). Figure 1d shows how an attractive or repulsive
interaction is mapped into the correlation function. For an attractive
interaction the magnitude of the correlation function will be above
unity for small values of k*, whereas for a repulsive interaction it will
be between zero and unity. In the former case, the presence of a bound
state would create a depletion of the correlation function with a depth
increasing with increasing binding energy.
Correlations can occur in nature from quantum mechanical inter-
ference, resonances, conservation laws or final-state interactions.
Here, it is the final-state interactions that contribute predominantly
at low relative momentum; in this work we focus on the strong and
Coulomb interactions in pairs composed of a proton and either a Ξ− or
a Ω− hyperon.
Protons do not decay and can hence be directly identified within the
ALICE detector, but Ξ− and Ω− baryons are detected through their weak
decays, Ξ− → Λ + π− and Ω− → Λ + Κ−. The identification and momentum
measurement of protons, Ξ−, Ω− and their respective antiparticles are
described in Methods. Figure 2 shows a sketch of the Ω− decay and the
invariant mass distribution of the ΛΚ− and ΛK ¯ + pairs. The clear peak
+
corresponding to the rare Ω− and Ω̄ baryons demonstrates the excel-
lent identification capability, which is the key ingredient for this meas-
urement. The contamination from misidentification is ≤5%. For the
+ associated with the source radius. The considered radius values are 1.02 ± 0.05
Ξ− ( Ξ̄ ) baryon the misidentification amounts to 8%11.
Once the p, Ω− and Ξ− candidates and charge conjugates are selected
and their 3-momenta measured, the correlation functions can be built.
Since we assume that the same interaction governs baryon–baryon
and antibaryon–antibaryon pairs8, we consider in the following the
+
direct sum (⊕) of particles and antiparticles ( p – Ξ − ⊕ p¯ – Ξ¯ ≡ p – Ξ −
+
− ¯ −
and p – Ω ⊕ p¯ – Ω ≡ p – Ω ). The determination of the correction ξ(k*)
and the evaluation of the systematic uncertainties are described in
Methods.
Comparison of the p–Ξ− and p–Ω− interactions
The obtained correlation functions are shown in Fig. 3a, b for the p–Ξ−
and p–Ω− pairs, respectively, along with the statistical and systematic
uncertainties. The fact that both correlations are well above unity
implies the presence of an attractive interaction for both systems. For
opposite-charge pairs, as considered here, the Coulomb interaction
3 0.8
100 200
2 k* (MeV/c)
1
0 100 200 300
k* (MeV/c)
Fig. 3 | Experimental p–Ξ− and p–Ω− correlation functions. a, b, Measured
p–Ξ− (a) and p–Ω− (b) correlation functions in high multiplicity p–p collisions at
s = 13 TeV . The experimental data are shown as black symbols. The black
vertical bars and the grey boxes represent the statistical and systematic
uncertainties. The square brackets show the bin width and the horizontal black
lines represent the statistical uncertainty in the determination of the mean k*
for each bin. The measurements are compared with theoretical predictions,
shown as coloured bands, that assume either Coulomb or Coulomb + strong
HAL QCD interactions. For the p–Ω− system the orange band represents the
prediction considering only the elastic contributions and the blue band
represents the prediction considering both elastic and inelastic contributions.
The width of the curves including HAL QCD predictions represents the
uncertainty associated with the calculation (see Methods section ‘Corrections
of the correlation function’ for details) and the grey shaded band represents, in
addition, the uncertainties associated with the determination of the source
radius. The width of the Coulomb curves represents only the uncertainty
fm for p–Ξ− and 0.95 ± 0.06 fm for p–Ω− pairs, respectively. The inset in b shows
an expanded view of the p–Ω− correlation function for C(k*) close to unity. For
more details see text.
is attractive and its effect on the correlation function is illustrated
by the green curves in both panels of Fig. 3. These curves have been
obtained by solving the Schrödinger equation for p–Ξ− and p–Ω− pairs
using the Correlation Analysis Tool using the Schrödinger equation
(CATS) equation solver39, considering only the Coulomb interaction and
assuming that the shape of the source follows a Gaussian distribution
with a width equal to 1.02 ± 0.05 fm for the p–Ξ− system and to 0.95 ±
0.06 fm for the p–Ω− system, respectively. The source-size values have
been determined via an independent analysis of p–p correlations15,
where modifications of the source distribution due to strong decays
of short-lived resonances are taken into account, and the source size
is determined as a function of the transverse mass mT of the pair, as

234 | Nature | Vol 588 | 10 December 2020

 100
0 prediction for the global p–Ξ− correlation function is lower than that
p–Ξ – HAL QCD, I = 0, S = 0
–100
p–Ω – HAL QCD, I = 1/2, S = 2
V(r) (MeV)
10
–200
p–Ξ – HAL QCD, I = 0, S = 0
p–Ξ – HAL QCD
–300 p–Ω – HAL QCD, I = 1/2, S = 2
C(k*)
5
–400
0 tions are accounted for within lattice calculations by exploiting the well
–500 0 50 100 150 200
k* (MeV)
0 1
r (fm)
Fig. 4 | Potentials for the p–Ξ− and p–Ω− interactions. p–Ξ− (pink) and p–Ω−
(orange) interaction potentials as a function of the pair distance predicted by
the HAL QCD collaboration13,14. Only the most attractive component, isospin
I = 0 and spin S = 0, is shown for p–Ξ−. For the p–Ω− interaction the I = 1/2 and spin
S = 2 component is shown. The widths of the curves correspond to the
uncertainties (see Methods section ‘Corrections of the correlation function’
for details) associated with the calculations. The inset shows the correlation
functions obtained using the HAL QCD strong interaction potentials for: (i) the
channel p–Ξ− with isospin I = 0 and spin S = 0, (ii) the channel p–Ξ− including all
allowed spin and isospin combinations (dashed pink), and (iii) the channel
p–Ω− with isospin I = 1/2 and spin S = 2. For details see text.
described in Methods. The average mT of the p–Ξ− and p–Ω− pairs are
1.9 GeV/c and 2.2 GeV/c, respectively. The difference in size between
the source of the p–Ξ− and p–Ω− pairs might reflect the contribution
of collective effects such as (an)isotropic flow. The width of the green
curves in Fig. 3 reflects the quoted uncertainty of the measured source
radius. The correlations obtained, accounting only for the Coulomb
interaction, considerably underestimate the strength of both measured
correlations. This implies, in both cases, that an attractive interaction
exists and exceeds the strength of the Coulomb interaction.
To discuss the comparison of the experimental data with the predic-
tions from lattice QCD, it is useful to first focus on the distinct charac-
teristics of the p–Ξ− and p–Ω− interactions. Figure 4 shows the radial
shapes obtained for the strong-interaction potentials calculated from
first principles by the HAL QCD (Hadrons to Atomic nuclei from Lat-
tice QCD) collaboration for the p–Ξ− (ref. 14) and the p–Ω− systems13,
see Methods for details. Only the most attractive (isospin I = 0 and
spin S = 0) of the four components14 of the p–Ξ− interaction and the
isospin I = 1/2 and spin S = 2 component of the p–Ω− interaction are
shown. Aside from an attractive component, we see that the interac-
tion contains also a repulsive core starting at very small distances,
below 0.2 fm. For the p–Ω− system no repulsive core is visible and the
interaction is purely attractive. This very attractive interaction can
accommodate a p–Ω− bound state, with a binding energy of about
2.5 MeV, considering the Coulomb and strong forces13. The p–Ξ− and
p–Ω− interaction potentials look very similar to each other above a
distance of 1 fm. This behaviour is not observed in phenomenologi-
cal models that engage the exchange of heavy mesons and predict a
quicker fall off of the potentials45.
The inset of Fig. 4 shows the correlation functions obtained using the
HAL QCD strong interaction potentials for: (i) the channel p–Ξ− with
isospin I = 0 and spin S = 0, (ii) the channel p–Ξ− including all allowed
spin and isospin combinations, and (iii) the channel p–Ω− with isospin
I = 1/2 and spin S = 2. The correlation functions are computed using the
experimental values for the p–Ξ− and p–Ω− source-size. Despite the fact
that the strong p–Ω− potential is more attractive than the p–Ξ− I = 0
and S = 0 potential, the resulting correlation function is lower. This is
due to the presence of the bound state in the p–Ω− case46. If we con-
sider all four isospin and spin components of the p–Ξ− interaction11 the
for p–Ω−. Experimentally, as shown in Fig. 3, the less attractive strong
p–Ξ− interaction translates into a correlation function that reaches
values of 3 in comparison with the much higher values of up to 6 that
are visible for the p–Ω− correlation. The theoretical predictions shown
in Fig. 3 also include the effect of the Coulomb interaction.
Regarding the p–Ξ− interaction, it should be considered that
strangeness-rearrangement processes can occur, such as pΞ− → ΛΛ, ΣΣ,
ΛΣ. This means that the inverse processes (for example, ΛΛ → pΞ−) can
also occur and modify the p–Ξ− correlation function. These contribu-
known quark symmetries14 and are found to be very small. Moreover,
the ALICE collaboration measured the Λ–Λ correlation in p–p and p–Pb
2
collisions10 and good agreement with the shallow interaction predicted
by the HAL QCD collaboration was found.
The resulting prediction for the correlation function, obtained by
solving the Schrödinger equation for the single p–Ξ− channel includ-
ing the HAL QCD strong and Coulomb interactions, is shown in Fig. 3a.
The first measurement of the p–Ξ− interaction using p–Pb collisions11
showed a qualitative agreement to lattice QCD predictions. The
improved precision of the data in the current analysis of p–p collisions
is also in agreement with calculations that include both the HAL QCD
and Coulomb interactions.
Detailed study of the p–Ω− correlation
Concerning the p–Ω− interaction, strangeness-rearrangement pro-
cesses can also occur47, such as pΩ− → ΞΛ, ΞΣ. Such processes might
affect the p–Ω− interaction in a different way depending on the relative
orientation of the total spin and angular momentum of the pair. Since
the proton has Jp = 1/2 and the Ω has JΩ = 3/2 and the orbital angular
momentum L can be neglected for correlation studies that imply low
relative momentum, the total angular momentum J equals the total
spin S and can take on values of J = 2 or J = 1. The J = 2 state cannot couple
to the strangeness-rearrangement processes discussed above, except
through D-wave processes, which are strongly suppressed. For the
J = 1 state only two limiting cases can be discussed in the absence of
measurements of the pΩ− → ΞΛ, ΞΣ cross-sections.
The first case assumes that the effect of the inelastic channels is
negligible for both configurations and that the radial behaviour of the
interaction is driven by elastic processes, following the lattice QCD
potential (see Fig. 4), for both the J = 2 and J = 1 channels. This results in
a prediction, shown by the orange curve in Fig. 3b, that is close to the
data in the low k* region. The second limiting case assumes, follow-
ing a previous prescription47, that the J = 1 configuration is completely
dominated by strangeness-rearrangement processes. The obtained
correlation function is shown by the blue curve Fig. 3b. This curve clearly
deviates from the data. Both theoretical calculations also include the
effect of the Coulomb interaction and they predict the existence of a
p–Ω− bound state with a binding energy of 2.5 MeV, which causes a deple-
tion in the correlation function in the k* region between 100 and 300
MeV/c, because pairs that form a bound state are lost to the correlation
yield. The inset of Fig. 3 shows that in this k* region the data are consist-
ent with unity and do not follow either of the two theoretical predictions.
At the moment, the lattice QCD predictions underestimate the data,
but additional measurements are necessary to draw a firm conclu-
sion on the existence of the bound state. Measurements of Λ–Ξ− and
Σ0–Ξ− correlations will verify experimentally the strength of possible
non-elastic contributions. Measurements of the p–Ω− correlation func-
tion in collision systems with slightly larger size (for example, p–Pb
collisions at the LHC)11 will clarify the possible presence of a deple-
tion in C(k*). Indeed, the appearance of a depletion in the correlation
function depends on the interplay between the average intra-particle
Nature | Vol 588 | 10 December 2020 | 235
Article
distance (source size) and the scattering length associated with the 18. Gebrerufael, E., Vobig, K., Hergert, H. & Roth, R. Ab initio description of open-shell nuclei:
merging no-core shell model and in-medium similarity renormalization group. Phys. Rev.
p–Ω− interaction47. Lett. 118, 152503 (2017).
19. Klijnsma, T., Bethke, S., Dissertori, G. & Salam, G. P. Determination of the strong coupling
constant αs(mZ) from measurements of the total cross section for top-antitop quark
Summary production. Eur. Phys. J. C 77, 778 (2017).
20. NPLQCD Collaboration. Two nucleon systems at mπ ~ 450 MeV from lattice QCD. Phys.
We have shown that the hyperon–proton interaction can be studied in Rev. D 92, 114512 (2015).
unprecedented detail in p–p collisions at s = 13 TeV at the LHC. We 21. Wagman, M. L. et al. Baryon–baryon interactions and spin-flavor symmetry from lattice
quantum chromodynamics. Phys. Rev. D 96, 114510 (2017).
have demonstrated, in particular, that even the as-yet-unknown p–Ω− 22. Hatsuda, T. Lattice quantum chromodynamics and baryon-baryon interactions. Front.
interaction can be investigated with excellent precision. The com- Phys. 13, 132105 (2018).
parison of the measured correlation functions shows that the 23. Hashimoto, O. & Tamura, H. Spectroscopy of Λ hypernuclei. Prog. Part. Nucl. Phys. 57,
564–653 (2006).
p–Ω− signal is up to a factor two larger than the p–Ξ− signal. This reflects 24. Nakazawa, K. et al. The first evidence of a deeply bound state of Xi−–14N system. Prog.
the large difference in the strong-attractive interaction predicted by Theor. Exp. Phys. 2015, 033D02 (2015).
the first-principle calculations by the HAL QCD collaboration. The 25. Nagae, T. et al. Search for a Ξ bound state in the 12C(K−, K+)X reaction at 1.8 GeV/c. PoS
(INPC2016) 038 (2017).
correlation functions predicted by HAL QCD are in agreement with the 26. Francis, A. et al. Lattice QCD study of the H dibaryon using hexaquark and two-baryon
measurements for the p–Ξ− interaction. For the p–Ω− interaction, interpolators. Phys. Rev. D 99, 074505, (2019).
the inelastic channels are not yet accounted for quantitatively within 27. Jaffe, R. L. Perhaps a stable dihyperon. Phys. Rev. Lett. 38, 195–198 (1977); erratum 38, 617
(1977).
the lattice QCD calculations. Additionally, the depletion in the correla- 28. Nagels, M. M., Rijken, T. A. & de Swart, J. J. Baryon–baryon scattering in a
tion function that is visible in the calculations around k* = 150 MeV/c, one-boson-exchange-potential approach. II. Hyperon–nucleon scattering. Phys. Rev. D
owing to the presence of a p–Ω− bound state, is not observed in the meas- 15, 2547–2564 (1977).
29. Nagels, M. M., Rijken, T. A. & de Swart, J. J. Baryon–baryon scattering in a
ured correlation. To draw quantitative conclusions concerning the exist- one-boson-exchange-potential approach. III. A nucleon–nucleon and hyperon–nucleon
ence of a p–Ω− bound state, we plan a direct measurement of the Λ–Ξ− and analysis including contributions of a nonet of scalar mesons. Phys. Rev. D 20, 1633–1645
Σ0−Ξ− correlations and a study of the p–Ω− correlation in p–Pb collisions (1979).
30. Gongyo, S. et al. Most strange dibaryon from lattice QCD. Phys. Rev. Lett. 120, 212001
in the near future. Indeed, with the upgraded ALICE apparatus48 and the (2018).
increased data sample size expected from the high luminosity phase of 31. ALICE Collaboration. Search for weakly decaying Λn and ΛΛ exotic bound states in
central Pb–Pb collisions at √sNN = 2.76 TeV. Phys. Lett. B 752, 267–277 (2016).
the LHC Run 3 and Run 449, the missing interactions involving hyperons
32. Belle Collaboration. Search for an H-dibaryon with mass near 2mΛ in ϒ(1S) and ϒ(2S)
will be measured in p–p and p–Pb collisions and this should enable us to decays. Phys. Rev. Lett. 110, 222002 (2013).
answer the question about the existence of a new baryon–baryon bound 33. Chrien, R. H particle searches at Brookhaven. Nucl. Phys. A 629, 388–397 (1998).
34. Yoon, C. et al. Search for the H-dibaryon resonance in 12C (K−, K+ΛΛX). Phys. Rev. C 75,
state. Since this method can be extended to almost any hadron–hadron
022201 (2007).
pair, an unexpected avenue for high-precision tests of the strong inter- 35. KEK-PS E224 Collaboration. Enhanced ΛΛ production near threshold in the 12C(K−, K+)
action at the LHC has been opened. reaction. Phys. Lett. B 444, 267–272 (1998).
36. Tolos, L. & Fabbietti, L. Strangeness in nuclei and neutron stars. Prog. Part. Nucl. Phys. 112,
103770 (2020).
37. HADES Collaboration. The Λp interaction studied via femtoscopy in p + Nb reactions at
Online content √sNN = 3.18 GeV. Phys. Rev. C 94, 025201 (2016).
38. ALICE Collaboration. Enhanced production of multi-strange hadrons in high-multiplicity
Any methods, additional references, Nature Research reporting sum-
proton–proton collisions. Nat. Phys. 13, 535–539 (2017).
maries, source data, extended data, supplementary information, 39. Mihaylov, D. et al. A femtoscopic correlation analysis tool using the Schrödinger equation
acknowledgements, peer review information; details of author con- (CATS). Eur. Phys. J. C 78, 394 (2018).
40. STAR Collaboration. ΛΛ correlation function in Au+Au collisions at √sNN = 200 GeV. Phys.
tributions and competing interests; and statements of data and code
Rev. Lett. 114, 022301 (2015).
availability are available at https://doi.org/10.1038/s41586-020-3001-6. 41. Morita, K., Furumoto, T. & Ohnishi, A. ΛΛ interaction from relativistic heavy-ion collisions.
Phys. Rev. C 91, 024916 (2015).
1. NPLQCD Collaboration. Nucleon–nucleon scattering parameters in the limit of SU(3) 42. STAR Collaboration. The proton–Ω correlation function in Au + Au collisions at
flavor symmetry. Phys. Rev. C 88, 024003 (2013). √sNN = 200 GeV. Phys. Lett. B 790, 490–497 (2019).
2. Epelbaum, E., Krebs, H., Lee, D. & Meissner, U.-G. Lattice effective field theory calculations 43. Pratt, S. Pion interferometry of quark–gluon plasma. Phys. Rev. D 33, 1314–1327 (1986).
for A = 3, 4, 6, 12 nuclei. Phys. Rev. Lett. 104, 142501 (2010). 44. Lisa, M. A., Pratt, S., Soltz, R. & Wiedemann, U. Femtoscopy in relativistic heavy ion
3. Eisele, F., Filthuth, H., Foehlisch, W., Hepp, V. & Zech, G. Elastic Σ±p scattering at low collisions. Annu. Rev. Nucl. Part. Sci. 55, 357–402 (2005).
energies. Phys. Lett. B 37, 204–206 (1971). 45. Haidenbauer, J. & Meißner, U.-G. Phenomenological view on baryon–baryon potentials
4. Alexander, G. et al. Study of the Λ–N system in low-energy Λ–p elastic scattering. Phys. from lattice QCD simulations. Eur. Phys. J. A 55, 70 (2019).
Rev. 173, 1452–1460 (1968). 46. Morita, K., Ohnishi, A., Etminan, F. & Hatsuda, T. Probing multistrange dibaryons with
5. Sechi-Zorn, B., Kehoe, B., Twitty, J. & Burnstein, R. Low-energy Λ–proton elastic scattering. proton–omega correlations in high-energy heavy ion collisions. Phys. Rev. C 94, 031901
Phys. Rev. 175, 1735–1740 (1968). (2016); erratum 100, 069902 (2019).
6. Muller, R. Observation of cascade hyperon interactions. Phys. Lett. B 38, 123–124 (1972). 47. Morita, K. et al. Probing ΩΩ and pΩ dibaryons with femtoscopic correlations in relativistic
7. Stoks, V. & de Swart, J. Comparison of potential models with the pp scattering data below heavy-ion collisions. Phys. Rev. C 101, 015201 (2020).
350 MeV. Phys. Rev. C 47, 761–767 (1993). 48. ALICE Collaboration. Upgrade of the ALICE experiment: letter of intent. J. Phys. G 41,
8. ALICE Collaboration. p–p, p–Λ and Λ–Λ correlations studied via femtoscopy in pp 087001 (2014).
reactions at √s = 7 TeV. Phys. Rev. C 99, 024001 (2019). 49. Citron, Z. et al. Report from Working Group 5: future physics opportunities for
9. ALICE Collaboration. Scattering studies with low-energy kaon–proton femtoscopy in high-density QCD at the LHC with heavy-ion and proton beams. CERN Yellow Rep.
proton–proton collisions at the LHC. Phys. Rev. Lett. 124, 092301 (2020). Monogr. 7, 1159–1410 (2019).
10. ALICE Collaboration. Study of the Λ–Λ interaction with femtoscopy correlations in pp and
p–Pb collisions at the LHC. Phys. Lett. B 797, 134822 (2019). Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in
11. ALICE Collaboration. First observation of an attractive interaction between a proton and a published maps and institutional affiliations.
cascade baryon. Phys. Rev. Lett. 123, 112002 (2019).
12. ALICE Collaboration. Investigation of the p–Σ0 interaction via femtoscopy in pp collisions. Open Access This article is licensed under a Creative Commons Attribution
Phys. Lett. B 805, 135419 (2020). 4.0 International License, which permits use, sharing, adaptation, distribution
13. HAL QCD Collaboration. NΩ dibaryon from lattice QCD near the physical point. Phys. and reproduction in any medium or format, as long as you give appropriate
Lett. B 792, 284–289 (2019). credit to the original author(s) and the source, provide a link to the Creative Commons license,
14. Sasaki, K. et al. ΛΛ and NΞ interactions from lattice QCD near the physical point. Nucl. and indicate if changes were made. The images or other third party material in this article are
Phys. A 998, 121737 (2020). included in the article’s Creative Commons license, unless indicated otherwise in a credit line
15. ALICE Collaboration. Search for a common baryon source in high-multiplicity pp to the material. If material is not included in the article’s Creative Commons license and your
collisions at the LHC. Phys. Lett. B 811, 135849 (2020). intended use is not permitted by statutory regulation or exceeds the permitted use, you will
16. Epelbaum, E., Hammer, H.-W. & Meissner, U.-G. Modern theory of nuclear forces. Rev. need to obtain permission directly from the copyright holder. To view a copy of this license,
Mod. Phys. 81, 1773–1825 (2009). visit http://creativecommons.org/licenses/by/4.0/.
17. Hebeler, K., Holt, J., Menendez, J. & Schwenk, A. Nuclear forces and their impact on
neutron-rich nuclei and neutron-rich matter. Annu. Rev. Nucl. Part. Sci. 65, 457–484 (2015). © The Author(s) 2020

236 | Nature | Vol 588 | 10 December 2020

ALICE Collaboration✉
S. Acharya1, D. Adamová2, A. Adler3, J. Adolfsson4, M. M. Aggarwal5, G. Aglieri Rinella6, M.
Agnello7, N. Agrawal8,9, Z. Ahammed1, S. Ahmad10, S. U. Ahn11, Z. Akbar12, A. Akindinov13, M.
Al-Turany14, S. N. Alam1,15, D. S. D. Albuquerque16, D. Aleksandrov17, B. Alessandro18, H. M.
Alfanda19, R. Alfaro Molina20, B. Ali10, Y. Ali21, A. Alici8,9,22, N. Alizadehvandchali23, A. Alkin6,24,
J. Alme25, T. Alt26, L. Altenkamper25, I. Altsybeev27, M. N. Anaam19, C. Andrei28, D. Andreou6, A.
Andronic29, M. Angeletti6, V. Anguelov30, C. Anson31, T. Antičić32, F. Antinori33, P. Antonioli9,
N. Apadula34, L. Aphecetche35, H. Appelshäuser26, S. Arcelli22, R. Arnaldi18, M. Arratia34, I. C.
Arsene36, M. Arslandok30, A. Augustinus6, R. Averbeck14, S. Aziz37, M. D. Azmi10, A. Badalà38, Y.
W. Baek39, S. Bagnasco18, X. Bai14, R. Bailhache26, R. Bala40, A. Balbino7, A. Baldisseri41, M.
Ball42, S. Balouza43, D. Banerjee44, R. Barbera45, L. Barioglio46, G. G. Barnaföldi47, L. S.
Barnby48, V. Barret49, P. Bartalini19, C. Bartels50, K. Barth6, E. Bartsch26, F. Baruffaldi51, N.
Bastid49, S. Basu52, G. Batigne35, B. Batyunya53, D. Bauri54, J. L. Bazo Alba55, I. G. Bearden56, C.
Beattie57, C. Bedda58, N. K. Behera59, I. Belikov60, A. D. C. Bell Hechavarria29, F. Bellini6, R.
Bellwied23, V. Belyaev61, G. Bencedi47, S. Beole46, A. Bercuci28, Y. Berdnikov62, D. Berenyi47, R.
A. Bertens63, D. Berzano18, M. G. Besoiu64, L. Betev6, A. Bhasin40, I. R. Bhat40, M. A. Bhat44, H.
Bhatt54, B. Bhattacharjee65, A. Bianchi46, L. Bianchi46, N. Bianchi66, J. Bielčík67, J. Bielčíková2,
A. Bilandzic43, G. Biro47, R. Biswas44, S. Biswas44, J. T. Blair68, D. Blau17, C. Blume26, G. Boca69, F.
Bock70, A. Bogdanov61, S. Boi71, J. Bok59, L. Boldizsár47, A. Bolozdynya61, M. Bombara72, G.
Bonomi73, H. Borel41, A. Borissov61, H. Bossi57, E. Botta46, L. Bratrud26, P. Braun-Munzinger14,
M. Bregant74, M. Broz67, E. Bruna18, G. E. Bruno75,76, M. D. Buckland50, D. Budnikov77, H.
Buesching26, S. Bufalino7, O. Bugnon35, P. Buhler78, P. Buncic6, Z. Buthelezi79,80, J. B. Butt21, S.
A. Bysiak81, D. Caffarri82, A. Caliva14, E. Calvo Villar55, J. M. M. Camacho83, R. S. Camacho84, P.
Camerini85, F. D. M. Canedo74, A. A. Capon78, F. Carnesecchi22, R. Caron41, J. Castillo
Castellanos41, A. J. Castro63, E. A. R. Casula86, F. Catalano7, C. Ceballos Sanchez53, P.
Chakraborty54, S. Chandra1, W. Chang19, S. Chapeland6, M. Chartier50, S. Chattopadhyay1, S.
Chattopadhyay87, A. Chauvin71, C. Cheshkov88, B. Cheynis88, V. Chibante Barroso6, D. D.
Chinellato16, S. Cho59, P. Chochula6, T. Chowdhury49, P. Christakoglou82, C. H. Christensen56,
P. Christiansen4, T. Chujo89, C. Cicalo86, L. Cifarelli8,22, L. D. Cilladi46, F. Cindolo9, M. R.
Ciupek14, G. Clai9,148, J. Cleymans90, F. Colamaria91, D. Colella91, A. Collu34, M. Colocci22, M.
Concas18,149, G. Conesa Balbastre92, Z. Conesa del Valle37, G. Contin85,93, J. G. Contreras67, T.
M. Cormier70, Y. Corrales Morales46, P. Cortese94, M. R. Cosentino95, F. Costa6, S. Costanza69,
P. Crochet49, E. Cuautle96, P. Cui19, L. Cunqueiro70, D. Dabrowski97, T. Dahms43, A. Dainese33, F.
P. A. Damas35,41, M. C. Danisch30, A. Danu64, D. Das87, I. Das87, P. Das98, P. Das44, S. Das44, A.
Dash98, S. Dash54, S. De98, A. De Caro99, G. de Cataldo91, J. de Cuveland100, A. De Falco71, D. De
Gruttola8, N. De Marco18, S. De Pasquale99, S. Deb101, H. F. Degenhardt74, K. R. Deja97, A.
Deloff102, S. Delsanto46,80, W. Deng19, P. Dhankher54, D. Di Bari75, A. Di Mauro6, R. A. Diaz103, T.
Dietel90, P. Dillenseger26, Y. Ding19, R. Divià6, D. U. Dixit104, Ø. Djuvsland25, U. Dmitrieva105, A.
Dobrin64, B. Dönigus26, O. Dordic36, A. K. Dubey1, A. Dubla14,82, S. Dudi5, M. Dukhishyam98, P.
Dupieux49, R. J. Ehlers70, V. N. Eikeland25, D. Elia91, B. Erazmus35, F. Erhardt106, A. Erokhin27, M.
R. Ersdal25, B. Espagnon37, G. Eulisse6, D. Evans107, S. Evdokimov108, L. Fabbietti43, M. Faggin51,
J. Faivre92, F. Fan19, A. Fantoni66, M. Fasel70, P. Fecchio7, A. Feliciello18, G. Feofilov27, A.
Fernández Téllez84, A. Ferrero41, A. Ferretti46, A. Festanti6, V. J. G. Feuillard30, J. Figiel81, S.
Filchagin77, D. Finogeev105, F. M. Fionda25, G. Fiorenza91, F. Flor23, A. N. Flores68, S. Foertsch79,
P. Foka14, S. Fokin17, E. Fragiacomo93, U. Frankenfeld14, U. Fuchs6, C. Furget92, A. Furs105, M.
Fusco Girard99, J. J. Gaardhøje56, M. Gagliardi46, A. M. Gago55, A. Gal60, C. D. Galvan83, P.
Ganoti109, C. Garabatos14, J. R. A. Garcia84, E. Garcia-Solis110, K. Garg35, C. Gargiulo6, A.
Garibli111, K. Garner29, P. Gasik14,43, E. F. Gauger68, M. B. Gay Ducati112, M. Germain35, J. Ghosh87,
P. Ghosh1, S. K. Ghosh44, M. Giacalone22, P. Gianotti66, P. Giubellino14,18, P. Giubilato51, A. M. C.
Glaenzer41, P. Glässel30, A. Gomez Ramirez3, V. Gonzalez14,52, L. H. González-Trueba20, S.
Gorbunov100, L. Görlich81, A. Goswami54, S. Gotovac113, V. Grabski20, L. K. Graczykowski97, K.
L. Graham107, L. Greiner34, A. Grelli58, C. Grigoras6, V. Grigoriev61, A. Grigoryan114, S.
Grigoryan53, O. S. Groettvik25, F. Grosa7,18, J. F. Grosse-Oetringhaus6, R. Grosso14, R.
Guernane92, M. Guittiere35, K. Gulbrandsen56, T. Gunji115, A. Gupta40, R. Gupta40, I. B.
Guzman84, R. Haake57, M. K. Habib14, C. Hadjidakis37, H. Hamagaki116, G. Hamar47, M. Hamid19,
R. Hannigan68, M. R. Haque58,98, A. Harlenderova14, J. W. Harris57, A. Harton110, J. A.
Hasenbichler6, H. Hassan70, Q. U. Hassan21, D. Hatzifotiadou8,9, P. Hauer42, L. B. Havener57, S.
Hayashi115, S. T. Heckel43, E. Hellbär26, H. Helstrup117, A. Herghelegiu28, T. Herman67, E. G.
Hernandez84, G. Herrera Corral118, F. Herrmann29, K. F. Hetland117, H. Hillemanns6, C. Hills50, B.
Hippolyte60, B. Hohlweger43, J. Honermann29, D. Horak67, A. Hornung26, S. Hornung14, R.
Hosokawa31,89, P. Hristov6, C. Huang37, C. Hughes63, P. Huhn26, T. J. Humanic119, H. Hushnud87,
L. A. Husova29, N. Hussain65, S. A. Hussain21, D. Hutter100, J. P. Iddon6,50, R. Ilkaev77, H. Ilyas21,
M. Inaba89, G. M. Innocenti6, M. Ippolitov17, A. Isakov2, M. S. Islam87, M. Ivanov14, V. Ivanov62, V.
Izucheev108, B. Jacak34, N. Jacazio6,9, P. M. Jacobs34, S. Jadlovska120, J. Jadlovsky120, S.
Jaelani58, C. Jahnke74, M. J. Jakubowska97, M. A. Janik97, T. Janson3, M. Jercic106, O. Jevons107,
M. Jin23, F. Jonas29,70, P. G. Jones107, J. Jung26, M. Jung26, A. Jusko107, P. Kalinak121, A. Kalweit6, V.
Kaplin61, S. Kar19, A. Karasu Uysal122, D. Karatovic106, O. Karavichev105, T. Karavicheva105, P.
Karczmarczyk97, E. Karpechev105, A. Kazantsev17, U. Kebschull3, R. Keidel123, M. Keil6, B.
Ketzer42, Z. Khabanova82, A. M. Khan19, S. Khan10, A. Khanzadeev62, Y. Kharlov108, A. Khatun10,
A. Khuntia81, B. Kileng117, B. Kim59, B. Kim89, D. Kim124, D. J. Kim125, E. J. Kim126, H. Kim127, J.
Kim124, J. S. Kim39, J. Kim30, J. Kim124, J. Kim126, M. Kim30, S. Kim128, T. Kim124, T. Kim124, S.
Kirsch26, I. Kisel100, S. Kiselev13, A. Kisiel97, J. L. Klay129, C. Klein26, J. Klein6,18, S. Klein34, C.
Klein-Bösing29, M. Kleiner26, A. Kluge6, M. L. Knichel6, A. G. Knospe23, C. Kobdaj130, M. K.
Köhler30, T. Kollegger14, A. Kondratyev53, N. Kondratyeva61, E. Kondratyuk108, J. Konig26, S. A.
Konigstorfer43, P. J. Konopka6, G. Kornakov97, L. Koska120, O. Kovalenko102, V. Kovalenko27, M.
Kowalski81, I. Králik121, A. Kravčáková72, L. Kreis14, M. Krivda107,121, F. Krizek2, K. Krizkova
Gajdosova67, M. Krüger26, E. Kryshen62, M. Krzewicki100, A. M. Kubera119, V. Kučera6,59, C.
Kuhn60, P. G. Kuijer82, L. Kumar5, S. Kundu98, P. Kurashvili102, A. Kurepin105, A. B. Kurepin105, A.
Kuryakin77, S. Kushpil2, J. Kvapil107, M. J. Kweon59, J. Y. Kwon59, Y. Kwon124, S. L. La Pointe100, P.
La Rocca45, Y. S. Lai34, M. Lamanna6, R. Langoy131, K. Lapidus6, A. Lardeux36, P. Larionov66, E.
Laudi6, R. Lavicka67, T. Lazareva27, R. Lea85, L. Leardini30, J. Lee89, S. Lee124, S. Lehner78, J.
Lehrbach100, R. C. Lemmon48, I. León Monzón83, E. D. Lesser104, M. Lettrich6, P. Lévai47, X. Li132,
X. L. Li19, J. Lien131, R. Lietava107, B. Lim127, V. Lindenstruth100, A. Lindner28, C. Lippmann14, M. A.
Lisa119, A. Liu104, J. Liu50, S. Liu119, W. J. Llope52, I. M. Lofnes25, V. Loginov61, C. Loizides70, P.
Loncar113, J. A. Lopez30, X. Lopez49, E. López Torres103, J. R. Luhder29, M. Lunardon51, G.
Luparello93, Y. G. Ma15, A. Maevskaya105, M. Mager6, S. M. Mahmood36, T. Mahmoud42, A.
Maire60, R. D. Majka57,153, M. Malaev62, Q. W. Malik36, L. Malinina53,150, D. Mal’Kevich13, P.
Malzacher14, G. Mandaglio38,133, V. Manko17, F. Manso49, V. Manzari91, Y. Mao19, M.
Marchisone88, J. Mareš134, G. V. Margagliotti85, A. Margotti9, A. Marín14, C. Markert68, M.
Marquard26, C. D. Martin85, N. A. Martin30, P. Martinengo6, J. L. Martinez23, M. I. Martínez84, G.
Martínez García35, S. Masciocchi14, M. Masera46, A. Masoni86, L. Massacrier37, E. Masson35, A.
Mastroserio91,135, A. M. Mathis43, O. Matonoha4, P. F. T. Matuoka74, A. Matyja81, C. Mayer81, F.
Mazzaschi46, M. Mazzilli91, M. A. Mazzoni136, A. F. Mechler26, F. Meddi137, Y. Melikyan61,105, A.
Menchaca-Rocha20, C. Mengke19, E. Meninno78,99, A. S. Menon23, M. Meres138, S. Mhlanga90, Y.
Miake89, L. Micheletti46, L. C. Migliorin88, D. L. Mihaylov43, K. Mikhaylov13,53, A. N. Mishra96, D.
Miśkowiec14, A. Modak44, N. Mohammadi6, A. P. Mohanty58, B. Mohanty98, M. Mohisin
Khan10,151, Z. Moravcova56, C. Mordasini43, D. A. Moreira De Godoy29, L. A. P. Moreno84, I.
Morozov105, A. Morsch6, T. Mrnjavac6, V. Muccifora66, E. Mudnic113, D. Mühlheim29, S. Muhuri1,
J. D. Mulligan34, A. Mulliri71,86, M. G. Munhoz74, R. H. Munzer26, H. Murakami115, S. Murray90, L.
Musa6 ✉, J. Musinsky121, C. J. Myers23, J. W. Myrcha97, B. Naik54, R. Nair102, B. K. Nandi54, R.
Nania8,9, E. Nappi91, M. U. Naru21, A. F. Nassirpour4, C. Nattrass63, R. Nayak54, T. K. Nayak98, S.
Nazarenko77, A. Neagu36, R. A. Negrao De Oliveira26, L. Nellen96, S. V. Nesbo117, G. Neskovic100,
D. Nesterov27, L. T. Neumann97, B. S. Nielsen56, S. Nikolaev17, S. Nikulin17, V. Nikulin62, F.
Noferini8,9, P. Nomokonov53, J. Norman50,92, N. Novitzky89, P. Nowakowski97, A. Nyanin17, J.
Nystrand25, M. Ogino116, A. Ohlson4,30, J. Oleniacz97, A. C. Oliveira Da Silva63, M. H. Oliver57, C.
Oppedisano18, A. Ortiz Velasquez96, A. Oskarsson4, J. Otwinowski81, K. Oyama116, Y.
Pachmayer30, V. Pacik56, S. Padhan54, D. Pagano73, G. Paić96, J. Pan52, S. Panebianco41, P.
Pareek1,101, J. Park59, J. E. Parkkila125, S. Parmar5, S. P. Pathak23, B. Paul71, J. Pazzini73, H. Pei19, T.
Peitzmann58, X. Peng19, L. G. Pereira112, H. Pereira Da Costa41, D. Peresunko17, G. M. Perez103, S.
Perrin41, Y. Pestov139, V. Petráček67, M. Petrovici28, R. P. Pezzi112, S. Piano93, M. Pikna138, P.
Pillot35, O. Pinazza6,9, L. Pinsky23, C. Pinto45, S. Pisano8,66, D. Pistone38, M. Płoskoń34, M.
Planinic106, F. Pliquett26, M. G. Poghosyan70, B. Polichtchouk108, N. Poljak106, A. Pop28, S.
Porteboeuf-Houssais49, V. Pozdniakov53, S. K. Prasad44, R. Preghenella9, F. Prino18, C. A.
Pruneau52, I. Pshenichnov105, M. Puccio6, J. Putschke52, S. Qiu82, L. Quaglia46, R. E. Quishpe23,
S. Ragoni107, S. Raha44, S. Rajput40, J. Rak125, A. Rakotozafindrabe41, L. Ramello94, F. Rami60, S.
A. R. Ramirez84, R. Raniwala140, S. Raniwala140, S. S. Räsänen141, R. Rath101, V. Ratza42, I.
Ravasenga82, K. F. Read63,70, A. R. Redelbach100, K. Redlich102,152, A. Rehman25, P. Reichelt26, F.
Reidt6, X. Ren19, R. Renfordt26, Z. Rescakova72, K. Reygers30, A. Riabov62, V. Riabov62, T.
Richert4,56, M. Richter36, P. Riedler6, W. Riegler6, F. Riggi45, C. Ristea64, S. P. Rode101, M.
Rodríguez Cahuantzi84, K. Røed36, R. Rogalev108, E. Rogochaya53, D. Rohr6, D. Röhrich25, P. F.
Rojas84, P. S. Rokita97, F. Ronchetti66, A. Rosano38, E. D. Rosas96, K. Roslon97, A. Rossi33,51, A.
Rotondi69, A. Roy101, P. Roy87, O. V. Rueda4, R. Rui85, B. Rumyantsev53, A. Rustamov111, E.
Ryabinkin17, Y. Ryabov62, A. Rybicki81, H. Rytkonen125, O. A. M. Saarimaki141, R. Sadek35, S.
Sadhu1, S. Sadovsky108, K. Šafařík67, S. K. Saha1, B. Sahoo54, P. Sahoo54, R. Sahoo101, S.
Sahoo142, P. K. Sahu142, J. Saini1, S. Sakai89, S. Sambyal40, V. Samsonov61,62, D. Sarkar52, N.
Sarkar1, P. Sarma65, V. M. Sarti43, M. H. P. Sas58, E. Scapparone9, J. Schambach68, H. S.
Scheid26, C. Schiaua28, R. Schicker30, A. Schmah30, C. Schmidt14, H. R. Schmidt143, M. O.
Schmidt30, M. Schmidt143, N. V. Schmidt26,70, A. R. Schmier63, J. Schukraft56, Y. Schutz60, K.
Schwarz14, K. Schweda14, G. Scioli22, E. Scomparin18, J. E. Seger31, Y. Sekiguchi115, D.
Sekihata115, I. Selyuzhenkov14,61, S. Senyukov60, D. Serebryakov105, A. Sevcenco64, A.
Shabanov105, A. Shabetai35, R. Shahoyan6, W. Shaikh87, A. Shangaraev108, A. Sharma5, A.
Sharma40, H. Sharma81, M. Sharma40, N. Sharma5, S. Sharma40, O. Sheibani23, K. Shigaki144, M.
Shimomura145, S. Shirinkin13, Q. Shou15, Y. Sibiriak17, S. Siddhanta86, T. Siemiarczuk102, D.
Silvermyr4, G. Simatovic82, G. Simonetti6, B. Singh43, R. Singh98, R. Singh40, R. Singh101, V. K.
Singh1, V. Singhal1, T. Sinha87, B. Sitar138, M. Sitta94, T. B. Skaali36, M. Slupecki141, N. Smirnov57,
R. J. M. Snellings58, C. Soncco55, J. Song23, A. Songmoolnak130, F. Soramel51, S. Sorensen63, I.
Sputowska81, J. Stachel30, I. Stan64, P. J. Steffanic63, E. Stenlund4, S. F. Stiefelmaier30, D.
Stocco35, M. M. Storetvedt117, L. D. Stritto99, A. A. P. Suaide74, T. Sugitate144, C. Suire37, M.
Suleymanov21, M. Suljic6, R. Sultanov13, M. Šumbera2, V. Sumberia40, S. Sumowidagdo12, S.
Swain142, A. Szabo138, I. Szarka138, U. Tabassam21, S. F. Taghavi43, G. Taillepied49, J. Takahashi16,
G. J. Tambave25, S. Tang19,49, M. Tarhini35, M. G. Tarzila28, A. Tauro6, G. Tejeda Muñoz84, A.
Telesca6, L. Terlizzi46, C. Terrevoli23, D. Thakur101, S. Thakur1, D. Thomas68, F. Thoresen56, R.
Tieulent88, A. Tikhonov105, A. R. Timmins23, A. Toia26, N. Topilskaya105, M. Toppi66, F.
Torales-Acosta104, S. R. Torres67, A. Trifiró38,133, S. Tripathy96,101, T. Tripathy54, S. Trogolo51, G.
Trombetta75, L. Tropp72, V. Trubnikov24, W. H. Trzaska125, T. P. Trzcinski97, B. A. Trzeciak58,67, A.
Tumkin77, R. Turrisi33, T. S. Tveter36, K. Ullaland25, E. N. Umaka23, A. Uras88, G. L. Usai71, M.
Vala72, N. Valle69, S. Vallero18, N. van der Kolk58, L. V. R. van Doremalen58, M. van Leeuwen58, P.
Vande Vyvre6, D. Varga47, Z. Varga47, M. Varga-Kofarago47, A. Vargas84, M. Vasileiou109, A.
Vasiliev17, O. Vázquez Doce43, V. Vechernin27, E. Vercellin46, S. Vergara Limón84, L. Vermunt58,
R. Vernet146, R. Vértesi47, L. Vickovic113, Z. Vilakazi80, O. Villalobos Baillie107, G. Vino91, A.
Vinogradov17, T. Virgili99, V. Vislavicius56, A. Vodopyanov53, B. Volkel6, M. A. Völkl143, K.
Voloshin13, S. A. Voloshin52, G. Volpe75, B. von Haller6, I. Vorobyev43, D. Voscek120, J.
Vrláková72, B. Wagner25, M. Weber78, S. G. Weber29, A. Wegrzynek6, S. C. Wenzel6, J. P.
Wessels29, J. Wiechula26, J. Wikne36, G. Wilk102, J. Wilkinson8,9, G. A. Willems29, E. Willsher107,
B. Windelband30, M. Winn41, W. E. Witt63, J. R. Wright68, Y. Wu147, R. Xu19, S. Yalcin122, Y.
Yamaguchi144, K. Yamakawa144, S. Yang25, S. Yano41, Z. Yin19, H. Yokoyama58, I.-K. Yoo127, J. H.
Yoon59, S. Yuan25, A. Yuncu30, V. Yurchenko24, V. Zaccolo85, A. Zaman21, C. Zampolli6, H. J. C.
Zanoli58, N. Zardoshti6, A. Zarochentsev27, P. Závada134, N. Zaviyalov77, H. Zbroszczyk97, M.
Zhalov62, S. Zhang15, X. Zhang19, Z. Zhang19, V. Zherebchevskii27, Y. Zhi132, D. Zhou19, Y. Zhou56,
Z. Zhou25, J. Zhu14,19, Y. Zhu19, A. Zichichi8,22, G. Zinovjev24 & N. Zurlo73
1
Variable Energy Cyclotron Centre, Homi Bhabha National Institute, Kolkata, India. 2Nuclear
Physics Institute of the Czech Academy of Sciences, Řež, Czech Republic. 3Institut für
Informatik, Fachbereich Informatik und Mathematik, Johann-Wolfgang-Goethe Universität
Frankfurt, Frankfurt am Main, Germany. 4Division of Particle Physics, Department of Physics,
Lund University, Lund, Sweden. 5Physics Department, Panjab University, Chandigarh, India.
6
European Organization for Nuclear Research (CERN), Geneva, Switzerland. 7Dipartimento
DISAT del Politecnico and Sezione INFN, Turin, Italy. 8Centro Fermi – Museo Storico della
Fisica e Centro Studi e Ricerche “Enrico Fermi”, Rome, Italy. 9INFN, Sezione di Bologna,
Bologna, Italy. 10Department of Physics, Aligarh Muslim University, Aligarh, India. 11Korea
Nature | Vol 588 | 10 December 2020 | 237
Article
Institute of Science and Technology Information, Daejeon, Republic of Korea. 12Indonesian
Institute of Sciences, Jakarta, Indonesia. 13NRC «Kurchatov» Institute – ITEP, Moscow, Russia.
14
Research Division and ExtreMe Matter Institute EMMI, GSI Helmholtzzentrum für
Schwerionenforschung, Darmstadt, Germany. 15Fudan University, Shanghai, China.
16
Universidade Estadual de Campinas (UNICAMP), Campinas, Brazil. 17National Research
Centre Kurchatov Institute, Moscow, Russia. 18INFN, Sezione di Torino, Turin, Italy. 19Central
China Normal University, Wuhan, China. 20Instituto de Fsica, Universidad Nacional Autónoma
de México, Mexico City, Mexico. 21COMSATS University Islamabad, Islamabad, Pakistan.
22
Dipartimento di Fisica e Astronomia dell’Università and Sezione INFN, Bologna, Italy.
23
University of Houston, Houston, TX, USA. 24Bogolyubov Institute for Theoretical Physics,
National Academy of Sciences of Ukraine, Kiev, Ukraine. 25Department of Physics and
Technology, University of Bergen, Bergen, Norway. 26Institut für Kernphysik, Johann Wolfgang
Goethe-Universität Frankfurt, Frankfurt am Main, Germany. 27St. Petersburg State University,
St. Petersburg, Russia. 28Horia Hulubei National Institute of Physics and Nuclear Engineering,
Bucharest, Romania. 29Westfälische Wilhelms – Universität Münster, Institut für Kernphysik,
Münster, Germany. 30Physikalisches Institut, Ruprecht-Karls-Universität Heidelberg,
Heidelberg, Germany. 31Creighton University, Omaha, NB, USA. 32Rudjer Bošković Institute,
Zagreb, Croatia. 33INFN, Sezione di Padova, Padova, Italy. 34Lawrence Berkeley National
Laboratory, Berkeley, CA, USA. 35SUBATECH, IMT Atlantique, Université de Nantes,
CNRS-IN2P3, Nantes, France. 36Department of Physics, University of Oslo, Oslo, Norway.
37
Laboratoire de Physique des 2 Infinis, Irène Joliot-Curie, Orsay, France. 38INFN, Sezione di
Catania, Catania, Italy. 39Gangneung-Wonju National University, Gangneung, Republic of
Korea. 40Physics Department, University of Jammu, Jammu, India. 41Départment de Physique
Nucléaire (DPhN), Université Paris-Saclay Centre d’Etudes de Saclay (CEA), IRFU, Saclay,
France. 42Helmholtz-Institut für Strahlen- und Kernphysik, Rheinische
Friedrich-Wilhelms-Universität Bonn, Bonn, Germany. 43Physik Department, Technische
Universität München, Munich, Germany. 44Bose Institute, Department of Physics and Centre
for Astroparticle Physics and Space Science (CAPSS), Kolkata, India. 45Dipartimento di Fisica e
Astronomia dell’Università degli Studi di Catania and Sezione INFN, Catania, Italy.
46
Dipartimento di Fisica dell’Università degli studi di Torino and Sezione INFN, Turin, Italy.
47
Wigner Research Centre for Physics, Budapest, Hungary. 48Nuclear Physics Group, STFC
Daresbury Laboratory, Daresbury, UK. 49Université Clermont Auvergne, CNRS/IN2P3, LPC,
Clermont-Ferrand, France. 50University of Liverpool, Liverpool, UK. 51Dipartimento di Fisica e
Astronomia dell’Università degli Studi di Padova and Sezione INFN, Padua, Italy. 52Wayne
State University, Detroit, MI, USA. 53Joint Institute for Nuclear Research (JINR), Dubna, Russia.
54
Indian Institute of Technology Bombay (IIT), Mumbai, India. 55Sección Fsica, Departamento
de Ciencias, Pontificia Universidad Católica del Perú, Lima, Peru. 56Niels Bohr Institute,
University of Copenhagen, Copenhagen, Denmark. 57Yale University, New Haven, CT, USA.
58
Institute for Subatomic Physics, Utrecht University/Nikhef, Utrecht, Netherlands. 59Inha
University, Incheon, Republic of Korea. 60Université de Strasbourg, CNRS, IPHC UMR 7178,
Strasbourg, France. 61NRNU Moscow Engineering Physics Institute, Moscow, Russia.
62
Petersburg Nuclear Physics Institute, Gatchina, Russia. 63University of Tennessee, Knoxville,
TN, USA. 64Institute of Space Science (ISS), Bucharest, Romania. 65Gauhati University,
Department of Physics, Guwahati, India. 66INFN, Laboratori Nazionali di Frascati, Frascati, Italy.
67
Faculty of Nuclear Sciences and Physical Engineering, Czech Technical University in Prague,
Prague, Czech Republic. 68The University of Texas at Austin, Austin, TX, USA. 69Università degli
Studi di Pavia, Pavia, Italy. 70Oak Ridge National Laboratory, Oak Ridge, TN, USA.
71
Dipartimento di Fisica dell’Università degli studi di Bari Aldo Moro and Sezione INFN,
Cagliari, Italy. 72Faculty of Science, P.J. Šafárik University, Košice, Slovakia. 73Università di
Brescia, Brescia, Italy. 74Universidade de São Paulo (USP), São Paulo, Brazil. 75Dipartimento
Interateneo di Fisica ‘M. Merlin’, Università degli studi di Bari Aldo Moro and Sezione INFN,
Bari, Italy. 76Politecnico di Bari, Bari, Italy. 77Russian Federal Nuclear Center (VNIIEF), Sarov,
Russia. 78Stefan Meyer Institut für Subatomare Physik (SMI), Vienna, Austria. 79iThemba LABS,
National Research Foundation, Somerset West, South Africa. 80University of the
Witwatersrand, Johannesburg, South Africa. 81The Henryk Niewodniczanski Institute of
Nuclear Physics, Polish Academy of Sciences, Krakow, Poland. 82Nikhef, National Institute for
Subatomic Physics, Amsterdam, Netherlands. 83Universidad Autónoma de Sinaloa, Culiacán,
238 | Nature | Vol 588 | 10 December 2020
Mexico. 84High Energy Physics Group, Universidad Autónoma de Puebla, Puebla, Mexico.
85
Dipartimento di Fisica dell’Università degli studi di Trieste and Sezione INFN, Trieste, Italy.
86
INFN, Sezione di Cagliari, Cagliari, Italy. 87Institute of Nuclear Physics, Homi Bhabha National
Institute, Kolkata, India. 88Université de Lyon, Université Lyon 1, CNRS/IN2P3, IPN-Lyon, Lyon,
France. 89University of Tsukuba, Tsukuba, Japan. 90University of Cape Town, Cape Town, South
Africa. 91INFN, Sezione di Bari, Bari, Italy. 92Laboratoire de Physique Subatomique et de
Cosmologie, Université Grenoble-Alpes, CNRS-IN2P3, Grenoble, France. 93INFN, Sezione di
Trieste, Trieste, Italy. 94Dipartimento di Scienze e Innovazione Tecnologica dell’Università del
Piemonte Orientale and INFN Sezione di Torino, Alessandria, Italy. 95Universidade Federal do
ABC, Santo Andre, Brazil. 96Instituto de Ciencias Nucleares, Universidad Nacional Autónoma
de México, Mexico City, Mexico. 97Warsaw University of Technology, Warsaw, Poland.
98
National Institute of Science Education and Research, Homi Bhabha National Institute, Jatni,
India. 99Dipartimento di Fisica ‘E.R. Caianiello’ dell’Università and Gruppo Collegato INFN,
Salerno, Italy. 100Frankfurt Institute for Advanced Studies, Johann Wolfgang
Goethe-Universität Frankfurt, Frankfurt am Main, Germany. 101Indian Institute of Technology
Indore, Indore, India. 102National Centre for Nuclear Research, Warsaw, Poland. 103Centro de
Aplicaciones Tecnológicas y Desarrollo Nuclear (CEADEN), Havana, Cuba. 104Department of
Physics, University of California, Berkeley, CA, USA. 105Institute for Nuclear Research,
Academy of Sciences, Moscow, Russia. 106Physics Department, Faculty of Science, University
of Zagreb, Zagreb, Croatia. 107School of Physics and Astronomy, University of Birmingham,
Birmingham, UK. 108NRC Kurchatov Institute IHEP, Protvino, Russia. 109Department of Physics,
School of Science, National and Kapodistrian University of Athens, Athens, Greece. 110Chicago
State University, Chicago, IL, USA. 111National Nuclear Research Center, Baku, Azerbaijan.
112
Instituto de Física, Universidade Federal do Rio Grande do Sul (UFRGS), Porto Alegre, Brazil.
113
Faculty of Electrical Engineering, Mechanical Engineering and Naval Architecture,
University of Split, Split, Croatia. 114A.I. Alikhanyan National Science Laboratory (Yerevan
Physics Institute) Foundation, Yerevan, Armenia. 115University of Tokyo, Tokyo, Japan.
116
Nagasaki Institute of Applied Science, Nagasaki, Japan. 117Faculty of Engineering and
Science, Western Norway University of Applied Sciences, Bergen, Norway. 118Centro de
Investigación y de Estudios Avanzados (CINVESTAV), Mexico City, Mexico. 119Ohio State
University, Columbus, OH, USA. 120Technical University of Košice, Košice, Slovakia. 121Institute
of Experimental Physics, Slovak Academy of Sciences, Košice, Slovakia. 122KTO Karatay
University, Konya, Turkey. 123Zentrum für Technologietransfer und Telekommunikation (ZTT),
Hochschule Worms, Worms, Germany. 124Yonsei University, Seoul, Republic of Korea.
125
University of Jyväskylä, Jyväskylä, Finland. 126Jeonbuk National University, Jeonju, Republic
of Korea. 127Department of Physics, Pusan National University, Pusan, Republic of Korea.
128
Department of Physics, Sejong University, Seoul, Republic of Korea. 129California
Polytechnic State University, San Luis Obispo, CA, USA. 130Suranaree University of
Technology, Nakhon Ratchasima, Thailand. 131University of South-Eastern Norway, Tonsberg,
Norway. 132China Institute of Atomic Energy, Beijing, China. 133Dipartimento di Scienze MIFT,
Università di Messina, Messina, Italy. 134Institute of Physics of the Czech Academy of Sciences,
Prague, Czech Republic. 135Università degli Studi di Foggia, Foggia, Italy. 136INFN, Sezione di
Roma, Rome, Italy. 137Dipartimento di Fisica dell’Università ‘La Sapienza’ and Sezione INFN,
Rome, Italy. 138Faculty of Mathematics, Physics and Informatics, Comenius University
Bratislava, Bratislava, Slovakia. 139Budker Institute for Nuclear Physics, Novosibirsk, Russia.
140
Physics Department, University of Rajasthan, Jaipur, India. 141Helsinki Institute of Physics
(HIP), Helsinki, Finland. 142Institute of Physics, Homi Bhabha National Institute, Bhubaneswar,
India. 143Physikalisches Institut, Eberhard-Karls-Universität Tübingen, Tübingen, Germany.
144
Hiroshima University, Hiroshima, Japan. 145Nara Women’s University (NWU), Nara, Japan.
146
Centre de Calcul de l’IN2P3, Lyon, France. 147University of Science and Technology of China,
Hefei, China. 148Present address: Italian National Agency for New Technologies, Energy and
Sustainable Economic Development (ENEA), Bologna, Italy. 149Present address: Dipartimento
DET, Politecnico di Torino, Turin, Italy. 150Present address: D.V. Skobeltsyn Institute of Nuclear
Physics, M.V. Lomonosov Moscow State University, Moscow, Russia. 151Present address:
Department of Applied Physics, Aligarh Muslim University, Aligarh, India. 152Present address:
Institute of Theoretical Physics, University of Wrocław, Wrocław, Poland. 153Deceased: R. D.
Majka. ✉e-mail: alice-publications@cern.ch
Methods
Event selection
Events were recorded from inelastic p–p collisions by ALICE50,51 at
the LHC. A trigger that requires the total signal amplitude measured
in the V0 detector52 to exceed a certain threshold was used to select
high-multiplicity (HM) events. The V0 detector comprises two plas-
tic scintillator arrays placed on both sides of the interaction point at
pseudorapidities 2.8 < η < 5.1 and −3.7 < η < −1.7. The pseudorapidity is
defined as η = −ln[tan(θ/2)], where θ is the polar angle of the particle
with respect to the proton beam axis.
At s = 13 TeV, in the HM events, 30 charged particles in the range |η|
< 0.5 are produced on average. This η range corresponds to the region
within 26 degrees of the transverse plane that is perpendicular to the
beam axis. The HM events are rare, constituting 0.17% of the p–p colli-
sions that produce at least one charged particle in the pseudorapidity
range |η| < 1.0. It was shown38 that HM events contain an enhanced yield
of hyperons, which facilitates this analysis. The yield of Ω− in HM events
is at least a factor 5 larger, on average, compared with that in total inelas-
tic collisions53. A total of 1 × 109 HM events were analysed. Additional
details on the HM event selection can be found in a previous work12.
Particle tracking and identification
For the identification and momentum measurement of charged par-
ticles, the Inner Tracking System (ITS)54, Time Projection Chamber
(TPC)55, and Time-Of-Flight (TOF)56 detectors of ALICE are used. All
three detectors are located inside a solenoid magnetic field (0.5 T)
leading to a bending of the trajectories of charged particles. The meas-
urement of the curvature is used to reconstruct the particle momenta.
Typical transverse momentum (pT) resolutions for protons, pions and
kaons vary from about 2% for tracks with pT = 10 GeV/c to below 1% for
pT < 1 GeV/c. The particle identity is determined by the energy lost per
unit of track length inside the TPC detector and, in some cases, by the
particle velocity measured in the TOF detector. Additional experimental
details are discussed in a previous work51.
Protons are selected within a transverse momentum range of 0.5
< pT < 4.05 GeV/c. They are identified requiring TPC information for
candidate tracks with momentum p < 0.75 GeV/c, whereas TPC and TOF
information are both required for candidates with p > 0.75 GeV/c. An
incorrect identification of primary protons occurs in 1% of the cases,
as evaluated by Monte Carlo simulations.
Direct tracking and identification is not possible for Ξ− and Ω− hyper-
ons and their antiparticles, because they are unstable and decay as a
result of the weak interaction within a few centimetres after their pro-
duction. The mean decay distances (evaluated as c × τ, where τ is the
+
¯+) → Λ(Λ¯) + K −(K +)
particle lifetime) of Ξ −(Ξ¯ ) → Λ(Λ¯) + π −(π +) and Ω−(Ω
57
are 4.9 and 2.5 cm, respectively . Both decays are followed by a second
decay of the unstable Λ(Λ¯) hyperon, Λ(Λ¯) → p(p¯) + π −(π +), with an aver-
age decay path of 7.9 cm (ref. 57). Consequently, pions (π±), kaons (Κ±)
+
and protons have to be detected and then combined to search for Ξ −(Ξ¯ )
− ¯+
and Ω (Ω ) candidates. Those secondary particles are identified by
+
the TPC information in the case of the reconstruction of Ξ −(Ξ¯ ), and in
− ¯+
the case of Ω (Ω ) it is additionally required that the secondary protons
+
and kaons are identified in the TOF detector. To measure the Ξ −(Ξ¯ )
− ¯+
and Ω (Ω ) hyperons, the two successive weak decays need to be recon-
structed. The reconstruction procedure is very similar for both hyper-
ons and is described in detail previously58. Topological selections
are performed to reduce the combinatorial background, evaluated
via a fit to the invariant mass distribution.
Determination of the source size
The widths of the Gaussian distributions constituting S(r*), and defin-
ing the source size, are calculated on the basis of the results of the
analysis of the p–p correlation function in p–p collisions at s = 13 TeV
by the ALICE collaboration15. Assuming a common source for all
baryons, its size was studied as a function of the transverse mass of the
2
baryon–baryon pair, mT = (k T + m 2)1/2, where m is the average mass and
kT = |pT,1 + pT,2|/2 is the transverse momentum of the pair. The source
size decreases with increasing mass, which could reflect the collective
evolution of the system. The average transverse mass ⟨mT⟩ for the p–Ξ−
and p–Ω− pairs differ and are equal to 1.9 GeV/c and 2.2 GeV/c, respec-
tively. To determine the source sizes for these values, the measurement
from p–p correlations (shown in figure 5 of ref. 15) is parameterized as
rcore = ambT + c, where rcore denotes the width of the Gaussian distribution
defining the source before taking into account the effect produced by
short lived resonances.
In p–p collisions at s = 13 TeV , Ξ− and Ω− baryons are produced
mostly as primary particles, but about 2/3 of the protons originate from
the decay of short-lived resonances with a lifetime of a few fm per c.
As a result, the effective source size of both p–Ξ− and p–Ω− is modified.
This effect is taken into account by folding the Gaussian source with
an exponential distribution following the method outlined previously15.
The resulting source distribution can be characterized by an effective
Gaussian source radius equal to 1.02 ± 0.05 fm for p–Ξ− pairs and to
0.95 ± 0.06 fm for p–Ω− pairs. The quoted uncertainties correspond to
variations of the parametrization of the p–p results according to their
systematic and statistical uncertainties.
Corrections of the correlation function
The correction factor ξ(k*) accounts for the normalization of the k*
distribution of pairs from mixed-events, for effects produced by finite
momentum resolution and for the influence of residual correlations.
The mixed-event distribution, Nmixed(k*), has to be scaled down,
because the number of pairs available from mixed events is much higher
than the number of pairs produced in the same collision used
in Nsame(k*). The normalization parameter N is chosen such that the
mean value of the correlation function equals to unity in a region of
k* values where the effect of final-state interactions are negligible,
500 < k* < 800 MeV/c.
The finite experimental momentum resolution modifies the meas-
ured correlation functions at most by 8% at low k*. A correction for this
effect is applied. Resolution effects due to the merging of tracks that
are very close to each other were evaluated and found to be negligible.
The two measured correlation functions are dominated by the con-
tribution of the interaction between p–Ξ− and p–Ω− pairs. Nevertheless,
other contributions also influence the measured correlation function.
They originate either from incorrectly identified particles or from par-
ticles stemming from other weak decays (such as protons from Λ → p +
π− decays) combined with primary particles. Because weak decays occur
typically some centimetres away from the collision vertex, there is no
final-state interaction between their decay products and the primary
particles of interest. Hence, the resulting correlation function either
will be completely flat or will carry the residual signature of the interac-
tion between the particles before the decay. A method to determine
the exact shape and relative yields of the residual correlations has been
previously developed8,59, and it is used in this analysis. Such contribu-
tions are subtracted from the measured p–Ξ− and p–Ω− correlations
to obtain the genuine correlation functions. The residual correlation
stemming from misidentification is evaluated experimentally11 and its
contribution is also subtracted from the measured correlation function.
The systematic uncertainties associated with the genuine correlation
function arise from the following sources: (i) the selection of the pro-
+
ton, Ξ −(Ξ¯ ) and Ω−(Ω¯+), (ii) the normalization of the mixed-event dis-
tributions, (iii) uncertainties on the residual contributions, and (iv)
uncertainties due to the finite momentum resolution. To evaluate the
associated systematic uncertainties: (i) all single-particle and topo-
logical selection criteria are varied with respect to their default values
and the analysis is repeated for 50 different random combinations of
such selection criteria so that the maximum change introduced in the
number of p–Ξ− and p–Ω− pairs is 25% and the changes in the purity of
Article
+ +
protons, Ξ −(Ξ¯ ) and Ω−(Ω¯ ) are kept below 3%; (ii) the k*-normalization
range of the mixed-events is varied, and a linear function of k* is also
used for an alternative normalization which results in an asymmetric
uncertainty; (iii) the shape of the residual correlations and its relative
contribution are altered; and (iv) the momentum resolution and the
used correction method are changed. The total systematic uncertain-
ties associated with the genuine correlation function are maximal at
low k*, reaching a value of 9% and 8% for p–Ξ− and p–Ω−, respectively.
HAL QCD potentials
Results from calculations by the HAL QCD Collaboration for the p–Ξ−14
and p–Ω−13 interactions are shown in Figs. 3, 4. Such interactions were
studied via (2 + 1)-flavor lattice QCD simulations with nearly physical
quark masses (mπ = 146 MeV/c2).
In Fig. 4, the p–Ξ− and p–Ω− potentials are shown for calculations
with t/a = 12, with t the Euclidean time and a the lattice spacing of the
calculations. The HAL QCD Collaboration provided 23 and 20 sets of
parameters for the description of the shape of the p–Ξ− and p–Ω− poten-
tials, respectively. Such parametrizations result from applying the
jackknife method, which takes into account the statistical uncertainty
of the calculations. The width of the curves in Fig. 4 corresponds to
the maximum variations observed in the potential shape by using the
different sets of parameters.
To obtain the correlation functions shown in Fig. 3 we consider the
calculations with t/a = 12, both for p–Ξ− and p–Ω−. The statistical uncer-
tainty of the calculations is evaluated using the jackknife variations,
and a systematic uncertainty is added in quadrature evaluated by con-
sidering calculations with t/a = 11 and t/a = 13.
Data availability
All data shown in histograms and plots are publicly available on the
HEPdata repository (https://hepdata.net).
Code availability
The source code used in this study is publicly available under the names
AliPhysics (https://github.com/alisw/AliPhysics) and AliRoot (https://
github.com/alisw/AliROOT). Further information can be provided by
the authors upon reasonable request.
50. ALICE Collaboration. The ALICE experiment at the CERN LHC. JINST 3, S08002 (2008).
51. ALICE Collaboration. Performance of the ALICE experiment at the CERN LHC. Int. J. Mod.
Phys. A 29, 1430044 (2014).
52. ALICE Collaboration. Performance of the ALICE VZERO system. JINST 8, P10016 (2013).
53. ALICE Collaboration. Multiplicity dependence of (multi-)strange hadron production in
proton-proton collisions at √s = 13 TeV. Eur. Phys. J. C 80, 167 (2020).
54. ALICE Collaboration. Alignment of the ALICE Inner Tracking System with cosmic-ray
tracks. JINST 5, P03003 (2010).
55. Alme, J. et al. The ALICE TPC, a large 3-dimensional tracking device with fast readout for
ultra-high multiplicity events. Nucl. Instrum. Meth. A 622, 316–367 (2010).
56. Akindinov, A. et al. Performance of the ALICE Time-Of-Flight detector at the LHC. Eur.
Phys. J. Plus 128, 44 (2013).
57. Particle Data Group Collaboration. Review of particle physics. Phys. Rev. D 98, 030001
(2018).
58. ALICE Collaboration. Strange particle production in proton–proton collisions at √s = 0.9
TeV with ALICE at the LHC. Eur. Phys. J. C 71, 1594 (2011).
59. Kisiel, A., Zbroszczyk, H. & Szymański, M. Extracting baryon–antibaryon
strong-interaction potentials from pΛ femtoscopic correlation functions. Phys. Rev. C 89,
054916 (2014).
Acknowledgements We are grateful to T. Hatsuda and K. Sasaki from the HAL QCD
Collaboration for their valuable suggestions and for providing the lattice QCD results
regarding the p–Ξ− and p–Ω− interactions. We are also grateful to A. Ohnishi, T. Hyodo, T. Iritani,
Y. Kamiya and T. Sekihara for their suggestions and discussions. We thank all the engineers and
technicians of the LHC for their contributions to the construction of the experiment and the
CERN accelerator teams for the performance of the LHC complex. We acknowledge the
resources and support provided by all Grid centres and the Worldwide LHC Computing Grid
(WLCG) collaboration. We acknowledge the following funding agencies for their support in
building and running the ALICE detector: A. I. Alikhanyan National Science Laboratory
(Yerevan Physics Institute) Foundation (ANSL), State Committee of Science and World
Federation of Scientists (WFS), Armenia; Austrian Academy of Sciences, Austrian Science Fund
(FWF): [M 2467-N36] and Nationalstiftung für Forschung, Technologie und Entwicklung,
Austria; Ministry of Communications and High Technologies, National Nuclear Research
Center, Azerbaijan; Conselho Nacional de Desenvolvimento Cientfico e Tecnológico (CNPq),
Financiadora de Estudos e Projetos (Finep), Fundação de Amparo à Pesquisa do Estado de São
Paulo (FAPESP) and Universidade Federal do Rio Grande do Sul (UFRGS), Brazil; Ministry of
Education of China (MOEC), Ministry of Science and Technology of China (MSTC) and National
Natural Science Foundation of China (NSFC), China; Ministry of Science and Education and
Croatian Science Foundation, Croatia; Centro de Aplicaciones Tecnológicas y Desarrollo
Nuclear (CEADEN), Cubaenerga, Cuba; Ministry of Education, Youth and Sports of the Czech
Republic, Czech Republic; The Danish Council for Independent Research | Natural Sciences,
the VILLUM FONDEN and Danish National Research Foundation (DNRF), Denmark; Helsinki
Institute of Physics (HIP), Finland; Commissariat à l’Energie Atomique (CEA) and Institut
National de Physique Nucléaire et de Physique des Particules (IN2P3) and Centre National de la
Recherche Scientifique (CNRS), France; Bundesministerium für Bildung und Forschung (BMBF)
and GSI Helmholtzzentrum für Schwerionenforschung GmbH, Germany; General Secretariat
for Research and Technology, Ministry of Education, Research and Religions, Greece; National
Research, Development and Innovation Office, Hungary; Department of Atomic Energy
Government of India (DAE), Department of Science and Technology, Government of India
(DST), University Grants Commission, Government of India (UGC) and Council of Scientific and
Industrial Research (CSIR), India; Indonesian Institute of Science, Indonesia; Centro Fermi –
Museo Storico della Fisica e Centro Studi e Ricerche Enrico Fermi and Istituto Nazionale di
Fisica Nucleare (INFN), Italy; Institute for Innovative Science and Technology, Nagasaki
Institute of Applied Science (IIST), Japanese Ministry of Education, Culture, Sports, Science
and Technology (MEXT) and Japan Society for the Promotion of Science (JSPS) KAKENHI,
Japan; Consejo Nacional de Ciencia (CONACYT) y Tecnología, through Fondo de Cooperación
Internacional en Ciencia y Tecnología (FONCICYT) and Dirección General de Asuntos del
Personal Academico (DGAPA), Mexico; Nederlandse Organisatie voor Wetenschappelijk
Onderzoek (NWO), Netherlands; The Research Council of Norway, Norway; Commission on
Science and Technology for Sustainable Development in the South (COMSATS), Pakistan;
Pontificia Universidad Católica del Perú, Peru; Ministry of Science and Higher Education,
National Science Centre and WUT ID-UB, Poland; Korea Institute of Science and Technology
Information and National Research Foundation of Korea (NRF), Republic of Korea; Ministry of
Education and Scientific Research, Institute of Atomic Physics and Ministry of Research and
Innovation and Institute of Atomic Physics, Romania; Joint Institute for Nuclear Research
(JINR), Ministry of Education and Science of the Russian Federation, National Research Centre
Kurchatov Institute, Russian Science Foundation and Russian Foundation for Basic Research,
Russia; Ministry of Education, Science, Research and Sport of the Slovak Republic, Slovakia;
National Research Foundation of South Africa, South Africa; Swedish Research Council (VR)
and Knut & Alice Wallenberg Foundation (KAW), Sweden; European Organization for Nuclear
Research, Switzerland; Suranaree University of Technology (SUT), National Science and
Technology Development Agency (NSDTA) and Office of the Higher Education Commission
under NRU project of Thailand, Thailand; Turkish Atomic Energy Agency (TAEK), Turkey;
National Academy of Sciences of Ukraine, Ukraine; Science and Technology Facilities Council
(STFC), United Kingdom; National Science Foundation of the United States of America (NSF)
and United States Department of Energy, Office of Nuclear Physics (DOE NP), United States of
America.
Author contributions All authors have contributed to the publication, being variously involved
in the design and the construction of the detectors, in writing software, calibrating subsystems,
operating the detectors and acquiring data, and finally analysing the processed data. The ALICE
Collaboration members discussed and approved the scientific results. The manuscript was
prepared by a subgroup of authors appointed by the collaboration and subject to an internal
collaboration-wide review process. All authors reviewed and approved the final version of the
manuscript.
Competing interests The authors declare no competing interests.
Additional information
Correspondence and requests for materials should be addressed to L.M.
Peer review information Nature thanks Kazuya Aoki, Andrzej Kupsc, Manuel Lorenz and the
other, anonymous, reviewer(s) for their contribution to the peer review of this work.
Reprints and permissions information is available at http://www.nature.com/reprints.
Article

Dipolar evaporation of reactive molecules to

below the Fermi temperature

https://doi.org/10.1038/s41586-020-2980-7 Giacomo Valtolina1,2 ✉, Kyle Matsuda1,2, William G. Tobias1,2, Jun-Ru Li1,2,

Luigi De Marco1,2 & Jun Ye1,2 ✉
Received: 20 July 2020

Accepted: 29 September 2020

The control of molecules is key to the investigation of quantum phases, in which rich
Published online: 9 December 2020
degrees of freedom can be used to encode information and strong interactions can be
Check for updates
precisely tuned1. Inelastic losses in molecular collisions2–5, however, have greatly
hampered the engineering of low-entropy molecular systems6. So far, the only
quantum degenerate gas of molecules has been created via association of two highly
degenerate atomic gases7,8. Here we use an external electric field along with optical
lattice confinement to create a two-dimensional Fermi gas of spin-polarized
potassium–rubidium (KRb) polar molecules, in which elastic, tunable dipolar
interactions dominate over all inelastic processes. Direct thermalization among the
molecules in the trap leads to efficient dipolar evaporative cooling, yielding a rapid
increase in phase-space density. At the onset of quantum degeneracy, we observe the
effects of Fermi statistics on the thermodynamics of the molecular gas. These results
demonstrate a general strategy for achieving quantum degeneracy in dipolar
molecular gases in which strong, long-range and anisotropic dipolar interactions can
drive the emergence of exotic many-body phases, such as interlayer pairing and
p-wave superfluidity.

The complex internal structure of molecules can be both useful and a
hindrance: it represents a key resource for the development of tunable
and programmable quantum devices1,9,10, but it is also responsible for
strong inelastic losses during collisions11–14. Despite recent advances in
molecular quantum science15–24, full control of elastic collisions between
molecules has not been achieved, making it very difficult to create the
low-entropy bulk molecular gases that are required for the exploration
of rich many-body physics and emergent quantum phenomena1,25.
Here, we report the realization of highly tunable elastic interactions
in a quantum gas of polar molecules through the application of an
external electric field along a stack of two-dimensional (2D) layers
generated with a one-dimensional optical lattice. The induced electric
dipole moment in the laboratory frame gives rise to repulsive dipolar
interactions that stabilize the molecular gas against reactive collisions
and formation of collisional complexes. These long-range interactions
provide a large elastic collision cross-section for identical ultracold
fermionic molecules, in contrast to contact interactions26. We dem-
onstrate the enhancement of dipolar interactions by several orders
of magnitude and achieve a ratio of elastic-to-inelastic collisions that
exceeds 100. This favourable interaction regime enables direct molecu-
lar thermalization and efficient evaporative cooling, allowing us to
bring the molecular temperature T below the Fermi temperature TF.
The onset of quantum degeneracy is signalled by deviations from the
classical expansion energy as the ratio T/TF is reduced below unity7,27.
Our strategy follows previous theory proposals28–30 and our earlier
experimental study on molecular reactions in quasi-two dimensions31.
This geometry allows us to take advantage of the anisotropic charac-
ter of the dipolar potential and retain only the repulsive side-to-side
dipole–dipole interactions within each 2D site, while preventing the
attractive head-to-tail interactions that facilitate losses at short range.
Our recent advances in the production of degenerate Fermi gases of
polar molecules7,8, combined with precise electric-field control using
in-vacuum electrodes32 (Fig. 1), allow us to perform a systematic char-
acterization of the properties of a 2D Fermi gas of polar molecules.
A long-lived 2D Fermi gas of polar molecules
The KRb 2D Fermi gas is created from an ultracold atomic mixture of
fermionic 40K and bosonic 87Rb atoms. The atomic mixture is initially
held in a crossed optical dipole trap (ODT) and then transferred into a
single layer of a large-spacing lattice (LSL) with an 8-μm spatial period,
which increases the mixture’s confinement along the vertical direction
(y). The mixture is then transferred into a vertical lattice (VL) with spac-
ing of 540 nm that confines it to a quasi-2D geometry. The intermediate
LSL transfer results in the Rb cloud populating a controllable number
of VL layers τ ranging between 5 and 15. We directly probe the number
of occupied 2D layers via a matter-wave focusing technique on the
Rb cloud (Fig. 1c)33,34. The measured τ is in excellent agreement with
theoretical modelling of the in situ cloud size (see Methods).
Magneto-optical association is used to pair roughly half of the initial
Rb atoms into ground-state KRb molecules35. This process is fast and
coherent, and the resulting molecular cloud populates the same lay-
ers originally occupied by the Rb cloud. The leftover K and Rb atoms
are selectively and quickly removed from the trap. In the VL, the trap
frequencies are set to (ωx, ωy, ωz) = 2π × (40, 17,000, 40) Hz. The quoted
trapping frequencies are for KRb throughout the paper unless otherwise

JILA, National Institute of Standards and Technology, Boulder, CO, USA. 2Department of Physics, University of Colorado, Boulder, CO, USA. ✉e-mail: giva4289@colorado.edu; ye@jila.colorado.edu
1

Nature | Vol 588 | 10 December 2020 | 239

Article
a
a 6

n (×107 cm–2)
4

0
0 2 4 6 8 10
Time (s)

g
y b EDC (kV cm–1)
x
0.0 2.1 4.7 9.1
z 12

b c 9

E (×10–8 cm2 s–1)

+V

+JV 6

EDC
3
–JV

y –V 0
0.0 0.1 0.2 0.3
x
z Dipole moment (D)

Fig. 1 | Experimental setup. a, The 2D molecular cloud is trapped at the centre c 300 4.5
of the electrode assembly (grey). 2D optical trapping is achieved with the VL 4.0
(green), which is loaded using the ODT (orange) and LSL (red). Absorption
3.5
images of molecules are collected through the same lens as that used to

Heating rate (μK s–1)

200 3.0
E (×10–8 cm2 s–1)

focus the LSL. b, Sketch of the experiment as seen down the z axis. The bias
electric field is generated along y, perpendicular to the 2D layers of the VL.
0.5
c, Matter-wave focusing data of the Rb layers in the VL, which have a spacing of
540 nm. 100

stated. We create a 2D gas with N ≈ 20,000 trapped molecules, a typical
temperature T ≈ 250 nK, and T/TF ranging from 1.5 to 3 depending on τ.
The 2D molecular cloud is at the centre of an in-vacuum six-electrode
assembly composed of two indium tin oxide (ITO)-coated glass plates
and four tungsten rods (Fig. 1a). With this, we generate a highly tunable
bias electric field EDC that induces strong dipolar interactions between
molecules (Fig. 1b). The ratio γ of the voltage of the rods to the voltage
of the plates can be used to cancel the curvature introduced by the
parallel plate edges (flat-field configuration) or to introduce additional
curvatures and gradients for molecule manipulation.
The chemically reactive KRb molecules suffer from inelastic
two-body losses2,12, which result in the average molecular density n
decaying over time t according to a two-body rate equation of the form:
0 0.0
7 10 13 16
Zy /(2π) (kHz)
Fig. 2 | Long-lived polar molecules in 2D. a, Time evolution of the molecular
density n at d = 0.2 D. b, Inelastic loss rate β as a function of dipole moment. All
error bars are 1 standard error of the mean (s.e.), determined from two-body
decay fits (equation (1)) The top x axis shows the bias electric field EDC at the
corresponding dipole moment . c, Both β (grey circles) and the heating rate
(orange squares) saturate at their minimum values near ω y = 2π × 7 kHz (vertical
grey bar indicates uncertainty in the molecule temperature), consistent with
the mechanism of quasi-2D dipolar scattering. Heating rate error bars are 1 s.e.,
determined from linear fits.

dn ∂n dT for several seconds (Fig. 2a). The evolution of β as a function of EDC is

= − βn 2 + , (1)
dt ∂T dt
where β is the two-body loss rate coefficient and the second term on
the right side of equation (1) accounts for temperature changes affect-
ing the density3.
In a three-dimensional (3D) harmonic trap, β increases sharply with EDC,
so that inelastic interactions dominate elastic ones3,36. However, strong
confinement along the direction of EDC suppresses this detrimental loss
increase31 by preventing head-to-tail collisions along EDC. Even though
n is large in the occupied layers, the molecular gas shows a remarkable
stability with repulsive interactions turned on. With an induced dipole
moment d = 0.2 D in the flat-field configuration, KRb molecules survive
shown in Fig. 2b. Close to EDC = 4.7 kV cm−1, β reaches a minimum of nearly
five times below the zero-field value. The increase of β for d > 0.2 D is
consistent with a quasi-2D picture of dipolar scattering37,38.
To understand the effect of optical confinement, we perform a thor-
ough characterization of the 3D-to-2D crossover by measuring β versus
ωy at EDC = 5 kV cm−1 (Fig. 2c). Here, β drops abruptly as the lattice vertical
confinement is increased and reaches a plateau near ωy = 2π × 7 kHz,
corresponding to the quasi-2D limit where kBT ≱ ħωy, with kB being the
Boltzmann constant and ħ the reduced Planck constant, and the mol-
ecules principally occupy the lowest band of the VL. In contrast to the
3D case, where the heating rate exceeds 3 µK s−1, in quasi-two dimen-
sions we do not record a substantial increase in temperature along the

240 | Nature | Vol 588 | 10 December 2020

a b We diabatically change the power in one of the ODT beams to suddenly
0.6
increase the energy along z. Elastic collisions between the molecules
100
then redistribute the excess energy from z onto x. The rate Γth of the
0.5 temperature equilibration between the two axes is proportional to the

Γth (s–1)
Tx (μK)

dipolar elastic collision rate28,39,40. We extract Γth by fitting the increase

10
0.4
1
0.3
0 200 400 600 800 1,000 0.03
Time (ms)
Fig. 3 | Tuning strong dipolar elastic interactions in a 2D molecular gas.
a, Cross-dimensional thermalization dynamics at d = 0.1 D (orange diamonds)
and d = 0.21 D (grey squares). Error bars are 1 s.e. of 5 independent
measurements. b, The trend of Γth extracted from cross-dimensional
thermalization dynamics as a function of d. The solid line is a power-law fit
for d ≥ 0.1 D that yields a power of 3.3(1.0). The filled grey circle corresponds
to the measurement at d = 0.0 D, and it is artificially placed at d = 0.03 D for
figure clarity. The dashed horizontal line at Γth = 2 s−1 is the background
cross-dimensional relaxation from trap anharmonicity. All error bars are 1 s.e.,
determined from exponential fits.
radial direction. The suppression of heating is due to cancellation of
anti-evaporation in quasi-two dimensions30 and represents another
advantage of this configuration.
Cross-dimensional thermalization
To characterize elastic interactions in our thermal molecular cloud,
we perform cross-dimensional thermalization at various values of EDC.
of T along x with an exponential curve.
With the loss suppressed in quasi-2D, we expect a substantial increase
of the thermalization rate Γth with a d4 dependence28,30. Comparing the
thermalization dynamics observed for d = 0.1 D and 0.21 D (Fig. 3a), we
0.1 0.3
Dipole moment (D) see Γth increase by a factor of 10. Over our investigated range of EDC, Γth
changes by two orders of magnitude (Fig. 3b), showing the extreme
tunability of elastic dipolar interactions in our system. We observe the
cross-dimensional thermalization dynamics at lower dipolar strength
(d < 0.1 D) being dominated by cross-dimensional relaxation owing to
trap anharmonicity, which limits the smallest Γth that we can measure.
For d ≥ 0.1 D, a fit to a power-law dependence of Γth on d yields a power of
3.3 ± 1.0, in good agreement with theoretical expectations28,30. For the
highest values of d we explored, the rate Γth is comparable to the radial
trapping frequency, opening the way for future studies of collective
dynamics in molecular gases41.
An estimate of the ratio of elastic-to-inelastic collisions is obtained by
comparing Γth to the initial rate of inelastic losses Γin, which is expressed
as Γin = βn0, with n0 the initial average density of the 2D gas. From the
data in Fig. 2b, at d = 0.2 D, we estimate a rate Γin = 0.83(5) s−1, whereas
Γth = 21(6) s−1 for the same dipole strength. In the temperature regime of
the cross-dimensional thermalization experiments, theoretical mod-
els30 predict that α = 8 elastic collisions are needed for each molecule to
reach thermal equilibrium. This indicates a ratio of elastic-to-inelastic
collisions α(Γth/Γin) = 200 ± 60, demonstrating that elastic processes
dominate in this system.

a b c
10
Trap depth (μK)

4
0 0
3
N (×103)

–2
Sevap

–2
2
–200 0 200 –200 0 200
Position (μm) Position (μm)
Optical trap 1
Electric field 1
Combined 0.05 0.10 0.15 0.20 0.0 0.1 0.2 0.3
T (μK) Dipole moment (D)

d 0.3
e f
4 0.4
Integrated OD

3 0.3
0.2
δU/U

2 0.2

y 1 0.1
0.1

0 0.0
x
0.0 –100 –50 0 50 100 0.0 0.5 1.0 1.5 2.0
OD Position (μm) T/TF

Fig. 4 | Evaporative cooling to the quantum degenerate regime. a, Cuts
along the x axis of the combined electro-optical potential for the flat-field
configuration (left panel) and at the end of the evaporation (right panel).
b, Evolution of N and T (orange squares) at different stages of the evaporation
trajectory at EDC = 6.5 kV cm−1 and d = 0.25 D. The power-law fit (orange line)
yields Sevap = 1.06(15). The dashed grey line is for a constant T/TF, corresponding
to Sevap = 2.0. Error bars are 1 s.e. of four independent measurements. c, Summary
of Sevap versus d. All error bars are 1 s.e., determined from power-law fits.
d, Average of 20 band-mapped absorption images of the molecular cloud in the
x–y plane after 5.84 ms of time of flight for T/TF = 2.0(1) (top) and T/TF = 0.81(15)
(bottom). e, Optical density (OD, dimensionless) profiles (orange circles for
T/TF = 2.0(1) and grey diamonds for T/TF = 0.81(15)) of the images in d
(integrated along y), together with the Fermi–Dirac fit to the whole cloud (grey
line) and the Gaussian fit to the outer wings (orange line). f, Measurement of
δU/U at different values of T/TF from the Fermi–Dirac fit to the entire cloud
(grey circles) and from the Gaussian fit to the outer wings of the cloud (orange
squares). The solid and dashed curves show δU/U for the 2D and 3D ideal Fermi
gases, respectively. All error bars are 1 s.e., determined from Gaussian or
polylogarithmic fits.

Nature | Vol 588 | 10 December 2020 | 241

Article
Electric-field-controlled evaporative cooling
The large elastic-to-inelastic collision ratio is an excellent setting for
evaporatively cooling to enhance the phase-space density of our molec-
ular cloud. For non-degenerate 2D fermionic gases, phase-space density
is inversely proportional to (T/TF)2, and phase-space density increases
only if the change of N versus T during evaporation fulfills the criterion:
∂logN
Sevap = < 2. (2)
∂logT
When Sevap = 2, the gas maintains a constant T/TF.
The efficiency of evaporative cooling relies on our ability to selectively
remove the hottest molecules from the trap and to let the remaining mol-
ecules re-thermalize to a lower temperature. Reducing the trap depth
by lowering the optical trap power for evaporation, as is routinely done
in ultracold atom experiments, is not a viable solution here because we
cannot lower the tight 2D confinement without affecting the stability of
the cloud (Fig. 2c). Instead, by increasing γ with respect to the flat-field
configuration, we introduce a tunable anti-trapping electric field along
the x direction to reduce the radial confinement experienced by the
molecules (Fig. 4a). By measuring the change of ωx as a function of γ,
we can directly reconstruct the combined electro-optical potential
and benchmark its theoretical modelling (see Extended Data Fig. 2).
For the evaporation measurement, we start with a 2D gas with layer
number τ = 5 ± 1, ωy = 2π × 17 kHz, and an average T/TF = 1.5(1). After
creating the molecules (see Methods), we ramp EDC to a target field
while keeping γ at the flat-field value. We trigger the evaporation by
increasing γ to reduce the trap depth. We do not observe any evapo-
ration until the truncation parameter η, defined as the ratio of trap
depth over thermal energy kBT, reaches a value of 4 (see Methods), in
good agreement with theoretical expectations30. We further increase
γ over a timescale of hundreds of milliseconds, which is long enough
for the molecules to efficiently re-thermalize at a lower T as the trap
depth is reduced. At the end of the evaporation ramp, we return to
the flat-field configuration and ramp EDC back to its initial value. We
coherently convert the ground state molecules back to the Feshbach
state and image the cloud of Feshbach molecules after band-mapping
from the VL. Detailed time sequences for the evolution of EDC, γ and trap
depth are shown in the Methods.
At EDC = 6.5 kV cm−1, the evolution of N and T at different stages of the
optimized evaporation sequence is shown in Fig. 4b. To characterize
the evaporation efficiency, we fit the N versus T dependence with a
power-law function to extract Sevap. For the data shown in Fig. 4b, we
obtain Sevap = 1.06(15), far below the threshold of 2 required to increase
phase-space density. The trend of Sevap versus d is plotted in Fig. 4c
and reaches a minimum (that is, maximum increase in phase-space
density) at d = 0.25 D, where the ratio of elastic-to-inelastic collisions
is the largest37,38.
When we cool molecules to T < TF (Fig. 4d), we witness the onset of
Fermi degeneracy, which is signalled by deviations from classical ther-
modynamics owing to the increasing role of the Pauli exclusion prin-
1/2
ciple. Here,TF =
ħω R
kB ( )
2N
τ
, with ωR = ωx ωz the geometric mean of the
radial trapping frequency. Our best result produced a 2D molecular
Fermi gas with N = 1.7(1) × 103 and T/TF = 0.6(2).
We extract T by using either a fit to the Fermi–Dirac distribution on
the entire expanded cloud or a Gaussian fit restricted to the cloud’s
outer wings (see Methods). As shown in Fig. 4e, for T/TF = 0.81(15) the
Gaussian fit to the outer wings of the time of flight density profile
overestimates the density at the centre. We quantify this through the
increasing difference δU = U − Ucl between the energy U of the fermionic
gas and the energy Ucl ∝ kBT, as T/TF decreases7,27. U is determined from
a Gaussian fit to the whole cloud (see Methods) and Ucl is calculated
based on the measured T. Owing to the different density of states in the
harmonic trap, the chemical potential crosses zero for T/TF = 0.78 for
242 | Nature | Vol 588 | 10 December 2020
the 2D case, in contrast to 0.57 for the 3D case. Correspondingly, the
2D Fermi gas shows a larger δU with respect to the 3D case at the same
T/TF (ref. 42). As we reach T < TF, in excellent agreement with theoreti-
cal expectations, we observe a large increase of δU (Fig. 4f). This is a
hallmark for the onset of quantum degeneracy in trapped Fermi gases27.
Conclusions
We have realized a 2D Fermi gas of reactive polar molecules where
precisely tunable elastic dipolar interactions dominate all inelastic
processes. This allowed us to perform evaporative cooling of mol-
ecules to the onset of Fermi degeneracy. We demonstrated a general
and robust scheme for ultracold gases of polar molecules to reach
quantum degeneracy. For example, using a strong 2D confinement and
large dipolar interactions, this method should enable Bose–Einstein
condensation in bosonic molecular gases. It has long been anticipated
that quantum gases of polar molecules in two dimensions would allow
access to strongly correlated many-body phases43–52. Our results set
the stage for exploration of these exotic regimes.
Online content
Any methods, additional references, Nature Research reporting sum-
maries, source data, extended data, supplementary information,
acknowledgements, peer review information; details of author con-
tributions and competing interests; and statements of data and code
availability are available at https://doi.org/10.1038/s41586-020-2980-7.
1. Bohn, J. L., Rey, A. M. & Ye, J. Cold molecules: progress in quantum engineering of
chemistry and quantum matter. Science 357, 1002–1010 (2017).
2. Ospelkaus, S. et al. Quantum-state controlled chemical reactions of ultracold
potassium-rubidium molecules. Science 327, 853–857 (2010).
3. Ni, K.-K. et al. Dipolar collisions of polar molecules in the quantum regime. Nature 464,
1324–1328 (2010).
4. Guo, M. et al. Dipolar collisions of ultracold ground-state bosonic molecules. Phys. Rev. X
8, 041044 (2018).
5. Gregory, P. D. et al. Sticky collisions of ultracold RbCs molecules. Nat. Commun. 10, 3104
(2019).
6. Moses, S. A. et al. Creation of a low-entropy quantum gas of polar molecules in an optical
lattice. Science 350, 659–662 (2015).
7. De Marco, L. et al. A degenerate Fermi gas of polar molecules. Science 363, 853–856 (2019).
8. Tobias, W. G. et al. Thermalization and sub-Poissonian density fluctuations in a
degenerate molecular Fermi gas. Phys. Rev. Lett. 124, 033401 (2020).
9. André, A. et al. A coherent all-electrical interface between polar molecules and
mesoscopic superconducting resonators. Nat. Phys. 2, 636–642 (2006).
10. DeMille, D., Doyle, J. M. & Sushkov, A. O. Probing the frontiers of particle physics with
tabletop-scale experiments. Science 357, 990–994 (2017).
11. Gregory, P. D., Blackmore, J. A., Bromley, S. L. & Cornish, S. L. Loss of ultracold 87Rb133Cs
molecules via optical excitation of long-lived two-body collision complexes. Phys. Rev.
Lett. 124, 163402 (2020).
12. Hu, M. G. et al. Direct observation of bimolecular reactions of ultracold KRb molecules.
Science 366, 1111–1115 (2019).
13. Kirste, M. et al. Quantum-state resolved bimolecular collisions of velocity-controlled OH
with NO radicals. Science 338, 1060–1063 (2012).
14. Christianen, A., Zwierlein, M. W., Groenenboom, G. C. & Karman, T. Photoinduced
two-body loss of ultracold molecules. Phys. Rev. Lett. 123, 123402 (2019).
15. Danzl, J. G. et al. Quantum gas of deeply bound ground state molecules. Science 321,
1062–1066 (2008).
16. Seeßelberg, F. et al. Extending rotational coherence of interacting polar molecules in a
spin-decoupled magic trap. Phys. Rev. Lett. 121, 253401 (2018).
17. Will, S. A., Park, J. W., Yan, Z. Z., Loh, H. & Zwierlein, M. W. Coherent microwave control of
ultracold 23Na40K molecules. Phys. Rev. Lett. 116, 225306 (2016).
18. Barry, J. F., McCarron, D. J., Norrgard, E. B., Steinecker, M. H. & Demille, D. Magneto-optical
trapping of a diatomic molecule. Nature 512, 286–289 (2014).
19. Truppe, S. et al. Molecules cooled below the Doppler limit. Nat. Phys. 13, 1173–1176 (2017).
20. Anderegg, L. et al. An optical tweezer array of ultracold molecules. Science 365,
1156–1158 (2019).
21. Ding, S., Wu, Y., Finneran, I. A., Burau, J. J. & Ye, J. Sub-Doppler cooling and compressed
trapping of YO molecules at μK temperatures. Phys. Rev. X 10, 021049 (2020).
22. Yang, H. et al. Observation of magnetically tunable Feshbach resonances in ultracold
23
Na40K + 40K collisions. Science 363, 261–264 (2019).
23. Son, H., Park, J. J., Ketterle, W. & Jamison, A. O. Collisional cooling of ultracold molecules.
Nature 580, 197–200 (2020).
24. Segev, Y. et al. Collisions between cold molecules in a superconducting magnetic trap.
Nature 572, 189–193 (2019).
25. Baranov, M. A., Dalmonte, M., Pupillo, G. & Zoller, P. Condensed matter theory of dipolar
quantum gases. Chem. Rev. 112, 5012–5061 (2012).
26. Aikawa, K. et al. Reaching Fermi degeneracy via universal dipolar scattering. Phys. Rev.
Lett. 112, 010404 (2014).
27. DeMarco, B. & Jin, D. S. Onset of Fermi degeneracy in a trapped atomic gas. Science 285,
1703–1706 (1999).
28. Quéméner, G. & Bohn, J. L. Dynamics of ultracold molecules in confined geometry and
electric field. Phys. Rev. A 83, 012705 (2011).
29. Micheli, A. et al. Universal rates for reactive ultracold polar molecules in reduced
dimensions. Phys. Rev. Lett. 105, 073202 (2010).
30. Zhu, B., Quéméner, G., Rey, A. M. & Holland, M. J. Evaporative cooling of reactive
polar molecules confined in a two-dimensional geometry. Phys. Rev. A 88, 063405
(2013).
31. de Miranda, M. H. G. et al. Controlling the quantum stereodynamics of ultracold
bimolecular reactions. Nat. Phys. 7, 502–507 (2011).
32. Covey, J. P. Enhanced Optical and Electric Manipulation of a Quantum Gas of KRb
Molecules. Thesis, Univ. Colorado, https://doi.org/10.1007/978-3-319-98107-9 (2018).
33. Shvarchuck, I. et al. Bose–einstein condensation into nonequilibrium states studied by
condensate focusing. Phys. Rev. Lett. 89, 270404 (2002).
34. Hueck, K. et al. Two-dimensional homogeneous Fermi gases. Phys. Rev. Lett. 120, 060402
(2018).
35. Ni, K.-K. et al. A high phase-space-density gas of polar molecules. Science 322, 231–235
(2008).
36. Quéméner, G. & Bohn, J. L. Strong dependence of ultracold chemical rates on electric
dipole moments. Phys. Rev. A 81, 022702 (2010).
37. Bohn, J. L., Cavagnero, M. & Ticknor, C. Quasi-universal dipolar scattering in cold and
ultracold gases. New J. Phys. 11, 055039 (2009).
38. Ticknor, C. Quasi-two-dimensional dipolar scattering. Phys. Rev. A 81, 042708 (2010).
39. Bohn, J. L. & Jin, D. S. Differential scattering and rethermalization in ultracold dipolar
gases. Phys. Rev. A 89, 022702 (2014).
40. Aikawa, K. et al. Anisotropic relaxation dynamics in a dipolar Fermi gas driven out of
equilibrium. Phys. Rev. Lett. 113, 263201 (2014).
41. Babadi, M. & Demler, E. Collective excitations of quasi-two-dimensional trapped dipolar
fermions: transition from collisionless to hydrodynamic regime. Phys. Rev. A 86, 063638
(2012).
42. Giorgini, S., Pitaevskii, L. P. & Stringari, S. Theory of ultracold atomic Fermi gases. Rev.
Mod. Phys. 80, 1215 (2008).
43. Büchler, H. P. et al. Strongly correlated 2D quantum phases with cold polar molecules:
controlling the shape of the interaction potential. Phys. Rev. Lett. 98, 060404 (2007).
44. Góral, K., Santos, L. & Lewenstein, M. Quantum phases of dipolar bosons in optical
lattices. Phys. Rev. Lett. 88, 170406 (2002).
45. Capogrosso-Sansone, B., Trefzger, C., Lewenstein, M., Zoller, P. & Pupillo, G. Quantum
phases of cold polar molecules in 2D optical lattices. Phys. Rev. Lett. 104, 125301 (2010).
46. Cooper, N. R. & Shlyapnikov, G. V. Stable topological superfluid phase of ultracold polar
fermionic molecules. Phys. Rev. Lett. 103, 155302 (2009).
47. Gorshkov, A. V. et al. Tunable superfluidity and quantum magnetism with ultracold polar
molecules. Phys. Rev. Lett. 107, 115301 (2011).
48. Potter, A. C., Berg, E., Wang, D. W., Halperin, B. I. & Demler, E. Superfluidity and
dimerization in a multilayered system of fermionic polar molecules. Phys. Rev. Lett. 105,
220406 (2010).
49. Yao, N. Y. et al. Many-body localization in dipolar systems. Phys. Rev. Lett. 113, 243002 (2014).
50. Barbiero, L., Menotti, C., Recati, A. & Santos, L. Out-of-equilibrium states and
quasi-many-body localization in polar lattice gases. Phys. Rev. B 92, 180406 (2015).
51. Zinner, N. T. & Bruun, G. M. Density waves in layered systems with fermionic polar
molecules. Eur. Phys. J. D 65, 133–139 (2011).
52. Peter, D., Müller, S., Wessel, S. & Büchler, H. P. Anomalous behavior of spin systems with
dipolar interactions. Phys. Rev. Lett. 109, 025303 (2012).
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in
published maps and institutional affiliations.
© The Author(s), under exclusive licence to Springer Nature Limited 2020
Nature | Vol 588 | 10 December 2020 | 243
各位同学：

我们于 2012 年推出外刊 VIP 终身会员服务，2015 年开始通过微信平台招募会员，并通过微信朋

友圈发布外刊更新提醒，在全网都属首创（欢迎查证） 。八年来，我们的服务惠及上万名会员，资
源更新从未间断（包括春节等节假日）。

现作以下说明（特别重要，请务必看完）：

外刊 VIP 终身会员服务的初衷是采用类似众筹的模式共享订阅资源，节省大家的时间和金钱成本。
外刊资源仅供个人学习，严禁用于个人学习以外的其他目的（譬如在网络上肆意传播、转卖等） 。

外刊 VIP 只通过如下 14 个微信号（官方正式途径）进行：

微信号 1：book408 微信号 8： book361
微信号 2：book608 微信号 9： book362
微信号 3：book208 微信号 10：book599
微信号 4：book308 微信号 11：book7749
微信号 5：book4008 微信号 12：book8848
微信号 6：book5008 微信号 13：book9669
微信号 7：book7008 微信号 14：MeijiEnglish

其他任何途径（包括淘宝、微店、微博、QQ 或其他任意微信号）都属假冒。最近几年网络上不
乏模仿甚至抄袭我们模式的。不夸张的说，网络上一半以上提供外刊资源的，都是直接或间接源
自我们这里。看的外刊多了不难发现这一点。选择最原始的源头，长期更新才最有保障。

对于淘宝平台上提供类似服务的店铺，资源也是多半来自我们。由于众所周知的原因，淘宝平台
上的此类商品和店铺一般几个月以内就会下架或者封禁。我们的一些会员，就经历过“在淘宝上
订阅然后遭遇店铺关闭跑路”，最终才选择加入我们。请大家避免陷入惯性思维（认为淘宝上更可
靠），因为淘宝上外刊资源类的商品和店铺几乎每天都在上演“下架和封禁”，所以淘宝卖家跑路
已是屡见不鲜。

1. 如果你是通过上述 14 个微信号之一加入 VIP 的，首先恭喜你选择了官方正式途径。其次，VIP

会员严禁再添加上述其他微信号（所有微信号信息发布都一样，添加多个微信号会造成信息混淆） ，
如发现有添加上述两个甚至两个以上微信号的，我们会取消会员资格（之前无意添加了多个微信
号的会员请主动告知，我们会定期清理重复好友） 。如有外刊资源相关的问题，直接联系你当时加
入 VIP 的那个微信号即可。我们是终身制的，永远无需任何续费。

2. 如果你是通过其他途径订阅该外刊的，很遗憾你选择了非官方途径。你可能已经区分出我们才
是资源最终的源头。其他很多途径，卖家自己可能根本不读外刊，做这个完全是为了一时有利可
图，随时可能会中断更新甚至蒸发消失（这种事情实在不胜枚举） 。因此我们强烈建议你联系我们
（请添加我们的特别微信号 1295512531），我们会视具体情况弥补你的损失、甚至免费注册你为
我们的正式终身会员！

3. 如果你是通过免费渠道获取到该外刊、同时又有外刊订阅需求的，欢迎加入我们（请添加我们
的特别微信号 1295512531）。也欢迎长期观察我们的微信朋友圈和资源更新情况。如你遇到模式
和我们很相似的其它渠道，欢迎仔细考察、查证哪一个才是最终源头。

再次强调，我们的外刊 VIP 终身会员服务，是提供 100 多种英文原版杂志终身更新的服务，以请

朋友吃顿饭的花费（或者买几斤猪肉的花费）即可终身订阅所有外刊并享受后续所有福利，具体
服务介绍请添加我们的特别微信号 1295512531 索要。
Article
Methods we extract τ = 4.9(0.2). For the T/TF data in the paper we thus use the
closest estimate τ = 5 ± 1, where the uncertainty accounts for possible
Experimental protocol systematic errors arising from non-uniform conversion of Rb to KRb
The experiment starts with an ultracold atomic mixture of 40K and and variation of evaporation efficiency between the layers (since the
87
Rb, held in the ODT at a magnetic field of 555 G. The trap frequencies density, and hence thermalization rate, varies between the layers).
for Rb in the ODT are (ωx, ωy, ωz) = 2π × (40, 180, 40) Hz. The atomic To determine the 2D density for the measurement of β, we need to use
mixture is then transferred into a single layer of the LSL. The LSL beams a time-averaged layer number that considers the density dependence
propagate at a shallow angle of 4 degrees along z, resulting in a lattice of the loss in each layer3. The layer-averaged 2D density is defined as
spacing of 8 µm along y. At the end of the LSL ramp, we decrease the n = N/(4πσ2τ), where σ is the root-mean-square cloud size in the radial
ODT power, so that the trap frequencies for Rb in the combined trap are direction. Through numerical simulation, we obtain the decay over time
(ωx, ωy, ωz) = 2π × (25, 600, 25) Hz. Typically, we have 4.1 × 105 K atoms for a cloud with the layer distribution plotted in Extended Data Fig. 1
and 7.0 × 104 Rb atoms at T = 115(10) nK. About 30% of Rb is condensed. and compare it with the decay of a gas with a uniform layer distribution
At this point, we load the mixture into the VL and adjust the ODT in and same number N. In this case, we define τ as the value for which β
order for the KRb molecules to experience radial trap frequencies in the uniform case matches β in the non-uniform case. For Extended
at zero electric field of (ωx, ωz) = 2π × (40, 40) Hz, with ωy/(2π) rang- Data Fig. 1, we obtain τ = 8 ± 1.
ing from a few kilohertz up to 20 kHz. To compensate for the limited
transmittivity of the ITO plates at the 1.064-µm VL wavelength and to Electric field potential
avoid spurious superlattices, the VL beams have a 11-degree tilt with The anti-trapping potential that is used for dipolar evaporation intro-
respect to y, resulting in a lattice spacing of 540 nm. duces an anti-curvature that changes the trap frequency ωx as a function
To create molecules, we first sweep the magnetic field adiabatically of γ. Owing to the geometry of our electrodes, when γ = 0.4225 the
through the KRb heteronuclear Feshbach resonance at 546.6 G. The electric field potential at the molecule position is as homogeneous
magnetic field is ramped from 555 G to 545.5 G in 4 ms, creating 2.5 × 104 as possible (fourth-order cancellation of the electric-field curvature).
Feshbach molecules that are subsequently transferred to the absolute By increasing (decreasing) γ, ωx decreases (increases), as shown in
KRb ground state by stimulated Raman adiabatic passage (STIRAP)1. Extended Data Fig. 2 at EDC = 5 kV cm−1. Our results follow the expected
By tuning the Raman lasers, we create KRb molecules at 0 kV cm−1 or trend from finite-element simulations of the combined electro-optical
4.5 kV cm−1. For molecule creation at 4.5 kV cm−1, the field is ramped to potential at different γ.
the target value 10 ms before the Feshbach sweep. The STIRAP one-way
transfer efficiency is 85(2)% at 0 kV cm−1 and 82(3)% at 4.5 kV cm−1. We Electric-field evaporation ramps
do not observe any dependence of β and Sevap on the initial value of For the evaporation experiments, we lower the trap depth by increas-
the electric field. ing γ over time. The trap depth at each time point of the evaporation
is estimated by simulations of the combined electro-optical potential,
which is benchmarked with the measurement of ωx versus γ displayed in
Matter-wave focusing and layer counting
Extended Data Fig. 2. For the data shown in Fig. 4, the evaporation ramp
The VL layer spacing of 540 nm is too small to resolve with conventional
takes about 800 ms and EDC, γ and trap depth evolve over time according
absorption imaging. To quantify the number of occupied layers τ, we
to the plots in Extended Data Fig. 3. We also plot the trend of the param-
use a matter-wave technique that maps the in situ density distribution
eter η and T/TF at different time points of the evaporation sequence.
onto the momentum distribution, which can then be imaged in time
of flight. To do so, we instantaneously release the cloud from the VL Thermometry of 2D Fermi gas
and the LSL into the ODT. The cloud expands into the ODT harmonic The 2D in situ density n of the molecular Fermi gas is given by the Fermi–
potential for a quarter of the oscillation period along y. This corre- Dirac distribution54:
sponds to a 90-degree rotation in phase space. As a result, after the
rotation, the momentum distribution along y in time of flight corre- 1   1 1 
sponds to the original in-situ density profile. Increasing the time of n(x , z) = − 2
Li1 −exp  − βth  m ω 2x x 2 + m ω 2z z 2 − µ , (3)
λ dB    2 2 
flight increases the layer separation until they can be resolved optically.
1
From a set of averaged matter-wave density profiles, we obtain a his- where m is the molecular mass, βth = k T , µ the chemical potential,
B
togram with the normalized particle number per layer from which we 2
λ dB = 2πħ βth /m the thermal de Broglie wavelength, and Lin the
extract the number of layers τ. We perform this analysis on a cloud of polylogarithmic function of order n. The column-integrated density
Rb atoms without K, eliminating the K–Rb interactions during the phase profile is:
space rotation time. The Rb cloud used for matter-wave amplification
imaging has the same trap parameters, number and temperature of 1 2π    1 
the Rb cloud used for the molecule experiment. When the K–Rb inter- nint(x) = − 2 Li3/2 −exp βth µ − mω 2x x 2 . (4)
λ dB βthmω 2z    2 
actions are removed by setting the magnetic field to the zero-crossing
of the Feshbach resonance, the full contrast is restored. For the data After a certain time of flight t, the x coordinate scales by a factor
in Fig. 1c, obtained by averaging 20 matter-wave images of the Rb cloud, 1/ 1 + ω 2x t 2. The density is consequently divided by 1 + ω 2x t 2 for pro­
the density histogram is shown in Extended Data Fig. 1. For a fixed mol- ­­per renormalization. The chemical potential µ is defined through the
ecule distribution, the definition of τ depends on the physical quantity relation:
being calculated. Using τ = N /⟨Ni ⟩ , where ⟨Ni ⟩ is the average particle
2
number per layer over the measured distribution, we extract τ = 4.6(2) N  kBT 
= Li (− e β th µ). (5)
for the data in Extended Data Fig. 1. Theoretical modelling53 for the Rb τ  ħωR  2
cloud in the same conditions yields a consistent value of τ = 5.1(2).
Our measurements of the molecular cloud thus involve averaging Combining equation (5) with the definition of TF in 2D, we obtain:
over layers that are not equally populated. To determine T/TF, we use
1/2 2
T 
the layer-averaged Fermi temperature TF =
2
ħω R
kB
2 Ni =
ħω R
kB ( )
2N
τ
, T  = −
 F 2 Li
1
(− e β th µ)
, (6)
2
where τ = N / Ni is an effective number of layers that accounts for the
nonlinear dependence of TF on Ni. For the data in Extended Data Fig. 1, which allows us to extract the ratio T/TF from the polylogarithmic fit.
 From the Gaussian fit to the whole cloud, we obtain the Gaussian
width σ and a release temperature Trel, defined as:
mω 2x σ 2
Trel = . (7)
kB(1 + ω 2x t 2)
The release temperature Trel is proportional to the energy density
U of the Fermi gas, U = 2kBTrel, which saturates to a non-zero value as
T → 0. In contrast, the energy density Ucl = 2kBT of a classical gas
approaches zero as T → 0.
When the Gaussian fit is constrained to only the outer wings (that is,
the high-momentum states) of the cloud, we can extract a new width
σout from which, using equation (7), we obtain a corrected temperature
Tout through the relation:
mω 2x σ out
2
Tout = . (8)
kB(1 + ω 2x t 2)
As the excluded region from the centre of the Gaussian fit is
expanded, Tout decreases from an initial value of Trel and approaches
T. This is shown in Extended Data Fig. 4, where we plot the ratio
Tout/Trel at different exclusion regions in units of σ. For the range of
T/TF studied here, we find that an exclusion region of 1.5σ leaves us
enough signal-to-noise ratio for the fit to properly converge and to
return a value of Tout that is only 5% higher than T. In the main text, the
Gaussian fit on the outer wings is performed with an excluded region
of 1.5σ.
Data availability
The data that support the findings of this study are available from the
corresponding author upon reasonable request. Source data are pro-
vided with this paper.
53. Dalfovo, F., Giorgini, S., Pitaevskii, L. P. & Stringari, S. Theory of Bose–Einstein
condensation in trapped gases. Rev. Mod. Phys. 71, 463–512 (1999).
54. Inguscio, M., Ketterle, W. & Salomon, C. Ultra-cold Fermi gases. In Proceedings of the
International School of Physics ‘Enrico Fermi’ (2007).
Acknowledgements We acknowledge funding from NIST, DARPA DRINQS, ARO MURI and NSF
Phys-1734006. We thank J. L. Bohn, A. M. Kaufman, and C. Miller for careful reading of the
manuscript and T. Brown for technical assistance.
Author contributions All authors contributed to carrying out the experiments, interpreting the
results, and writing the manuscript.
Competing interests The authors declare no competing interests.
Additional information
Correspondence and requests for materials should be addressed to G.V. or J.Y.
Peer review information Nature thanks Georgy Shlyapnikov and the other, anonymous,
reviewer(s) for their contribution to the peer review of this work.
Reprints and permissions information is available at http://www.nature.com/reprints.
Article

Extended Data Fig. 1 | Layer occupancy. Histogram of the average number per
layer (relative population) for the data shown in Fig. 1c.
Extended Data Fig. 2 | Trend of ω x /(2π) versus γ. Grey points are the
experimental measurements at EDC = 5 kV cm−1, the solid grey line is a linear fit to
guide the eye, and the dashed line is the prediction (Sim) from the
finite-element model. All error bars are 1 standard deviation of the mean.
Article

Extended Data Fig. 3 | Evaporation sequence. a, Ramp in EDC. b, Ramp in γ. temperature at each time point. e, Evolution of T/TF during the ramp. All error
c, Trap depth versus time from the finite-element model of electro-optical bars are 1 standard error of the mean.
potential. d, Evolution of η, calculated by taking the ratio of the trap depth and
Extended Data Fig. 4 | Fermi gas thermometry. Trend of Tout/Trel as a function
of the excluded region from the centre of the Gaussian fit for T/TF = 0.81(15)
(orange diamonds) and T/TF = 2.0(1) (black circles). Solid lines are Gaussian fits
to simulated density profiles for T/TF = 2.0 (black) and T/TF = 0.8 (orange). All
error bars are 1 standard error of the mean.
Article

Operation of an optical atomic clock with a

Brillouin laser subsystem

https://doi.org/10.1038/s41586-020-2981-6 William Loh1,3 ✉, Jules Stuart1,2,3, David Reens1, Colin D. Bruzewicz1, Danielle Braje1,
John Chiaverini1, Paul W. Juodawlkis1, Jeremy M. Sage1,2 & Robert McConnell1
Received: 16 January 2020

Accepted: 1 October 2020

Microwave atomic clocks have traditionally served as the ‘gold standard’ for precision
Published online: 9 December 2020
measurements of time and frequency. However, over the past decade, optical atomic
Check for updates
clocks1–6 have surpassed the precision of their microwave counterparts by two orders
of magnitude or more. Extant optical clocks occupy volumes of more than one cubic
metre, and it is a substantial challenge to enable these clocks to operate in field
environments, which requires the ruggedization and miniaturization of the atomic
reference and clock laser along with their supporting lasers and electronics4,7,8,9. In
terms of the clock laser, prior laboratory demonstrations of optical clocks have relied
on the exceptional performance gained through stabilization using bulk cavities,
which unfortunately necessitates the use of vacuum and also renders the laser
susceptible to vibration-induced noise. Here, using a stimulated Brillouin scattering
laser subsystem that has a reduced cavity volume and operates without vacuum, we
demonstrate a promising component of a portable optical atomic clock architecture.
We interrogate a 88Sr+ ion with our stimulated Brillouin scattering laser and achieve a
clock exhibiting short-term stability of 3.9 × 10−14 over one second—an improvement
of an order of magnitude over state-of-the-art microwave clocks. This performance
increase within a potentially portable system presents a compelling avenue for
substantially improving existing technology, such as the global positioning system,
and also for enabling the exploration of topics such as geodetic measurements of the
Earth, searches for dark matter and investigations into possible long-term variations
of fundamental physics constants10–12.

The ability to precisely measure time with a portable system has long
been essential to navigation. The necessity for accurate and portable
timekeeping inspired the development of Harrison’s marine chronom-
eter nearly 300 years ago and continues to this day, reflected in modern
societal reliance on the global positioning system (GPS). Recently,
optical atomic clocks, operating at frequencies of hundreds of tera-
hertz, have achieved performance far surpassing that of the best micro-
wave clocks and have substantially advanced the precision with which
time—and, equivalently, distance—can be measured. However, portable
implementations of these optical clocks will require substantial modi-
fications to the existing clock architecture, including miniaturization
of the clock laser whose performance is central to the operation of the
clock. A key challenge is to maintain the frequency stability of the clock
laser while reducing its size. This stability is required for the laser to
(1) remain within the vicinity of the atom’s narrow-linewidth transi-
tion during the period of time before the clock feedback is engaged
(minutes) and (2) remain locked to the atomic transition between feed-
back cycles (milliseconds). These stringent requirements have thus
far eliminated all possible candidates for clock interrogation apart
from bulk-cavity-stabilized (BCS) lasers13–15, which exhibit currently
unrivalled narrow linewidths of <1 Hz under high-vacuum operation,
but otherwise are unwieldy and prone to vibration4,7,16,17.
A promising portable alternative to these BCS lasers has recently
emerged via generation of stimulated Brillouin scattering (SBS) light
in an ultrahigh quality factor (Q) resonator18–27. In comparison to the
BCS laser, the SBS laser offers the advantages of reduced cavity volume,
operation without vacuum, the ability to be rigidly mounted to any flat
surface, and a potentially higher tolerance to vibration. Despite these
properties, the application of the SBS laser to state-of-the-art atomic
physics has yet to be demonstrated, primarily owing to the laser’s sub-
stantial frequency drift in response to temperature change21. Here, we
overcome the challenges of drift by applying a recently developed
technique28 to sense and control the SBS laser’s long-term frequency
fluctuations in the regime of a few hundred hertz. We utilize the stabi-
lized SBS light in the demonstration of a strontium-ion optical clock,
which breaks the long-standing clock paradigm of requiring a
BCS laser to serve as the master oscillator. Through a clock self-
comparison measurement, we achieve a fractional frequency stability
of 3.9 × 10−14 / τ (with the interrogation time τ in s), beyond the
short-term stability achievable by the best microwave clocks29 and

Lincoln Laboratory, Massachusetts Institute of Technology, Lexington, MA, USA. 2Department of Physics, Massachusetts Institute of Technology, Cambridge, MA, USA. 3These authors
1

contributed equally: William Loh, Jules Stuart. ✉e-mail: William.Loh@ll.mit.edu

244 | Nature | Vol 588 | 10 December 2020

 a Resonator input b Resonator output c Feedforward d Output after AOM e Lock to ion
88Sr+
~180 MHz ~180 MHz
Pol. beat Correction Stabilized SBS out
Pump 2 Pump 1 SBS 2 SBS 1 signal
×39
×
Stable SBS
f f f f
1,348.0 nm 1,348.1 nm 180
80 MH
MHz 1,348.1 nm

f SBS laser subsystem

PID
LO

PD 3
PID
PD 1 g
EOM

SBS 1
SOA
SO SBS 1 PBS

Pump SBS out Rot.

Rot
t.
nm)
(1,348.1 nm)
PD 5
AOM SOA R
Res.
SBS 2
SBS 2
Rot.
EOM

LO PID
PD 4 h Feedforward Frequency
PD 2 SBS out AOM SOA doubler To Sr ion
(1,348.1 nm)

SBS 1 To measurement
PID
Rot. 200 MHz
Servo 180 MHz 30 MHz
PLL
PD 5 Servo out LO
SBS 2 PID 30 MHz
VCO
LO
LPF

30 MHz VCO Ramp

DC amp.

Fig. 1 | SBS laser setup and stabilization scheme. a–e, Illustration of the steps
(a–e) required for SBS stabilization. Two orthogonal polarization SBS signals
are generated whose beat note is applied to an acousto-optic modulator (AOM)
to compensate for the SBS laser’s frequency drift. See text for details.
f, Diagram of the SBS laser setup. The AOM, electro-optic modulator (EOM),
semiconductor optical amplifier (SOA), polarization rotator (Rot.), polarization
beam splitter (PBS), photodetector (PD), local oscillator (LO) and proportional-
integral-derivative (PID) controller enable the independent locking of a pump
laser to two orthogonal polarization modes of the resonator (Res.). Two
counter-propagating SBS lasers (SBS 1 and SBS 2) are created from the pump
lasers. The blue shaded area designates the region where feedforward
stabilization is applied to the SBS laser and will be continued further in h.
g, Photograph of the SBS resonator resting within a copper enclosure and
placed next to a quarter dollar. The SBS resonator including the enclosure and
lid occupies a total space of 3.1 inch × 3.5 inch × 1 inch. h, Diagram of the SBS
stabilization circuitry. The beat note of two SBS lasers is first converted to a
voltage signal through a voltage-controlled oscillator (VCO) operating in a
phase-locked loop (PLL). Afterwards, a linear ramp subtraction, low-pass filter
(LPF) and factor of 39 multiplication is applied before this signal is used for
correcting the SBS laser drift. The dashed box indicates the elements that
comprise the PLL.

close to an order of magnitude away from state-of-the-art laboratory
single-ion clocks3.
For our clock demonstration, we interrogate a 88Sr+ ion using a
fibre-cavity SBS laser with radiation at 1,348 nm that is frequency dou-
bled to reach the narrow-linewidth 88Sr+ clock transition at 674 nm
(ref. 30). Figure 1a–e depicts our overall strategy for achieving a stable SBS
optical atomic clock. Starting from two orthogonally polarized pump
lasers input into a high-Q optical-fibre resonator (Fig. 1a), the resona-
tor generates two orthogonally polarized SBS outputs that are Stokes
shifted from their respective pumps by about 12.5 GHz (Fig. 1b). The two
SBS lasers interfere on a photodetector to produce a 180-MHz beat note
(Fig. 1c), the frequency deviation of which predominantly corresponds
to the temperature drift of the SBS resonator28. This method of tem-
perature sensing is based on techniques that connect the differential
temperature sensitivity between two orthogonal polarization resonator
modes to a measurement of temperature change31–33. However, owing to
issues associated with both the detection and control of small shifts in
temperature, the best prior stabilization efforts have so far been limited
to a range of about 10 μK (ref. 34). Here, as a consequence of the excep-
tionally narrow SBS lasing linewidth, the resolution of our temperature
sensor reaches below 100 nK. In order to circumvent the need for direct
control of temperature, we apply the dual-polarization beat note (Pol.
beat) as a feedforward correction to the SBS laser’s frequency (Fig. 1d).
This correction brings the SBS laser’s stability to a regime where it can
be used to interrogate a 88Sr+ ion optical clock (Fig. 1e).
The SBS laser (Fig. 1f) consists of a single pump laser whose output at
1,348 nm is split into two separate paths (see Methods section ‘SBS laser
setup’). On one path, an acousto-optic modulator (AOM) is used both
to shift the light by approximately 180 MHz and to enable independent
control of its frequency. Afterwards, the polarization of one of the pump
beams is rotated by 90°, and the two pump beams are then sent into the
SBS resonator in opposite directions to accomplish Pound–Drever–Hall
(PDH) locking35 to the resonator’s two orthogonal polarization modes.
The SBS resonator itself exhibits a loaded Q of 1.9 × 108 and consists
of 2 m of optical fibre wound around a 2-inch-diameter mandrel that
rests within a 3.1 inch × 3.5 inch × 1 inch temperature-controlled cop-
per enclosure (Fig. 1g). The two pump lasers each generate their own
counter-propagating SBS beams, which upon leaving the resonator are
both coupled out of the system through a pair of circulators. A portion
of the SBS light is also monitored and used to stabilize the amplitude
of the SBS laser. Past the circulator, the outputs of both orthogonal
polarization SBS lasers are interfered on a photodetector, which serves
as our method for sensing temperature change.
Figure 1h presents the feedforward circuitry used in the stabilization
of the SBS laser (see Methods section ‘Feedforward implementation’).
After the 180-MHz SBS temperature error signal is generated, a series

Nature | Vol 588 | 10 December 2020 | 245

Article
a Feedforward AOM
Free-running SBS
Subtraction
SBS 1

×39
SBS 2 Pol. beat ×39
Res. PD 5

b 2.0 c 20
Free-running SBS Free-running SBS
1.5 Pol. beat ×39 Pol. beat ×39
Frequency drift (MHz)

10

Frequency drift (kHz)

Subtraction
1.0
0
0.5

–10
0

–0.5 –20
0 4 8 12 0 4 8 12
Time (min) Time (min)
d 20 e 10–8 10–3
Free-runniing SBS
Free-running
F
10–9 Drift-subtracted
D
Drift-subtr
rracted SBS 10–4
10
Frequency drift (kHz)

10–10 Drift regime 10–5

Drift-subtracted SBS

ΔT (K)
0 10–11
Δf/f

10–6
220 Hz
10–12 10–7
–10
10–13 22 Hz 10–8
τ
Ideal 1/√
–20 10–14
0 4 8 12 10–3 10–2 10–1 100 101 102
Time (min) Time (s)

Fig. 2 | SBS laser drift cancellation procedure. a, Simplified diagram
illustrating the correction of the SBS frequency drift via subtraction of the
free-running SBS laser and a frequency multiplied (×39) polarization beat
signal (Pol. beat ×39). b, Experimental time series of the free-running SBS laser
(red) and the polarization beat-note (blue) drift. A numerical subtraction of the
two (black) produces a residual linear drift. c, Numerical removal of the linear
drift from the free-running SBS and polarization beat-note time series
revealing the underlying temperature-induced frequency drift. The excellent
agreement between the two demonstrates that the correction signal
accurately tracks the SBS temperature drift. d, Subtraction of the SBS and
polarization beat-note signals in Fig. 2c. The drift is numerically cancelled
through the linear ramp and temperature correction procedure. e, Fractional
frequency (Δf/f ) noise measurements and corresponding extracted
temperature deviation (ΔT) of the SBS laser. The green shaded section
indicates the region where the SBS laser noise is gradually averaging down.
We lock the SBS laser to the ion at these timescales for the demonstration
of the optical atomic clock. The blue section indicates where the SBS laser
experiences drift that can be mostly compensated for via subtraction. The
red shaded section indicates where the SBS laser drift dominates. After
subtraction, the resulting noise of the SBS laser reveals a short-term frequency
variation of 22 Hz and a long-term frequency drift of 220 Hz, which can be
further reduced by locking to the ion.

of operations is applied to prepare this signal to serve as a correction
to the SBS laser. Using a phase-locked loop (PLL), the temperature
error is first converted to a voltage that is low-pass filtered and ampli-
fied before it is converted back again to a radio-frequency (RF) signal.
A net frequency multiplication factor of 39 is achieved through this
process, which enables the temperature error to precisely match and
cancel the SBS frequency drift. The subtraction of a ramp signal also
enables the ability to compensate for a residual linear SBS drift. After
correction, the SBS laser output is frequency doubled to reach 674 nm
for interrogation of the 88Sr+ ion. Together, Fig. 1f–h combined comprise
the SBS laser subsystem whose output is directed onto a 88Sr+ ion for
operation of an optical atomic clock.
Figure 2a demonstrates the feedforward procedure for correcting the
SBS laser’s drift. In comparison to the feedback approach of ref. 28, the
use of feedforward enables direct control of the SBS laser’s frequency
with advantages in avoiding the slow servo response of controlling
temperature at the fibre core and also in circumventing unintentional
length shifts that arise from thermal expansion in the underlying copper
plate. The beat note between two orthogonal polarization SBS lasers,
which is used as a sensor for temperature change, is subtracted from
one free-running SBS laser to stabilize its frequency drift. To dem-
onstrate this procedure, both the frequency of the free-running SBS
laser’s beat note with a reference BCS laser and the frequency of the
dual-polarization beat note (see Methods section ‘Measurement of
the SBS laser noise’) are experimentally measured at time intervals of
1 ms and plotted for the wavelength of 1,348 nm over a period of 12 min
(Fig. 2b). The polarization beat is also numerically multiplied by the
empirically determined correction factor of 39 in order to account
for the common-mode suppression of the extracted temperature
shifts. The resulting multiplied signal serves as the feedforward cor-
rection to temperature drift, which on numerical subtraction from the
free-running SBS laser yields a residual linear drift of 160 kHz min−1.
This linear drift results from the individual linear drifts of the SBS
and polarization beat signals, which are in opposing directions for
the case of Fig. 2b, and represents a parasitic frequency shift that our
temperature sensing technique cannot account for. We attribute the

246 | Nature | Vol 588 | 10 December 2020

a b
500
Feedforward on
400
Frequency drift (kHz)
300
200
100
Free
Stabilized
running
0 spectrum is taken with a sweep time of
0 5 10
Time (min)
c 10
0 drift-cancelled SBS laser of Fig. 2e. The
Δf = 50 Hz
Normalized power (dB)
–10
–20
–30
–40
–50
–1.0 –0.5 0.0 0.5
Offset frequency (kHz)
observed linear drift to a slow relaxation of the SBS resonator over
time36 (see Methods section ‘SBS laser residual linear drift’), which is
projected to equilibrate on a timescale of months. Rather than per-
forming this subtraction step first, we instead numerically remove the
linear drift from the SBS and polarization-beat traces, which reveals
the underlying temperature-induced drift of the SBS laser frequency
(Fig. 2c). The correction signal, which now occupies a span of 30 kHz
at the wavelength of 1,348 nm, shows excellent agreement with the
remaining SBS laser drift for cancellation. With the linear drift removed,
a subtraction of the polarization-beat frequency from the SBS laser
frequency yields a stabilized SBS laser frequency with about 220-Hz
frequency fluctuations (Fig. 2d).
Figure 2e shows the fractional frequency noise, that is, the
root-mean-square (r.m.s.) frequency shift of the SBS laser normalized
to the laser’s centre frequency, derived from the time series traces of
Fig. 2b–d. The free-running 1,348-nm SBS laser reaches a minimum
noise level of 1.4 × 10−13 (corresponding to a linewidth of 30 Hz) at
10 ms but becomes unbounded at longer timescales and increases to
5.8 × 10−10 at 100 s. When the measured drift is subtracted to account for
both the linear and temperature-induced drift, the SBS laser maintains
its performance of 1.0 × 10−13 (22 Hz) at short timescales and experi-
ences a >500-fold reduction in drift over the long term. The SBS laser
frequency excursions become bounded at the value of 1.0 × 10−12 or
220 Hz, which corresponds to temperature stabilization of the SBS
laser at a level below 120 nK.
We experimentally implement the numerical feedforward procedure
of Fig. 2 and apply the corrections onto the SBS laser’s frequency via
an AOM (Fig. 3a). The linear drift is accounted for by a ramp generator,
and the factor of 39 multiplication is implemented through a PLL. The
stabilized output is frequency doubled to 674 nm and is measured
against a reference BCS laser for characterization of the SBS laser drift.
For the remainder of this Article, we will refer to the frequency-doubled
and amplified SBS laser as the ‘SBS laser subsystem’, whose output at
674 nm is used to interrogate the clock ion. The SBS laser subsystem,
operating at 674 nm, starts initially in a free-running state for the first
6.5 min of measurement and drifts by 400 kHz within this time. Once
Fig. 3 | SBS laser subsystem
stabilization. a, Measurement of the
440 SBS laser’s frequency drift (solid line)
Frequency drift (kHz)
with and without feedforward
420
stabilization. The measurement interval
400 was 1 ms. b, Zoomed-in plot of the SBS
frequency drift. The average long-term
380
frequency deviation is 310 Hz. c, Measured
360 lineshape of the stabilized SBS laser
7 9 11 13 15 (red solid line) and corresponding Voigt
Time (min)
fit with R 2 = 0.93 (dashed line). The
measured linewidth (Δf ) is 50 Hz The
15
225 ms and a resolution bandwidth of
20 Hz. d, Fractional frequency(Δf/f )
d 10–11 noise comparison between the stabilized
SBS laser (red line) and the ideal
170 nK
stabilized SBS exhibits a short-term
10–12
frequency variation of 48 Hz. Long-term
temperature stabilization at the level of
Δf/f
170 nK is also experimentally achieved.
10–13 48 Hz
Stabilized SBS
Fig. 2e comparison
10–14
1.0 10–3 10–2 10–1 100 101 102
Time (s)
the feedforward correction is applied, the SBS frequency drift flattens
to a value near zero, as indicated by the horizontal dashed line guide of
Fig. 3a. A zoomed-in trace of the stabilized SBS frequency drift (Fig. 3b)
shows close agreement to Fig. 2d (accounting for the frequency dou-
bling), and indicates that our experimental implementation of the
feedforward stabilization accurately compensates for both linear and
temperature-induced SBS laser drift.
The lineshape of the SBS laser subsystem (Fig. 3c), measured at
674 nm via beating with a BCS laser, demonstrates further the laser’s
exceptional short-term noise. The spectrum exhibits a linewidth of
50 Hz. This value of linewidth is confirmed by the measured fractional
frequency noise of the stabilized SBS laser (Fig. 3d), which reaches a
minimum of 1.1 × 10−13 at 60 ms and corresponds to a frequency devia-
tion of 48 Hz at 674 nm. At long timescales, the noise level becomes
1.4 × 10−12 (620 Hz), which is slightly larger than the frequency excur-
sions found with the numerical drift-cancelling procedure of Fig. 2e.
We attribute this difference in drift to noise in the electronics used
for feedforward stabilization. At the level of 620 Hz, our achieved fre-
quency drift corresponds to temperature stabilization of the SBS laser
subsystem at a level below 170 nK.
To demonstrate experimentally the practical capability of the SBS
laser, we use the subsystem to run an atomic clock, stabilizing the laser
to the narrow-linewidth S1/2 ↔ D5/2 quadrupole clock transition in 88Sr+
(0.4 Hz natural linewidth). We interrogate a single strontium ion con-
fined 50 μm above the surface of a microfabricated surface-electrode
trap within a cryogenic ultrahigh-vacuum apparatus37. As shown in
Fig. 4a, the clock interrogation light is amplified through a series of
injection-locked lasers followed by tapered amplifiers, with fibre-noise
cancellation stages to mitigate phase noise picked up as the clock laser
is routed between and across rooms within optical fibre; a single flipper
mirror allows us to select either an existing BCS laser (see Methods sec-
tion ‘Characteristics of the BCS laser’) or the SBS laser subsystem as the
initial seed for the injection stages. To maximize the coherence time of
the ion’s optical transition, we employ a system of passive magnetic field
stabilization using persistent superconducting currents38 and active
laser-vibration compensation using an interferometric scheme similar
Nature | Vol 588 | 10 December 2020 | 247
Article
a b
BCS
AOM ×2
SBS subsystem
c d 10–12
1.0
Ground state probability
0.9
0.8
0.7
–2 –1
Detuning (kHz)
Fig. 4 | SBS laser optical clock. a, Schematic of laser beam path to the clock
chamber. A flipper mirror allows the SBS laser subsystem to be interchanged
with a BCS laser as the master frequency source. The injected laser and tapered
amplifier represent a series of two injection locking and amplification stages,
respectively. An AOM is used to finely adjust the laser frequency before the
beam is focused onto an ion inside the cryogenic vacuum chamber used as the
clock chamber. b, Measured Ramsey fringes on the clock transition. Two π/2
pulses are applied around a τ = 1 ms interrogation period. The ground state
probability is measured as a function of the phase of the probe pulse.
c, Spectroscopy of the |5S1/2, m J = −1/2⟩ → |4D5/2, m J = −3/2⟩ transition in 88Sr+
measured with the SBS laser subsystem. Spectroscopic measurements are
to fibre noise cancellation39. With these techniques, we perform a series
of Ramsey experiments and deduce a coherence time of 2.9 ms with the
BCS laser, which we take as an upper bound for the performance of the
clock experiment. Assuming that frequency fluctuations in the laser
are the dominant decoherence mechanism, we infer an effective laser
linewidth of 1/(2π × 2.9 ms) = 55 Hz, which is close to the SBS laser sub-
systems’s linewidth in Fig. 3c. Thus we conclude that uncompensated
noise in our experiment (for example, from optics subject to acoustic
vibrations or magnetic field instability at the ion) contributes at about
the same level as noise from the SBS laser subsystem (see Methods
section ‘Limits of the optical clock measurement’ for discussion of
noise sources). All together, these noise sources limit the clock inter-
rogation time we achieve, and therefore set the ultimate limit in clock
performance.
An AOM is used both for scanning the SBS laser subsystem over the
clock transition and for locking the laser’s frequency to the atomic reso-
nance. To discipline the SBS laser subsystem to the atomic resonance
frequency, we create an error signal via a Ramsey experiment consisting
of two π/2 pulses on the clock transition, separated by an interroga-
tion time τ = 1 ms. As the phase of the second π/2 pulse is varied, the
population in the lower clock state traces out a sine curve (Fig. 4b).
The slope of this signal reaches a maximum when the second pulse is
90° out of phase with the first pulse. Variations in the laser frequency
248 | Nature | Vol 588 | 10 December 2020
0.9
Ground state probability
0.7
0.5
Injected laser
Tapered
amplifier
0.3
AOM 0.1
Cryogenic 0 5 10
vacuum chamber Probe phase (rad)
10–13
3.9 × 10–14/√
W
Δf/f
Δf = 370 Hz 10–14
10–15
10–16
0 1 2 10–1 100 101 102 103
Time (s)
interleaved with frequency corrections to keep the SBS laser subsystem locked
to the atomic transition. The ground state probability is measured as a function
of frequency detuning relative to the clock transition. The vertical blue error
bars represent 1σ error derived from photon counting statistics. d, Fractional
frequency (Δf/f ) noise of the difference frequency between interleaved clocks.
The measured fractional frequency noise is divided by 2 to estimate the error
of a single clock, assuming even distribution of error in the correction signals.
The blue points represent the frequency noise at a selection of averaging times,
and the vertical blue bars indicate the 1σ error in this calculation46. From a fit to
these data (dashed red line) assuming a purely white noise spectrum, we obtain
the function 3.9 × 10−14/ τ .
during the interrogation time result in an additional phase shift and
are thus mapped to the ion’s state distribution. Between interrogation
cycles, the state of the ion must be detected and then re-initialized
into the lower clock level; this leads to a 1.85-ms dead time, during
which the system is insensitive to frequency fluctuations of the laser
(see Methods section ‘Clock protocol and simulation’ for more details
of the lock procedure). As a demonstration of the efficacy of the atomic
lock, we perform Rabi spectroscopy of the clock transition using the
SBS laser subsystem, interleaved with frequency corrections follow-
ing the protocol above (Fig. 4c). These spectroscopic measurements
are performed with pulse time of 10 ms (to avoid Fourier broadening
of the line), which leads to a longer clock dead time of about 11 ms for
these measurements. The width of the measured feature, 370 Hz, is
in close correspondence with the simulated value of 250 Hz, derived
from numerical application of the clock protocol to the free-running
SBS laser subsystem noise measured in Fig. 3d, given the effective dead
time of 11 ms. This provides further evidence that noise mechanisms
downstream from the laser source do not substantially affect measure-
ments with the SBS laser subsystem.
In order to assess the stability of the clock when running with the SBS
laser subsystem, we perform a self-comparison measurement via two
independently operated clock signals generated by interleaving dis-
tinct sets of correction signals applied to the laser (see Methods section
‘Clock protocol and simulation’). This technique has been previously 11.
used to characterize clock performance when two independent clocks 12.
are not available40,41 and is known to accurately capture the short-term
clock stability, including ion projection noise and the Dick effect due 13.
to dead time in the clock protocol. The self-comparison technique is 14.
insensitive to long-term frequency drifts of the ion, which if present
would be common to both clock signals, but at the same time under- 15.
estimates the performance of the clock owing to the extra dead time 16.
incurred through the interleaving operation. We note that our clock
is interrogated on the magnetically sensitive |5S1/2, mJ = −1/2⟩ → |4D5/2, 17.
mJ = −3/2⟩ transition, which would limit long-term stability if magnetic 18.
field drifts were not cancelled; however, a simple protocol for 88Sr+
clocks has been developed42 that eliminates both first-order Zeeman 19.
and electric quadrupole shifts of the clock transition and can be imple- 20.
mented in future measurements.
For a 1-ms interrogation time followed by a 1.85-ms recooling and 21.
state preparation period, the effective dead time for each clock is 4.7 ms 22.
and results in a measured clock stability of 3.9 × 10−14 / τ (Fig. 4d). This
value of frequency stability agrees well with numerical simulations we 23.
performed of the optical clock using the measured SBS laser subsystem 24.
noise, which predict the clock to operate at the level of 4.1 × 10−14 / τ
for 4.7 ms of dead time. The good correspondence between measure- 25.
ment and simulation suggests that under normal clock operation with
1.85 ms of dead time, the clock stability would reach 2.5 × 10−14 / τ . 26.
The advances made here to the SBS laser bring that laser’s frequency
fluctuations into a new regime that approaches the level of performance 27.
only currently achievable by BCS lasers. When the SBS laser subsystem
28.
is used to interrogate an atomic system, the combination offers the
potential for creating portable optical atomic clocks with stability 29.
surpassing that of state-of-the-art microwave clocks by an order of
magnitude or more. For the realization of a truly portable clock, future
30.
effort is required in miniaturizing the atomic physics package and in
maturing the technology of compact frequency combs43 and SBS lasers.
31.
Additional benefits in size and vibration tolerance may be gained by
replacing the fibre SBS resonator with an integrated resonator19,23 pend-
ing further improvements to the linewidth and stability of chip-based 32.
SBS lasers. Although not directly explored in this work, other promising
33.
architectures for ultralow-noise lasers44,45 may also benefit from the
techniques of temperature stabilization developed here. 34.
35.
Online content 36.
Any methods, additional references, Nature Research reporting sum- 37.
maries, source data, extended data, supplementary information,
acknowledgements, peer review information; details of author con- 38.
tributions and competing interests; and statements of data and code
availability are available at https://doi.org/10.1038/s41586-020-2981-6. 39.
1. Hinkley, N. et al. An atomic clock with 10−18 instability. Science 341, 1215–1218 (2013). 40.
2. Bloom, B. J. et al. An optical lattice clock with accuracy and stability at the 10−18 level.
Nature 506, 71–75 (2014). 41.
3. Brewer, S. M. et al. 27Al+ quantum-logic clock with a systematic uncertainty below 10−18.
Phys. Rev. Lett. 123, 033201 (2019). 42.
4. Koller, S. B. et al. Transportable optical lattice clock with 7 × 10−17 uncertainty. Phys. Rev.
Lett. 118, 073601 (2017). 43.
5. Godun, R. M. et al. Frequency ratio of two optical clock transitions in 171Yb+ and constraints
on the time variation of fundamental constants. Phys. Rev. Lett. 113, 210801 (2014). 44.
6. Huntemann, N., Sanner, C., Lipphardt, B., Tamm, Chr. & Peik, E. Single-ion atomic clock
with 3 × 10−18 systematic uncertainty. Phys. Rev. Lett. 116, 063001 (2016). 45.
7. Leibrandt, D. R., Thorpe, M. J., Bergquist, J. C. & Rosenband, T. Field-test of a robust,
portable, frequency-stable laser. Opt. Express 19, 10278–10286 (2011). 46.
8. Liu, L. et al. In-orbit operation of an atomic clock based on laser-cooled 87Rb.
Nat. Commun. 9, 2760 (2018).
9. Takamoto, M. et al. Test of general relativity by a pair of transportable optical lattice Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in
clocks. Nat. Photon. 14, 411–415 (2020). published maps and institutional affiliations.
10. Maleki, L. & Prestage, J. Applications of clocks and frequency standards: from the routine
to tests of fundamental models. Metrologia 42, S145–S153 (2005). © The Author(s), under exclusive licence to Springer Nature Limited 2020
Ludlow, A. D., Boyd, M. M., Ye, J., Peik, E. & Schmidt, P. O. Optical atomic clocks. Rev. Mod.
Phys. 87, 637–701 (2015).
McGrew, W. F. et al. Atomic clock performance enabling geodesy below the centimetre
level. Nature 564, 87–90 (2018).
Young, B. C., Cruz, F. C., Itano, W. M. & Bergquist, J. C. Visible lasers with subhertz
linewidths. Phys. Rev. Lett. 82, 3799–3802 (1999).
Jiang, Y. Y. et al. Making optical atomic clocks more stable with 10−16-level laser
stabilization. Nat. Photon. 5, 158–161 (2011).
Kessler, T. et al. A sub-40-mHz-linewidth laser based on a silicon single-crystal optical
cavity. Nat. Photon. 6, 687–692 (2012).
Davila-Rodriguez, J. et al. Compact, thermal-noise-limited reference cavity for
ultra-low-noise microwave generation. Opt. Lett. 42, 1277–1280 (2017).
Didier, A. et al. Ultracompact reference ultralow expansion glass cavity. Appl. Opt. 57,
6470–6473 (2018).
Grudinin, I. S., Matsko, A. B. & Maleki, L. Brillouin lasing with a CaF2 whispering gallery
mode resonator. Phys. Rev. Lett. 102, 043902 (2009).
Lee, H. et al. Chemically etched ultrahigh-Q wedge-resonator on a silicon chip.
Nat. Photon. 6, 369–373 (2012).
Kabakova, I. V. et al. Narrow linewidth Brillouin laser based on chalcogenide photonic
chip. Opt. Lett. 38, 3208–3211 (2013).
Loh, W. et al. Dual-microcavity narrow-linewidth Brillouin laser. Optica 2, 225–232
(2015).
Otterstrom, N. T., Behunin, R. O., Kittlaus, E. A., Wang, Z. & Rakich, P. T. A silicon Brillouin
laser. Science 360, 1113–1116 (2018).
Gundavarapu, S. et al. Sub-hertz fundamental linewidth photonic integrated Brillouin
laser. Nat. Photon. 13, 60–67 (2019).
Geng, J. et al. Highly stable low-noise Brillouin fiber laser with ultranarrow spectral
linewidth. IEEE Photonics Technol. Lett. 18, 1813–1815 (2006).
Shee, Y. G., Al-Mansoori, M. H., Ismail, A., Hitam, S. & Mahdi, M. A. Multiwavelength
Brillouin-erbium fiber laser with double-Brillouin-frequency spacing. Opt. Express 19,
1699–1706 (2011).
Tow, K. H. et al. Linewidth-narrowing and intensity noise reduction of the 2nd order
Stokes component of a low threshold Brillouin laser made of Ge10As22Se68 chalcogenide
fiber. Opt. Express 20, B104–B109 (2012).
Debut, A., Randoux, S. & Zemmouri, J. Linewidth narrowing in Brillouin lasers: theoretical
analysis. Phys. Rev. A 62, 023803 (2000).
Loh, W., Yegnanarayanan, S., O’Donnell, F. & Juodawlkis, P. W. Ultra-narrow linewidth
Brillouin laser with nanokelvin temperature self-referencing. Optica 6, 152–159 (2019).
Jefferts, S. R., Meekhof, D. M., Shirley, J. H., Stepanovic, M. & Parker, T. E. Accuracy results
from NIST-F1 laser-cooled cesium primary frequency standard. In Proc. IEEE/EIA Int.
Frequency Control Symposium and Exhibition 714–717 (IEEE, 2000).
Madej, A. A., Dubé, P., Zhou, Z., Bernard, J. E. & Gertsvolf, M. 88Sr+ 445-THz single-ion
reference at the 10−17 level via control and cancellation of systematic uncertainties and its
measurement against the SI second. Phys. Rev. Lett. 109, 203002 (2012).
Strekalov, D. V., Thompson, R. J., Baumgartel, L. M., Grudinin, I. S. & Yu, N. Temperature
measurement and stabilization in a birefringent whispering gallery mode resonator.
Opt. Express 19, 14495–14501 (2011).
Fescenko, I. et al. Dual-mode temperature compensation technique for laser stabilization
to a crystalline whispering gallery mode resonator. Opt. Express 20, 19185–19193 (2012).
Weng, W. et al. Nanokelvin thermometry and temperature control: beyond the thermal
noise limit. Phys. Rev. Lett. 112, 160801 (2014).
Lim, J. et al. Probing 10 μK stability and residual drifts in the cross-polarization dual-mode
stabilization of single-crystal ultrahigh-Q optical resonators. Light Sci. Appl. 8, 1 (2019).
Drever, R. W. P. et al. Laser phase and frequency stabilization using an optical resonator.
Appl. Phys. B 31, 97–105 (1983).
Storz, R., Braxmaier, C., Jäck, K., Pradl, O. & Schiller, S. Ultrahigh long-term dimensional
stability of a sapphire cryogenic optical resonator. Opt. Lett. 23, 1031–1033 (1998).
Bruzewicz, C. D., McConnell, R., Chiaverini, J. & Sage, J. M. Scalable loading of a
two-dimensional trapped-ion array. Nat. Commun. 7, 13005 (2016).
Wang, S. X., Labaziewicz, J., Ge, Y., Shewmon, R. & Chuang, I. L. Demonstration of a
quantum logic gate in a cryogenic surface-electrode ion trap. Phys. Rev. A 81, 062332
(2010).
Ma, L. S., Junger, P., Ye, J. & Hall, J. L. Delivering the same optical frequency at two places:
accurate cancellation of phase noise introduced by an optical fiber or other time-varying
path. Opt. Lett. 19, 1777–1779 (1994).
Nicholson, T. L. et al. Comparison of two independent Sr optical clocks with 1 × 10−17
stability at 103 s. Phys. Rev. Lett. 109, 230801 (2012).
Nicholson, T. L. et al. Systematic evaluation of an atomic clock at 2 × 10−18 uncertainty.
Nat. Commun. 6, 6896 (2015).
Dubé, P. et al. Electric quadrupole shift cancellation in single-ion optical frequency
standards. Phys. Rev. Lett. 95, 033001 (2005).
Spencer, D. T. et al. An optical frequency synthesizer using integrated photonics.
Nature 557, 81–85 (2018).
Liang, W. et al. Ultralow noise miniature external cavity semiconductor laser.
Nat. Commun. 6, 7371 (2015).
Zhang, W. et al. Ultranarrow linewidth photonic-atomic laser. Laser Photonics Rev. 14,
1900293 (2020).
Howe, D. A. The total deviation approach to long-term characterization of frequency
stability. IEEE Trans. Ultrason. Ferroelectr. Freq. Control 47, 1102–1110 (2000).
Nature | Vol 588 | 10 December 2020 | 249
Article
Methods
SBS laser setup
A 1,348-nm external-cavity diode laser with an output of 30 mW serves as
the pump of the SBS laser. The pump light is split in a 90:10 ratio between
two separate paths that are used to probe the two orthogonal polariza-
tion modes of the optical resonator. The 90% path passes through an
additional AOM, which equalizes the power between the two paths.
Each path is also phase modulated and amplified such that the power
reaches about 10 mW at the resonator input, which produces approxi-
mately 5 mW of SBS light at the output. It is essential to the operation
of this SBS laser that the counter-propagating SBS light generated by
one pump travels alongside the opposite polarization pump that passes
through the resonator. The PBS that follows after the resonator then
separates the SBS light from that of the orthogonal polarization pump.
The pump light is demodulated and used for PDH locking to the cavity
resonance. 10% of the output SBS power is tapped off and used to servo
the SBS laser’s amplitude, while the remaining 90% is redirected out by
a circulator. The majority of the passive optical fibre components for
both polarization paths, including circulators, couplers and polarization
beam splitters, are spliced together and contained on a 5 inch × 5 inch
plate. The resonator itself sits within a copper enclosure comprising a
copper platform that rests on a layer of Viton. The enclosure isolates
the resonator from the environment but permits the ability to control
the resonator’s temperature through the copper.
The output of the circulator is amplified by an SOA and then fre-
quency doubled through a fibre-pigtailed potassium titanyl phos-
phate (KTP) waveguide doubler for interrogation of the 88Sr+ ion at
674 nm. For an amplified SBS power of about 60 mW at the input of the
frequency doubler, the 674-nm output after doubling reaches about
60 μW, corresponding to an efficiency of 1.7% W−1. In order to achieve
stable injection locking of a subsequent slave laser, the 674-nm SBS
power should exceed 10 μW. Through possible future integration of the
SBS laser with the frequency doubler, the coupling loss into the doubler
may be avoided thereby eliminating the need for one or more of the
amplification and/or injection locking stages. This would require the
union of an on-chip SBS laser, which has recently been demonstrated
in a low-confinement Si3N4 platform23, with a chip-based frequency
doubler, which necessitates the use of a χ(2) material such as LiNbO3.
The integration of these two material systems may be accomplished
through wafer bonding techniques using tapers that transition the
optical mode from Si3N4 to LiNbO3 (ref. 47).
Feedforward implementation
In order to stabilize the SBS laser via feedforward, a frequency scaling
factor of 39 must be applied to the dual-polarization beat note. This
factor is tuned to match the long-term frequency movement of the
polarization beat to the drift of the SBS laser (see Extended Data Fig. 2).
Although multiplication by 39 appears to be optimal, the accuracy
of the correction signal does not degrade substantially for scale fac-
tors of 39 ± 2. In addition to a frequency scaling, the polarization beat
must undergo (1) a ramp subtraction to cancel a linear component
of the SBS drift and (2) a 0.03 Hz low-pass filtering step that removes
excess noise at short timescales. These steps are accomplished by first
phase-locking an RF waveform generator (FM deviation = 30 kHz for
Vin = ±2.5 V) to the polarization beat (FM, frequency modulation). Owing
to the intrinsically lower noise of the waveform generator at micro-
wave frequencies, the servo output becomes a voltage replica of the
information contained in the polarization beat note. After subtracting
a ramp and low-pass filtering the result, the voltage signal is sent to a
tunable amplifier with its amplification factor set to 2. The output of the
amplifier is used to control the frequency of a second signal generator
operating with FM deviation = 1.075 MHz and Vin = ±5 V. The combina-
tion of the amplification factor along with the chosen FM deviation
values produces the scale factor of 39 applied to the polarization beat.
Measurement of the SBS laser noise
The SBS laser’s frequency drift, noise and spectrum are all measured
by interfering the frequency-doubled SBS output with an independent
bulk-cavity-stabilized laser operating at 674 nm. After photodetec-
tion, a 0−200 MHz signal is generated whose noise characteristics are
assumed to all be attributed to the SBS laser. In contrast, the 180-MHz
dual-polarization beat note is directly produced by separately inter-
fering two orthogonal-polarization SBS signals. For measuring drift,
the frequencies of the SBS laser and dual-polarization beat note are
tracked across two frequency counters that are simultaneously trig-
gered. Alternatively, for measuring the spectrum, the SBS signal is
mixed to 5 MHz and directly detected on a spectrum analyser.
The frequency noise of the SBS laser, before and after feedforward, is
depicted in Extended Data Fig. 1. The feedforward system adds a slight
amount of additional noise to the SBS laser including spurious peaks
across the spectrum, but otherwise does not substantially degrade the
SBS laser’s performance. The white-noise floor is slightly increased to
7 Hz2 Hz−1. The feedforward-enabled SBS laser linewidth is estimated to
be 51 Hz, which is found by integrating the noise spectrum48, and agrees
with other linewidth estimations based on a direct spectrum measure-
ment and on a fractional frequency noise measurement. After account-
ing for the frequency doubling, this measured SBS laser frequency
noise is similar in spectral shape and white-noise floor as compared
to the 20-Hz-linewidth SBS laser reported in ref. 28.
SBS laser residual linear drift
The residual linear drift of the SBS laser operating at 1,348 nm is tracked
across two separate but otherwise identical resonators in order to
monitor how the amplitude of this drift changes with time (Extended
Data Fig. 3). The first resonator starts at a linear drift of 1.6 kHz s−1 and
reaches a drift of 0.2 kHz s−1 by the end of 79 days. The second resonator,
which is used for the datasets contained in this work, starts at a linear
drift of 2.8 kHz s−1 and reaches 0.6 kHz s−1 after 13 days. An exponential
fit to the plotted data points across both resonators reveals a 1/e time
constant of 6.9 days. In all of the cases reported, the temperature cor-
rection factor remained constant at 39.
In our initial experiments with the SBS laser subsystem, we used a
BCS laser as a stable reference to monitor the linear drift of the SBS
laser by tracking the beat note between the two optical signals. This
use of the BCS laser was merely a convenience, and is not required for
the clock application. When nothing about the drift of the SBS laser
is known, the atom itself can be used as a stable frequency reference.
The frequency of the full SBS laser subsystem must first be brought
to within a few tens of megahertz of the atomic resonance frequency;
this degree of frequency determination is afforded by commercial
wavelength meters based on Fizeau interferometers. When the laser
frequency is close to the clock transition, we may perform Rabi spec-
troscopy by scanning the frequency of the laser using an AOM and
search for atomic transitions. There are multiple transitions in the
neighbourhood of our chosen clock transition, due to the number of
distinct Zeeman sublevels and the fact that each transition is dressed
with sidebands displaced by multiples of the ion’s trap frequency. With
our knowledge of the magnetic field, the laser polarization and the
frequency of the ion’s motion—all of which can be determined without
using the laser—we are able to uniquely identify our clock transition.
Next, we perform repeated spectroscopy across the clock transition
(see Extended Data Fig. 4); since the drift of the SBS laser’s frequency
is the dominant source of frequency error, any apparent shift in the
clock transition is attributed to the SBS laser. See also ref. 49 where a
similar method is used to track the drift of a BCS laser.
An alternative to this method is to simply wait for a long enough time
that the laser drift decreases to an acceptable level, as suggested by
the decay discussed above. In fact, at the time that the measurements
in Extended Data Fig. 4 were taken, the drift of the SBS resonator had
decreased to approximately 30 Hz s−1 (at 1,348 nm). We have found that
this level of drift does not preclude the operation of the clock—even
without additional corrections. To obtain the results in Extended Data
Fig. 4, we intentionally applied a linear drift to the SBS laser at a level
approximately equal to the highest drift measured in Extended Data
Fig. 3, to demonstrate how one would be able to proceed if this resona-
tor were being used closer to its original construction date.
Finally, once the linear drift is brought down to an acceptable level, the
error signal of the clock itself can be used to track the frequency of the
SBS laser subsystem with respect to the frequency of the atomic tran-
sition. By accumulating the frequency corrections applied to the final
frequency-shifting AOM, the frequency difference between the SBS laser
subsystem and the clock transition can be monitored in real time (see
Extended Data Fig. 4d). In Extended Data Fig. 4d, we run our clock protocol
with a reduced interrogation time, which leads to a decreased frequency
resolution but an increased capture range for the lock signal. For future
operations, we may construct our control software such that a limited ver-
sion of the clock protocol runs between all other calibration routines and
experiments. The frequency difference could thus be constantly tracked
and programmatically fed forward into each experimental operation.
Characteristics of the BCS laser
In our laboratory, we have access to an ultra-stable reference cav-
ity, consisting of two highly reflective mirrors with a spacer made of
ultra-low-expansion (ULE) glass. The cavity is housed in a high vacuum
environment inside a chamber that has two independently controlled
radiation shields to minimize temperature fluctuations. The vacuum
chamber sits on an active vibration cancellation stage on an optical
table in the same room as the SBS laser setup. The mirrors are separated
by 77.57 mm, corresponding to a free spectral range of 1.932 GHz. The
cavity finesse is approximately 105 at 674 nm, which results in a linewidth
of about 20 kHz. From independent measurements of the 88Sr+ qubit
transition with a laser locked to the ULE cavity, we have determined
that the length of the cavity varies slowly, resulting in a drift of about
60 mHz s−1 or 1.3 × 10−16 at one second.
Limits of the optical clock measurement
Several factors limit the linewidth with which we are able to interrogate
the ion. Experiments are performed in a room that is subject to acoustic
noise from four cryogenic vacuum apparatuses in constant operation,
and, while acoustic noise cancellation is performed in the optical fibre
pathways, there are still sections of free-space optics (approximately
4 m total path length) that can vibrate and thus write noise onto the
optical signal. Two tapered amplifiers are in the free-space path, which
are both driven into saturation. Acoustic interference can modulate
the coupling into the amplifier chip, which can then appear as phase
modulation on the laser. Our ability to stabilize the magnetic field in the
cryogenic chamber—and thus the frequency of the ion’s clock transi-
tion—also limits the atomic coherence time and effectively increases the
linewidth of the measured transition. As a check to establish a baseline
level of performance given these noise sources, we use the BCS laser
as the local oscillator in the same clock protocol described in the main
text (see Extended Data Fig. 5). We find the stability of this clock to be
close to, but slightly better than, the stability of the SBS clock, which
suggests that noise in the experiment is not a substantial limitation in
our ability to evaluate the performance of the SBS as the local oscillator.
Clock protocol and simulation
Our clock protocol consists of a Ramsey measurement, with the initial
π/2 pulse driven by the SBS laser, an interrogation time of 1 ms, and a
final π/2 pulse 90° out of phase with the first pulse. Both pulses are
amplitude corrected using a composite pulse sequence50 to mitigate
amplitude fluctuations. After completing this Ramsey sequence, we
measure the ion state via state-dependent scattering of photons on
the cycling S1/2 → P1/2 ion transition. After determining which of the two
clock states the ion is in, we update a frequency correction applied to
the ion as follows:
1 αmi
Δfi +1 = Δfi + , (1)
2π τ
where τ = 1 ms is the interrogation time, α ≈ 0.2 is a gain term optimized
to achieve the best stability, and mi represents the measurement result,
with mi = +1 (−1) corresponding to finding the ion in the D5/2 (S1/2) state.
In order to implement our interleaved clock self-comparison, we run
two independent control loops according to the above protocol. During
odd-numbered interrogations, we use and update the first frequency
(1)
correction Δfi ; during even-numbered interrogations we use and
(2)
update the second frequency correction Δfi . See Extended Data Fig. 6
for an illustration of the timing of the interleaved clocks. After the exper-
iment is over, the two frequency series Δf(1) and Δf(2) are compared against
one another to compute the Allan deviation of the interleaved clock.
The performance of the interleaved clock is compared to a numeri-
cal simulation we developed. This simulation models noise of the SBS
laser using measured data for timescales of 1 ms and above (taken from
comparisons with a BCS laser) to which is added a white noise term
with spectral density of 7 Hz2 Hz−1 to simulate short-timescale fluctua-
tions of the laser frequency. In the simulation, the laser frequency is
updated once per clock cycle based on ion measurements that take
into account quantum projection noise and dead time effects. Extended
Data Fig. 7 shows the measured Allan deviation of the unlocked SBS
laser, the simulated clock operation in self-comparison mode, and the
equivalent Allan deviation that the simulation predicts if a single clock
were run with an interrogation time of 1 ms and dead time of 1.85 ms.
The simulated clock self-comparison finds Allan deviation of 4.1 × 10−14
at 1 s, in excellent agreement with our experimental result of
3.9 × 10−14 / τ. The same model predicts that a single clock interrogated
by the SBS laser would achieve stability of 2.5 × 10−14 in 1 s.
Data availability
The datasets that support this study are available from the correspond-
ing author on reasonable request.
Code availability
The codes used for analysis and simulations are available from the
corresponding author on reasonable request.
47. Chang, L. et al. Heterogeneous integration of lithium niobate and silicon nitride
waveguides for wafer-scale photonic integrated circuits on silicon. Opt. Lett. 42,
803–806 (2017).
48. Hjelme, D. R., Mickelson, A. R. & Beausoleil, R. G. Semiconductor laser stabilization by
external optical feedback. IEEE J. Quantum Electron. 27, 352–372 (1991).
49. Akerman, N., Navon, N., Kotler, S., Glickman, Y. & Ozeri, R. Universal gate-set for
trapped-ion qubits using a narrow linewidth diode laser. New J. Phys. 17, 113060 (2015).
50. Cummins, H. K., Llewellyn, G. & Jones, J. A. Tackling systematic errors in quantum logic
gates with composite rotations. Phys. Rev. A 67, 042308 (2003).
Acknowledgements We thank A. Libson, G. N. West and I. L. Chuang for helpful discussions.
This work was sponsored by the Under Secretary of Defense for Research and Engineering
under Air Force contract number FA8721-05-C-0002. Opinions, interpretations, conclusions
and recommendations are those of the authors and are not necessarily endorsed by the US
Government.
Author contributions W.L., J.S. and R.M. conceived, designed and carried out the experiments
with the SBS laser. W.L., J.S., D.R. and R.M. conceived, designed and carried out the experiments
with the clock protocol. All authors discussed the results and contributed to the manuscript.
Competing interests The authors declare no competing interests.
Additional information
Correspondence and requests for materials should be addressed to W.L.
Peer review information Nature thanks Clément Lacroûte and the other, anonymous,
reviewer(s) for their contribution to the peer review of this work.
Reprints and permissions information is available at http://www.nature.com/reprints.
Article

Extended Data Fig. 1 | SBS laser noise. Shown is the measurement of the
674-nm SBS laser subsystem’s frequency noise before and after the application
of feedforward stabilization. The SBS laser subsystem features a white noise
floor of 3 Hz 2 Hz−1 before feedforward and a gradual increase in noise at lower
offset frequencies. An integration over the noise spectral density yields a
full-width at half-maximum linewidth of 45 Hz, while a calculation of the
white-noise limited linewidth at large offset frequencies yields 9 Hz. With
feedforward turned on, the SBS laser subsystem’s noise increases slightly and
exhibits additional noise peaks that arise from the RF signal generator used for
feedforward correction. The integrated linewidth increases to 51 Hz.
Extended Data Fig. 2 | Optimization of feedforward stabilization.
a, Measurement of the differential temperature sensitivity of the SBS
resonator’s two orthogonal-polarization modes. The linewidths of the modes
are measured to be 1.2 MHz (blue arrows). For an applied temperature shift of
ΔT = −0.25 °C, the centre frequency at 1,348 nm changes by 490 MHz, while the
mode separation (black arrows) changes by 11.9 MHz. This corresponds to a
feedforward correction ratio of 40:1. b, Time series trace of the 1,348-nm
SBS laser’s frequency with the linear drift removed and the SBS amplitude
unservoed. The lack of correspondence between the free-running SBS
(red trace) and the polarization beat note (‘Pol. beat ×39’; blue trace) indicates
the inability to cancel frequency drift when amplitude noise is present.
Article

Extended Data Fig. 3 | Linear drift decay of SBS resonator. Record of the
SBS laser’s linear drift at 1,348 nm for two resonators. The first resonator
(red circles) is tracked over 79 days, and its linear drift decreases to a value
of 200 Hz s−1 at the end of the elapsed period of time. The second resonator
(blue squares) is tracked over 260 days and reaches a minimum of 30 Hz s−1.
Extended Data Fig. 4 | Determining and tracking linear drift. a, Rabi
spectroscopy of the |5S1/2, m J = −1/2⟩ → |4D5/2, m J = −3/2⟩ clock transition and the
first-order motional sidebands at νclock ± ν trap (blue and red sideband,
respectively) is taken at regular intervals of approximately 25 s. After
performing fits to the symmetric sidebands, we can average the two
frequencies to obtain an accurate measure of the frequency of the central
feature. Only the spectroscopic data for the final experiment (black line) is
shown; for all other datasets, the Gaussian peak fits to the sidebands (red and
blue curves) are shown with progressively darkening colour to illustrate the
movement of these features over time. For these data, an intentional linear
drift of 5 kHz s−1 (at 674 nm; equivalently 2.5 kHz s−1 at 1,348 nm) was applied to
demonstrate the efficacy of this method in cases of high drift, as in the initial
few points shown in Extended Data Fig. 3. b, Rabi spectroscopy of the clock
transition and sidebands after applying a linear drift correction to null out the
natural drift of the resonator’s frequency. Over the course of 20 min of
measurements, very little deviation in the centre frequency is observed.
c, Linear drift determined from the data presented in b and c. The linear drift
can be obtained from a fit (lines) to the apparent frequency of the clock
transition as a function of time (data points). In the first case, with the large
drift intentionally applied to the laser frequency, we obtain a drift of 5.2 kHz s−1
(at 674 nm) from the fit (green line). After a few iterations of applying a
correction and measuring the resulting drift, the drift is driven down to 17 Hz s−1
(blue line). d, Integrated clock correction signal applied to the laser to keep the
frequency resonant with the atom’s transition. In this case, we use a simplified
clock protocol with an interrogation time of τ = 100 μs and no interleaving.
Article

Extended Data Fig. 5 | BCS laser 88Sr+ ion clock. Measured interleaved clock
performance comprising a BCS laser locked to a 88Sr+ ion operating with 1-ms
interrogation time (blue points). The effective dead time is 4.7 ms. The blue
points represent the frequency noise at a selection of averaging times, and the
vertical blue bars indicate 1σ error. A fit (dashed line) to the data yields a
stability of 3.1 × 10−14/ τ , which is slightly lower than the same clock operated
with a SBS laser.
Extended Data Fig. 6 | Schematic of interleaved clock protocol. A pictorial
representation of the interleaved clock procedure is shown. Here the Doppler
segments represent the 700-μs duration in which the ion is Doppler cooled.
During the OP segments, the ion undergoes 450 μs of optical pumping in order
to prepare the electron in the lower level of the clock transition. The
‘Interrogate’ segments are each 1 ms of interrogation time, bounded by
composite π/2 pulses. Last, the ‘Detect’ segments are 700 μs of detection time,
during which the photons emitted by the ion are detected on a photomultiplier
tube and counted by our timing controller. During the ‘Update’ segment, and
depending on the number of photons collected, the state of the ion is
determined, and the frequency of the clock is either increased or decreased. As
discussed in the text, two separate clock signals, f(1) and f(2), are maintained;
here these are indicated as Clock 1 and Clock 2. While the frequency of either
clock is updated, the experiment begins to prepare the state for the next
measurement, as indicated by the black arrows. Each of these clocks is sensitive
to laser frequency fluctuation only during the 1 ms interrogation period of the
total 5.7 ms cycle time; during all other times, the frequency of the laser must
stay within the capture range of the lock.
Article

Extended Data Fig. 7 | Numerical simulation of clock performance. The

measured performance of the stabilized SBS laser (Allan deviation, red curve)
is used as an input into a clock protocol simulation incorporating projection
noise and dead time. The simulation accurately predicts the measured clock
performance via the interleaved self-comparison (green squares) and predicts
a single-clock stability of 2.5 × 10−14/ τ (blue circles).
Article

Capillary condensation under atomic-scale

confinement

https://doi.org/10.1038/s41586-020-2978-1 Qian Yang1,2 ✉, P. Z. Sun1,2, L. Fumagalli2, Y. V. Stebunov2, S. J. Haigh3, Z. W. Zhou4,

I. V. Grigorieva1,2, F. C. Wang5 ✉ & A. K. Geim1,2 ✉
Received: 29 April 2020

Accepted: 21 September 2020

Capillary condensation of water is ubiquitous in nature and technology. It routinely
Published online: 9 December 2020
occurs in granular and porous media, can strongly alter such properties as adhesion,
Check for updates
lubrication, friction and corrosion, and is important in many processes used by
microelectronics, pharmaceutical, food and other industries1–4. The century-old
Kelvin equation5 is frequently used to describe condensation phenomena and has
been shown to hold well for liquid menisci with diameters as small as several
nanometres1–4,6–14. For even smaller capillaries that are involved in condensation
under ambient humidity and so of particular practical interest, the Kelvin equation is
expected to break down because the required confinement becomes comparable to
the size of water molecules1–22. Here we use van der Waals assembly of two-dimensional
crystals to create atomic-scale capillaries and study condensation within them. Our
smallest capillaries are less than four ångströms in height and can accommodate just a
monolayer of water. Surprisingly, even at this scale, we find that the macroscopic
Kelvin equation using the characteristics of bulk water describes the condensation
transition accurately in strongly hydrophilic (mica) capillaries and remains
qualitatively valid for weakly hydrophilic (graphite) ones. We show that this
agreement is fortuitous and can be attributed to elastic deformation of capillary
walls23–25, which suppresses the giant oscillatory behaviour expected from the
commensurability between the atomic-scale capillaries and water molecules20,21. Our
work provides a basis for an improved understanding of capillary effects at the
smallest scale possible, which is important in many realistic situations.

The Kelvin equation predicts that capillaries become spontaneously scale, the Kelvin equation is usually modified to account for ‘wetting
filled with water at the relative humidity films’ that are adsorbed on internal surfaces before the condensation
transition and effectively narrow the capillaries. For the smallest capil-
RHK = exp(−2σ /kBTdρN) (1) laries, the thickness of the wetting films was used as a free parameter. In
the real world, pores, cracks and cavities obviously do not terminate at
where σ ≈ 73 mJ m−2 is the surface tension of water at room tempera- the scale of several nanometres but extend even below 1 nm or 2σ/kBTρN,
ture T, ρN ≈ 3.3 × 1028 m−3 is the number density of water, kB is the Boltz- the fact that makes condensation phenomena omnipresent under ambi-
mann constant and d is the diameter of the meniscus curvature. For a ent conditions. The latter scale is comparable to the diameter of water
two-dimensional (2D) confinement created by parallel walls separated molecules, which makes it challenging to study experimentally because
by a distance h, d = h/cos θ where θ is the contact angle of water on the of difficulties in creating the required atomic-scale confinement1,10,12,
walls’ material. For capillary condensation to occur at relative humid- the varying thickness of wetting films1,2,7–13,17 and huge capillary pres-
ity (RH) considerably below 100%, equation (1) dictates that d must be sures that can cause considerable deformations13,23–25. As for theory,
comparable to 2σ/kBTρN ≈ 1.1 nm. For example, under typical ambient the Kelvin equation is also believed to reach its applicability limit for
RH of 40–50%, water is expected to condense in slits with h < 1.5 nm and confinement containing a few molecular layers because, at this smallest
cylindrical pores with diameters <3 nm, if θ is close to zero. Even stronger scale, the properties of water notably change2,3,12,15,16 and the description
confinement is required for capillaries involving less hydrophilic materi- in terms of homogeneous macroscopic thermodynamics becomes
als. So far, a broad consensus has been reached that the Kelvin equation questionable1–4,16–20, leaving aside the fact that such quantities as d and
remains accurate for menisci with d ≥ 8 nm (refs. 1–4,6–11) and can also θ in equation (1) can no longer be defined1–3,18–20.
describe condensation phenomena in hydrophilic pores as small as The capillary devices that we studied are shown schematically in
4 nm in diameter12–14. To achieve agreement with the experiments at this Fig. 1a. Their most important part is atomically flat 2D channels made
1
National Graphene Institute, University of Manchester, Manchester, UK. 2Department of Physics and Astronomy, University of Manchester, Manchester, UK. 3School of Materials, University of
Manchester, Manchester, UK. 4Key Laboratory of Advanced Technologies of Materials, School of Materials Science and Engineering, Southwest Jiaotong University, Chengdu, China. 5Chinese
Academy of Sciences Key Laboratory of Mechanical Behavior and Design of Materials, Department of Modern Mechanics, University of Science and Technology of China, Hefei, China.
✉e-mail: qian.yang-2@manchester.ac.uk; wangfc@ustc.edu.cn; geim@manchester.ac.uk

250 | Nature | Vol 588 | 10 December 2020

 a AFM chamber
Top crystal
Graphene spacers
l
sta
cry
ne
H e
a
m
r
mb
tto
e 0
me
Na
Bo
w
fer
SiN
wa
Water
Si
molecules
Chamber with
controlled humidity
Sagging depth (Å)
c d
G 80%
h 85%
5Å
500 nm
Fig. 1 | Atomic-scale capillaries and water condensation inside. a, Schematic
of the capillary devices studied here. b, Cross-sectional imaging of a four-layer
graphite capillary by scanning transmission electron microscopy (STEM). The
top layer was more than 100 nm thick in this case. c, d, AFM imaging of the same
mica capillary (N = 11) exposed to 30% and 95% relative humidity, respectively.
In the dry state, the top crystal sagged by ∼5 Å, but it became flat at high RH, as
illustrated in the corresponding schematics above the images. The black
dotted lines indicate the edge of the top crystal (compare with a). In the upper
part of the AFM images, the colour scale is given by the observed sagging (grey
curves). The bottom part shows graphene spacers without the top crystal cover
by van der Waals (vdW) assembly following the fabrication procedures
described in the Methods. In brief, two atomically flat crystals were
exfoliated from bulk muscovite mica or graphite to become the top and
bottom walls of our capillaries. Separately, narrow strips of multilayer
graphene were fabricated to serve as spacers between the two mica
or graphite crystals. Stacking the crystals and spacers on top of each
other resulted in the 2D channels shown in Fig. 1 and Extended Data
Fig. 1. We used graphene spacers between N = 2 and about 10 layers
thick so that the capillaries had the designated height Na (see Fig. 1a),
where a ≈ 3.35 Å is the effective thickness of monolayer graphene26,27.
Examples of transmission electron microscopy imaging of our capillar-
ies are provided in Fig. 1b and Extended Data Fig. 1d. Mica and graphite
were chosen as archetypal strongly and weakly hydrophilic materials.
Their contact angles are known to be in the range of 0–20° and 55–85°,
respectively16,28,29. For surfaces exposed to air, θ is close to the above
upper bounds28,29 (Methods).
To detect RH at which capillary condensation occurred in the 2D cap-
illaries described above, we exploited the fact26,30 that the suspended
thin crystals exhibited noticeable sagging caused by their vdW adhe-
sion to sidewalls (Fig. 1c). In our experiments we found that, when the
capillaries became filled with water, the sagging depth δ diminished
(Fig. 1d), presumably because intercalating water molecules ‘screen’ the
adhesion27,30. To make the resulting changes in δ detectable by atomic
force microscopy (AFM), it was important to choose the thickness H
of the top crystal and the channel width w carefully (see Fig. 1a). As
described in Methods (‘Remnant sagging above the condensation
b STEM bright field
Four graphene layers
20 nm
50%
3Å
60%
–2
70%
75%
90%
100 200
Distance (nm)
–4
40 60 80 100
Relative humidity (%)
(dark-to-bright scale, 40 Å). e, Sagging depth δ as a function of RH for a graphite
capillary with N = 4. Coloured symbols, AFM measurements. The grey symbol
with error bars indicates our experimental accuracy. The two solid curves in
orange indicate the constant sagging δ0 below the condensation transition and
the ln(RH) dependence above it. The transition is marked by the dashed vertical
line. Inset, AFM profiles (averaged over 100 nm along the channel) of the top
crystal for several values of RH. All the AFM measurements were carried out in
the non-contact PeakForce mode (Methods, ‘AFM topography under
controlled humidity’).
transition’), these two parameters define the stiffness of the top crystal
and, hence, how deeply it bends inwards. We found that, for w ≈ 150 nm,
the top crystal should be ∼50–70 nm thick to exhibit a sagging depth δ
of several ångströms. If either w or H were changed only by a factor of 2,
the strong dependence δ ∝ w4/H3 resulted in either collapsed channels
(the top crystal attached to the bottom one) or such a small δ (<1 Å)
that the condensation transition was impossible to discern by AFM.
The capillaries studied here were typically 5–10 μm long.
As shown in Fig. 1a and Extended Data Fig. 1c, our capillary devices
were assembled on top of a silicon nitride membrane. It had a rectangu-
lar opening that was extended into the bottom crystal by dry etching.
The Si chip supporting the entire assembly was used to separate two
miniature gas chambers that were integrated into an AFM set-up as
shown in Extended Data Fig. 2a. The bottom chamber provided variable
humidity so that one entrance of the 2D capillaries was exposed to a
chosen RH. The opposite entrance was facing the top chamber, which
enclosed the AFM scanning head and was usually kept at low humidity.
The two-chamber configuration allowed us to avoid the influence of
RH on measurements of the top crystal’s topography (for example, no
condensation occurred at the AFM tip during scanning)31. Examples of
AFM imaging for mica and graphite devices are given in Fig. 1c, d and
in Extended Data Fig. 3, respectively. They reveal pronounced sagging
under dry conditions, which disappeared in high humidity. Typical
evolution of the top crystal’s profiles with changing RH is shown in
Fig. 1e and Extended Data Figs. 4 and 5. In these measurements, we
increased RH inside the bottom chamber in steps of 5%, waited for an
Nature | Vol 588 | 10 December 2020 | 251
Article
a b
100
1 2 3 4
50
80 0 80
Relative humidity RHC and RHK (%)
–50
Energy
(mJ m–2)
60
T = 20°
40
20
0 0
0 5 10 15
Channel height (Å)
Fig. 2 | Condensation transition under extreme 2D confinement. a, Relative
humidity RHC required for water condensation in mica channels of different
heights h. Blue circles indicate experimental observations, their size reflects
the 3.5% experimental uncertainty in determining RHC (Methods, ‘AFM
topography under controlled humidity’). Two solid curves indicate RHK given
by equation (1) with bulk water’s characteristics for the range of possible θ for
mica (colour-coded). The upper curve (open black circles), with its own y axis
and the common x axis, shows our MD calculations for changes in γSL caused by
restructuring of water inside 2D channels (θ ≈ 10°). The arrows mark the energy
hour for the system to stabilize and then recorded AFM images. The
temperature was kept at 294 ± 1 K. For the device in Fig. 1e, the sagging
remained practically constant for RH ≤ 75% and then exhibited a pro-
nounced jump at RHC, which we attribute to the condensation transition
(another example is shown in Extended Data Fig. 5). Further increase
in RH led to a gradual decrease in δ such that the top crystal became
practically flat at RH > 95% (Fig. 1). The remnant sagging at RH > RHC is
well described by the negative capillary pressure which keeps the top
crystal bent inwards even after water has filled the 2D channels, sup-
pressing the adhesion of the top crystal to the sidewalls. Indeed, for
RH > RHC, δ is expected to evolve proportionally to ln(RH) and reach
zero at 100% humidity23,24, in agreement with the observed behaviour
in Fig. 1e and Extended Data Fig. 6 (Methods, ‘Remnant sagging above
the condensation transition’). If we repeated the same measurements
but with decreasing RH, a reverse jump occurred at the same RHC, that
is, the condensation transition was non-hysteretic (Extended Data
Fig. 4a; Methods, ‘Non-hysteretic behaviour of the condensation transi-
tion’). Note, however, that it could take up to several days for capillaries
exposed to high RH to dry out completely and return to their original
state (Extended Data Fig. 4b). On the other hand, for measurements
with increasing RH, no difference in RHC was observed after either an
hour or days of equilibration. Accordingly, our experiments were nor-
mally carried out with increasing rather than decreasing RH, as in Fig. 1e.
Figure 2 summarizes our results for the condensation transitions
observed in mica and graphite 2D capillaries. To allow more accurate
comparison between data collected from different devices, we have
accounted for the fact that capillaries with the same N often exhibited
different sagging in their dry state, δ0. For capillaries with large δ0,
we observed consistently lower RHC than for those with small initial
sagging and same N. Moreover, comparing capillaries with different
N but similar channel heights h = Na − δ0, we found close values of RHC
(Extended Data Fig. 5). This implies that it was the narrowest, central
region of the 2D channels that determined the onset of condensation,
in agreement with general expectations32 (Methods, ‘Effect of initial
252 | Nature | Vol 588 | 10 December 2020
100
T = 85°
Relative humidity RHC and RHK (%)
T = 75°
60
T = 0°
40
20
Mica channels Graphite channels
20 25 30 35 0 5 10 15 20
Channel height (Å)
minima that correspond to the integer number of water monolayers that can fit
inside the 2D capillaries. Red symbols (connected by the dashed curve) are the
expected behaviour calculated using the oscillating γSL shown in the upper
curve and equation (2). Black dashed curve, same analysis but assuming fully
flexible capillary walls allowing relaxation into the energy minima at
commensurate h. Green filled circles, same analysis but for a finite rigidity of
the confining walls. b, Same as a, but for graphite capillaries. The simulated
curves are for θ ≈ 85°.
sagging’). Accordingly, to account for the effect of different δ0, Fig. 2
plots RHC as a function of h rather than of N. For mica capillaries, the
experimental data are well described by equation (1) using θ and σ of
bulk water. Because RHK(h) depends little on the exact value of θ for
strongly hydrophilic capillaries (Fig. 2a), the comparison of RHC for
mica with equation (1) is straightforward. This is not the case for weakly
hydrophilic graphite, for which relatively small variations in θ lead to
considerable changes in RHK(h) as per equation (1). Nonetheless, the
values of RHC observed for our graphite capillaries fall well within the
range expected from the Kelvin equation using the contact angles
θ = 80 ± 5°, typical for graphite surfaces under ambient conditions29.
It is surprising that the macroscopic Kevin equation using the char-
acteristics of bulk water describes condensation in our mica capillar-
ies so well and also provides qualitative agreement for the graphite
capillaries. As mentioned in the introduction, strong discrepancy is
expected for the ångström-scale confinement where only one or two
layers of water fit inside capillaries. Before trying to explain the unex-
pected agreement between the experiment and the macroscopic Kelvin
equation, we note that RHC values in Fig. 2a are notably lower than the
RH values required to achieve condensation in the previous studies
for d ≥ 8 nm. At our low RH, no continuous wetting layer is expected
even on fresh mica surfaces12,33, and a partial coverage by monolayer
water is probably suppressed further by adsorbates from air, which
are responsible for the relatively large θ close to 20°. The same con-
sideration about the apparent absence of wetting films also applies
for the graphite capillaries in which the wetting transition is even less
likely1,2,28. Second, to avoid the macroscopic variables σ and θ that are
poorly defined under our extreme confinement, the Kelvin equation
can be rewritten as1,2,18
RHK = exp[−2(γSV − γSL)/hkBTρN] (2)
where γSV and γSL are the surface energies for solid–vapour and solid–
liquid interfaces, respectively, and γSV − γSL = σ cos θ. The energy γSV
is largely independent of h because the interaction of gas molecules
with surfaces should depend little on confinement. Also, ρN changes
relatively little for nearly incompressible water20,34. Therefore, the
dominant effect of extreme confinement is likely to come from γSL(h),
which is governed by vdW interactions of liquid water with solid sur-
faces. Because these interactions are short-range, it is predominantly
the first near-surface layer of water that determines γSL. If this layer
changes little under confinement, then Δγ = γSL(h) – γSL(∞) ≈ 0, and
capillary condensation should closely follow equation (1) even at the
nanoscale1,2,18. Substantial changes in γSL and, hence, RHC are expected
only in the limit of few-layer water where its near-surface structure is
notably altered20,34 (Extended Data Fig. 7).
For further analysis, we used molecular dynamics (MD) simulations
(Methods) to evaluate Δγ and the resulting corrections to the macro-
scopic Kelvin equation, which are given by the factor exp(2Δγ/hkBTρN)
according to equation (2). Examples of the calculated Δγ(h) are shown
in Fig. 2a and Extended Data Fig. 8. There are pronounced commensu-
rability oscillations20,21,34 in γSL(h) such that energy minima appear if 2D
channels accommodate exactly one, two, three or four molecular layers
of water. The oscillations practically disappear for h > 15 Å where Δγ
becomes almost zero, which also implies that the macroscopic Kelvin
equation should be valid in this regime. For smaller h, changes in γSL
are comparable to σ, which means that the above correction factor is
comparable to RHK itself. Consequently, the simulated RHC(h) depend-
ences shown in Fig. 2 (red dotted curves and symbols) exhibit giant
oscillations such that, for incommensurate h, water condensation
becomes unfavourable and should not occur even at 100% humidity.
No such oscillatory behaviour could be detected in our experiments.
We attribute its absence to elastic adjustment such that 2D channels
tend to accommodate an integer number of molecular layers of water20.
Indeed, the energy minimization should be applied to the entire system,
including the elastic energy of confining walls23–25. For an extremely
soft confinement, 2D channels would adjust their h to reach the com-
mensurate states at minima of Δγ. The condensation behaviour in this
case should follow the step-like black dashed curves shown in Fig. 2.
A finite rigidity pushes the equilibrium conditions away from the
commensurability minima. To estimate a likely elastic response of
our 2D channels, note that the capillary pressure above RHC, which is
defined by σ, keeps the top crystal bent inwards typically by several
ångströms (Fig. 1e; Extended Data Figs. 4–6). Similar elastic adjust-
ments can be expected in our capillaries because changes in Δγ are
comparable to the absolute value of σ. Accordingly, our confinement
should be considered as rather soft. To illustrate the likely conden-
sation behaviour in such a case, the green curves in Fig. 2 show the
RHC(h) dependences expected if the walls’ finite rigidity allows their
deformations to reach within 0.5 Å of the commensurability minima
in Δγ. The latter curves are in good agreement with the experiment
and, in the case of graphite capillaries, also exhibit the same trend
towards lower RH for h < 10 Å as observed experimentally in Fig. 2b. In
principle, the elastic response of 2D confinement could be included in
the simulations self-consistently, but the scatter in the experimental
data and different H used for different capillary devices make this effort
beyond the rationale of the present study.
Finally, we note that elastic adjustments should also play an impor-
tant role in real-life capillaries responsible for condensation phe-
nomena under ambient humidity. Indeed, capillary pressures at 1-nm
scale typically exceed 1 kbar, and the resulting elastic response of even
infinitely thick walls can exceed 1 Å for the case of 2D confinement
(Methods, ‘Remnant sagging above the condensation transition’). This
should force atomic-scale capillaries to elastically adjust their geom-
etry13,23,24, suppressing commensurability oscillations and resulting in
the condensation transition at RH close to the values prescribed for a
soft confinement. Accordingly, capillary condensation under ambient
conditions can be expected to qualitatively follow the macroscopic
Kelvin equation, as happened for the reported capillaries.
Online content
Any methods, additional references, Nature Research reporting sum-
maries, source data, extended data, supplementary information,
acknowledgements, peer review information; details of author con-
tributions and competing interests; and statements of data and code
availability are available at https://doi.org/10.1038/s41586-020-2978-1.
1. Charlaix, E. & Ciccotti, M. Capillary Condensation in Confined Media (CRC 2010).
2. van Honschoten, J. W., Brunets, N. & Tas, N. R. Capillarity at the nanoscale. Chem. Soc.
Rev. 39, 1096–1114 (2010).
3. Malijevský, A. & Jackson, G. A perspective on the interfacial properties of nanoscopic
liquid drops. J. Phys. Condens. Matter 24, 464121 (2012).
4. Barsotti, E., Tan, S. P., Saraji, S., Piri, M. & Chen, J.-H. A review on capillary condensation in
nanoporous media: implications for hydrocarbon recovery from tight reservoirs. Fuel 184,
344–361 (2016).
5. Thomson, W. On the equilibrium of vapour at a curved surface of liquid. Proc. R. Soc.
Edinb. 7, 63–68 (1872).
6. Aukett, P. N., Quirke, N., Riddiford, S. & Tennison, S. R. Methane adsorption on
microporous carbons—a comparison of experiment, theory, and simulation. Carbon 30,
913–924 (1992).
7. Fisher, L. R., Gamble, R. A. & Middlehurst, J. The Kelvin equation and the capillary
condensation of water. Nature 290, 575–576 (1981).
8. Kohonen, M. M. & Christenson, H. K. Capillary condensation of water between rinsed
mica surfaces. Langmuir 16, 7285–7288 (2000).
9. Mitropoulos, A. Ch. The Kelvin equation. J. Colloid Interface Sci. 317, 643–648 (2008).
10. Zhong, J. et al. Capillary condensation in 8 nm deep channels. J. Phys. Chem. Lett. 9,
497–503 (2018).
11. Yang, G., Chai, D., Fan, Z. & Li, X. Capillary condensation of single- and multicomponent
fluids in nanopores. Ind. Eng. Chem. Res. 58, 19302–19315 (2019).
12. Kim, S., Kim, D., Kim, J., An, S. & Jhe, W. Direct evidence for curvature-dependent surface
tension in capillary condensation: Kelvin equation at molecular scale. Phys. Rev. X 8,
041046 (2018).
13. Gruener, S., Hofmann, T., Wallacher, D., Kityk, A. V. & Huber, P. Capillary rise of water in
hydrophilic nanopores. Phys. Rev. E 79, 067301 (2009).
14. Vincent, O., Marguet, B. & Stroock, A. D. Imbibition triggered by capillary condensation in
nanopores. Langmuir 33, 1655–1661 (2017).
15. Shin, D., Hwang, J. & Jhe, W. Ice-VII-like molecular structure of ambient water
nanomeniscus. Nat. Commun. 10, 286 (2019).
16. Verdaguer, A., Sacha, G. M., Bluhm, H. & Salmeron, M. Molecular structure of water at
interfaces: wetting at the nanometer scale. Chem. Rev. 106, 1478–1510 (2006).
17. Matsuoka, H., Fukui, S. & Kato, T. Nanomeniscus forces in undersaturated
vapors: observable limit of macroscopic characteristics. Langmuir 18, 6796–6801 (2002).
18. Powles, J. G. On the validity of the Kelvin equation. J. Phys. Math. Gen. 18, 1551–1560 (1985).
19. Walton, J. P. R. B. & Quirke, N. Capillary condensation: a molecular simulation study. Mol.
Simul. 2, 361–391 (1989).
20. Cheng, S. & Robbins, M. O. Nanocapillary adhesion between parallel plates. Langmuir 32,
7788–7795 (2016).
21. Knežević, M. & Stark, H. Capillary condensation in an active bath. EPL 128, 40008 (2019).
22. Sing, K. S. W. & Williams, R. T. Historical aspects of capillarity and capillary condensation.
Microporous Mesoporous Mater. 154, 16–18 (2012).
23. Schoen, M. & Günther, G. Phase transitions in nanoconfined fluids: synergistic coupling
between soft and hard matter. Soft Matter 6, 5832–5838 (2010).
24. Gor, G. Y., Huber, P. & Bernstein, N. Adsorption-induced deformation of nanoporous
materials—a review. Appl. Phys. Rev. 4, 011303 (2017).
25. Altabet, Y. E., Haji-Akbari, A. & Debenedetti, P. G. Effect of material flexibility on the
thermodynamics and kinetics of hydrophobically induced evaporation of water. Proc.
Natl Acad. Sci. USA 114, E2548–E2555 (2017).
26. Radha, B. et al. Molecular transport through capillaries made with atomic scale precision.
Nature 538, 222–225 (2016).
27. Gopinadhan, K. et al. Complete steric exclusion of ions and proton transport through
confined monolayer water. Science 363, 145–148 (2019).
28. Drelich, J., Chibowski, E., Meng, D. D. & Terpilowski, K. Hydrophilic and superhydrophilic
surfaces and materials. Soft Matter 7, 9804–9828 (2011).
29. Mücksch, C., Rösch, C., Müller-Renno, C., Ziegler, C. & Urbassek, H. M. Consequences of
hydrocarbon contamination for wettability and protein adsorption on graphite surfaces.
J. Phys. Chem. C 119, 12496–12501 (2015).
30. Fumagalli, L. et al. Anomalously low dielectric constant of confined water. Science 360,
1339–1342 (2018).
31. Weeks, B. L. & Vaughn, M. W. Direct imaging of meniscus formation in atomic force
microscopy using environmental scanning electron microscopy. Langmuir 21,
8096–8098 (2005).
32. Malijevský, A. & Parry, A. O. Modified Kelvin equations for capillary condensation in
narrow and wide grooves. Phys. Rev. Lett. 120, 135701 (2018).
33. Christenson, H. K. & Thomson, N. H. The nature of the air-cleaved mica surface. Surf. Sci.
Rep. 71, 367–390 (2016).
34. Neek-Amal, M., Peeters, F. M., Grigorieva, I. V. & Geim, A. K. Commensurability effects in
viscosity of nanoconfined water. ACS Nano 10, 3685–3692 (2016).
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in
published maps and institutional affiliations.
© Crown 2020
Nature | Vol 588 | 10 December 2020 | 253
Article
Methods
Fabrication procedures
The capillary devices studied here were fabricated following the pro-
cedures described in refs. 26,27 and shown in the flow chart of Extended
Data Fig. 1a. First, a large crystal of multilayer graphene (number of
layers N) was prepared on an oxidized Si wafer by mechanical exfolia-
tion. Using electron-beam lithography and oxygen plasma etching, we
patterned the crystal into a set of parallel narrow strips that had a width
of ∼150 nm and were separated by approximately the same distance
(Extended Data Fig. 1b). These spacers were then put on top of a mica
or graphite crystal of typically 20 nm in thickness. The latter crystal
was prepared on a separate Si wafer and, in this work, is referred to as
the bottom crystal (Extended Data Fig. 1a, stage 1).
In parallel, we prepared a suspended silicon nitride (SiN) membrane
with a rectangular hole in the centre (Extended Data Fig. 1a, stage 2). To
this end, we used commercial Si wafers with 500 nm of SiN deposited
on both sides. Using photolithography and reactive ion etching (RIE),
we made a window of about 750 × 750 μm2 in size in one of the SiN lay-
ers. The wafer was then placed in hot KOH to etch through the entire
Si thickness and obtain a freestanding SiN membrane of ∼70 × 70 μm2
in size. After that, a rectangular hole (∼3 × 20 μm2) was plasma-etched
in the SiN membrane by means of another round of photolithography
and RIE (Extended Data Fig. 1a, stage 2).
The two-layer assembly consisting of the bottom crystal and
graphene spacers (Extended Data Fig. 1a, stage 1) was transferred on top
of the SiN membrane in such a way that graphene strips were aligned
perpendicular to the long edge of the rectangular hole. This step was
followed by RIE from the backside of the Si wafer to extend the hole
through the bottom crystal (Extended Data Fig. 1a, stage 3). Finally,
another mica (or graphite) crystal was placed on top of the two-layer
assembly to form 2D channels (Extended Data Fig. 1a stage 4, Extended
Data Fig. 1c).
After each crystal transfer, samples were cleaned in acetone, deion-
ized water and isopropanol. This was followed by annealing at 400 °C
in a hydrogen–argon atmosphere for 3 h. Such thorough cleaning was
essential to remove polymer residues and other possible contamination,
which could otherwise block the capillaries. In our experiments, we
used only those capillaries that exhibited uniform sagging along their
entire length, such as those shown in Fig. 1c and Extended Data Fig. 3.
The contact angle for muscovite mica and natural graphite used for
making our devices was measured by Drop Shape Analyzer 100S (Krüss).
We found θ ≈ 80–85° for graphite and 15–20° for mica after exposure
to ambient air for a few days, in agreement with previous reports (see,
for example, refs. 28,29).
AFM topography under controlled humidity
Our set-up for AFM measurements is shown in Extended Data Fig. 2a.
The SiN wafer containing a capillary device such as that shown in
Extended Data Fig. 1c was placed to seal an airtight metal chamber
with a volume of about 1 cm3. A continuous flow of nitrogen gas into this
chamber was provided through miniature inlets. The humidity was con-
trolled by mixing dry nitrogen with nitrogen bubbled through deion-
ized water, using a computer-controlled flowmeter. RH was measured
by a humidity sensor (Sensirion), which was mounted inside the bottom
chamber. The sensor was calibrated with three different saturated salt
solutions (lithium chloride, magnesium nitride and potassium chloride)
to ensure readings of RH with an accuracy of ±1%. A commercial silicone
rubber enclosure (Bruker) was attached to the AFM head (Extended
Data Fig. 2a). When it was lowered for taking AFM images, the enclosure
edges sealed the space above the devices being studied. Fresh silica
gel granules were usually placed inside the enclosure to provide low
humidity in the top chamber.
All AFM images in our experiments were taken in the PeakForce
mode using Dimension FastScan (Bruker) and analysed with WSxM
software35. Selected capillaries were imaged at regular RH intervals of
5%, after stabilizing humidity in the bottom chamber for approximately
an hour. As an example of our AFM measurements, Extended Data
Fig. 3 shows three critical steps in the evolution of the topography for
an N = 3 graphite capillary. The initial sagging in this case was ∼4 Å as
seen for RH = 55% in Extended Data Fig. 3a. With RH increasing in 5%
steps, no change in the sagging profile was observed until we reached
70% RH. At the latter humidity, the top crystal was found to sag notably
less (Extended Data Fig. 3b). On increasing RH further, the top crystal
gradually lifted and became practically flat at 95% RH (Extended Data
Fig. 3c). The flattening process is described in detail below. Because the
rapid change in sagging happened somewhere between 65% and 70%
RH, we assigned the condensation transition to RHC = 67.5 ± 2.5% with
an additional error of ±1% because of the humidity sensor’s accuracy
(as indicated by the symbol size in Fig. 2).
The use of low humidity in the top chamber notably improved the
stability of AFM imaging but was not essential. Indeed, if we simply con-
nected the top and bottom chambers so that both sides of the studied
capillaries were exposed to the same RH, the condensation transition
was found to occur at the same RHC as in the case of low RH in the top
chamber. This shows that the condensation transition is determined
by the RH at the entrance side (that is, the highest RH) (Extended Data
Fig. 2b). This observation is consistent with the fact that no difference
in sagging was observed along the entire length of the capillaries, even
when their exits were at low RH (Fig. 1d; Extended Data Fig. 3b), which
indicates no detectable gradient in negative capillary pressure along
the 2D channels. The constant capillary pressure can be attributed to
a very fast flow of liquid water through our atomically flat capillaries26,
which allows essentially the same meniscus curvature at both entrance
and exit sides, as shown schematically in Extended Data Fig. 2b. If the
top chamber is kept at low humidity, such an equilibrium state is stabi-
lized by a slightly retracted exit meniscus and slow Knudsen diffusion
of water vapour near the capillary exit26, which provide the required
RH gradient. This is different from the case of nanoporous media with
rough internal surfaces and tortuous capillaries, where both liquid and
vapour transport are slow, allowing large pressure gradients to build
up along the liquid flow direction14.
Non-hysteretic behaviour of the condensation transition
Capillary condensation in nanochannels is often accompanied by
hysteresis such that RHC required to reach the transition depends on
whether external RH is increased or decreased1,4,11,13,23,24,36. This was not
the case for our capillaries, which exhibited no hysteresis within our
experimental accuracy. This behaviour is illustrated in Extended Data
Fig. 4a, which shows the profiles of the top crystal for a four-layer graph-
ite capillary where RH was changed in a small loop around the transition
observed at RHC = 77.5 ± 2.5%. The capillary’s sagging was constant for
RH ≤ 75%. Then we increased RH to 80% and equilibrated for 1 h, fol-
lowing the experimental procedures described above. The 5% change
in RH led to a pronounced jump in the sagging depth δ, indicating the
condensation transition (compare black and red curves in Extended
Data Fig. 4a). The capillary profile remained stable while RH was main-
tained at 80%. When we decreased RH back to 75%, the capillary did
not return to the initial dry state after 4 h (blue curve). Nonetheless,
the top crystal continued to sag gradually with time (Extended Data
Fig. 4a). The dry state was eventually reached (after more than 9 h but
less than 16 h). Therefore, the condensation transition occurred at the
same RHC with either increasing or decreasing humidity, although long
equilibration times were needed for 2D channels to dry up.
The slow recovery of the initial dry state is further exemplified by
Extended Data Fig. 4b. It shows the case of a graphite capillary with N = 6,
where the condensation transition was found to occur between 60%
and 65% RH. In Extended Data Fig. 4b, we first increased RH from 60%
directly to 95%, well above the transition (black and red curves, respec-
tively). Then RH was reduced to ∼30%, well below RHC = 62.5 ± 2.5%.
As seen in Extended Data Fig. 4b, the top crystal regained its original
profile very slowly, and the capillary returned to its dry state only after
several days. After this the sagging remained stable. The reason for
such a slow drying process remains to be understood.
It is also worth mentioning that our capillary devices did not show
any discernible change in sagging with changing RH below the con-
densation transition, as seen, for example, in Fig. 1e. This is in contrast
to the usual elasto-capillary response of nanoporous media, in which
adsorption of water molecules on internal surfaces leads to notable
strain, usually referred to as the Bangham effect37. Its apparent absence
in our experiment is perhaps not surprising. First, as discussed in the
main text, we expect only a small partial coverage of internal walls by
adsorbed water molecules before the transition. Second, if there were
notable adsorption, the adsorption-induced strain is typically of the
order of 10−4 for materials with high Young’s moduli13,23,24,36. Therefore,
for our top crystals with H and w of ∼100 nm, this strain would translate
into sagging of the order of 0.1 Å, below our experimental accuracy.
Effect of initial sagging
For 2D channels with the same N, their heights h = Na − δ0 could vary
considerably because of different H and slightly different w, which
control the initial sagging δ0. This resulted in different RHC observed
for capillaries with the same N. This behaviour is illustrated in Extended
Data Fig. 5, which shows the condensation transition in two capillaries
with N = 5 but different δ0. The capillary in Extended Data Fig. 5a had
a top layer with H ≈ 70 nm and exhibited initial sagging of ∼4 Å. The
transition in this device occurred at RHC = 82.5%. The other capillary
(Extended Data Fig. 5b) had a thinner (∼45 nm) top crystal and, accord-
ingly, its δ0 was larger (∼8.5 Å). The latter device exhibited RHC = 72.5%,
considerably lower than that in Extended Data Fig. 5a. This implies that
N was not the characteristic determining the onset of water condensa-
tion. The importance of h rather than N is even better exemplified by the
results of Extended Data Fig. 4: the capillary with N = 4 in Extended Data
Fig. 4a exhibited the condensation transition at 77.5% RH, whereas the
nominally larger capillary with N = 6 in Extended Data Fig. 4b showed the
transition at lower RHC = 62.5%. This obviously contradicts the expecta-
tion that the smaller-N capillary should exhibit lower RHC. However,
because of different δ0, the smaller (N = 4) capillary in Extended Data
Fig. 4a had h ≈ 7 Å, which was larger than h ≈ 3.5 Å for the larger (N = 6)
capillary in Extended Data Fig. 4b. The smaller RHC for 2D channels
with smaller h agrees with the general expectations.
The above observations strongly suggest that h is the size parameter
that best describes the condensation transition in our atomic-scale cap-
illaries. This means that it is the central region between the sagged-top
and flat-bottom crystals where the condensation process is effectively
initiated. This is not entirely unexpected because MD simulations have
previously shown that corner menisci in narrow channels were unfa-
vourable for condensation32. Furthermore, note that the mean free
path of water molecules in air is about 65 nm, which is comparable
to the channel width w ≈ 150 nm. This implies that the entire chan-
nel should present a single entity from the standpoint of thermody-
namics, allowing only one condensation transition over the channel’s
cross-section. To this end, it is important to note that although our
capillaries contained only a few monolayers of water, there were still mil-
lions of molecules inside each capillary, which should be sufficient for
the thermodynamic description, unlike in the case of nanometre-scale
droplets containing a small number of molecules3,16,18.
Remnant sagging above the condensation transition
Nanometre-thick suspended crystals are known to exhibit considerable
sagging, which is believed to be caused by their vdW attraction to side-
walls26,30,38,39. As water spontaneously condensed inside the 2D channels,
the sagging was found to decrease suddenly (see the jump-like changes
in Fig. 1e and in Extended Data Figs. 4a and 5), which we attribute to
suppression of the vdW attraction by intercalating water molecules.
This explanation is supported by MD simulations27. They showed that
capillaries with monolayer spacers (N = 1) always collapsed (indepen-
dently of H) because of vdW interaction between the top and bottom
crystals. The collapsed capillaries could be opened by intercalating
water, because the attraction rapidly diminishes at distances more than
a few ångströms so that even a monolayer of water provided sufficient
‘screening’ of the vdW attraction26,27.
In all our measurements, the sagging depth δ0 became abruptly
smaller at the condensation transition but did not completely dis-
appear. Only if RH was increased further did the remnant sagging
gradually decrease, approaching zero at 100% RH, so that the top layer
became essentially flat (Fig. 1d, e; Extended Data Fig. 3). The remnant
sagging at the transition and its gradual changes with further increases
in RH can be explained by the negative pressure P caused by the con-
densed water meniscus1,40. Let us consider our typical mica capillary
with w ≈ 150 nm and top-layer thickness H ≈ 50–60 nm (Extended Data
Fig. 6a). After the condensation transition occurred at a certain RH
(which depended on channel height h), the top layer remained sagged
typically by several ångströms. At the condensation transition, the
capillary pressure is given by P ≈ 2σ cos θ/h. Using the contact angle
θ ≈ 20° for mica and the surface tension of bulk water, σ ≈ 73 mN m−1,
the Young–Laplace equation yields P ≈ 700 bar for h ≈ 2 nm.
The negative pressure P forces the top crystal to bend downwards,
resulting in its sagging given by41
5Pw 4
δ= , (3)
32EH 3
where E ≈ 60 GPa is Young’s modulus of mica in the out-of-plane direc-
tion42. Equation (3) yields δ ≈ 4−7 Å, in agreement with our observa-
tions. This substantiates our model that the remnant sagging above the
condensation transition is caused by the negative capillary pressure.
Similar agreement is found for graphite capillaries, although there is
a larger uncertainty in the estimates (a factor of 2) because P strongly
depends on θ for the large contact angles exhibited by water on graph-
ite. As RH is increased beyond the condensation point, the meniscus
extends outside capillaries, and its curvature becomes progressively
smaller to match the external RH. Accordingly, the negative capillary
pressure above RHC evolves with RH and is given by the other Kelvin
equation as1,23,24,36,40
P = kBTρNln RH. (4)
According to this equation, the pressure that bends the top crystal
should decrease logarithmically with RH, in good agreement with our
observations (Extended Data Fig. 6b). Note that, close to 100% RH,
δ ∝ P ∝ ln(RH) ≈ (RH – 1) is expected to approach zero linearly, as indeed
observed in Fig. 1e and Extended Data Fig. 6b.
The condition of partially sagged but open capillaries (that is, few
ångströms < δ0 < Na, as in our devices) is rather difficult to satisfy experi-
mentally. Indeed, if we were to decrease H or increase w by only a factor
of 2 with respect to the optimal design found, δ in equation (3) would
increase by an order of magnitude because of the high powers. On the
other hand, the capillary pressure P in equation (4) depends on RH
only logarithmically, which means that even at very low humidity (for
example, 5%), it would be thermodynamically favourable for the top
layer with the non-optimal w or H to bend all the way down and reach
the channel’s bottom. Therefore, such non-optimized 2D channels
are unstable with respect to spontaneous water condensation under
low-humidity conditions. If we were to do the opposite and increase
the top crystal’s stiffness (by halving w or doubling H), δ in equation (3)
becomes so small (<1 Å) that changes in the sagging would be impos-
sible to detect by AFM. The above consideration shows that there is a
subtle interplay between materials parameters and the design of 2D
channels, and stringent rules should be followed in order to detect the
Article
condensation transition in experiment. Following this insight, we usu-
ally increased H by ∼50% for our smallest 2D channels with N = 2 and 3,
which ensured that they remained open. Also, when making graphite
capillaries, we used top crystals slightly (∼20%) thicker than in the case
of mica capillaries with the same N because mica has a higher Young’s
modulus than graphite43.
For nanoscale 2D capillaries such as cracks or slits inside bulk materi-
als (H → ∞), their elastic deformations caused by large capillary pres-
sures can notably shift the condensation transition with respect to that
expected for the rigid confinement20,21. To estimate the magnitude
of such adjustments, let us consider the deformation of a half-space
elastic medium subject to the uniform load p over a suspended strip
with the width w = 2a in the range of −a ≤ x ≤ a. The vertical deforma-
tion is given by44
(1 − ν)p  x x 
uz ( x ) = − a  + 1ln|x + a| −  − 1ln|x − a| − 2,
πG  a  a  
where ν is Poisson’s ratio and G is the shear modulus. This equation
yields the sagging
ln(2)(1 − ν)pw
δ = uz (0) − uz ( ± a) = . (6)
πG
If we take as an example the elastic properties of graphite with
G ≈ 10 GPa and ν ≈ 0.343, equation (6) yields δ ≈ 2.3 Å for capillary pres-
sures of about 1,000 bar. Such P are typical for cavities of 1–2 nm in
height (see above). This indicates that elastic deformations can not
only be a contributing factor during the condensation transition23,24,36
but also allow atomic-scale cavities in bulk materials to adjust their size
so that an integer number of water layers can fit inside, similar to our
case where the top crystal was intentionally made sufficiently flexible.
Molecular dynamics simulations of water–surface interaction
under strong confinement
To investigate the dependence of the solid–liquid surface energy γSL on
h, MD simulations were performed using LAMMPS simulation code45
and the SPC/E model for water molecules46. The interaction between
water and confining walls was modelled by the Lennard–Jones potential
with parameters taken from ref. 47. Flat rigid graphene sheets were used
to mimic the confining walls. For simplicity, to account for surfaces with
different θ, we varied the interaction energies of carbon with hydrogen,
εHC, and oxygen, εOC. These energies47 were multiplied by a factor of k
that was varied from 0.7 to 1.3 in steps of 0.2 to find the water–wall
interaction that would approximate the experimental contact angles.
The MD angles θ were estimated by using water droplets containing
4,000 molecules. Our simulations yielded θ ≈ 85°, 63°, 30° and 11° for
k = 0.7, 0.9, 1.1 and 1.3, respectively. The insets of Extended Data Fig. 7
show the profiles for the water droplets found in the case of θ ≈ 11°
and 85°. We used these two θ and the corresponding k to model γSL(h)
for our mica and graphite capillaries, respectively. Note that the for-
mer value lies in the middle of the contact-angle interval observed for
mica28 and, importantly, our MD results exhibited little sensitivity to
the exact θ for strongly hydrophilic capillaries, as expected from the
cos θ dependence.
Having established parameters for the desired contact angles, we
proceeded to another simulation set-up that consisted of two flat
four-layer graphite sheets immersed in a water box containing 40,000
molecules. The dimension of each graphene sheet was 102.2 × 100.9 Å2
whereas the water box was 140.0 × 140.0 Å2 in size, which allowed water
molecules confined between the rigid graphite sheets to exchange eas-
ily with outside molecules. After an equilibration run of 1.0 ns, the two
sheets were brought progressively closer in steps of 0.2 Å. Each time
the system was equilibrated for 0.1 ns and its total potential energy
was calculated for further analysis. Periodic boundary conditions were
imposed in all three directions. All the simulations were carried out with
the isothermal–isobaric ensemble at 298 K. The density profiles found
in our simulations are shown in Extended Data Fig. 7. The confined
water exhibits a pronounced layered structure that extends over two
intermolecular distances from each surface, before the water density
converges to its bulk value, in agreement with the earlier literature
(see, for example, refs. 19,20,48,49).
The deviations Δγ in the solid–liquid surface energy γSL from its bulk
value may be considered as extra work spent to rearrange water mol-
ecules into the strongly layered structures shown in Extended Data Fig. 7.
If h is sufficiently large, the extra work is negligible because the opposite
surfaces do not ‘feel’ each other, and their near-surface water structures
remain unchanged with respect to the case of infinite h. However, as the
walls are getting closer, the layered structures overlap (see the density
profiles for h ≱ 10 Å in Extended Data Fig. 7). As a result, the total energy
(5) and, hence, Δγ exhibit pronounced oscillations (Extended Data Fig. 8).
Using equation (2) and the numerically found Δγ, it is straightforward
to calculate the RH required for water condensation inside atomic-scale
capillaries. The results are plotted in Fig. 2 of the main text and reveal
giant oscillations in RHC which emerge when the structured layers of
water near the two confining surfaces start overlapping. Note that the
confining walls in the MD simulations were made rigid, disallowing
elastic deformations considered separately in our analysis in Fig. 2.
Data availability
All the mentioned data to support this study and its conclusions are
available upon request from Q.Y.
35. Horcas, I. et al. WSXM: a software for scanning probe microscopy and a tool for
nanotechnology. Rev. Sci. Instrum. 78, 013705 (2007).
36. Gor, G. Y. et al. Elastic response of mesoporous silicon to capillary pressures in the pores.
Appl. Phys. Lett. 106, 261901 (2015).
37. Bangham, D. H. & Fakhoury, N. The expansion of charcoal accompanying sorption of
gases and vapours. Nature 122, 681–682 (1928).
38. Li, T. & Zhang, Z. Substrate-regulated morphology of graphene. J. Phys. D 43, 075303 (2010).
39. Scharfenberg, S., Mansukhani, N., Chialvo, C., Weaver, R. L. & Mason, N. Observation of a
snap-through instability in graphene. Appl. Phys. Lett. 100, 021910 (2012).
40. Israelachvili, J. N. Intermolecular and Surface Forces (Academic, 2011).
41. Hibbeler, R. C. Mechanics of Materials 814 (Pearson, 2015).
42. McNeil, L. E. & Grimsditch, M. Elastic moduli of muscovite mica. J. Phys. Condens. Matter
5, 1681–1690 (1993).
43. Cost, J. R., Janowski, K. R. & Rossi, R. C. Elastic properties of isotropic graphite. Phil. Mag.
17, 851–854 (1968).
44. Ling, F. F., Lai, W. M. & Lucca, D. A. Fundamentals of Surface Mechanics: With Applications
96–97 (Springer, 2012).
45. Plimpton, S. Fast parallel algorithms for short-range molecular dynamics. J. Comput.
Phys. 117, 1–19 (1995).
46. Berendsen, H. J. C., Grigera, J. R. & Straatsma, T. P. The missing term in effective pair
potentials. J. Phys. Chem. 91, 6269–6271 (1987).
47. Wu, Y. & Aluru, N. R. Graphitic carbon–water nonbonded interaction parameters. J. Phys.
Chem. B 117, 8802–8813 (2013).
48. Cicero, G., Grossman, J. C., Schwegler, E., Gygi, F. & Galli, G. Water confined in nanotubes
and between graphene sheets: a first principle study. J. Am. Chem. Soc. 130, 1871–1878
(2008).
49. Sendner, C., Horinek, D., Bocquet, L. & Netz, R. R. Interfacial water at hydrophobic and
hydrophilic surfaces: slip, viscosity, and diffusion. Langmuir 25, 10768–10781 (2009).
Acknowledgements This work was funded by Lloyd’s Register Foundation, the European
Research Council, Graphene Flagship and the Royal Society. Q.Y. acknowledges support from
the Leverhulme Early Career Fellowship, and F.C.W. from the Strategic Priority Research
Program of the Chinese Academy of Sciences (XDB22040402) and the CAS Youth Innovation
Promotion Association.
Author contributions A.K.G. suggested the project and directed it together with Q.Y. Q.Y. and
P.Z.S. fabricated devices. Q.Y. performed measurements and carried out data analysis with
help from L.F., Y.V.S., S.J.H. and Z.W.Z. F.C.W. provided theoretical support. A.K.G., Q.Y., F.C.W.
and I.V.G. wrote the manuscript. All authors contributed to discussions.
Competing interests The authors declare no competing interests.
Additional information
Correspondence and requests for materials should be addressed to Q.Y., F.C.W. or A.K.G.
Peer review information Nature thanks Patrick Huber and the other, anonymous, reviewer(s)
for their contribution to the peer review of this work.
Reprints and permissions information is available at http://www.nature.com/reprints.
Extended Data Fig. 1 | Nanofabrication of 2D channels. a, Simplified flow
chart for our fabrication procedures. (1) Graphene spacers and the bottom
crystal of either mica or graphite (shown in yellow) were assembled on top of an
oxidized Si wafer. (2) A suspended SiN membrane with a rectangular hole was
prepared separately. (3) The two-layer assembly was transferred from the Si
oxide wafer onto the SiN membrane. The opening was extended through the
assembly by RIE. (4) The top crystal of mica or graphite was placed on top of
graphene spacers. b, AFM micrograph of graphene spacers with N = 5. The
colour scale is given by the height profile (blue curve). c, Optical image of a final
mica device used in our experiments. The bottom mica crystal shows up in
purple on top of the square SiN membrane. Graphene spacers (N = 3) and the
top mica layer are outlined in blue and yellow, respectively. d, Cross-sectional
scanning transmission electron microscopy image of a graphite channel with
N = 2. The blue ticks mark the channel’s edges.
Article

Extended Data Fig. 2 | Measurements of capillary condensation. a, Our AFM

set-up. Humidified nitrogen gas flows through the bottom chamber made from
an aluminium alloy. A silicon wafer of 15 × 15 mm2 in size is seen to cover the
chamber, flush with its top surface. The white rubber gasket was lowered
during AFM measurements to seal the space above the Si wafer. Inset,
cross-sectional schematic showing how capillary devices were mounted during
AFM measurements. b, Schematic of a water plug inside our capillaries. For
brevity, the layered structure of water is ignored in this sketch. When the top
chamber is at low RH, the meniscus slightly retracts inside the capillary to
create a vapour pressure gradient. The RH gradient stabilizes two menisci with
the same curvature at both exit and entrance. The distance from the exit
meniscus to the opening is expected to be short because, in our atomically flat
capillaries, water moves much faster as liquid than vapour26.
Extended Data Fig. 3 | Visualization of the condensation transition using
AFM. a–c, Images of a graphite capillary with N = 3 at RH of 55%, 70% and 95%
(a, b and c, respectively). The upper part of each image shows sagging of the
top graphite crystal (H ≈ 80 nm) into the 2D channel. The lower part shows the
area immediately outside the channel, which is not covered by the top graphite.
The black dotted lines mark a border between the two regions (edge of the top
crystal). The colour scales for the lower and upper parts of the AFM images are
given by the green and black curves, respectively. The profiles are averaged
over ∼100 nm along the y direction, and the curves in the upper parts of all the
panels are provided on the same scale given by the black arrows in panel a. A
small number of horizontal scanning lines (x direction) around the black-dot
dividing lines were removed for clarity because they contained numerous
instabilities caused by the AFM tip moving along the edge of the top crystal and
jumping up and down. Such instabilities are typical for AFM scanning close to
edges.
Article
Extended Data Fig. 4 | Non-hysteretic capillary condensation with slow
dynamics. a, Sagging profiles for a graphite capillary (N = 4) with increasing
and decreasing RH between 75% and 80%. Black curve, initial dry-state profile.
Red curve, RH was increased to 80%. Then, RH was returned to 75% and
maintained at this humidity. AFM profiles were taken after 4 h, 9 h and 16 h
(colour coded). b, The N = 6 graphite capillary was brought from the dry state
(black curve) into the state filled with water and kept for an hour at 95% RH
(red). The humidity was then decreased to ∼30%, well below the condensation
transition observed at 62.5 ± 2.5% for this device. The colour-coded curves
show the time evolution towards the original dry state. Note that the sagging
depths δ for such hysteretic loops were highly reproducible but details of
sagging profiles could differ in different RH cycles. For example, the top
crystal’s adhesion to the right wall was different in the original and final dry
states, as seen in a (compare black and purple curves). This hysteresis is
attributed to irreproducible vdW attachments of top crystals to channel
sidewalls.
Extended Data Fig. 5 | Capillary condensation in 2D channels with different condensation transition occurred between 80% and 85% RH in a and between
initial sagging. a, b, Sagging profiles for two N = 5 graphite capillaries with 70% and 75% in b. The difference in RHC for the same N is attributed to different
different δ0. RH was increased in 5% steps (colour coded). The water h in the two cases.
Article

Extended Data Fig. 6 | Remnant sagging above the condensation

transition. a, Schematic of top crystal sagging. b, Typical behaviour observed
for the sagging depth δ as a function of RH, after the condensation transition
occurred at RH < 60%. Symbols: Measurements for two different mica
capillaries with N = 8. The solid curves are best fits using equations (3) and (4)
(colour-coded). The grey symbol with error bars indicates the experimental
accuracy.
Extended Data Fig. 7 | MD simulations of strongly confined water. a, Its channels. Water exhibits a pronounced layered structure near each surface,
density profiles at different distances h between two rigid capillary walls with and the structures start to overlap for h < 15 Å. Top insets, cross-sectional
the contact angle θ ≈ 11°. b, Same calculations but for contact angle 85°. The profiles for water droplets placed on the surfaces with the given θ.
orange dashed lines mark positions of the surfaces that defined the 2D
Article

Extended Data Fig. 8 | Changes in the solid–liquid surface energy caused by

atomic-scale confinement. Calculated Δγ(h) for several characteristic θ. The
arrows indicate the number of molecular layers of water that fit inside the 2D
channels.
Article

Catalytic asymmetric addition of an amine

N–H bond across internal alkenes

https://doi.org/10.1038/s41586-020-2919-z Yumeng Xi1,2, Senjie Ma1,2 & John F. Hartwig1,2 ✉

Received: 26 November 2019

Accepted: 27 October 2020 Hydroamination of alkenes, the addition of the N–H bond of an amine across an
Published online: 3 November 2020 alkene, is a fundamental, yet challenging, organic transformation that creates an
alkylamine from two abundant chemical feedstocks, alkenes and amines, with full
Check for updates
atom economy1–3. The reaction is particularly important because amines, especially
chiral amines, are prevalent substructures in a wide range of natural products and
drugs. Although extensive efforts have been dedicated to developing catalysts for
hydroamination, the vast majority of alkenes that undergo intermolecular
hydroamination have been limited to conjugated, strained, or terminal alkenes2–4;
only a few examples occur by the direct addition of the N–H bond of amines across
unactivated internal alkenes5–7, including photocatalytic hydroamination8,9, and no
asymmetric intermolecular additions to such alkenes are known. In fact, current
examples of direct, enantioselective intermolecular hydroamination of any type of
unactivated alkene lacking a directing group occur with only moderate
enantioselectivity10–13. Here we report a cationic iridium system that catalyses
intermolecular hydroamination of a range of unactivated, internal alkenes,
including those in both acyclic and cyclic alkenes, to afford chiral amines with
high enantioselectivity. The catalyst contains a phosphine ligand bearing
trimethylsilyl-substituted aryl groups and a triflimide counteranion, and the reaction
design includes 2-amino-6-methylpyridine as the amine to enhance the rates of
multiple steps within the catalytic cycle while serving as an ammonia surrogate.
These design principles point the way to the addition of N–H bonds of other
reagents, as well as O–H and C–H bonds, across unactivated internal alkenes to
streamline the synthesis of functional molecules from basic feedstocks.

Chiral amines are essential structural motifs in numerous active phar-
maceutical ingredients and in many agrochemicals and materials. They
also serve as chiral catalysts, resolving reagents and chiral auxiliaries14.
Thus, efficient methods to prepare chiral amines have been long sought.
Traditional approaches15 include chemical16,17 and enzymatic18 reduc-
tive amination, hydrogenation19, nucleophilic addition to imines20 and
nucleophilic substitution21,22. However, these methods require starting
materials containing reactive functionality that is often derived from
feedstock alkenes. Thus, hydroamination of alkenes is the most direct
method to construct chiral amines from a functional group that is present
in basic feedstocks and is typically unprotected. Asymmetric additions
to conjugated alkenes, such as dienes23–26 and vinylarenes27,28, have been
reported, but the scope of the direct addition of N–H bonds to more
common and less reactive unconjugated alkenes is severely limited, and
the enantioselectivities of asymmetric processes are far lower than those
that would enable applications for the synthesis of chiral amines. Formal
hydroaminations provide an alternative approach to this problem, but
the use of silanes with electrophilic aminating reagents29 and even metal
reductants30, instead of amines, undermines the atom economy of the
hydroamination reaction. Direct N–H additions of unactivated internal
alkenes that occur with high enantioselectivity are unknown.
There are considerable challenges facing the development of cata-
lytic hydroaminations of unactivated alkenes that bear more than one
substituent (Fig. 1a). Both experiments and theoretical studies have
shown that the thermodynamic driving force is weak and the kinetic
barrier to combining two nucleophiles is high1,31. Moreover, catalysts
for hydroamination often catalyse undesirable, competing alkene
isomerization, and isomerization is typically faster than addition of
the N–H bond during many metal-catalysed hydroaminations (Fig. 1b).
Such relative rates lead to a mixture of isomeric products32,33. Many cata-
lysts for alkene hydroamination also promote formation of oxidative
amination products by β-hydrogen elimination of β-aminoalkylmetal
intermediates34,35. These isomerization and oxidative side reactions
must be suppressed to achieve hydroamination of unactivated internal
alkenes. Finally, because hydroamination is usually almost ergoneutral,
isomerization and racemization of the products during the reaction
can erode regioselectivity and enantioselectivity36.
To address these challenges, we modified a neutral iridium complex
containing a bisphosphine ligand, which was previously shown to cata-
lyse the formation of hydroamination and oxidative amination products
from the reaction of terminal alkenes with amides and indoles34,37–40.
We have previously shown that these aminations occur by a mechanism

Department of Chemistry, University of California, Berkeley, CA, USA. 2Division of Chemical Sciences, Lawrence Berkeley National Laboratory, Berkeley, CA, USA. ✉e-mail: jhartwig@berkeley.edu
1

254 | Nature | Vol 588 | 10 December 2020

 a NH2 Challenges
R1 H-NH2 NH2
or R2 or • Weak thermodynamic driving force • Alkene isomerization and chain walking
R2 R1 • Low reactivity of internal alkenes • Competing oxidative amination
cat.

b Me Me NHR
Me Selectivity?
Me
or cat. Me cat., RNH2 Me
NHR Me
Me
Me (Reacts much faster)
How to suppress Me Me Me NHR
isomerization? Me
Me Mixtures Mixtures

c Removable SiMe 3
Enantioselective O
regioselective NH2 SiMe 3
PyR removal
R + Ir+ 2
R N NH R′
R N NH2 R
O P
R′ R′ Ir NTf2–
Ammonia O P
R Primary amines
surrogate
O SiMe3
Design elements: cationic iridium, bidendate amine with adjacent substituent 2
SiMe3
Dissociates R′
readily R R′ Migratory
R H R′ R
LIr + H H insertion C–H RE
N R P H
P P
Ir Ir
R N NH2 N–H OA * P Ir NH NH – PyRNH2 *P NH
*P NH R N NH
2 N
N N R R′
R R R
e–-deficient Ir e–-deficient Ir Rigid iridacyle
Weakens Extra N-coordination accelerates accelerates prevents
coordination facilitates N–H OA I alkene insertion II reductive elimination III β-H elimination

Fig. 1 | Catalytic asymmetric hydroamination of unactivated internal iridium catalyst and an ammonia surrogate based on 2-aminopyridine to
alkenes. a, Long-standing challenges and previous strategies for catalytic achieve asymmetric hydroamination of internal alkenes. OA, oxidative
hydroamination of unactivated internal alkenes. b, Alkene isomerization that addition. RE, reductive elimination.
leads to a mixture of constitutional isomeric products. c, Design of a cationic

comprising oxidative addition of N–H bonds, migratory insertion of
alkenes and reductive elimination of C–H bonds34. We envisioned that
switching the iridium catalyst from neutral to cationic would lead to
the formation of cationic iridium intermediates that would undergo
migratory insertion of the alkene more rapidly41, a step that has been
shown to be rate-limiting in the reaction of terminal alkenes with the
neutral catalyst34 (Fig. 1c). If binding and insertion of the alkene were
sufficiently enhanced, then the reaction scope might include internal
alkenes. In addition, we envisioned that coordinating groups adja-
cent to the amine (for example, 2-aminopyridine derivatives) could
facilitate and make more thermodynamically favourable the initial
N–H oxidative addition, and such enhancements of this step could be
important because the rate of oxidative addition of electron-deficient
metal complexes is generally slower than that of electron-rich ones42.
If properly placed, the coordinating groups of the amine could also
form a rigid six-membered iridacycle upon insertion of the alkene; this
geometry could suppress β-hydrogen elimination to form enamines.
Such a combination of 2-aminopyridine and a cationic iridium cata-
lyst was evaluated12 for the hydroamination of terminal vinylarenes,
but only briefly for the hydroamination of unactivated α-alkenes. Low
enantioselectivities (≤11% enantiomeric excess, e.e.) were observed,
and the reactions of unstrained internal alkenes were not examined.
We reasoned that a substituent near the binding group of the amine
could modulate the strength of the coordination of the pyridine, lead-
ing to potential enhancement of the activity of the iridium catalyst,
due to weakened coordination (Fig. 1c). Finally, the pyridyl group of
the product could be cleaved by known methods to reveal the cor-
responding primary amine43.
Figure 2 summarizes our studies on the development of a cati-
onic iridium catalyst and an ammonia surrogate for the asymmetric
hydroamination of unactivated internal alkenes. These experiments
revealed the effect of the reaction components of the system on the
reaction yield, level of alkene isomerization and enantioselectivity.
To identify a suitable ammonia surrogate for this reaction, we surveyed
a series of heteroaromatic amines and one control arylamine that
possess varying structural properties. The hydroamination products
from these reactions consisted of three isomers (denoted A, B and
C) that probably resulted from competing alkene isomerization and
hydroamination. A larger amount of A reflects a faster rate for direct
addition of the amine to the starting alkene than alkene isomerization
before addition.
As shown in Fig. 2a, hydroamination of cis-4-octene in the presence of
[Ir(coe)2Cl]2, (S)-DTBM-SEGPHOS ((S)-(+)-5,5′-bis[di(3,5-di-tert-butyl-
4-methoxyphenyl)phosphino]-4,4′-bi-1,3-benzodioxole) and NaBARF
(sodium tetrakis[3,5-bis(trifluoromethyl)phenyl]borate) occurred
only when 6-substituted 2-aminopyridine derivatives were used as
the amine source. The hydroamination of 2-amino-6-methylpyridine
formed amines A, B and C in a combined 73% yield, with A constitut-
ing 28% of the amine products (defined as A/(A + B + C)). The reaction
of 2-amino-6-trifluoromethylpyridine afforded only 5% of isomer A.
Other amines tested, including the parent 2-aminopyridine, did not
undergo this hydroamination. These results suggest that the substitu-
ent in the 6-position of the pyridine ring is essential to promote the
hydroamination.
To suppress competing alkene isomerization and improve the
reaction yield, we examined catalysts containing a series of bis-
phosphine ligands and counteranions for hydroamination with
2-amino-6-methylpyridine as the amine. Studies of the electronic
and steric properties of ligands indicated that the ligands that are
less electron-rich (Fig. 2b, entry 2) and more sterically encumbered
(Fig. 2b, entries 3–5) than (S)-DTBM-SEGPHOS form catalysts that
are more reactive and selective for direct addition to form amine
1. The observed higher reactivity of catalysts generated from more
electron-poor ligands probably results from a greater rate of migratory
insertion of the alkene (see below) into the Ir–N bond of complexes
containing L2–L5 lacking the methoxy group than into the Ir–N bond
of the complex ligated by (S)-DTBM-SEGPHOS possessing the p-OMe
group44. The higher selectivity from the more hindered ligands could

Nature | Vol 588 | 10 December 2020 | 255

Article
a 2.5 mol% [Ir(coe)2Cl]2
NHAr NHAr NHAr
6 mol%
Me (S)-DTBM-SEGPHOS Me + Me + Me
ArNH2 + Me Me Me
Me 6 mol% NaBARF, 120 °C
(10 equiv.) 4-amine (A) 3-amine (B) 2-amine (C)

Me

N NH2 Me N NH2 F 3C N NH2 N NH2 N

Me NH2
73%a (28%) c 51%a (5%) c NH2
0%a,b 0%a,b 0%a,b 0%a,b

b NHPyMe Ligands used in this study

Me tBu OMe R
Me NHPyMe
4-amine (1) Me O tBu O R
Me 5 mol% L*Ir +X – Me 2 2
+
3-amine (1′) O P O
100 °C or P
Me N NH2 NHPyMe
(10 equiv.) 120 °C O P P
O
Me Me
Me
O tBu O R
PyMe 2-amine (1″) 2 2

tBu OMe R
Yield rac-DTB-SEGPHOS
Entry L* X 4-Selectivity d (1+1′+1″) (S)-DTBM-SEGPHOS (R = tBu) (L2)
(L1) rac-TMS-SEGPHOS
1e (S)-DTBM-SEGPHOS (L1) BARF 48% 16% (R = SiMe 3 ) (L3)
2e rac-DTB-SEGPHOS (L2) BARF 50% 42% SiMe3 SiMe3
100 °C

3e rac-TMS-SEGPHOS (L3) BARF 83% 34%

O SiMe3 SiMe3
4e (R)-TMS-SYNPHOS (L4) BARF 83% 26% 2 2

5e (R)-TMS-MeOBIPHEP (L5) BARF 83% 22% O P MeO P

6e (S)-DTBM-SEGPHOS (L1) BARF 28% O P MeO P
73%
7f (S)-DTBM-SEGPHOS (L1) OTf 83% 9% SiMe3 SiMe3
O 2 2
8f (S)-DTBM-SEGPHOS (L1) BF4 67% 9%
120 °C

SiMe3 SiMe3
9f (S)-DTBM-SEGPHOS (L1) NTf2 63% 60% (–95% e.e.)
10 (S)-DTBM-SEGPHOS (L1) Cl n/a 0% (R)-TMS-SYNPHOS (L4) (R)-TMS-MeOBIPHEP
(new ligand) (L5)
11f,g (R)-TMS-SYNPHOS (L4) NTf2 89% 78% (97% e.e.)

Fig. 2 | Development of asymmetric hydroamination of unactivated
internal alkenes with 2-amino-6-methylpyridine as an ammonia
surrogate. a, Identification of suitable ammonia surrogates to enable
hydroamination of unactivated internal alkenes. b, Identification of reaction
conditions to achieve asymmetric hydroamination and to suppress alkene
isomerization. aCombined yield. bNo reaction. cDefined as A/(A + B + C).
d
4-Selectivity defined as 1/(1 + 1′ + 1″). eConditions: 2.5 mol% [Ir(coe)2Cl]2,
6 mol% ligand, 6 mol% NaBARF, toluene. fConditions: 5 mol% [L*Ir(COD)]X,
toluene. gIn 2-MeTHF.

result from a greater sensitivity of the rate of insertion of the alkene
into the metal–amido bond (Fig. 1c, structure II) to steric effects than
the rate of insertion into the metal–hydride bond, which probably
leads to alkene isomerization45,46. Studies with various counterions
revealed that reactions with triflimide as the counterion of the cata-
lyst occurred with the highest selectivity at high conversion (Fig. 2b,
entries 6–9). The origin of the effects of the counterions is difficult to
ascertain, but an effect is clear. Control experiments showed that the
reaction catalysed by a neutral iridium complex under otherwise iden-
tical conditions (Fig. 2b, entry 10) resulted in the exclusive formation
of oxidative amination products. By combining the chiral ligand that
led to the highest selectivity with the triflimide anion in the form of
[((R)-TMS-SYNPHOS)Ir(COD)]NTf2, the model hydroamination reaction
formed the 4-aminooctane (1) in high yield; this reaction also occurred
with remarkably high enantioselectivity (Fig. 2b, entry 11). Thus, the
substituent on the amine, the new phosphine ligand and the use of a
triflimide counterion all led to the high activity, chemoselectivity and
enantioselectivity of the reaction.
With this catalyst and reagent, we examined the scope of alkenes that
underwent hydroamination (Fig. 3). Both symmetrical internal alkenes
and unsymmetrical internal alkenes underwent hydroamination with
2-amino-6-methylpyridine. The reactions all proceeded in greater
than 90% e.e. Hydroamination of symmetrical alkenes afforded prod-
ucts containing linear alkyl groups (1), aryl-substituted alkyl groups
(2), branched alkyl groups (3), silyloxy groups (4), alkoxy groups (5)
and alkoxycarbonyl groups (6) in good to high yields. Reactions of
unsymmetrical internal alkenes bearing polar functional groups at the
homoallylic position occurred with synthetically useful regioselectiv-
ity (2:1 to 10:1). These functional groups include phthalimidyl groups
(7, 8), sulfonamido groups (9), silyloxy groups (10), bis(ethoxycarbonyl)
methyl groups (11) and (hetero)aryloxy groups (12–14). These groups
presumably influence the regioselectivity by inductive effects that
are similar to those observed for other classes of functionalization of
unsymmetrical internal alkenes47–49.
The reactivity of the system for Z-alkenes enabled the hydroamina-
tion of cyclic alkenes, and these reactions also occurred with high enan-
tioselectivity. The combination of [Ir(coe)2Cl]2, (S)-DTBM-SEGPHOS
and NaBARF was used as the catalyst because it was more reactive than
[((R)-TMS-SYNPHOS)Ir(COD)]NTf2 for the hydroamination of cyclic
alkenes. The cyclic hydrocarbons cyclopentene, cyclohexene, cyclo-
heptene and cyclooctene all underwent hydroamination in high yields
(15–18). The hydroamination of a series of substituted cyclopentenes
formed chiral amine products with high enantioselectivity (19–23,
90–92% e.e.). The reaction to form the 1,3-substituted cyclopentane 23
from the 4-methoxycarbonyl-substituted cyclopentene occurred with
high diastereoselectivity for the trans product. In addition, cyclohex-
ene derivatives with 3,3-substituents underwent hydroamination to
afford two products with regioselectivity of approximately 1:3 (24/24′,
25/25′). Although the major isomer is achiral, the chiral minor isomer
was formed with good to high enantioselectivity. We observed in some
cases that the hydroaminations of substituted cycloalkenes with rings
larger than that of cyclopentene occurred to high conversion, but were

256 | Nature | Vol 588 | 10 December 2020

 5 mol% = Py Me
R [(R)-TMS-SYNPHOSIr(COD)]NTf2 HN N Me
+
Me N NH2 R′ R
120 °C, 2-MeTHF R′

a NHPyMe NHPyMe Me NHPyMe NHPyMe

Me Ph Me TBSO
Me Ph Me OTBS
1a 2a 3 Me 4
52%, 97% e.e. 82%, 92% e.e. 76%, 98% e.e. 77%, 97% e.e.
O
NHPyMe Me Me NHPyMe O NHPyMe
MeO MeO N
OMe OMe OTBS
O Me Me O 7
5b 6 63%, 95% e.e.
55%, 97% e.e. 85%, 98% e.e. regioselectivity 2.9:1

O NHPyMe NHPyMe Me NHPyMe

NHPyMe Tf
N Me TBSO Me EtO 2C Me
N Me
EtO 2C
O 8 9 10 11
64%, 94% e.e. 62%, >99% e.e. 37%, 97% e.e. 62%, 90% e.e.
regioselectivity 2.9:1 regioselectivity 10:1 regioselectivity 2.5:1 regioselectivity 2.2:1

NHPyMe
NHPyMe NHPyMe
O Me
F 3C O Me F 3C N O Me
H H
12 13c 14
56%, >99% e.e. 60% (13+13′), 99% e.e. H 60%, >20:1 d.r.
regioselectivity 8:1 regioselectivity 2.8:1
regioselectivity 4:1
CF 3 O
O

bd
NHPyMe NHPyMe NHPyMe NHPyMe NHPyMe O NHPyMe
MeO2C Me

MeO2C Me
O

15 16 17 18 19 20
84% 81% 90% 82% 60%, 92% e.e. 58%, 90% e.e.

trans
NHPyMe R
TBSO AcO NHPyMe NHPyMe NHPyMe NHPyMe
R
MeO2C + R
TBSO AcO
R
R = CO 2Et 24, 24%, 92% e.e. 24′, 48%
21 22 23
64%, 90% e.e. 53%, 90% e.e. 55%, 91% e.e. R = Me 25, 69% e.e. 25′
10:1 d.r. 52% (25+25′), regioselectivity 1:3.5

ce,f O
Me
NH2 NH2 Me Me HN Me
Tf
Me N Me MeO
Me
26 27 O 28
85% GC yield, >93% e.e. 71% yield, >96% e.e. 79% yield
(35% yield isolated as Boc amide)

Fig. 3 | Scope of internal alkenes that undergo hydroamination. a, Scope of
asymmetric hydroamination of acyclic internal alkenes. b, Scope of
hydroamination of simple cycloalkenes and of asymmetric hydroamination of
substituted cyclic alkenes. c, Products from the removal of the 2-(6-methyl)
pyridyl group. a2.5 mol% catalyst. b7.5 mol% catalyst. c20 mol% catalyst.
d
Conditions: 2.5 mol% [Ir(coe)2Cl]2, 7.5 mol% (S)-DTBM-SEGPHOS, 6 mol%
NaBARF, 1,4-dioxane, 120 °C. eConditions: PtO2, HCl, H2 (1 atm); NaBH4,
THF/EtOH. fEnantioselectivities were determined after conversion to the
original hydroamination product by palladium-catalysed cross-coupling of the
primary amine with 2-bromo-6-methylpyridine.

complicated by competing alkene isomerization, which led to mixtures
of isomers that were difficult to separate.
The pyridyl group of the hydroamination products (1, 9) was cleaved
by a short sequence that consisted of protonation, hydrogenation
and borohydride reduction. The corresponding primary amines
(26, 27) formed in 71–85% yield with little or no erosion in enantio-
meric excess (see Supplementary Information sections VI and X). By
the same sequence, hydroamination product 6 was converted to the
corresponding δ-lactam (28) in 79% yield.
To understand how the combination of a cationic iridium catalyst
and 2-amino-6-methylpyridine enabled the hydroamination of unac-
tivated internal alkenes, we investigated the reaction mechanism.
The reaction of a substituted cyclopentene with N,N-dideuterio-
2-amino-6-methylpyridine showed that the addition occurred in
a syn fashion. This stereochemistry is consistent with a mecha-
nism that involves migratory insertion of an alkene, rather than
nucleophilic attack on a metal-bound alkene complex (Fig. 4a).
Kinetic experiments showed that the reaction is first-order in

Nature | Vol 588 | 10 December 2020 | 257

Article
CO2Me 2.5 mol% [Ir(coe)2Cl]2 R′
a NHPyMe D
6 mol% (S)-DTBM-SEGPHOS
P R
6 mol% NaBARF MeO2C
+ *P Ir
Me N ND2 Probably via ND
Dioxane, 120 °C, 36 h D
44% isolated yield 40% D N
90% D 5 equiv. Me
23-D

NHPyMe
b Me [(R)-TMS-SYNPHOSIr(COD)]NTf2
+ Me
Me 2-MeTHF, 120 °C Me
Me N NH2 1

4.5 1.7 25
y = 5.7x + 1.4 1.5 y = 2.5x + 0.14 y = 2.6 × 103x – 4.5
1/(Initial rate) (×10–6 s M–1)

1/(Initial rate) (×10–6 s M–1)

4.0 R2 = 0.99 20

Initial rate × 107 (M s–1)

R2 = 0.99 R2 = 0.99
1.3
3.5 1.1 15

3.0 0.9 10
0.7
2.5 5
0.5
2.0 0.3 0
0.1 0.2 0.3 0.4 0.5 0.05 0.15 0.25 0.35 0.45 0.55 0.65 0 0.002 0.004 0.006 0.008 0.010 0.012
1/[Alkene] (M–1) [Amine] (M) [lr] (M)

c 5 mol%
Me [(R)-TMS-SYNPHOSIr(COD)]NTf2
+ +
Me No reaction
Me N NH2 N NH2 2-MeTHF, 120 °C, 48 h
1 equiv. 1 equiv. 10 equiv.

H
H
Ir
Ir

C C
Npy C
Nam
Nam C

Npy

H
P NH
H Ir
P TS-1a TS-2a *
P
Ir
* Me Ir-Npy (Å) 2.14 2.42 Me N
Side view P N
HN Ir-Nam (Å) 2.32 2.14

TS-1a TS-2a
ΔΔG‡ = 0 kcal mol–1 ΔΔ G‡ = 4.1 kcal mol–1

Fig. 4 | Mechanistic study of the hydroamination. a, Deuterium-labelling (S)-TMS-SEGPHOS. The hydride is located trans to the amido ligand in the
experiments. b, Experiments to reveal kinetic orders of each reaction lowest-energy transition state (TS-1a) that leads to the (R)-enantiomer but is
component. c, Competition experiments using 2-amino-6-methylpyridine and located trans to the pyridyl group in the lowest-energy transition state (TS-2a)
2-aminopyridine. d, Transition-state structures of alkene migratory insertion that leads to the (S)-enantiomer. Error bars in b correspond to those in the
computed by DFT. Single-point energies were computed at the M06/6- initial rates that result from errors in the integration of peaks in the gas
311+G(d,p)/SDD/SMD(1,4-dioxane) level of theory with structures optimized at chromatogram.
the B3LYP/6-31G(d)/SDD level. The ligand used for the calculations is

[iridium catalyst], positive-order in [cis-4-octene] and inverse-order reaction, we conducted hydroamination of cis-4-octene with equi-
in [2-amino-6-methylpyridine] (Fig. 4b). These data suggest that a molar amounts of 2-amino-6-methylpyridine and 2-aminopyridine.
molecule of 2-amino-6-methylpyridine dissociates reversibly from Whereas the reaction of 2-amino-6-methylpyridine alone afforded
iridium in the catalyst resting state before rate-limiting insertion of the hydroamination product in high yield, the reaction that contained
the alkene, presumably into the metal–amido bond41. To elucidate both 2-amino-6-methylpyridine and 2-aminopyridine provided neither
why the methyl group of 2-amino-6-methylpyridine is essential in this hydroamination product (Fig. 4c). This result implies that stronger

258 | Nature | Vol 588 | 10 December 2020

binding of the 2-aminopyridine than of 2-amino-6-methylpyridine
inhibits the catalyst. It also implies that the methyl group of
2-amino-6-methylpyridine weakens the binding to the iridium centre,
thereby allowing alkene binding and insertion.
To investigate the origin of the high levels of enantioselectivity in the
hydroamination, we conducted calculations on the alkene migratory
insertion using density functional theory (DFT). On the basis of the
kinetic experiments, this step is probably the rate-limiting and enantio-
determining step of the catalytic cycle. We calculated 12 possible iso-
meric transition states of migratory insertion that would lead to either
enantiomer of the products with cis-2-butene as the model alkene.
The structures of the lowest-energy transition states that lead to
the major and minor enantiomers are illustrated in Fig. 4d. The
two transition states of many metal-catalysed enantioselective
hydrofunctionalization reactions of alkenes to form two enantiomers
differ principally by the face of the alkene to which the metal is bound.
In the current system, the orientation of ancillary ligands around the
metal, in addition to the face of the alkene, differs greatly in the two
transition states. The geometry of TS-1a, the transition state leading
to the major enantiomer, contains meridionally oriented hydride,
pyridine and amido ligands with the hydride trans to the amido group.
By contrast, these three ligands in TS-2a, which is the lowest-energy
transition state leading to the minor enantiomer, are arranged with
the hydride trans to the pyridine donor. A geometry analogous to TS-
1a that would form the opposite enantiomer by orienting the methyl
groups away from the pyridine ligand (TS-2c and TS-2d; Supplemen-
tary Information Scheme S4) is higher in energy than is TS-2a. TS-1a
is probably the lowest-energy transition state for several reasons.
First, the Ir–Nam bond to the amido group, which is trans to a hydride in
TS-1a, is elongated (2.32 Å), thereby leading to higher reactivity
towards insertion. Second, the alkene is perpendicular to the P–Ir–P
plane in TS-1a, whereas the alkene is almost co-planar with the P–Ir–P
plane in TS-2a. These orientations place the substituents on the alkene
in TS-1a farther from the phosphine ligand than those on the alkene
in TS-2a, leading to less steric hindrance in TS-1a than in TS-2a. This
analysis suggests that electronic and steric effects together impart
high enantioselectivity to the hydroamination.
Our work demonstrates that the direct N–H addition of amines
to unactivated internal alkenes can occur with high enantioselec-
tivity under thermal conditions, without the need for strategies
involving formal hydroamination. Despite the typically high barri-
ers and weak thermodynamic driving force for hydroamination,
the use of cationic bisphosphine-ligated iridium as the catalyst and
2-amino-6-methylpyridine as the amine led to enhancements of the
rates of multiple steps within the catalytic cycle and to suppression of
alkene isomerization and oxidative amination, enabling the hydroami-
nation of unactivated internal alkenes in high yield and with high enan-
tioselectivity. The hydroamination products can be converted to the
corresponding primary amines readily with preservation of the high
enantiomeric excess of the hydroamination products. These design
principles should provide a starting point to address the long-standing
challenge of applying hydroamination of unactivated internal alkenes
to the synthesis of chiral amines and inspire advances in other asym-
metric hydrofunctionalizations of internal alkenes.
Online content
Any methods, additional references, Nature Research reporting sum-
maries, source data, extended data, supplementary information,
acknowledgements, peer review information; details of author con-
tributions and competing interests; and statements of data and code
availability are available at https://doi.org/10.1038/s41586-020-2919-z.
1. Müller, T. E. & Beller, M. Metal-initiated amination of alkenes and alkynes. Chem. Rev. 98,
675–704 (1998).
2. Müller, T. E., Hultzsch, K. C., Yus, M., Foubelo, F. & Tada, M. Hydroamination: direct
addition of amines to alkenes and alkynes. Chem. Rev. 108, 3795–3892 (2008).
3. Huang, L., Arndt, M., Gooßen, K., Heydt, H. & Gooßen, L. J. Late transition metal-catalyzed
hydroamination and hydroamidation. Chem. Rev. 115, 2596–2697 (2015).
4. Reznichenko, A. L. & Hultzsch, K. C. in Organic Reactions 1–554 (Wiley, 2015).
5. Gurak, J. A., Yang, K. S., Liu, Z. & Engle, K. M. Directed, regiocontrolled hydroamination of
unactivated alkenes via protodepalladation. J. Am. Chem. Soc. 138, 5805–5808 (2016).
6. Karshtedt, D., Bell, A. T. & Tilley, T. D. Platinum-based catalysts for the hydroamination of
olefins with sulfonamides and weakly basic anilines. J. Am. Chem. Soc. 127, 12640–12646
(2005).
7. Zhang, J., Yang, C.-G. & He, C. Gold(i)-catalyzed intra- and intermolecular hydroamination
of unactivated olefins. J. Am. Chem. Soc. 128, 1798–1799 (2006).
8. Musacchio, A. J. et al. Catalytic intermolecular hydroaminations of unactivated olefins
with secondary alkyl amines. Science 355, 727–730 (2017).
9. Nguyen, T. M., Manohar, N. & Nicewicz, D. A. anti-Markovnikov hydroamination of alkenes
catalyzed by a two-component organic photoredox system: direct access to
phenethylamine derivatives. Angew. Chem. Int. Ed. 53, 6198–6201 (2014).
10. Zhang, Z., Lee, S. D. & Widenhoefer, R. A. Intermolecular hydroamination of ethylene and
1-alkenes with cyclic ureas catalyzed by achiral and chiral gold(i) complexes. J. Am.
Chem. Soc. 131, 5372–5373 (2009).
11. Reznichenko, A. L., Nguyen, H. N. & Hultzsch, K. C. Asymmetric intermolecular
hydroamination of unactivated alkenes with simple amines. Angew. Chem. Int. Ed. 49,
8984–8987 (2010).
12. Pan, S., Endo, K. & Shibata, T. Ir(i)-catalyzed intermolecular regio- and enantioselective
hydroamination of alkenes with heteroaromatic amines. Org. Lett. 14, 780–783 (2012).
13. Vanable, E. P. et al. Rhodium-catalyzed asymmetric hydroamination of allyl amines.
J. Am. Chem. Soc. 141, 739–742 (2019).
14. Nugent, T. C. Chiral Amine Synthesis: Methods, Developments and Applications (Wiley
VCH, 2010).
15. Trowbridge, A., Walton, S. M. & Gaunt, M. J. New strategies for the transition-metal
catalyzed synthesis of aliphatic amines. Chem. Rev. 120, 2613–2692 (2020).
16. Wang, C. & Xiao, J. in Stereoselective Formation of Amines (eds Li, W. & Zhang, X.) 261–282
(Springer, 2014).
17. Nugent, T. C. & El-Shazly, M. Chiral amine synthesis – recent developments and trends for
enamide reduction, reductive amination, and imine reduction. Adv. Synth. Catal. 352,
753–819 (2010).
18. Patil, M. D., Grogan, G., Bommarius, A. & Yun, H. Oxidoreductase-catalyzed synthesis of
chiral amines. ACS Catal. 8, 10985–11015 (2018).
19. Xie, J.-H., Zhu, S.-F. & Zhou, Q.-L. Transition metal-catalyzed enantioselective
hydrogenation of enamines and imines. Chem. Rev. 111, 1713–1760 (2011).
20. Ellman, J. A., Owens, T. D. & Tang, T. P. N-tert-Butanesulfinyl imines: versatile intermediates
for the asymmetric synthesis of amines. Acc. Chem. Res. 35, 984–995 (2002).
21. You, S.-L., Zhu, X.-Z., Luo, Y.-M., Hou, X.-L. & Dai, L.-X. Highly regio- and enantioselective
Pd-catalyzed allylic alkylation and amination of monosubstituted allylic acetates with
novel ferrocene P,N-ligands. J. Am. Chem. Soc. 123, 7471–7472 (2001).
22. Ohmura, T. & Hartwig, J. F. Regio- and enantioselective allylic amination of achiral allylic
esters catalyzed by an iridium−phosphoramidite complex. J. Am. Chem. Soc. 124,
15164–15165 (2002).
23. Löber, O., Kawatsura, M. & Hartwig, J. F. Palladium-catalyzed hydroamination of
1,3-dienes: a colorimetric assay and enantioselective additions. J. Am. Chem. Soc. 123,
4366–4367 (2001).
24. Adamson, N. J., Hull, E. & Malcolmson, S. J. Enantioselective intermolecular addition of
aliphatic amines to acyclic dienes with a Pd–PHOX catalyst. J. Am. Chem. Soc. 139,
7180–7183 (2017).
25. Long, J., Wang, P., Wang, W., Li, Y. & Yin, G. Nickel/Brønsted acid-catalyzed chemo- and
enantioselective intermolecular hydroamination of conjugated dienes. iScience 22,
369–379 (2019).
26. Tran, G., Shao, W. & Mazet, C. Ni-catalyzed enantioselective intermolecular
hydroamination of branched 1,3-dienes using primary aliphatic amines. J. Am. Chem. Soc.
141, 14814–14822 (2019).
27. Kawatsura, M. & Hartwig, J. F. Palladium-catalyzed intermolecular hydroamination of
vinylarenes using arylamines. J. Am. Chem. Soc. 122, 9546–9547 (2000).
28. Utsunomiya, M. & Hartwig, J. F. Intermolecular, Markovnikov hydroamination of
vinylarenes with alkylamines. J. Am. Chem. Soc. 125, 14286–14287 (2003).
29. Yang, Y., Shi, S.-L., Niu, D., Liu, P. & Buchwald, S. L. Catalytic asymmetric hydroamination
of unactivated internal olefins to aliphatic amines. Science 349, 62–66 (2015).
30. Gui, J. et al. Practical olefin hydroamination with nitroarenes. Science 348, 886–891 (2015).
31. Johns, A. M., Sakai, N., Ridder, A. & Hartwig, J. F. Direct measurement of the
thermodynamics of vinylarene hydroamination. J. Am. Chem. Soc. 128, 9306–9307
(2006).
32. Liu, Z. & Hartwig, J. F. Mild, rhodium-catalyzed intramolecular hydroamination of
unactivated terminal and internal alkenes with primary and secondary amines. J. Am.
Chem. Soc. 130, 1570–1571 (2008).
33. Huang, J.-M., Wong, C.-M., Xu, F.-X. & Loh, T.-P. InBr3 catalyzed intermolecular
hydroamination of unactivated alkenes. Tetrahedron Lett. 48, 3375–3377 (2007).
34. Sevov, C. S., Zhou, J. & Hartwig, J. F. Iridium-catalyzed intermolecular hydroamination of
unactivated aliphatic alkenes with amides and sulfonamides. J. Am. Chem. Soc. 134,
11960–11963 (2012).
35. Utsunomiya, M., Kuwano, R., Kawatsura, M. & Hartwig, J. F. Rhodium-catalyzed
anti-Markovnikov hydroamination of vinylarenes. J. Am. Chem. Soc. 125, 5608–5609
(2003).
36. Pawlas, J., Nakao, Y., Kawatsura, M. & Hartwig, J. F. A general nickel-catalyzed
hydroamination of 1,3-dienes by alkylamines: catalyst selection, scope, and mechanism.
J. Am. Chem. Soc. 124, 3669–3679 (2002).
37. Casalnuovo, A. L., Calabrese, J. C. & Milstein, D. Rational design in homogeneous
catalysis. Iridium(i)-catalyzed addition of aniline to norbornylene via nitrogen-hydrogen
activation. J. Am. Chem. Soc. 110, 6738–6744 (1988).
Nature | Vol 588 | 10 December 2020 | 259
Article
38. Dorta, R., Egli, P., Zürcher, F. & Togni, A. The [IrCl(diphosphine)]2/fluoride system.
Developing catalytic asymmetric olefin hydroamination. J. Am. Chem. Soc. 119,
10857–10858 (1997).
39. Zhou, J. & Hartwig, J. F. Intermolecular, catalytic asymmetric hydroamination of bicyclic
alkenes and dienes in high yield and enantioselectivity. J. Am. Chem. Soc. 130,
12220–12221 (2008).
40. Sevov, C. S., Zhou, J. & Hartwig, J. F. Iridium-catalyzed, intermolecular hydroamination of
unactivated alkenes with indoles. J. Am. Chem. Soc. 136, 3200–3207 (2014).
41. Hanley, P. S. & Hartwig, J. F. Migratory insertion of alkenes into metal–oxygen and metal–
nitrogen bonds. Angew. Chem. Int. Ed. 52, 8510–8525 (2013).
42. Thompson, W. H. & Sears, C. T. Kinetics of oxidative addition to iridium(i) complexes.
Inorg. Chem. 16, 769–774 (1977).
43. Smout, V. et al. Removal of the pyridine directing group from α-substituted N-(pyridin-
2-yl)piperidines obtained via directed Ru-catalyzed sp3 C–H functionalization. J. Org.
Chem. 78, 9803–9814 (2013).
44. Hanley, P. S. & Hartwig, J. F. Intermolecular migratory insertion of unactivated olefins into
palladium–nitrogen bonds. Steric and electronic effects on the rate of migratory
insertion. J. Am. Chem. Soc. 133, 15661–15673 (2011).
260 | Nature | Vol 588 | 10 December 2020
45. Zhang, M., Hu, L., Lang, Y., Cao, Y. & Huang, G. Mechanism and origins of regio- and
enantioselectivities of iridium-catalyzed hydroarylation of alkenyl ethers. J. Org. Chem.
83, 2937–2947 (2018).
46. Xing, D., Qi, X., Marchant, D., Liu, P. & Dong, G. Branched-selective direct α-alkylation of
cyclic ketones with simple alkenes. Angew. Chem. Int. Ed. 58, 4366–4370 (2019).
47. Xi, Y., Butcher, T. W., Zhang, J. & Hartwig, J. F. Regioselective, asymmetric formal
hydroamination of unactivated internal alkenes. Angew. Chem. Int. Ed. 55, 776–780 (2016).
48. Mei, T.-S., Werner, E. W., Burckle, A. J. & Sigman, M. S. Enantioselective redox-relay
oxidative heck arylations of acyclic alkenyl alcohols using boronic acids. J. Am. Chem.
Soc. 135, 6830–6833 (2013).
49. Morandi, B., Wickens, Z. K. & Grubbs, R. H. Regioselective Wacker oxidation of internal
alkenes: rapid access to functionalized ketones facilitated by cross-metathesis. Angew.
Chem. Int. Ed. 52, 9751–9754 (2013).
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in
published maps and institutional affiliations.
© The Author(s), under exclusive licence to Springer Nature Limited 2020

Data availability
The data that support the findings of this study are available within the
article and its Supplementary Information.
Acknowledgements The enantioselective aspects of the work were supported by the National
Institutes of Health under grant R35GM130387 and the catalyst development was supported
by the Director, Office of Science, of the US Department of Energy under contract number
DE-AC02-05CH11231. Calculations were performed at the Molecular Graphics and
Computation Facility at UC Berkeley funded by the NIH (S10OD023532). We gratefully
acknowledge Takasago for gifts of (S)-DTBM-SEGPHOS, and H. Celik for assistance with
nuclear magnetic resonance (NMR) experiments. Instruments in the College of Chemistry
NMR facility are supported in part by NIH S10OD024998. We thank R. G. Bergman, B. Su and
T. Butcher for discussions. Y.X. thanks Bristol-Myers Squibb for a graduate fellowship,
S. Pedram for supply of NaBARF and D. Small for assistance with DFT calculations.
Author contributions Y.X. and J.F.H. conceived the project. Y.X. discovered the reaction and
performed experiments and DFT calculations. S.M. performed experiments for revision. Y.X.
and J.F.H. wrote the manuscript.
Competing interests The authors declare no competing interests.
Additional information
Supplementary information is available for this paper at https://doi.org/10.1038/s41586-020-
2919-z.
Correspondence and requests for materials should be addressed to J.F.H.
Peer review information Nature thanks the anonymous reviewer(s) for their contribution to the
peer review of this work.
Reprints and permissions information is available at http://www.nature.com/reprints.
Article

Quantification of an efficiency–sovereignty
trade-off in climate policy

https://doi.org/10.1038/s41586-020-2982-5 Nico Bauer1 ✉, Christoph Bertram1, Anselm Schultes1, David Klein1, Gunnar Luderer1,2,
Elmar Kriegler1, Alexander Popp1 & Ottmar Edenhofer1,2,3
Received: 13 March 2020

Accepted: 29 September 2020

The Paris Agreement calls for a cooperative response with the aim of limiting global
Published online: 9 December 2020
warming to well below two degrees Celsius above pre-industrial levels while
Check for updates
reaffirming the principles of equity and common, but differentiated responsibilities
and capabilities1. Although the goal is clear, the approach required to achieve it is not.
Cap-and-trade policies using uniform carbon prices could produce cost-effective
reductions of global carbon emissions, but tend to impose relatively high mitigation
costs on developing and emerging economies. Huge international financial transfers
are required to complement cap-and-trade to achieve equal sharing of effort, defined
as an equal distribution of mitigation costs as a share of income2,3, and therefore the
cap-and-trade policy is often perceived as infringing on national sovereignty2–7. Here
we show that a strategy of international financial transfers guided by moderate
deviations from uniform carbon pricing could achieve the goal without straining
either the economies or sovereignty of nations. We use the integrated assessment
model REMIND–MAgPIE to analyse alternative policies: financial transfers in uniform
carbon pricing systems, differentiated carbon pricing in the absence of financial
transfers, or a hybrid combining financial transfers and differentiated carbon prices.
Under uniform carbon prices, a present value of international financial transfers of
4.4 trillion US dollars over the next 80 years to 2100 would be required to equalize
effort. By contrast, achieving equal effort without financial transfers requires carbon
prices in advanced countries to exceed those in developing countries by a factor of
more than 100, leading to efficiency losses of 2.6 trillion US dollars. Hybrid solutions
reveal a strongly nonlinear trade-off between cost efficiency and sovereignty:
moderate deviations from uniform carbon prices strongly reduce financial transfers
at relatively small efficiency losses and moderate financial transfers substantially
reduce inefficiencies by narrowing the carbon price spread. We also identify risks and
adverse consequences of carbon price differentiation due to market distortions that
can undermine environmental sustainability targets8,9. Quantifying the advantages
and risks of carbon price differentiation provides insight into climate and
sector-specific policy mixes.

The Paris Agreement (PA) set out ambitious targets for climate change
stabilization that require a coordinated effort to limit and reduce green-
house gas (GHG) emissions. Negotiations of an ambitious international
emissions reductions framework building on the PA need to consider
at least three criteria. First, fair effort sharing requires that unfair
outcomes are avoided, such as regressive policy effects that imply
higher relative effort by low-income countries. Second, cost efficiency—
minimizing the global aggregate mitigation costs—requires equaliza-
tion of marginal abatement costs across countries for an incremental
ton of carbon dioxide (CO2) avoided. Finally, in this study, national sov-
ereignty focuses on nation states’ aim to maintain governing control of
economic resources by limiting international transfer payments6,7,10,11.
In view of the deep differences in economic, structural and technologi-
cal development across countries1,12,13 (Fig. 1a–d), international climate
policies have to balance these criteria to achieve the PA targets.
Sovereignty versus equity trade-off
So far, most analyses of the PA climate targets have applied cost-efficient
policy frameworks, implying uniform carbon prices across regions, at
least in the medium term. Although the uniform price sets the upper
bound for the marginal abatement costs of mitigation measures that
are undertaken as equal for each region, the resulting relative emis-
sions reductions and mitigation costs vary across regions (Fig. 1e, f).

Potsdam Institute for Climate Impact Research (PIK), Member of the Leibniz Association, Potsdam, Germany. 2Technical University of Berlin, Berlin, Germany. 3Mercator Institute on Global
1

Commons and Climate Change (MCC), Berlin, Germany. ✉e-mail: nico.bauer@pik-potsdam.de

Nature | Vol 588 | 10 December 2020 | 261

Article
a e 175
40 Regional emissions reduction
on

emissions reduction (global = 100)

Ratio of regional to global relative
stronger than global

Income per capita

(US$1,000 (2010))
30 150

20
125
10
100
0
15
b 75
Regional emissions reduction
on
(billion tCO2)
10
Emissions

50 weaker than global

CD

ia

Am atin

d A st

on ng
ca

ies
As

fric
an e Ea
5

ec ormi
eri
OE

om
L

l
dd

f
Re
Mi
0 f
c 12.5

Difference of regional to global mitigation costs

20

(percentage-point differences, global = 0)

10.0
Regional mitigation cost
Emissions per
capita (tCO2)

7.5 higher than global

5.0
10
2.5

0
d 4,000
0
(US$ (2010) per tCO2)
Carbon productivity

3,000 Regional mitigation cost

lower than global
–10
2,000

CD

ia

Am atin

d A st

on ng
ca

ies
As

fric
an e Ea

ec ormi
1,000

eri
OE

om
L

l
dd

f
Re
Mi
1960 1980 2000 2020
Year

Region Models
AIM/CGE IMAGE TIAM
Asia OECD
BET MESSAGE–GLOBIOM WITCH–GLOBIOM
Latin America Reforming economies GCAM POLES
Middle East and Africa GRAPE REMIND–MAgPIE

Fig. 1 | World economic development and emission mitigation projections. (e) and regional mitigation costs relative to GDP (f) in 2040 used by the IPCC for
a–d, Global socioeconomic developments across five world regions: income mitigation scenarios. The boxplots show the median, and the 25% and the 75%
per capita (a), annual CO2 emissions (b), annual CO2 emissions per capita (c) and quartiles; whiskers show 1.5 times the interquartile range. For details on data
carbon productivity (d). e, f, Model results for regional emissions reductions sources and regional aggregates, see Methods and Extended Data Table 1.

Low-income countries with low carbon and energy productivities tend carbon prices in 2030 ranging from nearly zero to more than US$100 per
to reduce emissions by a larger fraction and carry relatively higher tCO2 (refs. 15,21–23). This implies more equitable burden sharing, but
mitigation costs than economies with high carbon and energy produc- fragmented policy implementation compromises cost efficiency22–25.
tivities14,15. In particular, Asian countries show systematically higher In addition, aggregated globally, NDCs fall short of cost-effective emis-
relative mitigation costs and stronger relative emissions reductions sions reductions required in 2030 to achieve the long-term PA climate
than Organisation for Economic Co-operation and Development targets15,26,27.
(OECD) countries. The aggregate region of Middle East and Africa shows So far, studies on effort sharing have assumed full integration of
higher mitigation costs, but weaker emissions reduction rates because an international cap-and-trade system and varied permit allocations,
of high energy and low carbon intensity. Therefore, without transfers, which either results in regressive outcomes or implies large transfers.
cost-efficient international climate policies tend to cause regressive However, in these studies, the permit allocation does not affect the
income effects that deepen economic inequality. cost-optimal energy and land-use transformations. The evaluation of
The Kyoto Protocol negotiations in 1998 inspired research into cost real-world fragmented policy regimes such as the NDCs shows that
reductions by trading emissions permits16. Various subsequent stud- they fall short of required emissions reductions at relatively high
ies investigated the trade-off between fair effort sharing and national global costs that are not shared equitably15. This study investigates
sovereignty of a broad range of initial permit allocations derived from the trade-off between cost efficiency and sovereignty by exploring
different equity and sovereignty principles17–19. In cap-and-trade sys- the policy options between idealized cap-and-trade systems and frag-
tems, permit allocations following the equal per-capita allocation mented systems of real-world policies while pursuing the PA to limit
principle imply international transfers2–6 that can reach US$6 trillion warming to well below 2 °C above pre-industrial levels with equitable
present value globally until 21004. The equal-effort-sharing criterion effort sharing.
can require even larger transfers to avoid regressive income effects2.
In general, changes in permit allocation do not affect the modelled
energy and land-use sector transformation because the carbon price REMIND–MAgPIE and mitigation policies
does not change in a globally integrated cap-and-trade system if a given We develop a policy analysis framework that consistently varies
number of permits is allocated differently20. regional policy strength and international transfers in scenarios that
The PA, with its system of nationally determined contributions comply with the equal-effort-sharing principle and the well below 2 °C
(NDCs), moved regionally fragmented climate policies to the centre target. The policy framework is evaluated using the long-term multi-
of analysis. The policy ambition of the NDCs corresponds to regional regional integrated assessment model (IAM) REMIND–MAgPIE, which

262 | Nature | Vol 588 | 10 December 2020

 a 3,000 b
Carbon pricing
Uniform
Cost efficiency

Carbon price in 2030 (US$ per tCO2, log scale)

Differentiated √ √
1,000 Sovereignty √ √
Equity/fairness √ √

15,000 Region
Reforming economies
300
Middle East and North Africa

(billion US$, NPV 2020–2100)

Sub-Saharan Africa

Cumulative income loss

10,000 Other Asia
India
100
China
Latin America
5,000 Canada, Australia, New Zealand
Japan
30 United States
Other Europe
EU
0
EU

pe

n
nd

ca
ina

ia
sia

hA a
a
ies

e
tes

te d
tr a r m

h t orm
fer

r
ic
fric

rag
pa

Ind

sfe
eri
ur o

ala

rA

om
d N n Afr
Ch
Sta

o
ns
Ja

tia
ave

ho Unif

wit Unif
ran
Am
Ze

he
rE

on

en
ed

ara
ort
Ot

r ld
he

ec
w

fer
tin

ut
nit

Ne

ah
Ot

Wo

Dif
La

ng
U

b-S
lia,

wit
an

mi
tra

for
Su
st
Ea
us

Re
,A

le
dd
da

Mi
na
Ca

Fig. 2 | Carbon pricing and regional effort. a, Regional CO2 prices in 2030 on a which two out of three criteria are met in each scenario. The differences
log scale. The world average depicts the global average price weighted with the between components of the uniform pricing schemes represent regional
regional CO2 emissions in 2030. b, Total mitigation effort including discounted transfers, showing that the six regions at the bottom are donors and the six
income losses in NPV and eventual transfers compared with a no-policy regions at the top are recipients.
baseline for 2020–2100; the discount rate is 5% yr−1. The table summarizes

links the regional model of investment and development REMIND with mitigation costs by US$2.6 trillion or 21%. Owing to the nonlinearly
the model of agricultural production and its impact on the environment shaped regional mitigation cost functions32, the relative increase is
MAgPIE67. This model has been used to analyse transition pathways28, much lower than the quadrupling of the average carbon price.
effort-sharing schemes4 and NDCs21,25. REMIND–MAgPIE integrates
the macroeconomy with the land use and the energy sector in a gen-
eral equilibrium framework including energy and food trade29. The Deriving a trade-off curve
baseline scenario quantifies a middle-of-the-road scenario (Shared Each of the three corner solutions leads to extreme outcomes in
Socioeconomic Pathway 2 (SSP2)) with medium economic growth and regional mitigation costs, transfers or carbon price differentiation.
partial income convergence (Extended Data Fig. 3)29,30. For instance, Therefore, we explore the trade-off between sovereignty and cost
measured in market exchange rates, US per-capita income in 2020 efficiency by gradually compressing the carbon price spread, moving
is 28 times that of India declining to 13 by 2050 and 5 by 2100. In this from the no-transfer case with differentiated carbon prices towards
study, a carbon budget of 1,300 GtCO2 for the period 2011–2100 is the uniform carbon price case with transfers. We apply an exponential
assumed31. Short-term policies derived from NDCs are applied in 2020; function to adjust pairs of regional prices (pi/pj)α of the no-transfer
carbon prices thereafter increase by 5% yr−1 until 2060 and continue case for regions i and j. The compression parameter α is varied
growing linearly thereafter. The flattened time profile limits overshoot between zero (uniform prices case) and one (differentiated prices
of the carbon budget. Regional variations in climate policy strength case). The compressed carbon price set is used in REMIND–MAgPIE
are implemented by varying carbon prices; international transfers and jointly rescaled to meet the carbon budget, which leads to vari-
are implemented as direct payments. Benefits from avoided climate ations of global mitigation costs and global residual transfers. This
damages and adaptation costs are not considered in these scenarios approach does not necessarily determine the frontier of smallest
(see Methods). efficiency losses for varying transfer volumes, but derives a trade-off
curve that is economically feasible and complies with the PA climate
objective and the equal-effort-sharing criterion (see Methods for
Corner solutions of a policy trilemma detail).
Figure 2 shows that only two of the three criteria of cost efficiency, fair-
ness and sovereignty can be fulfilled simultaneously. The cost-efficient
policy with a uniform carbon price (US$56 per tCO2 in 2030) but no Relaxing uniform carbon pricing policies
transfers leads to mitigation costs of US$12.6 trillion that are distrib- The resulting trade-off curve is strongly nonlinear (Fig. 3a). Starting
uted regressively across regions (0.30% relative income loss for the at the solution with uniform carbon prices, the steep drop in trans-
European Union (EU) and 3.4% for India). A total of US$4.4 trillion in fers follows directly from the cost-efficient solution that requires
international transfers is required to neutralize the regressive income equalization of marginal abatement costs. If non-OECD economies
effects and achieve equal effort sharing. Alternatively, achieving equal lower their carbon taxes, their mitigation costs decrease by an amount
effort sharing without transfers requires regionally differentiated car- slightly smaller than the cost increase in OECD countries that need
bon prices. The ratio between the highest and lowest price is 130. The to impose higher carbon taxes to balance the global carbon budget.
global average carbon price weighted with regional emissions increases Hence, relatively large transfer reductions cause only small inefficien-
to US$225 per tCO2, with a weighted standard deviation of US$400 per cies, if the carbon tax spread remains sufficiently small (Methods,
tCO2. These deviations from uniform carbon pricing increase global Extended Data Fig. 1).

Nature | Vol 588 | 10 December 2020 | 263

Article
a
56;0
Uniform Global carbon price
with transfer Weighted
4,000

Interregional transfer (NPV billion US$)

average
59;14 63;32
60;19 Standard
3,000 deviation
63;32

68;49
2,000
77;72

88;102
1,000 Differentiated
106;146
133;203
Uniform 170;286
195;339
without transfer 216;383
0 225;402

13,000 14,000 15,000

Global mitigation costs (billion US$, NPV)

b c
2,000 Coal power capacity shutdown exceeds 50%
OECD
Cumulative CO2 emissions 2020–2100 (GtCO2)

Non-OECD
Net emissions
Non-OECD Electricity share exceeds 30%
OECD
OECD
Non-OECD
1,000 Carbon removal using BECCS exceeding 250 Mt yr–1
Emissions and removals OECD
D Non-OECD
EC
n -O CD
No OE Fossil primary energy share falls below 50%
OECD
Emissions deforestation
Non-OECD
Emissions FFI
0
Removals afforestation Total CO2 emissions turn negative
Removals BECCS OECD
Removals DAC Non-OECD

2020 2040 2060 2080 2100

Year threshold is reached
58.9
60.6
63.2
67.3
76.8
87.6
106
136
173
190
215
226
5
.
55

Global average carbon price in 2030 Scenarios

(US$ per tCO2) Differentiated Paritally differentiated Uniform Baseline

Fig. 3 | Sovereignty versus cost-efficiency trade-off and consequences of countries differentiated by emissions sources and carbon removals. FFI, fossil
differentiated carbon prices. a, The trade-off curve, including the three fuel and industry; DAC, direct air capture; BECCS, bioenergy with carbon
corner solutions (marked with red circles). The numbers indicated the 2030 capture and sequestration. c, Different timing of mitigation measures in OECD
global average carbon price and the standard deviation using the regional CO2 and non-OECD regions. ‘Partially differentiated’ is the case with an average
emissions as weights. The costs and transfers are NPVs for the period 2020– carbon price of US$63.3 per tCO2. In some scenarios, the threshold is not
2100. The time path of transfers is shown in Extended Data Fig. 5 and discussed reached before 2100 and, therefore, no marker is shown.
in Methods. b, The cumulative net carbon emissions in OECD and non-OECD

Fig. 5). As the carbon tax spread grows, changes in fossil fuel use become
Relaxing the no-transfer constraint exhausted, while carbon removal is intensified in OECD countries.
Starting at the opposite end of the trade-off curve with full price differ- Hence, untapped abatement potentials in low-carbon-price countries
entiation, the nonlinear shape highlights the effect of limited transfers. are largely offset by costly abatement options in high-income countries
Reducing the price spread by three quarters lessens the global ineffi- as their ability to reduce fossil fuel use is exhausted. This interacts with
ciency by 56%, but requires only 21% of transfers of the uniform price land market distortions: emissions in non-OECD countries increase
case. Owing to the strongly increasing mitigation costs, the marginal due to land-use extensification by deforestation to export bioenergy,
and total costs for OECD countries decline rapidly as their emissions whereas OECD countries reduce agricultural land for afforestation
reductions are relaxed, whereas increments of non-OCED countries while importing biomass that is used in the energy sector combined
are smaller, as their emissions constraints need to be tightened. Hence, with carbon capture and storage (BECCS)8,9. Therefore, market distor-
starting from the solution of full price differentiation makes transfers tions caused by regional policy differentiation risk detrimental impacts
highly effective with respect to reducing inefficiency while maintaining on environmental sustainability. See also Extended Data Figs. 5, 6.
the equal-effort-sharing criterion. The efficiency–sovereignty trade-off also interacts with the timing
of mitigation measures across regions. Figure 3c shows for selected
indicators of the energy sector transformation the year a threshold is
Unintended effects of distorted markets reached. Under uniform carbon pricing, mitigation measures proceed
The solution to the efficiency–sovereignty trade-off has broader impli- at a similar speed in different regions33. For example, BECCS deploy-
cations. Increasingly differentiated carbon prices lead to a reallocation ment exceeds 250 MtCO2 yr−1 shortly after 2040 in OECD and non-OECD
of regional emissions and multiple market distortions. The regional regions. Spreading carbon prices leads OECD countries to front-load
reallocation of gross emissions and fossil fuel market distortions are and tighten the timing of measures, whereas non-OECD countries delay
largest for relatively small carbon price spreads (Fig. 3b, Extended Data and stretch the timing. For instance, non-OECD countries exceed the

264 | Nature | Vol 588 | 10 December 2020

BECCS threshold nearly 10 years later, whereas OECD countries pass it
more than 10 years earlier. Reliance on fossil fuels changes as well: even
at moderate carbon price differentiation, fossil fuel use (including coal)
in some developing countries peaks only in 2035, while OECD countries
phase-out all fossil fuels more rapidly. Moreover, the accelerated energy
transition and the amplified carbon removal deployment in OECD
countries demand huge energy-sector investments that increasingly
crowd-out overall investments in non-OECD countries. Hence, relatively
low energy prices in non-OECD countries do not necessarily attract
investments to support economic development (Extended Data Fig. 6).
The shape of the trade-off matters
The core finding of this research is the strongly nonlinear trade-off
between cost-efficiency and sovereignty in achieving the long-term PA
climate target in an equitable way. The cost-efficient, yet idealized, pol-
icy framework of uniform carbon prices requires international transfers
that correspond to 35% of global mitigation costs to avoid unequitable
effort sharing. These huge transfers confirm previous cap-and-trade
studies2–5,10,15,16, but are unlikely to be agreed on internationally. We find
that modest deviations from uniform carbon prices allow to strongly
reduce transfers. Carbon price differentiation, however, causes market
distortions, sustainability risks and asynchronous timing of mitigation
measures. Real-world climate policy agreements, such as the NDCs,
do not reach the emissions reductions necessary to achieve the PA tar-
gets and are internationally fragmented, implying a broad variation of
regional carbon prices. Previous studies have quantified the resulting
inefficiency by comparing fragmented with idealized cap-and-trade
systems, but relatively large transfers would still be required15,16.
This study expands the analysis: if sovereignty is prioritized, the
implementation of the PA targets with equitable effort sharing would
require higher and more differentiated carbon prices. However, mod-
erate transfers allow for convergence of regional carbon prices and
substantial reductions of inefficiencies. Full convergence towards
uniform carbon prices leads to only small margins in reducing inef-
ficiency, but require increasing transfers. The high sensitivity close to
the corner solution follows from the optimality condition of uniform
carbon prices and the nonlinearity of abatement cost functions. These
features and, thus, the general results are not unique to the modelling
tool used in this study.
The efficiency–sovereignty trade-off and other consequences of
carbon price differentiation and transfers for sustainability and mitiga-
tion timing inform the discourse about Article 6 of the PA.
Sensitivity of the trade-off curve
The distributional challenge, reflected by the trade-off curve, depends
on various assumptions. As a sensitivity analysis, we varied the assump-
tion of socioeconomic drivers of economic inequality. Faster income
convergence that reflects more inclusive growth (we use the SSP1
demography and economy projections) softens the distributional
challenge by reducing the ratio of global transfers to global mitiga-
tion costs from 35% to 31%. Furthermore, the global carbon budget is
important: increasing it to 1,600 GtCO2 reduces the maximum carbon
prices spread to 40, whereas a tighter budget of 950 GtCO2 widens
the spread to 500. The metric chosen to operationalize burden shar-
ing is also crucial. In this analysis, we used relative income reduction
for the period 2020–2100, which emphasizes the potential losses of
developing countries (for example, India). The alternative metric of
consumption losses puts stronger emphasis on losses of countries well
endowed with fossil fuels. Shortening the time horizon of the metric to
2050 reduces the spread factor to 11. Delaying the deployment of CCS
until 2050 or slowing the maximum annual rate of abandoning fossil
fuel infrastructure to 6% increases minimum mitigation costs and also
intensifies the distributional challenge requiring either more transfers
or inducing higher efficiency losses. Imposing a trade ban on bioenergy
increases mitigation costs in the uniform policy regime, while eliminat-
ing carbon leakage via bioenergy trade and substantially reducing the
economic efficiency loss (Extended Data Table 2, Extended Data Fig. 7).
Inequality and fair climate policies
Equal effort sharing could be considered too weak a fairness crite-
rion for two reasons. First, even under PA warming limits, damages
are disproportionally more severe in developing countries34,35. Hence,
a more progressive effort-sharing criterion can be justified in a policy
framework based only on mitigation costs36. Second, different socio-
economic capabilities and inequality aversion suggest a more progres-
sive effort-sharing scheme for managing commons37. It is also argued
that OECD countries bear greater responsibility due to their historic
emissions. However, for OECD countries, equal effort and historic
responsibility allocation show only small differences, but allocations
are substantially shifted between non-OECD countries depending on
their fossil fuel endowments4.
Transfers and sovereignty
Equitable effort sharing in international climate policy requires a com-
promise of the sovereignty–efficiency trade-off. Recent research on
international climate policies and cooperation, for example, by carbon
tax harmonization, has made progress towards negotiating coopera-
tion with uniform carbon pricing, but has not addressed issues of equity
and transfers6,11,38. Transfers might infringe on sovereignty, but do not
necessarily run counter to national welfare. Especially in light of the
heterogeneity between countries, transfers facilitate agreement in
international environmental policies that improve the welfare of all
nations by increasing commitments and reducing free-rider behav-
iour39–41. Following previous research on permit allocation principles,
this study refers to sovereignty only as limits on ceding control over
economic resources by transfers within the regime of the PA4,10,18,19. In
a broader perspective, climate change also relates to the sovereignty
of self-governing as well as the sovereignty of territory, if sea-level rise
is considered.
New perspective on policy mixes
Large carbon price spreads induce market distortions in energy
and land-use transformations, thus undermining climate policy
cost-efficiency, effectiveness and broader sustainability targets. Addi-
tional sector-specific land-use and energy policies could correct associ-
ated market distortions, thereby complementing a non-optimal carbon
pricing regime, as exemplified by the bioenergy trade restriction men-
tioned above. Other complementary policies are coal phase-out, fossil
fuel subsidy reform, forest protection and international technology
transfers15,42–46. In the context of differentiated carbon prices, these
policies could reduce distributional challenges and attenuate market
distortions. Moreover, the inclusion of climate change damages is a
promising future line of research direction47.
Online content
Any methods, additional references, Nature Research reporting sum-
maries, source data, extended data, supplementary information,
acknowledgements, peer review information; details of author con-
tributions and competing interests; and statements of data and code
availability are available at https://doi.org/10.1038/s41586-020-2982-5.
1. Okereke, C. & Coventry, P. Climate justice and the international regime: before, during,
and after Paris: climate justice and the international regime. Wiley Interdiscip. Rev. Clim.
Change 7, 834–851 (2016).
Nature | Vol 588 | 10 December 2020 | 265
Article
2. Tavoni, M. et al. Post-2020 climate agreements in the major economies assessed in the 25.
light of global models. Nat. Clim. Change 5,119–126 (2015).
3. Tavoni, M. et al. The distribution of the major economies’ effort in the Durban Platform 26.
scenarios. Clim. Change Econ. 04, 1340009 (2013).
4. Leimbach, M. & Giannousakis, A. Burden sharing of climate change mitigation: global and 27.
regional challenges under shared socio-economic pathways. Climatic Change 155, 28.
273–291 (2019).
5. Lüken, M. et al. The role of technological availability for the distributive impacts of 29.
climate change mitigation policy. Energy Policy 39, 6030–6039 (2011).
6. Aldy, J. E., Krupnick, A. J., Newell, R. G., Parry, I. W. H. & Pizer, W. A. Designing climate 30.
mitigation policy. J. Econ. Lit. 48, 903–934 (2010).
7. Victor, V. The Collapse of the Kyoto Protocol and the Struggle to Slow Global Warming 31.
(Princeton Univ. Press, 2001).
8. González-Eguino, M., Capellán-Pérez, I., Arto, I., Ansuategi, A. & Markandya, A. Industrial
and terrestrial carbon leakage under climate policy fragmentation. Clim. Policy 17, 32.
S148–S169 (2017).
9. Otto, S. A. C. et al. Impact of fragmented emission reduction regimes on the energy 33.
market and on CO2 emissions related to land use: a case study with China and the
European Union as first movers. Technol. Forecast. Soc. Change 90, 220–229 (2015). 34.
10. Böhringer, C. & Welsch, H. Burden sharing in a greenhouse: egalitarianism and
sovereignty reconciled. Appl. Econ. 38, 981–996 (2006). 35.
11. Nordhaus, W. Climate clubs: overcoming free-riding in international climate policy.
Am. Econ. Rev. 105, 1339–1370 (2015). 36.
12. Csereklyei, Z. & Stern, D. I. Global energy use: decoupling or convergence? Energy Econ.
51, 633–641 (2015).
13. International Comparison Program Purchasing Power Parities and Real Expenditures of 37.
World Economies: Summary of Results and Findings of the 2011 International Comparison
Program (World Bank, 2014). 38.
14. Stern, D. I., Pezzey, J. C. V. & Lambie, N. R. Where in the world is it cheapest to cut carbon
emissions? Aust. J. Agric. Resour. Econ. 56, 315–331 (2012). 39.
15. Fujimori, S. et al. Will international emissions trading help achieve the objectives of the
Paris Agreement? Environ. Res. Lett. 11, 104001 (2016). 40.
16. Weyant, J. P. & Hill, J. Introduction and overview. The costs of the Kyoto Protocol: a
multi-model evaluation. Energy J. (Spec. Issue) vii–xliv (1999). 41.
17. Zhou, P. & Wang, M. Carbon dioxide emissions allocation: a review. Ecol. Econ. 125, 47–59
(2016). 42.
18. van den Berg, N. J. et al. Implications of various effort-sharing approaches for national
carbon budgets and emission pathways. Climatic Change https://doi.org/10.1007/s10584- 43.
019-02368-y (2019).
19. Höhne, N., den Elzen, M. & Escalante, D. Regional GHG reduction targets based on effort 44.
sharing: a comparison of studies. Clim. Policy 14, 122–147 (2014).
20. Manne, A. S. & Stephan, G. Global climate change and the equity–efficiency puzzle. 45.
Energy 30, 2525–2536 (2005).
21. Kriegler, E. et al. Making or breaking climate targets: the AMPERE study on staged 46.
accession scenarios for climate policy. Technol. Forecast. Soc. Change 90, 24–44 (2015).
22. Aldy, J. et al. Economic tools to promote transparency and comparability in the Paris 47.
Agreement. Nat. Clim. Change 6, 1000–1004 (2016).
23. Jacoby, H. D., Chen, Y.-H. H. & Flannery, B. P. Informing transparency in the Paris
Agreement: the role of economic models. Clim. Policy 17, 873–890 (2017). Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in
24. Vandyck, T., Keramidas, K., Saveyn, B., Kitous, A. & Vrontisi, Z. A global stocktake of the published maps and institutional affiliations.
Paris pledges: implications for energy systems and economy. Glob. Environ. Change 41,
46–63 (2016). © The Author(s), under exclusive licence to Springer Nature Limited 2020
266 | Nature | Vol 588 | 10 December 2020
Vrontisi, Z. et al. Enhancing global climate policy ambition towards a 1.5 °C stabilization:
a short-term multi-model assessment. Environ. Res. Lett. 13, 044039 (2018).
Rogelj, J. et al. Paris Agreement climate proposals need a boost to keep warming well
below 2 °C. Nature 534, 631–639 (2016).
The Emissions Gap Report 2019 (UNEP, 2019).
Luderer, G. et al. Residual fossil CO2 emissions in 1.5–2 °C pathways. Nat. Clim. Change 8,
626–633 (2018).
Kriegler, E. et al. Fossil-fueled development (SSP5): an energy and resource intensive
scenario for the 21st century. Glob. Environ. Change 42, 297–315 (2017).
Dellink, R., Chateau, J., Lanzi, E. & Magné, B. Long-term economic growth projections in
the shared socioeconomic pathways. Glob. Environ. Change 42, 200–214 (2017).
Rogelj, J., Forster, P. M., Kriegler, E., Smith, C. J. & Séférian, R. Estimating and tracking the
remaining carbon budget for stringent climate targets. Nature 571, 335–342 (2019);
correction 580, E4 (2020).
Luderer, G. et al. Economic mitigation challenges: how further delay closes the door for
achieving climate targets. Environ. Res. Lett. 8, 034033 (2013).
Bauer, N. et al. Shared socio-economic pathways of the energy sector—quantifying the
narratives. Glob. Environ. Change 42, 316–330 (2017).
Diffenbaugh, N. S. & Burke, M. Global warming has increased global economic inequality.
Proc. Natl Acad. Sci. USA 116, 9808–9813 (2019).
Burke, M., Hsiang, S. M. & Miguel, E. Global non-linear effect of temperature on economic
production. Nature 527, 235–239 (2015).
De Cian, E., Hof, A. F., Marangoni, G., Tavoni, M. & van Vuuren, D. P. Alleviating inequality
in climate policy costs: an integrated perspective on mitigation, damage and adaptation.
Environ. Res. Lett. 11, 074015 (2016).
Evans, D. J. & Sezer, H. Social discount rates for member countries of the European Union.
J. Econ. Stud. 32, 47–59 (2005).
Weitzman, M. L. Can negotiating a uniform carbon price help to internalize the global
warming externality? J. Assoc. Environ. Resour. Econ. 1, 29–49 (2014).
Barrett, S. in Conflicts and Cooperation in Managing Environmental Resources (ed. Pethig,
R.) 11–35 (Springer, 1991).
Carraro, C. & Siniscalco, D. in The Economics of Sustainable Development (eds Goldin, I. &
Winters, L. A.) 264–288 (Cambridge Univ. Press, 1995).
Kornek, U. & Edenhofer, O. The strategic dimension of financing global public goods.
Eur. Econ. Rev. 127, 103423 (2020).
Lazarus, M. & van Asselt, H. Fossil fuel supply and climate policy: exploring the road less
taken. Climatic Change 150, 1–13 (2018).
Canadell, J. G. & Raupach, M. R. Managing forests for climate change mitigation. Science
320, 1456–1457 (2008).
Glachant, M. & Dechezleprêtre, A. What role for climate negotiations on technology
transfer? Clim. Policy 17, 962–981 (2017).
Schultes, A. et al. Optimal international technology cooperation for the low-carbon
transformation. Clim. Policy 18, 1165–1176 (2018).
Paroussos, L. et al. Climate clubs and the macro-economic benefits of international
cooperation on climate policy. Nat. Clim. Change 9, 542–546 (2019).
Chichilnisky, G. & Heal, G. Who should abate carbon emissions? An international
viewpoint. Econ. Lett. 44, 443–449 (1994).
Methods
Modelling framework
The analysis is based on an integrated model framework that couples a
model of the energy sector and macroeconomic growth with a land-use
sector model. It covers all GHG emissions including CO2 emissions from
energy conversion, industry and land-use change, which are limited by
the global carbon budget. The other GHGs are priced based on conver-
sion factors assumed to equal the global warming potentials that the
Intergovernmental Panel on Climate Change (IPCC) Fifth Assessment
Report (AR5) documented.
The REMIND model and the MAgPIE model divide the world into
12 regions (Extended Data Fig. 8). Both models are coupled using a
soft-link approach29. For REMIND, version 2.0 is applied, which is avail-
able open access. For MAgPIE, version 4.1 is used. Both models are open
access on GitHub (see Code availability).
The REMIND model is solved by finding the international prices for
traded goods and energy carriers using a fixed-point iteration based
on a Nash-type algorithm. It is known that this algorithm replicates
the Negishi approach, but is computationally more efficient45. The
CO2 prices are adjusted between each iteration to comply with the
prescribed global carbon budget. The precision for the carbon budget is
better than 1%. For the scenarios in this study, the international spillover
of technology learning is not internalized45.
The MAgPIE model is run subject to regional food demands as well
as regional carbon and non-CO2 GHG emission prices and regional
biomass feedstock production requirements. The present version
includes the carbon removal option of afforestation. It is assumed that
the carbon removals are remunerated with only half the CO2 price in
each region. This discount reflects precautions related to monitoring,
reporting and verification as well as issues related to the permanency
of the integrity of the carbon stores.
Net present values (NPVs) are used for monetary values that are
aggregates of annual values over a time horizon, such as mitigation
costs, income and transfers. For the discounting of these future values,
we apply a 5% annual discount rate.
Implementation of uniform and differentiated carbon policies
In this study, we prescribe carbon prices as taxes that develop over time
and that are recycled on a lump-sum basis domestically. Over time, the
tax paths start after 2020 and increase with an annual growth rate of
5% until 2060. Afterwards, the tax paths continue to increase linearly
until 2100. The flattening of the carbon price trajectories is motivated
by the potentially huge carbon budget overshoots that result if the
exponential growth is continued by the end of the century48,49.
The carbon budgets are not implemented directly as quantity targets
to derive the resulting carbon prices. This would assume a globally
integrated emissions permit market. The implementation of uniform
carbon tax paths and the adjustments to meet the carbon budget is an
approximation to the cost-optimal solution. This approach allows the
implementation of the uniform carbon price policy as well as the poli-
cies with differentiated carbon prices, which are of particular interest in
this study. The quality of the approximation to the social optimal result
becomes weaker if the economic value of the flexibility to temporarily
overshoot the carbon budget increases.
The case with uniform carbon prices is implemented by assum-
ing a parameterized shape of the carbon price trajectory that starts
after 2020. This carbon tax path is shifted up until the endogenous
cumulative CO2 emissions up until 2100 are sufficiently close (<1%)
to the exogenously assumed carbon budget. The transfers necessary
to achieve the equal-effort-sharing criterion are computed ex-post.
This approach assumes that international transfers do not change the
global carbon price and the emissions in each region. This is fulfilled
by the REMIND–MAgPIE model and is the basis of analysis of varying
international carbon permit allocations, that is, the analysis of the
sovereignty–equity trade-off mentioned in the main text2,4,5,10. Fulfilling
the separability property requires the assumption that the transfers do
not change the carbon intensity of the goods bundles that are produced
in donor countries and transferred to recipient countries. Hence, given
income reductions ΔYr in region r compared with its baseline income
Yr, the equal-effort-sharing criterion is implemented by choosing a set
of transfers Tr such that the equal-effort-sharing criterion is fulfilled,
which equals the global relative income loss on the right-hand side:
ΔYr − Tr ∑ r ΔYr
= , ∀ r.
Yr ∑ r Yr
For the case of regionally differentiated carbon taxes without trans-
fers (that is, Tr = 0), the carbon price trajectories are varied (1) for each
region individually to fulfil the equal-effort-sharing criterion for all
regions and (2) the global set of carbon prices is adjusted to comply
with the global carbon budget. These two adjustments are imple-
mented sequentially. First, for each region, the carbon price trajec-
tory is increased (decreased), if the regional effort is smaller (larger)
than the global effort, measured as the mitigation costs relative to Yr.
Second, this updated set of carbon prices is shifted upwards (down-
wards), if the endogenous cumulative emissions are below (above) the
exogenously assumed carbon budget. Both adjustments are gradual
in each iteration step so that the fixed-point iteration converges in a
computationally efficient and robust.
For the derivation of the trade-off curve, the starting point is the set
of regional carbon prices p ∼ with full differentiation and no transfers.
r
The derivation requires a series of model runs with adjusted carbon
prices and derivation of residual transfers. Each model run comprises
three steps (see also Extended Data Fig. 2).
First, the spread of carbon prices in 2030 is compressed by the fol-
lowing function:
∼ α
∼min  pr 
pr = p
 ∼min 
p 
For each region, the carbon price is normalized to the smallest
regional carbon price p ∼min. This is the carbon price spread shown in
Fig. 2a. The exponent α is varied between zero and one, which com-
presses the carbon price spread for each region. The multiplication of
the compressed ratio (p ∼ /p
∼min)α with p∼min gives the updated carbon
r
price pr. Applying this formula to all regional carbon prices results in
the compressed set of carbon prices. Varying α leads to different degree
of compression. When α = 1, the compression is neutral, thus pr = p ∼.
r
When α = 0, the compression is complete and the result is the uniform
carbon price case.
Second, the set of compressed regional carbon prices is jointly
rescaled in a new REMIND–MAgPIE run so that the global cumulative
CO2 emissions comply with the exogenous carbon budget. This step is
necessary because the set of regional carbon prices from the first step
does not comply with the carbon budget. The adjustment of the carbon
price set maintains the price spread, but adjusts the overall level so that
the CO2 emissions comply with the carbon budget.
Finally, the residual transfers are computed based on this result so
that the equal-effort-sharing criterion is fulfilled.
The trade-off curve derived with this methodology does not rep-
resent a frontier in the space of global transfers and mitigation costs.
Such a frontier associates to every global transfer volume the mini-
mal additional global mitigation costs, while observing the carbon
budget and the equal-effort-sharing criterion. It cannot be excluded
that an improved method allows the identification of combinations
that allow a stronger reduction of transfers for each increase of global
mitigation costs; that is, points below the trade-off curve in Fig. 3a
cannot be excluded. Although such a frontier exists, its numerical
derivation is computationally very demanding. The reason is that
for each transfer level, all regional price trajectories (here 12) need
Article
to be varied independently to find the minimal additional mitigation
costs and simultaneously the carbon budget, the equal-effort-sharing
criterion as well as the global transfer volume are held constant.
The compression function avoids the independent price variations and
is therefore computationally more efficient, but only approximates
the frontier. Alternative compression functions perform differently.
We identified the exponential function performing best. Extended
Data Fig. 7d shows the comparison of the exponential with a linear
compression function. The linear function is inferior because the result-
ing trade-off curve is located above the trade-off curve derived with
the exponential function. This means that for every transfer level, the
exponential compression function derives a set of regional carbon
prices that leads to smaller additional mitigation costs.
The differentiation of the carbon prices to fit with the equal-effort-
sharing criterion could call into question the uniqueness of the price
vector that solves the problem. Given that the solution of uniform
carbon prices is unique, we do not identify a convincing argument that
the differentiated carbon price vector fulfilling the equal-effort-sharing
criterion is not unique. The numerical solution behaviour and the
sensitivity analysis did support this finding. Advanced future research
into the analytics of the uniqueness of differentiated price vectors
could lead to new and more general insights.
Analytical treatment of trade-off curve
The nonlinearity of the trade-off curve follows directly from the neces-
sary condition for the optimality of the uniform price solution and the
shape of the marginal abatement cost (MAC) functions that are consistent
with standard assumptions in models of environmental economics. The
basic argument is that around the cost-optimal solution, the distribu-
tional aspect (orange rectangles in Extended Data Fig. 1) is substantially
larger than the efficiency aspect (red triangles in Extended Data Fig. 1).
The REMIND–MAgPIE model does not explicitly comprise MACs,
but they can be derived using a series of model runs that vary carbon
prices to derive carbon emissions. That the conditions for the implicit
MACs are fulfilled has been published in the literature using REMIND–
MAgPIE and similar investigations have been performed with other
models as well32,50.
Research note on transfers
The transfers mentioned and discussed in the main text are computed
as NPVs over the time horizon 2020–2100. There is no explicit state-
ment about the timing of the transfer payments. This requires addi-
tional assumptions. Here we investigate the transfers over time in the
uniform carbon price case assuming that all regions experience the
same relative effort in each period. Extended Data Fig. 5 summarizes
the results for the OECD and the non-OECD regions. Note that differ-
ences in the percentage figures in both regions are due to differences
in gross domestic product (GDP) development.
The transfers from the OECD region increase rapidly over time and
stabilize at around 1% GDP in 2050. The reverse flow in 2020 is due to the
NDCs that lead to higher mitigation costs in the OECD region than in the
non-OECD region. This reverse financial flow considers only the actual
emissions limitations according to the NDCs, but does not account for
differences in historic emissions. Given the assumption of equal effort
in all periods and regions, the NDCs imply that the region of developing
and emerging countries needs to be the donor. Starting in 2025, the
uniform carbon pricing policy is applied, which leads to the expected
direction transfers from the OECD region to the non-OECD region.
For orientation purposes, we also added the line of 0.22% to the
Extended Data Fig. 5, which is the percentage share of GDP in 2020 con-
sidering the US$100 billion target by 2020 pledged by donor countries
after negotiations over the course of various Conferences of the Parties
under the United Nations Framework Convention on Climate Change.
The US$100 billion pledge was first introduced in the Copenhagen
Accord of 2009. Although this pledge was somewhat more clearly
defined in the 2010 Cancun Agreement, it still is not a precise com-
parison owing to open questions regarding definitions, accounting,
additionality, finance origin and the objectives of the available funds51,52.
The controversies about these open questions make it very difficult
to give a concluding statement on whether OECD countries are on
track to fulfil this pledge by 2020, but these controversies all the more
strengthen the argument that achieving the high rates of transfers
required in future decades with uniform carbon prices are a serious
hurdle for implementing such an international policy architecture. To
be sure: the numbers shown in Extended Data Fig. 5 refer only to trans-
fers to equate the burden sharing with respect to mitigation costs, so
financing for adaptation and loss and damage (itself a very thorny and
controversially debated topic)53 would need to come on top. As another
point of reference, OECD countries since 1970 also have a target of con-
tributing 0.7% of their GDP to official development assistance; though
here again, the accounting and track record is ambiguous, and further
complicated by the overlap with climate finance52.
Historic and scenario data
Figure 1a–d uses data from the World Bank’s World Development Indi-
cators and BP’s World Energy Statistics54,55.
Figure 1e, f uses data provided by the IPCC’s latest reports in the
scenario databases. These databases in turn rely on scenarios that
have been derived in various projects that applied a broad set of IAM
models under different boundary conditions and policies. Extended
Data Table 1 summarizes the selection of scenarios that have been used
for the analysis. It is worth noting that the databases include more sce-
narios for these projects, but these are usually only sensitivity cases.
Including the sensitivity cases could lead to biased results that arise
only from the different project designs and their number of sensitivity
cases. The scenarios were provided by the following projects: EMF2756,
AMPERE57, RoSE58, ADVANCE28, SSP33 CD-LINKS59,60, EMF3361 and GEA62.
The relative mitigation costs are derived as differences between a
policy case and the corresponding no policy baseline case. The differ-
ent models shown in Fig. 1f report different metrics of mitigation costs.
Where possible, we used differences in GDP. If this indicator was not
available, we used additional energy system costs or the area under
the marginal abatement cost curve63.
Data availability
Scenario data have been uploaded in Zenodo with the identifier https://
doi.org/10.5281/zenodo.4010426. Source data are provided with this
paper.
Code availability
The model codes of REMIND (identifier: ab2c995116e7fb402f6d-
d0183724496373af996e) and MAgPIE (identifier: 950bc7a08fd0e6c8f-
790c1399c7837133233e2fc) are open source (https://github.com/
remindmodel/remind and https://github.com/magpiemodel/magpie).
48. Obersteiner, M. et al. How to spend a dwindling greenhouse gas budget. Nat. Clim.
Change 8, 7–10 (2018).
49. Realmonte, G. et al. An inter-model assessment of the role of direct air capture in deep
mitigation pathways. Nat. Commun. 10, 3277 (2019).
50. Azar, C., Johansson, D. J. A. & Mattsson, N. Meeting global temperature targets—the role
of bioenergy with carbon capture and storage. Environ. Res. Lett. 8, 034004 (2013).
51. Lundsgaarde, E., Dupuy, K. & Persson, A. Coordination Challenges in Climate Finance
(Danish Institute for International Studies, 2018); https://www.econstor.eu/
bitstream/10419/204624/1/1042180393.pdf
52. Motty, M. & Ackom, E. K. in Climate Action (eds Leal Filho, W. et al.) 1–11 (Springer
International Publishing, 2019); https://doi.org/10.1007/978-3-319-71063-1_104-2
53. Sharma, A. Precaution and post-caution in the Paris Agreement: adaptation, loss and
damage and finance. Clim. Policy 17, 33–47 (2017).
54. BP Statistical Review of World Energy 2020 (BP, 2020); https://www.bp.com/en/global/
corporate/energy-economics/statistical-review-of-world-energy/downloads.html
55. World Development Indicators, DataBank (World Bank, 2020); https://databank.
worldbank.org/source/world-development-indicators#
56. Kriegler, E. et al. The role of technology for achieving climate policy objectives: overview
of the EMF 27 study on global technology and climate policy strategies. Climatic Change
123, 353–367 (2014).
57. Riahi, K. et al. Locked into Copenhagen pledges—implications of short-term emission
targets for the cost and feasibility of long-term climate goals. Technol. Forecast. Soc.
Change 90, 8–23 (2015).
58. Kriegler, E. et al. Will economic growth and fossil fuel scarcity help or hinder climate
stabilization? Climatic Change 136, 7–22 (2016).
59. Roelfsema, M. et al. Taking stock of national climate policies to evaluate implementation
of the Paris Agreement. Nat. Commun. 11, 2096 (2020).
60. McCollum, D. L. et al. Energy investment needs for fulfilling the Paris Agreement and
achieving the Sustainable Development Goals. Nat. Energy 3, 589–599 (2018); correction
3, 699 (2018).
61. Bauer, N. et al. Global energy sector emission reductions and bioenergy use: overview of
the bioenergy demand phase of the EMF-33 model comparison. Climatic Change https://
doi.org/10.1007/s10584-018-2226-y (2018).
62. Riahi, K. et al. in Global Energy Assessment—Toward a Sustainable Future 1203–1306
(Cambridge Univ. Press, 2012).
63. Krey, V. et al. in IPCC Climate Change 2014: Mitigation of Climate Change (eds Edenhofer,
O. et al.) 1281–1328 (Cambridge Univ. Press, 2014).
64. Huppmann, D., Rogelj, J., Krey, V., Kriegler, E. & Riahi, K. A new scenario resource for
integrated 1.5 °C research. Nat. Clim. Change 8, 1027–1030 (2018)
65. KC, S. & Lutz, W. The human core of the shared socioeconomic pathways: population
scenarios by age, sex and level of education for all countries to 2100. Glob. Environ.
Change 42, 181–192 (2017).
66. South, A. rworldmap: a new R package for mapping global data. The R Journal 3, 35–43
(2011).
67. Bauer, N. et al. Bio-energy and CO2 emission reductions: an integrated land-use and
energy sector perspective. Clim. Change https://doi.org/10.1007/s10584-020-02895-z
(2020).
Acknowledgements We acknowledge funding from the German Federal Ministry of
Education and Research (BMBF) in the Funding Priority ‘Economics of Climate Change’
(DIPOL: 01LA1809A). This work was supported by the European Union’s Horizon 2020
research and innovation programme under grant agreement numbers 730403
and 821124.
Author contributions N.B., C.B., D.K. and A.S. developed the policy analysis framework and
designed the experiments. N.B. and A.S. implemented the policy framework. N.B. wrote the
manuscript (with assistance from C.B. and G.L.). N.B. led the analysis and writing of the
manuscript with contributions from all authors.
Competing interests The authors declare no competing interests.
Additional information
Supplementary information is available for this paper at https://doi.org/10.1038/s41586-020-
2982-5.
Correspondence and requests for materials should be addressed to N.B.
Peer review information Nature thanks Hancheng Dai, Ritu Mathur, Wei Peng and the other,
anonymous, reviewer(s) for their contribution to the peer review of this work. Peer reviewer
reports are available.
Reprints and permissions information is available at http://www.nature.com/reprints.
Article
Extended Data Fig. 1 | Graphical illustration of the distributional effects
between advanced economy A and developing economy B for different
policy frameworks characterized by different marginal abatement cost
functions fA and fB . The case ‘Uniform price w/o transfers’ with carbon prices
equal to p in both regions implies different mitigation costs. In the case
‘Uniform price w/ transfers’ these differences are neutralized by transfers T.
Alternatively, in the case ‘No transfer’ the differentiation of policies leads to
equal costs; this case is not depicted explicitly. The case ‘Hybrid’ differentiates
carbon prices to reduce T. The change of global mitigation cost is only the
difference between the regional mitigation costs ΔAC a − ΔACb (indicated by the
red triangles), whereas the changes of transfers are represented by the orange
rectangles pΔR. As long as the differentiation of prices is relatively small, the
decline of transfers exceeds the increase in global mitigation costs.
Extended Data Fig. 2 | Illustration of the exponential compression regions and the light grey line highlights the compression for Latin America
function. The x axis shows the 2030 carbon prices in the full differentiation and the EU. We note that the set of compressed carbon prices is scaled to
case (see Fig. 2a). The example applies the parameter α = 0.5 to the compression comply with the global carbon budget (that is, the relative differences of the
∼min(p
function pr = p ∼ /p
∼min )α that has been introduced in the Methods. The y axis carbon price spread remain constant).
r
shows the carbon prices after compression. The figure shows a subset of
Article

Extended Data Fig. 3 | Socioeconomic drivers and CO2 emissions in the c, d, Energy and industry CO2 emissions (c) as well as total CO2 emissions (d).
no-policy baseline scenario. a, b, Population (a) and economic growth (b) See ‘Data availability’ section for more details.
from history 1990–2015 and in the SSP2 baseline scenario 2015–210030,65.
Extended Data Fig. 4 | Time path of transfers in the default case with
uniform carbon prices. The transfers are expressed as percentages of GDP in
the OECD and the non-OECD regions. The dashed line serves as a point of
orientation. It represents the share of the US$100 billion relative to the OECD’s
GDP in 2020.
Article
Extended Data Fig. 5 | Effect of carbon price differentiation on primary and
final energy use. a, b, Changes in the global energy mixes distinguished by
energy carriers and regions. The figure depicts differences compared with the
uniform carbon tax case for the 1,300 GtCO2 carbon budget. Primary energy (a)
is measured according to the direct equivalence principle; final energy (b) is
measured as delivered to final consumer. This means that fossil fuels, biomass
and geothermal energy are measured in primary energy input, whereas
renewables (hydro, wind and solar) as well as nuclear energy are measured by
their electricity output. Notable results are as follows. First, the total amount of
energy use increases. Second, OECD countries mostly reduce residual oil and
gas consumption, but non-OECD countries mostly increase the use of coal;
therefore, the total consumption of coal increases, whereas the global use of oil
and gas decreases. Third, the total use of biomass increases due to increasing
demand in OCED countries. Fourth, OECD countries accelerate modernization
of final energy use by mainly reducing the use of liquids and gases, but
increasing electricity and hydrogen. Finally, non-OECD countries delay
modernization of energy use by mainly increasing the use of solids, liquids and
gases with corresponding implications for air pollution and so on.
Extended Data Fig. 6 | Effect of carbon price differentiation on land-use
change and investments. a, b, Changes in global land use (a) and investment
and regions (b). The figure depicts differences compared with the uniform
carbon tax case for the 1,300 GtCO2 carbon budget. The two regions show
opposite changes in the variables; for example, OECD countries convert
agricultural land into forests to remove carbon by afforestation, whereas
non-OECD countries convert forests into cropland to grow biomass that is
exported to OECD countries. Moreover, OECD countries increase investments
in the energy sector substantially to facilitate the transition to low-carbon
technologies and to invest into carbon-removal technologies (which are
counted as part of the energy sector). The higher OECD investments crowd-out
the macroeconomic investments in OECD countries and energy sector
investments in non-OECD countries.
Article

Extended Data Fig. 7 | Sensitivity analysis of the trade-off curve. a, The technologies that rely on underground geological storage of CO2 (that is, CCS
sensitivity of the ban on bioenergy trade. b, The sensitivity with respect to the including direct air capture). d, The application of the linear compression
carbon budget. c, The variation of the maximum annual retirement rate of function; in this sensitivity analysis the case of uniform taxes and the
fossil-fuelled infrastructure from 9% to 6% and the delayed availability of no-transfer case are identical.
Extended Data Fig. 8 | Regional aggregates used in REMIND and MAgPIE. Regions and countries belonging to the OECD region are coloured in blue tones.
See ‘Data availability’ section for more details. World map based on rworldmap package66.
Article
Extended Data Table 1 | Overview of data sources for model comparisons

Scenario data in Fig. 1e, f were taken from the IPCC databases for the Fifth Assessment Report (AR5) and the Special Report on 1.5 °C warming (SR15)64. The carbon budgets are computed for the
period 2011–2100. Figure 1e, f shows relative differences between the ‘Policy case’ and the ‘Base case’.
Extended Data Table 2 | Overview of selected sensitivity cases

The mitigation costs and transfers are NPVs applying a 5% annual discount rate. The inefficiency is the difference between the mitigation costs in the differentiated and the uniform case. The
bold numbers are quoted in the text. The carbon prices in the default case are shown in Fig. 2a. The income growth sensitivity case uses per-capita GDP and population projections of SSP1 (but
the demands for final energy and food are the same as in the default case, which relies on SSP2). Note that in case of faster income growth and convergence (right-hand column), the carbon
price in the uniform case as well as the absolute mitigation costs increase, but the ratio of transfers to mitigation costs decreases. If the metric of consumption losses is used to measure effort,
the losses of fossil fuel-rich countries are more severe in the uniform price case and, consequently, in the differentiated case the carbon prices in these countries are lower.
Article

Clustered versus catastrophic global

vertebrate declines

https://doi.org/10.1038/s41586-020-2920-6 Brian Leung1,2 ✉, Anna L. Hargreaves1, Dan A. Greenberg3, Brian McGill4,5, Maria Dornelas6
& Robin Freeman7
Received: 28 January 2020

Accepted: 4 September 2020

Recent analyses have reported catastrophic global declines in vertebrate populations1,2.
Published online: 18 November 2020
However, the distillation of many trends into a global mean index obscures the
Check for updates
variation that can inform conservation measures and can be sensitive to analytical
decisions. For example, previous analyses have estimated a mean vertebrate decline
of more than 50% since 1970 (Living Planet Index2). Here we show, however, that this
estimate is driven by less than 3% of vertebrate populations; if these extremely
declining populations are excluded, the global trend switches to an increase. The
sensitivity of global mean trends to outliers suggests that more informative indices
are needed. We propose an alternative approach, which identifies clusters of extreme
decline (or increase) that differ statistically from the majority of population trends. We
show that, of taxonomic–geographic systems in the Living Planet Index, 16 systems
contain clusters of extreme decline (comprising around 1% of populations; these
extreme declines occur disproportionately in larger animals) and 7 contain extreme
increases (around 0.4% of populations). The remaining 98.6% of populations across
all systems showed no mean global trend. However, when analysed separately, three
systems were declining strongly with high certainty (all in the Indo-Pacific region) and
seven were declining strongly but with less certainty (mostly reptile and amphibian
groups). Accounting for extreme clusters fundamentally alters the interpretation of
global vertebrate trends and should be used to help to prioritize conservation efforts.

Rapid global change is threatening species across the globe1. The quan- one population declined by 99%. Even if a second population increased
tification of biodiversity trends is important to assess whether current 50-fold or 393 populations increased by 1% (that is, a large net increase),
investment is slowing or reversing declines, and to identify regions a geometric mean would show a catastrophic 50% decline. Thus, a
and taxa of concern. Although distilling disparate population trends geometric mean decline of 50% could arise from substantial, wide-
into a single global index can focus attention on biodiversity trends2–4, spread loss that is occurring across many populations (we term this the
simple metrics can distort the full picture. ‘catastrophic declines’ hypothesis) or from a few extremely declining
Estimates of global biodiversity trends vary depending on their populations (we term this the ‘clustered declines’ hypothesis). Both
data and mathematical model. The most apocalyptic models gather scenarios involve important conservation issues, but suggest vastly
extensive press coverage, even when based on controversial data (for different underlying problems and require different mitigation strate-
example, ‘biological annihilation’5, which described trend estimates gies14, thus distinguishing between them is of real-world importance.
based largely on expert opinion; or ‘insect Armageddon’, which is based We derive a Bayesian hierarchical mixture (BHM) model to distin-
on data disputed by the original collectors6). However, even analyses of guish between the catastrophic and clustered declines hypotheses. The
the best available data reach conflicting conclusions. An analysis of a model statistically separates population trends into extreme declines,
global dataset of abundance time series of vertebrates estimated that, typical trends and extreme increases (Fig. 1), while accounting for
on average, vertebrate populations have declined by more than 50% time-series size, within-population fluctuations, number of popula-
since 1970 (Living Planet Index2 (LPI)); however, other global analyses tions and among-population variance. We test declines in abundance
found that the mean population size7,8 and species richness9,10 have for more than 14,000 vertebrate populations (from the LPI)15. We chose
remained stable over similar timeframes. Explanations for the discrep- LPI data because of its large scope, because the data and analytical
ancies have been proposed8,11–13, but not resolved. details were publicly available, and because previous analyses of these
One crucial consideration is that summary indices may be easily data suggested widespread, global declines2.
misinterpreted. Calculating the geometric mean across populations We first examined whether the previous estimate2 of a mean decline
is the most common and straightforward approach, but is strongly of more than 50% was sensitive to extreme populations: robust declines
influenced by extremes. To illustrate, imagine an ecosystem in which would support the catastrophic declines hypothesis, whereas high
1
Department of Biology, McGill University, Montreal, Quebec, Canada. 2Bieler School of Environment, McGill University, Montreal, Quebec, Canada. 3Department of Biological Sciences, Simon
Fraser University, Burnaby, British Columbia, Canada. 4School of Biology and Ecology, University of Maine, Orono, ME, USA. 5Mitchell Center for Sustainability Solutions, University of Maine,
Orono, ME, USA. 6Centre for Biological Diversity, University of St Andrews, St Andrews, UK. 7Indicators and Assessments Unit, Institute of Zoology, Zoological Society of London, London, UK.
✉e-mail: brian.leung2@mcgill.ca

Nature | Vol 588 | 10 December 2020 | 267

Article
a All populations (n = 14,700)
–0.1 0 0.1
b 1.0
–0.1 0 0.1
c
–0.1 0 0.1
d
–0.1 0 0.1
e
–0.1 0 0.1
log(mean growth rate)
Fig. 1 | Stylized patterns of system-wide growth rates. a–e, Similar geometric
mean population growth rates (log(N t + 1/N t)) can reflect contrasting systems.
c, As a null model, systems can be stable (log-transformed growth rates centred
around zero). Deviations can occur in multiple ways. a, b, Most populations in a
system can be in substantial decline (catastrophic declines hypothesis) (a) or
the system can have multiple clusters, in which the majority of populations
show a distribution of growth rates centred around zero but with a small cluster
of populations experiencing extreme declines (clustered declines hypothesis)
(b). Each has the same metric of mean decline (vertical red line indicates a 1.5%
annual decline, corresponding to a 50% loss over 50 years), even though most
populations in b are stable. The converse can also happen; systems in which a
small cluster of populations shows an extreme increase, but that show an
otherwise stable distribution (d) or systems in which most populations
increase (e) can also occur (vertical blue line indicates a 1.5% annual increase,
corresponding to a doubling over 50 years).
sensitivity to a few populations would support the clustered declines
hypothesis (Fig. 1). We then applied our BHM model to assess the evi-
dence for catastrophic or clustered declines globally and by region and
taxonomy. Finally, we explore two additional conservation issues. First,
we test whether declines occur disproportionately in larger animals
(large animals tend to have lower reproductive rates), which might
release small animals from predation16. Second, previous analyses
often excluded time series with few data points10,12,17, but small time
series make up most of the available data. We test the effects of their
exclusion18.
Sensitivity of geometric mean to extreme populations
The geometric mean index that underlies the LPI analysis was highly sen-
sitive to extreme populations. Excluding only the 2.4% most-strongly
declining populations (354 out of 14,700 populations) reversed the
estimate of global vertebrate trends from a loss of more than 50%
to a slightly positive growth (Fig. 2). Similarly, excluding 2.4% of the
most-strongly increasing populations strengthened the mean decline
to 71%. High sensitivity suggests that extreme populations are dispro-
portionately affecting global trend estimates, such that clusters of
extreme population decline should be considered explicitly.
268 | Nature | Vol 588 | 10 December 2020
238 populations removed
120 populations removed 356 populations removed
Geometric growth index
0.8
0.6
1970 1980 1990 2000 2010
Year
Fig. 2 | Effect of extreme populations on the global growth index. Removing
a small fraction of extreme populations strongly influences the geometric
growth index, using the LPI dataset. Each line represents a different number of
removed populations, ranging from no removals (red line; all 14,700
populations, which show a >50% mean decline) to removing 356 populations
(yellow line, the removal of <2.4% of populations switches the global trend from
negative to positive). A geometric growth index of 1 indicates no change
(dashed horizontal black line).
Evidence for clustered declines
Among the 57 domain–realm–taxon systems of the LPI, 16 systems con-
tained clusters of extreme decline and 8 contained clusters of extreme
growth (of those, 3 systems are repeated, as they had both clusters of
extreme decline and growth) (Fig. 3 and Supplementary Table 2). Together,
clusters of extreme decline accounted for only 1% of populations across
systems (2% of populations in the 16 systems in which they occurred).
The mean population trend for extremely declining clusters across the 16
systems was θ2 = −3.94, or approximately 98% loss per year, and deviated
substantially from the mean trend of the primary cluster in those systems.
Clusters of extreme growth accounted for 0.4% of populations across
systems (2.4% in the 8 systems in which they occurred), with θ2 = 3.51, that
is, an explosive 33× growth per year (Fig. 3 and Supplementary Table 2).
Extreme clusters showed some taxonomic and geographic patterns.
The largest cluster of extreme declines was in Arctic marine mammals,
accounting for 7.6% of populations in that system. However, mam-
malian systems generally had the fewest clusters of extreme decline
(19% of 16 systems), followed by reptile–amphibian systems (21% of 14
systems), whereas bird and fish systems had more clusters of extreme
declines (31% of 16 and 45% of 11 systems, respectively) (Fig. 3). Clus-
ters of extreme decline occurred throughout the world, half of which
occurred in marine realms, whereas extreme increases occurred more
in temperate regions or terrestrial realms (Fig. 3).
Extreme population trends occurred predominantly in small time
series. Excluding time series with fewer than 10 points not only removed
all but two extreme clusters, but also removed 52% of the data (Sup-
plementary Table 3). The higher frequency of extreme trends among
small time series was also apparent in the raw data (Fig. 4). Thus the
decision of whether to include small time series will have large effects
on the resulting estimates of global trends.
Body size was related to population trends. Larger species had three
times more extreme declines than increases (15 compared with 5 clusters
of extreme declines compared with extreme increases). Comparatively,
smaller species had half as many (8) extremely declining and dispro-
portionately more (7) extremely increasing clusters (Supplementary
 a Palearctic
Nearctic
Neotropical
b Nearctic
Neotropical
c Tropical and subtropical
Indo-Pacific
Atlantic tropical
and subtropical
Fig. 3 | Population trends by taxonomic groups and realms. a, The terrestrial
realm. b,The freshwater realm. c, The marine realm. Red and blue asterisks
indicate the occurrence of extremely declining clusters (16 systems) and
increasing clusters (8 systems), respectively. Distributions show the primary
cluster in each system. Red, significant declines; blue, significant increases;
Table 4). Although size-specific models included fewer populations,
especially for smaller species, the number of clusters was not uniformly
lower (as might be expected given a reduction in power); therefore, the
differential occurrence of extremely declining versus increasing clusters
suggests that large animals are more vulnerable to extreme declines.
Evidence for catastrophic declines
In contrast to the extreme clusters, the primary clusters accounted for
the vast majority (98.6%) of populations across the 57 LPI systems. The
Indo-Malayan
Afrotropical
Palearctic
Afrotropical Indo-Malayan
Atlantic north Arctic
temperate Pacific north
temperate
*
*
South temperate
and Antarctic
orange, strong non-significant declines; green, strong non-significant
increases; yellow, weak changes). Maps were created using ArcGIS software by
Esri (ArcGIS and ArcMap are the intellectual property of Esri and are used
herein under licence. Copyright © Esri. All rights reserved. For more
information about Esri software, please visit https://www.esri.com).
overall growth rate of primary clusters was close to zero: θ1 = −0.00035,
corresponding to around 1.7% loss over 50 years, given a constant
rate across populations and time (Fig. 5). In addition, in contrast to
extreme clusters, primary cluster trends were robust to time-series
size, as excluding series with fewer than 10 data points yielded a similar
overall global trend (θ1 = 0.0043) (Extended Data Fig. 3).
Although the global BHM model reveals considerably more nuance
than a geometric mean index, analysing across systems still masked
important patterns. When systems were analysed separately (Supple-
mentary Table 2), primary population clusters were strongly declining
Nature | Vol 588 | 10 December 2020 | 269
Article
log(mean growth rate)
−3
−6
0 10 20 30
Number of data points in time series
Fig. 4 | Effect of the size of the time series. The number of data points in the
time series versus the mean log-transformed value of the geometric mean
growth rate.
(θ1 < −0.015) with high certainty (95% credible intervals not overlapping
zero) in three systems, all of which occurred in the Indo-Pacific realm
(freshwater mammals, freshwater birds and terrestrial birds) (Fig. 3).
This suggests that this region has the highest risk of system-wide
declines and should be a conservation priority. By contrast, the pri-
mary cluster was increasing with high certainty in seven systems, six
of which were in temperate regions. In addition, seven additional sys-
tems had strongly declining primary population clusters but with less
certainty (95% credible intervals overlapped zero), four of which were
amphibian or reptile groups. Finally, 14 systems showed strong but
low-certainty increases, with no obvious taxonomic nor geographic
patterns (Fig. 3).
Each primary cluster also contained variation among populations.
In the 10 systems with significant or non-significant mean declines
where θ1 < −0.015, 87% of the individual populations showed strong
declines (Fig. 5). These 10 systems accounted for around 20% of the total
global vertebrate populations, but for around 61% of strong declines.
The multimodality observed in Fig. 5 was an outcome of aggregating
unimodal primary clusters across systems, and suggests that there are
heterogeneous stressor levels among systems (that is, similar principles
to those that cause extreme clusters within systems). The remaining
approximately 11% of strongly declining populations were distributed
across 47 out of 57 systems; it is unclear whether they represent a devia-
tion from the natural dynamics that are expected to occur in any natu-
rally variable system.
Primary cluster trends were related to body size, but not as predicted.
In comparison to the overall patterns for larger animals, the same sys-
tems showed significant declines and increases, but two additional
temperate systems showed significant increases (Extended Data Fig. 4
and Supplementary Table 4). Smaller species also appeared to decline
more than larger species; there were 27 systems in which smaller spe-
cies had more-negative growth rates than larger species, compared
with 18 systems in which the reverse was true. However, analyses of
the smaller species were based on substantially fewer populations,
and trends were generally not significant (Supplementary Table 4),
so patterns remain tentative.
270 | Nature | Vol 588 | 10 December 2020
Significant decline
Strong non-significant decline
Weak change or increase
Frequency
–0.1 0 0.1
log(annual growth rate)
Fig. 5 | Populations in the primary clusters across all systems, after removal
of extreme clusters. The primary cluster of each system is unimodal, but
because systems are experiencing decline (or growth) heterogeneously,
plotting distributions across systems shows multimodality. Histograms show
significantly declining systems (red), strongly but not significantly declining
40
systems (orange) and weak changes or increases (yellow). Vertical lines show
thresholds for strongly declining (−0.015) and strongly increasing (+0.015)
growth rates, corresponding to an approximate 50% loss or a doubling (over 50
years), respectively. Distributions of primary clusters were calculated based on
the mean and s.d. from the hierarchical model, and using the system-specific
weights to adjust for species richness.
Discussion
By re-analysing a comprehensive dataset of global wildlife population
trends, we show that previously estimated global declines are driven by
a few extremely declining populations. Removing only 2.4% of declining
populations reversed the estimated global trends from more than 50%
mean decline since 1970 to a slightly positive growth. Our BHM model
revealed that clusters of extreme decline are widespread and occur
disproportionately in larger species, and that a few clusters of extreme
increase also exist and occur disproportionately in smaller species.
This is consistent with previous arguments of ‘trophic downgrading’16.
Clusters of extreme declines were largely due to small time-series
datasets. However, neither random sampling error nor ‘saw tooth’
population dynamics (in which ultimately stable populations experi-
ence sudden declines followed by gradual increases) can fully explain
this association (see Supplementary Information for a full discussion).
Additional explanations are needed. Extreme trends could reflect tran-
sient populations that naturally leave or enter a survey area19, which
could represent natural dynamics. Alternatively, researchers may stop
sampling after populations become (close to) extirpated, although the
converse has also been suggested20. A third possibility is that some
regions experience both lower sampling effort and greater declines,
such that poorly sampled datasets correlate with factors linked to
vulnerability, such as lower national wealth or conservation invest-
ment. Understanding why small time series contain so many extreme
declines is particularly important given that studies that did not find
widespread declines often excluded short time series7,10,12, potentially
reconciling divergent findings among studies.
Once extreme clusters were statistically separated, no global trend
remained across typical populations (that is, primary clusters; 98.6%
of populations). However, aggregating systems into one global trend
hid important variation. Three systems, all of which occurred in the
Indo-Pacific realm, showed widespread vertebrate declines across
typical populations. Moreover, among typical populations smaller
species may be faring worse than larger ones. Although these results
were tentative given lower sample sizes and high uncertainty, this trend
is contrary to common conservation assumptions and so merits addi-
tional research.
Our results emphasize an important point: biodiversity trends within
and across regions and taxa are highly disparate. This probably reflects
differences in both susceptibility and exposure to anthropogenic
environmental change21–23. Unravelling this variation is imperative to
understand in which regions biodiversity is threatened the most24 and
which conservation actions promote stability or recovery. A productive
global conversation about conservation requires that both scientists
and media pay more attention to variation and resist the temptation
of simple summary indices.
Shifting the message from ubiquitous catastrophe to foci of concern,
also touches on human psychology. Continual negative and guilt-ridden
messaging can cause despair, denial and inaction25,26. If everything is
declining everywhere, despite the expansion of conservation measures
in recent decades, it would be easy to lose hope. Our results identify
not only regions that need urgent action to ameliorate widespread
biodiversity declines, but also many systems that appear to be gener-
ally stable or improving, and thus provide a reason to hope that our
actions can make a difference.
Online content
Any methods, additional references, Nature Research reporting sum-
maries, source data, extended data, supplementary information,
acknowledgements, peer review information; details of author con-
tributions and competing interests; and statements of data and code
availability are available at https://doi.org/10.1038/s41586-020-2920-6.
1. IUCN. The IUCN Red List of Threatened Species. version 2019-3 http://www.iucnredlist.
org (2019).
2. WWF. Living Planet Report 2018: Aiming Higher (eds. Grooten, N. & Almond, R. E. A.)
(WWF, 2018).
3. Rosenberg, K. V. et al. Decline of the North American avifauna. Science 366, 120–124
(2019).
4. Sánchez-Bayo, F. & Wyckhuys, K. A. G. Worldwide decline of the entomofauna: a review of
its drivers. Biol. Conserv. 232, 8–27 (2019).
5. Ceballos, G., Ehrlich, P. R. & Dirzo, R. Biological annihilation via the ongoing sixth mass
extinction signaled by vertebrate population losses and declines. Proc. Natl Acad. Sci.
USA 114, E6089–E6096 (2017).
6. Willig, M. R. et al. Populations are not declining and food webs are not collapsing at the
Luquillo Experimental Forest. Proc. Natl Acad. Sci. USA 116, 12143–12144 (2019).
7. Daskalova, G. N., Myers-Smith, I. H. & Godlee, J. L. All is not decline across global
vertebrate populations. Preprint at https://doi.org/10.1101/272898 (2018).
8. Dornelas, M. et al. A balance of winners and losers in the Anthropocene. Ecol. Lett. 22,
847–854 (2019).
9. Vellend, M. et al. Global meta-analysis reveals no net change in local-scale plant
biodiversity over time. Proc. Natl Acad. Sci. USA 110, 19456–19459 (2013).
10. Dornelas, M. et al. Assemblage time series reveal biodiversity change but not systematic
loss. Science 344, 296–299 (2014).
11. Gonzalez, A. et al. Estimating local biodiversity change: a critique of papers claiming no
net loss of local diversity. Ecology 97, 1949–1960 (2016).
12. Leung, B., Greenberg, D. A. & Green, D. M. Trends in mean growth and stability in
temperate vertebrate populations. Divers. Distrib. 23, 1372–1380 (2017).
13. McGill, B. J., Dornelas, M., Gotelli, N. J. & Magurran, A. E. Fifteen forms of biodiversity
trend in the Anthropocene. Trends Ecol. Evol. 30, 104–113 (2015).
14. Anderson, S. C., Branch, T. A., Cooper, A. B. & Dulvy, N. K. Black-swan events in animal
populations. Proc. Natl Acad. Sci. USA 114, 3252–3257 (2017).
15. LPI. Living Planet Index. www.livingplanetindex.org/ (2016).
16. Estes, J. A. et al. Trophic downgrading of planet Earth. Science 333, 301–306 (2011).
17. Connors, B. M., Cooper, A. B., Peterman, R. M. & Dulvy, N. K. The false classification of
extinction risk in noisy environments. Proc. R. Soc. Lond. B 281, 20132935 (2014).
18. Hanks, E. M., Hooten, M. B. & Baker, F. A. Reconciling multiple data sources to improve
accuracy of large-scale prediction of forest disease incidence. Ecol. Appl. 21, 1173–1188
(2011).
19. Youngflesh, C. & Lynch, H. J. Black-swan events: population crashes or temporary
emigration? Proc. Natl Acad. Sci. USA 114, E8953–E8954 (2017).
20. Fournier, A. M. V., White, E. R. & Heard, S. B. Site-selection bias and apparent population
declines in long-term studies. Conserv. Biol. 33, 1370–1379 (2019).
21. Newbold, T. et al. Ecological traits affect the response of tropical forest bird species to
land-use intensity. Proc. R. Soc. Lond. B 280, 20122131 (2013).
22. Venter, O. et al. Sixteen years of change in the global terrestrial human footprint and
implications for biodiversity conservation. Nat. Commun. 7, 12558 (2016).
23. Allan, J. R. et al. Hotspots of human impact on threatened terrestrial vertebrates. PLoS
Biol. 17, e3000158 (2019).
24. Blowes, S. A. et al. The geography of biodiversity change in marine and terrestrial
assemblages. Science 366, 339–345 (2019).
25. O’Neill, S. & Nicholson-Cole, S. “Fear won’t do it”: promoting positive engagement with
climate change through visual and iconic representations Sci. Commun. 30, 355–379 (2009).
26. Brennan, L. & Binney, W. Fear, guilt, and shame appeals in social marketing. J. Bus. Res.
63, 140–146 (2010).
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in
published maps and institutional affiliations.
© The Author(s), under exclusive licence to Springer Nature Limited 2020
Nature | Vol 588 | 10 December 2020 | 271
Article
Methods
Dataset
The publically available LPI dataset includes 15,241 vertebrate
populations from 3,510 species15. When a species contained both
finer-resolution estimates within a country (2,593 entries) and a
country-wide aggregate, we excluded the country-wide aggregate
(537 entries), yielding 14,700 populations. LPI groups species into 57
systems defined by a combination of habitat domain (terrestrial, fresh-
water or marine), biogeographical realm (terrestrial/freshwater realms,
Afrotropical, Nearctic, Neotropical, Palearctic, Indo-Pacific; marine,
Arctic, Atlantic north temperate, Atlantic tropical/sub-tropical, Pacific
north temperate, Indo-Pacific tropical/sub-tropical, South-temperate/
Antarctic) and taxonomic grouping (fish, Actinopterygii, Elasmo-
branchii, Holocephali, Myxini, Chondrichthyes, Sarcopterygii, Ceph-
alaspidomorphi; birds, Aves; mammals, Mammalia; herps, Amphibia,
Reptilia) (Extended Data Figs. 5–8).
To analyse the effect of body size, we obtained information on each
taxonomic group. Given the diversity of vertebrate groups in this
dataset and the different conventions across groups, we used differ-
ent measures of body size for each taxonomic class on the basis of
data availability. For birds (n = 1,397), mammals (n = 534) and reptiles
(Squamata, n = 132; Testudines, n = 44; and Crocodylia, n = 16) we used
estimates of the mass of the species (in grams) collated in an extensive
comparative dataset27. When mass data were missing for a species (n = 14
birds; n = 1 mammal; n = 25 reptiles), we estimated body mass as the geo-
metric mean of available mass estimates for species in that genus. For
fishes (Chondrichthyes, Osteichthyes and Agnatha; n = 1,211), estimates
of mass were scarce for most species, so we instead used estimates of
total length or standard length (in centimetres), both of which were
extracted from FishBase28 using the rfishbase R package29. These length
estimates are an imperfect proxy for size (in terms of mass) given the
variability in body plans across groups, but given the large amount of
variation across these groups it suffices as a way to broadly categorize
species into distinct size classes. For amphibians, we used estimates
of snout–vent length (in millimetres) as our proxy for body size, as
this is the most widely available metric of size across species. Data on
snout–vent length for amphibian species (n = 175) were extracted from
a comprehensive ecological trait dataset: AmphiBio30.
Sensitivity of the geometric indices to extreme population
trends
The LPI analysis was based on a geometric mean approach, calculated
by summing across log-transformed growth rates31. We recreated the
geometric-mean-based analyses (see Supplementary Information 1a
for full details and model formulation) and examined the sensitivity of
the global estimate to extreme populations. We ordered populations
and sequentially removed the largest observed decline, determin-
ing the effect of each removal on the global estimate of biodiversity
loss. Low sensitivity would indicate that many or most populations
are declining, supporting the catastrophic declines hypothesis. High
sensitivity—that is, if removal of relatively few populations switched
the strongly negative global trend to neutral or positive—would sup-
port the clustered declines hypothesis. For balance, we also examined
sensitivity to sequential removal of the greatest increasing populations.
Catastrophic versus clustered declines approach
We developed an approach to separate extreme population clusters the
growth or decline of which statistically deviated from typical popula-
tion trends, such that a small number of extreme populations would
no longer mask trends of the majority of populations (Fig. 1). Although
some summarization is needed to understand global trends, hetero-
geneous growth rates and potentially multimodal distributions could
be expected, given multiple stressors with diverse effects, and differ-
ences in species vulnerabilities. We used a BHM model as our statistical
architecture, as it has several desirable properties: (1) it can represent
the null model and assess deviations from it; (2) it enables testing for
both negative and positive extremes (sometimes both existed in the
same system); (3) it quantifies the magnitude and proportion of those
extremes; (4) it provides a coherent way to separate extreme popula-
tions from the majority of populations (the primary cluster), which
enables tests of the clustered and catastrophic declines hypotheses;
(5) it provides a measure of uncertainty as a direct outcome of analysis
(through the posterior distribution); and (6) it accounts for population
fluctuations and adjusts for the number of data points in the time series.
First, we specify the null model. Even in a system with no overall
trend, we expect stochastic fluctuations in population size. We also
expect some populations to be increasing or decreasing during any
time interval, given complex, real-world ecological dynamics. Thus,
the null model should include among-population heterogeneity, and
therefore consists of a distribution of growth rates (Fig. 1c). Statistical
deviations from this null model could be caused by a shift in the overall
distribution, in which a system-wide mean growth <0 (that is, decline)
could indicate a risk to the entire system, which would support the
catastrophic declines hypothesis (Fig. 1a). Alternatively, statistical
deviation from the null model could be caused by a few populations
that experience extreme declines, which is consistent with the clustered
declines hypothesis (Fig. 1b).
To specify our model, we begin with a standard Bayesian hierarchical
formulation (that is, it does not yet contain mixtures of distributions).
We define θ and τ as the system-wide mean and variance, respectively,
of log-transformed growth rates across all populations in the system
(that is, hyperparameters in Bayesian terminology). θ and τ determine
the distribution of the log-transformed population trends (μi) and
define the properties of the overall system. However, within-population
dynamics are also occurring, and the log-transformed growth rates
for population i at time t are modelled as a population trend (μi) and
within-population fluctuations (σ) (see Supplementary Information
1b for full details and model formulation).
Using a standard Bayesian hierarchical model, we can test the cata-
strophic declines hypothesis by determining the probability that a
system-wide mean value of θ < 0. Testing the clustered declineshypoth-
esis, however, requires a mixture model to assess the evidence for the
occurrence of clusters. Thus, we define K as the number of clusters in
the mixture, fk is the fraction of populations in the kth cluster, and θ, τ
and f denote the vectors of the parameters for the K clusters.
To test the clustered declines hypothesis, we modelled three clus-
ters: a primary cluster, corresponding to the typical trend; a negative
extreme cluster; and a positive extreme cluster (Fig. 1). Although our
main interest was in the mechanisms behind apparent global popula-
tion declines (that is, catastrophic versus clustered declines hypoth-
eses), we also assayed positive extreme clusters so that analyses were
not biased to find only negative population trends. We considered four
cluster combinations: (1) a single distribution; (2) a primary distribution
and a negative extreme distribution; (3) a primary distribution and a
positive extreme distribution; or (4) a primary distribution and both
positive and negative extreme distributions (Fig. 1). For referencing
purposes, we denote k = 1 as the primary cluster, k = 2 as the negative
extreme cluster, and k = 3 as the positive extreme cluster. Reality need
not be bi-modal (or tri-modal), but exploring generalities in trends
necessitates some aggregation. Nonetheless, the extreme clusters
identified by the mixture model could contain multiple extreme modes
in the data (or even result from a skewed distribution). With any of these
deviations, model selection would still choose the mixture model as
explaining the data better than a single normal distribution (see Sup-
plementary Information 1c for full details and model formulation).
We used the (lowest) deviance information criterion value to select
the mixture model with the strongest statistical evidence32. The cata-
strophic declines hypothesis would be supported by a mean decline
of the primary population cluster (θ1 < 0 and credible intervals did not
overlap zero), and would be particularly severe if the mean θ1 was also
strongly negative (for example, θ1 = −0.015 would correspond to >50%
loss over 50 years). The clustered declines hypothesis would be sup-
ported if the deviance information criterion selected a mixture with a
negative extreme cluster (combinations 2 or 4 above). The catastrophic
and clustered declines hypotheses are not mutually exclusive, as a sys-
tem could have both a negative extreme cluster and declining primary
cluster. A large fraction of populations in the negative extreme cluster
(f2) could also be interpreted as widespread catastrophic declines, but
this did not occur in our results. Although our hypotheses focus on
understanding declining trends, our model will also detect increases
in abundances.
To estimate the model parameters, we used Bayesian analyses and
the Markov chain Monte Carlo algorithm, which simultaneously esti-
mated uncertainty. For each Bayesian analysis, we ran 3 chains, each
with 10,000 iterations (3,000 used for burn-in). Convergence was
determined using R̂ ≈ 1. Values for all parameters across all systems
ranged from (0.999 < Rˆ < 1.005). Bayesian analyses were conducted
using the STAN language33, and processed and analysed in R34.
Additionally, we explored the theoretical behaviour of each model,
including the geometric mean model, in the presence of clustered
declines (Supplementary Information 1d, 2a), and our catastrophic and
clustered declines approach given our selection of priors, application
of constraints and other modelling choices; these simulation analyses
showed that our approach yielded appropriate theoretical behaviour
(Extended Data Fig. 1 and Supplementary Information 1e, 2b). Finally,
we conducted sensitivity analyses and showed that results were robust
to modelling choices (Extended Data Fig. 2, Supplementary Informa-
tion 2c and Supplementary Table 1).
Application of the catastrophic and clustered approach to LPI
data
We tested for extreme clusters in each of the 57 domain–realm–taxon
systems of the LPI, by choosing the mixture model with the lowest
deviance information criterion value. We also examined the number of
populations in each cluster, as a fraction of the total number of popula-
tions, scaled using LPI system-specific weightings35 (see Supplementary
Information 1f for more details).
Next, we examined evidence for the catastrophic declines hypothesis
in each system by searching for negative mean growth rates in the pri-
mary cluster (θ1). We defined ‘high certainty’ of decline (or increase) as
95% credible intervals that did not overlap zero, and ‘strong’ decline as
θ1 < −0.015, corresponding to a ~50% decline if it persisted for 50 years
(θ1 > 0.015 was used for a strong positive relations, corresponding to a
doubling over 50 years).
We assessed the effect of small time series on both extreme clusters
and trends in primary clusters, by omitting all data with fewer than
10 points, as has often been done in other studies12. These small time
series accounted for 52% of the population estimates (7,110 populations
remained in the analysis).
Finally, we examined whether trends differed between large- versus
small-bodied animals. Within each class (but with Agnatha lumped
with Osteichthyes), we scaled body size as standard deviations on the
natural log scale—thus creating an index of relative species size within a
taxonomic group. In two cases, we separated out different groups within
a class that had relatively distinct body plans that would influence this
size scaling. We scaled size within the superorder Batoidea (Rajiformes,
Myliobatiformes and Torpediniformes) and separately scaled size for
the rest of the chondrichthyans (Selachimorpha and Holocephali). For
the amphibians, we separated out the orders Caudata and Anura and
scaled size within each of these groups. For each taxonomic group,
we scaled body size and separated species into larger-than-average
(hereafter ‘larger’) versus smaller-than-average (hereafter ‘smaller’)
species. This yielded 9,596 populations from 1,765 larger species, and
5,103 populations from 1,745 smaller species. We then reran the BHM
model for larger animals and again for smaller animals. Body sizes
were divided unevenly among habitat domains and realms; 12 domain–
realm–taxon systems contained ≤1 smaller species so were excluded
from the small-animal model.
Reporting summary
Further information on research design is available in the Nature
Research Reporting Summary linked to this paper.
Data availability
Data can be obtained from the LPI database (www.livingplanetin-
dex.org), AmphiBio30 (https://figshare.com/articles/Oliveira_et_al_
AmphiBIO_v1/4644424), FishBase (www.fishbase.org)28 and life-history
traits can be obtained from the amniote life-history database27 (https://
doi.org/10.6084/m9.figshare.c.3308127.v1).
Code availability
Code for the BHM model is available at: https://doi.org/10.5281/
zenodo.3901586.
27. Myhrvold, N. P. et al. An amniote life-history database to perform comparative analyses
with birds, mammals, and reptiles. Ecology 96, 3109 (2015).
28. Froese, R. & Pauly, D. FishBase version 12/2019 www.fishbase.org (2019).
29. Boettiger, C., Lang, D. T. & Wainwright, P. C. rfishbase: exploring, manipulating and
visualizing FishBase data from R. J. Fish Biol. 81, 2030–2039 (2012).
30. Oliveira, B. F., São-Pedro, V. A., Santos-Barrera, G., Penone, C. & Costa, G. C. AmphiBIO, a
global database for amphibian ecological traits. Sci. Data 4, 170123 (2017).
31. Collen, B. et al. Monitoring change in vertebrate abundance: the living planet index.
Conserv. Biol. 23, 317–327 (2009).
32. Gelman, A., Hwang, J. & Vehtari, A. Understanding predictive information criteria for
Bayesian models. Stat. Comput. 24, 997–1016 (2014).
33. Carpenter, B. et al. Stan: a probabilistic programming language. J. Stat. Softw. 76, 1–32
(2017).
34. R Core Team. R: A Language and Environment for Statistical Computing http://
www.R-project.org/ (R Foundation for Statistical Computing, 2016).
35. McRae, L., Deinet, S. & Freeman, R. The diversity-weighted living planet index: controlling
for taxonomic bias in a global biodiversity indicator. PLoS ONE 12, e0169156 (2017).
Acknowledgements We thank E. Hudgins, D. Nguyen, S. Varadarajan and A. Jones for
discussions, T. Coulson for comments and S. Varadarajan and F. Moyes for help creating the
figures. This work was supported by a Natural Sciences and Engineering Research Council
(NSERC) Discovery grant to B.L.
Author contributions Authors are listed in order of their contributions. B.L. formulated the
BHM model, conducted analyses and wrote the majority of the paper. A.L.H. discussed and
clarified the ideas and had a central role in the writing of the paper. D.A.G. discussed and
clarified the ideas, synthesized the data and contributed to the writing of the paper. B.M.
discussed and clarified the ideas, and commented on the manuscript. M.D. discussed and
clarified the ideas, commented on and improved the presentation of the manuscript. R.F.
discussed and clarified the ideas, and provided insight into the LPI data and analyses.
Competing interests The authors declare no competing interests.
Additional information
Supplementary information is available for this paper at https://doi.org/10.1038/s41586-020-
2920-6.
Correspondence and requests for materials should be addressed to B.L.
Peer review information Nature thanks Tim Coulson and the other, anonymous, reviewer(s) for
their contribution to the peer review of this work.
Reprints and permissions information is available at http://www.nature.com/reprints.
Article

Extended Data Fig. 1 | Theoretical analyses of BHM model. The p–p plots the fraction in each cluster ( f 1, f 2 = 1 − f 1). The 1:1 line is the theoretic
show that the posterior distributions for each estimated parameter are expectation, indicating that the true parameter value falls below the 0.01
unbiased and largely follow a 1:1 line for each hyper parameter (σ, τ) as well as quantile 1% of the time, the 0.02 quantile 2% of the time, and so on.
Extended Data Fig. 2 | Sensitivity analyses of primary cluster trends. The
trends of the primary clusters (θ1), for the main analysis (x axis) versus the
sensitivity analysis (y axis) for the threshold for extreme clusters (top) and the
offset when n = 0 was observed (bottom).
Article

Extended Data Fig. 3 | Effect of small time series on primary cluster trends.
Each point represents a trend estimate for the primary cluster of a system, with
the full dataset (x axis) versus data excluding time series with less than 10 data
points (y axis). The red dot indicates the freshwater Indo-Pacific mammals,
which was reduced from 22 populations (full) to 2 populations (only data with
at least 10 data points).
Extended Data Fig. 4 | Mean trends of primary clusters across systems
calculated using the BHM model. Top, all species (14,700 populations).
Middle, only large species (9,596 populations). Bottom, only small species
(5,103 populations). The small species appear to be declining more than large
species, although this finding needs to be interpreted with caution, as most
primary distributions did not significantly deviate from zero for small species.
Article
Extended Data Fig. 5 | Histograms of observed growth rates and output of
the BHM model for systems 1–16. Blue line, primary cluster; red line, extreme
cluster(s) from the model. Grey vertical lines show the range of observed
values. In comparing the model output to the data we show the following. (1)
The variation of the BHM primary cluster (blue line) is much lower than the raw
data, because the BHM separates variation in among-population trends from
variation due to within-population fluctuations. (2) The BHM model identifies
evidence for extreme clusters in both directions (for example, terrestrial
Indo-Pacific birds) or only one direction (for example, terrestrial Neotropical
mammals), but not for other apparent clusters (for example, terrestrial
Indo-Pacific herps). The BHM integrates the magnitude of within-population
fluctuations, time-series sizes, number of populations, among-population
variance, and the magnitude and frequency of the extreme populations in
determining whether additional (extreme) clusters are needed to account for
the observations.
Extended Data Fig. 6 | Histograms of observed growth rates and output of the BHM model for systems 17–32. Blue line, primary cluster; red line, extreme
cluster(s) from the model. Grey vertical lines show the range of observed values. For further information, see Extended Data Fig. 5.
Article

Extended Data Fig. 7 | Histograms of observed growth rates and output of the BHM model for systems 33–48. Blue line, primary cluster; red line, extreme
cluster(s) from the model. Grey vertical lines show the range of observed values. For further information, see Extended Data Fig. 5.
Extended Data Fig. 8 | Histograms of observed growth rates and output of the BHM model for systems 49–57. Blue line, primary cluster; red line, extreme
cluster(s) from the model. Grey vertical lines show the range of observed values. For further information, see Extended Data Fig. 5.
 nature research | reporting summary
Corresponding author(s): Brian Leung
Last updated by author(s): Aug 27, 2020

Reporting Summary
Nature Research wishes to improve the reproducibility of the work that we publish. This form provides structure for consistency and transparency
in reporting. For further information on Nature Research policies, see Authors & Referees and the Editorial Policy Checklist.

Statistics
For all statistical analyses, confirm that the following items are present in the figure legend, table legend, main text, or Methods section.
n/a Confirmed
The exact sample size (n) for each experimental group/condition, given as a discrete number and unit of measurement
A statement on whether measurements were taken from distinct samples or whether the same sample was measured repeatedly
The statistical test(s) used AND whether they are one- or two-sided
Only common tests should be described solely by name; describe more complex techniques in the Methods section.

A description of all covariates tested

A description of any assumptions or corrections, such as tests of normality and adjustment for multiple comparisons
A full description of the statistical parameters including central tendency (e.g. means) or other basic estimates (e.g. regression coefficient)
AND variation (e.g. standard deviation) or associated estimates of uncertainty (e.g. confidence intervals)

For null hypothesis testing, the test statistic (e.g. F, t, r) with confidence intervals, effect sizes, degrees of freedom and P value noted
Give P values as exact values whenever suitable.

For Bayesian analysis, information on the choice of priors and Markov chain Monte Carlo settings
For hierarchical and complex designs, identification of the appropriate level for tests and full reporting of outcomes
Estimates of effect sizes (e.g. Cohen's d, Pearson's r), indicating how they were calculated
Our web collection on statistics for biologists contains articles on many of the points above.

Software and code

Policy information about availability of computer code
Data collection Data from the Living Planet Index database. <www.livingplanetindex.org/>. (2016) was scraped in R 3.6.3. Data was extracted from
Fishbase using rfishbase 3.0.4.

Data analysis Bayesian analyses were conducted using the STAN 2.14 language, and processed and analyzed in R 3.6.3. The lme4 1.1-23 package was
referenced in the text. Custom code from this article can be obtained at: https://doi.org/10.5281/zenodo.3901586
For manuscripts utilizing custom algorithms or software that are central to the research but not yet described in published literature, software must be made available to editors/reviewers.
We strongly encourage code deposition in a community repository (e.g. GitHub). See the Nature Research guidelines for submitting code & software for further information.

Data
Policy information about availability of data
All manuscripts must include a data availability statement. This statement should provide the following information, where applicable:
- Accession codes, unique identifiers, or web links for publicly available datasets
- A list of figures that have associated raw data
- A description of any restrictions on data availability
October 2018

Data can be obtained from the Living Planet Index database. <www.livingplanetindex.org/>. (2016), AmphiBIO database from <https://figshare.com/articles/
Oliveira_et_al_AmphiBIO_v1/4644424>, Fishbase database <www.fishbase.org>, and mammal, bird and reptile life history traits from <https://doi.org/10.6084/
m9.figshare.c.3308127.v1>

1
 nature research | reporting summary
Field-specific reporting
Please select the one below that is the best fit for your research. If you are not sure, read the appropriate sections before making your selection.
Life sciences Behavioural & social sciences Ecological, evolutionary & environmental sciences
For a reference copy of the document with all sections, see nature.com/documents/nr-reporting-summary-flat.pdf

Ecological, evolutionary & environmental sciences study design

All studies must disclose on these points even when the disclosure is negative.
Study description The study re-examined previous estimates of global vertebrate declines from the Living Planet Index (LPI), and demonstrates that it is
driven by a small fraction of extreme population trends. The study then goes on to use a Bayesian Hierarchical Mixture Model to
separate the extreme clusters and primary cluster. The LPI dataset that was publicly available consists of 15241 vertebrate
populations from 3510 species, grouped into 57 domain-realm-taxon systems (hereafter just ‘systems’) (LPI 2016). Each system was
analyzed separately, and also was combined into a global metric using weighting factors (adjusted for species richness in each
system) from LPI, for comparability.

Research sample The data was obtained from the Living Planet Index database. <www.livingplanetindex.org/>. (2016), and consisted of 15241
vertebrate populations. To avoid double counting, when a species contained both finer resolution estimates within a country (2593
entries) as well as a country-wide aggregate, we excluded the country-wide aggregate (537 entries). This resulted in 14700
populations remaining in our analysis. Each system was defined by a combination of habitat domain (terrestrial, freshwater and
marine), biogeographic realm, and taxonomic grouping (Fish=Actinopterygii, Elasmobranchii, Holocephali, Myxini, Chondrichthyes,
Sarcopterygii, Cephalaspidomorphi; Birds=Aves, Mammals=Mammalia, Herps = Amphibia, Reptilia). Terrestrial and freshwater habitat
domains were separated into five realms (Afrotropical, Nearctic, Neotropical, Palearctic, and Indo-Pacific), whereas the marine
domain was separated into six realms (Arctic, Atlantic north temperate, Atlantic tropical/sub-tropical, Pacific north temperate, Indo-
Pacific tropical/sub-tropical, and South-temperate/Antarctic).

Sampling strategy All population time-series data in the LPI dataset were used. To avoid double counting, when a species contained both finer
resolution estimates within a country (2593 entries) as well as a country-wide aggregate, we excluded the country-wide aggregate
(537 entries). This resulted in 14700 populations remaining in our analysis.

Data collection The data was obtained by Dan Greenberg, and downloaded from publicly available databases identified in the data availability
statement

Timing and spatial scale Data were analyzed from 1970-2014, as these coincided with the analyses from the Living Planet Index. The spatial scale for the
analysis was global. The data was comprised of 14700 populations across many studies, and thus was measured at many scales. Thus,
relative changes per population was used.

Data exclusions To avoid double counting, when a species contained both finer resolution estimates within a country (2593 entries) as well as a
country-wide aggregate, we excluded the country-wide aggregate (537 entries). This resulted in 14700 populations remaining in our
analysis.

Reproducibility This is not relevant, as the existing LPI database was used. The purpose of the study was not an experiment, but instead to re-analyze
the available information on vertebrate trends, to evaluate whether previous estimates of decline (>50%) were due to clusters of
extremely declining populations, and to separate and analyze extreme clusters and primary clusters separately.

Randomization This is not relevant, as the existing LPI database was used, chosen for its impressive size and geographic coverage, and because
previous analyses of these data suggested broad-scale average vertebrate.

Blinding This is not relevant. Blinding as done in clinical trials, where group assignments of individuals is hidden from some researchers.
Primary data collection and experiments were not conducted in this study.

Did the study involve field work? Yes No

Reporting for specific materials, systems and methods

We require information from authors about some types of materials, experimental systems and methods used in many studies. Here, indicate whether each material,
system or method listed is relevant to your study. If you are not sure if a list item applies to your research, read the appropriate section before selecting a response.
October 2018

2
Materials & experimental systems Methods

nature research | reporting summary

n/a Involved in the study n/a Involved in the study
Antibodies ChIP-seq
Eukaryotic cell lines Flow cytometry
Palaeontology MRI-based neuroimaging
Animals and other organisms
Human research participants
Clinical data

October 2018

3
Article

Late Cretaceous bird from Madagascar

reveals unique development of beaks

https://doi.org/10.1038/s41586-020-2945-x Patrick M. O’Connor1,2,3 ✉, Alan H. Turner4, Joseph R. Groenke1, Ryan N. Felice5,

Raymond R. Rogers3,6, David W. Krause3,4 & Lydia J. Rahantarisoa7
Received: 30 April 2020

Accepted: 14 September 2020

Mesozoic birds display considerable diversity in size, flight adaptations and feather
Published online: 25 November 2020
organization1–4, but exhibit relatively conserved patterns of beak shape and
Check for updates
development5–7. Although Neornithine (that is, crown group) birds also exhibit
constraint on facial development8,9, they have comparatively diverse beak
morphologies associated with a range of feeding and behavioural ecologies, in
contrast to Mesozoic birds. Here we describe a crow-sized stem bird, Falcatakely
forsterae gen. et sp. nov., from the Late Cretaceous epoch of Madagascar that
possesses a long and deep rostrum, an expression of beak morphology that was
previously unknown among Mesozoic birds and is superficially similar to that of a
variety of crown-group birds (for example, toucans). The rostrum of Falcatakely is
composed of an expansive edentulous maxilla and a small tooth-bearing premaxilla.
Morphometric analyses of individual bony elements and three-dimensional rostrum
shape reveal the development of a neornithine-like facial anatomy despite the
retention of a maxilla–premaxilla organization that is similar to that of nonavialan
theropods. The patterning and increased height of the rostrum in Falcatakely reveals
a degree of developmental lability and increased morphological disparity that was
previously unknown in early branching avialans. Expression of this phenotype
(and presumed ecology) in a stem bird underscores that consolidation to the
neornithine-like, premaxilla-dominated rostrum was not an evolutionary prerequisite
for beak enlargement.

Our understanding of the evolution of Mesozoic birds continues to falls within a critical spatiotemporal gap. Very few avialans are known
improve, driven predominantly by discoveries from the Early Creta- from the entire Cretaceous period of Afro-Madagascar. The specimen
ceous epoch of China1–3,6. Although these specimens show considerable expands our knowledge of realized cranial shape disparity, in terms of
variation in body size, soft-tissue anatomy and inferred ecologies2–4,10,11, both morphological details and the proportions of elements, within
the disparity in Mesozoic avialan cranial shape remains restricted to the enantiornithine radiation and Mesozoic birds as a whole.
a relatively limited number of forms that are considered to be either
generalists or substrate-probing specialists5,6,12–15 and represent groups
that are only distantly related to crown birds. The Late Cretaceous Systematic palaeontology
(about 100–66 million years ago) chapter of avialan evolution remains Theropoda Marsh, 1881
relatively incomplete owing to a paucity of new fossil discoveries Paraves Sereno, 1997
(although see recent studies on birds such as Ichthyornis16 and Aste- Avialae Gauthier, 1986
riornis17). Thus, new fossils of Late Cretaceous birds are essential for Ornithothoraces Chiappe, 1995
refining hypotheses that relate to the morphological evolution and Enantiornithes Walker, 1981
diversification of avialans. Falcatakely forsterae gen. et sp. nov.
The phylogenetic diversity of early branching (non-neornithine)
Mesozoic birds is dominated by enantiornithines, which have been Etymology. ‘Falcata’ (from Latin falcatus), meaning armed with a
heralded as the first diversification of avialans and are characterized scythe, in reference to the shape of the rostrum; ‘kely’ (Malagasy),
by a range of body sizes and inferred habits2,14,18–21. This radiation is meaning small; ‘forsterae’, in recognition of Catherine A. Forster’s
notable for its apparent near-global distribution throughout most of contributions to work on Madagascan paravians.
the Cretaceous period. An exceptionally well-preserved partial cranium Holotype. Partial cranium (University of Antananarivo, UA 10015),
of a previously unknown enantiornithine (University of Antananarivo which consists of the rostrum, palate and periorbital regions (Fig. 1,
[UA] 10015) from the latest Cretaceous (Maastrichtian) of Madagascar Extended Data Figs. 1, 2 and Supplementary Videos 1–8).
1
Department of Biomedical Sciences, Heritage College of Osteopathic Medicine, Ohio University, Athens, OH, USA. 2Ohio Center for Ecological and Evolutionary Studies, Ohio University,
Athens, OH, USA. 3Department of Earth Sciences, Denver Museum of Nature & Science, Denver, CO, USA. 4Department of Anatomical Sciences, Stony Brook University, Stony Brook, NY, USA.
Centre for Integrative Anatomy, Department of Cell and Developmental Biology, University College London, London, UK. 6Geology Department, Macalester College, St Paul, MN, USA.
5

Département de Sciences de la Terre et de l’Environnement, Université d’Antananarivo, Antananarivo, Madagascar. ✉e-mail: oconnorp@ohio.edu
7

272 | Nature | Vol 588 | 10 December 2020

Locality and horizon. Locality MAD05-42, Berivotra Study Area, Upper
Cretaceous (Maastrichtian; 72.1–66 million years ago) Anembalemba
Member, Maevarano Formation, Mahajanga Basin, northwestern
Madagascar22.
Diagnosis. Differs from other paravians on the basis of the following
combination of features (*indicates autapomorphies): extended, high
maxilla that forms the dorsal contour of the rostrum*; dimpled texture
on the nasal and lacrimal, particularly on the triangular caudodorsal
process of the latter*; lacrimal with caudally expanded ventral pro-
cess*; large, flat jugal process of the postorbital*. Further differs from
most avialans by: a long, straight quadratojugal process of the jugal;
antorbital fenestra nearly as long as tall. Further differs from other
enantiornithines by: premaxilla slots into an extended V-shaped sulcus
of the maxilla*; narrow rostrum (width at premaxilla–maxilla junction
estimated at around 15% maximum width at rostral margin of orbit); a
nasal with distinct fossa near the rostral end*.
Remarks. Avialae and Neornithes are used herein to delimit increas-
ingly less inclusive monophyletic assemblages of theropod dinosaurs.
Avialae (that is, birds) refers to all theropods closer to living birds than
to dromaeosaurids and troodontids (that is, a stem-based monophy-
letic group containing Passer domesticus and all theropods closer
to it than to Dromaeosaurus or Troodon). Neornithes represents the
crown group of birds and is the equivalent of Aves (sensu ref. 23); see
Supplementary Information for additional phylogenetic definitions.
Further supporting information (such as interactive PDFs, matrices
and executable files) is available on DRYAD (https://doi.org/10.5061/
dryad.mkkwh70wg).
Cranial osteology
UA 10015 pertains to an enantiornithine bird (estimated cranial length,
8.5 cm) with a high, but extremely narrow, preorbital region (Fig. 1).
The lightly built face consists of a long, high edentulous maxilla and a
short, tooth-bearing premaxilla, forming a rostrum unlike that of any
known bird. The external nares are rostrally positioned and widely
separated from a large, parallelogram-shaped antorbital fenestra. The
premaxillae are fused rostrally and exhibit a short frontal process as in
some other enantiornithines5,24,25. The maxillary process is short and
slots into a V-shaped concavity on the maxilla (Fig. 1b, c and Extended
Data Figs. 1f, 2c). A single, conical, unserrated tooth is preserved in
the left premaxilla; the presence of additional premaxillary teeth is
uncertain owing to incomplete preservation.
The maxilla of Falcatakely is less than 1 mm thick and unique among
avialans in being extremely high and long, and forming at least 90%
of the reconstructed pre-orbital rostrum height (Fig. 1c). It is unfen-
estrated, a condition shared with Enantiornithes (Supplementary
Information), and lacks an antorbital fossa; it also preserves detailed
vascular sulci over its surface that indicate the presence of an expansive
keratinous rhamphotheca (beak) in life26 (Fig. 1a; interactive PDFs can
be found at https://doi.org/10.5061/dryad.mkkwh70wg). The premax-
illary process is well-developed and forms part of the ventral border
of the external naris. An elongate, tapering caudoventral projection
contributes to the jugal bar, delimiting the ventral border of the orbit
where it underlies the lacrimal boot.
The elongate nasals expand in width caudally and are unique in
possessing dorsomedially positioned fossae near the external nares
(Extended Data Fig. 1b). The mid-portion of the nasals are broad and
vaulted, as in bohaiornithid enantiornithines5,27 and express surface
dimpling on the lateral margin near the articulation with the lacrimal
(Extended Data Fig. 1b–d). The reconstruction of Falcatakely was gener-
ated using microcomputed tomography and reveals a nearly complete
right lacrimal (Fig. 1c). The dorsal half of the element is T-shaped, as
in avialans such as Archaeopteryx, Pengornis and Parapengornis2,27.
The rostrodorsal process is considerably longer than the caudo-
dorsal process and is most similar to the condition in pengornithid
a
b
jpmx
po (r) mpmx
pmx (l)
ect (r)
pter (r)
pal (l) pal (r) mx (r) to
mx (l)
pmx (r)
sr na (r)
c
na (l)
po (r) mpmx
Orbit lc (r)
AOF EN
ITF
qj (r) pal (r) mx (r)
ju (r) jpmx pmx (r)
ect (r)
d to
Fig. 1 | Cranium of the Cretaceous enantiornithine bird Falcatakely
forsterae (UA 10015, holotype). a, Photograph of the specimen, with a right
lateral view of the pre-orbital region (right side of the image) and a ventral view
of the palatal region (left side of the image). b, Digital polygon reconstruction
from the microcomputed tomography scan of the specimen shown in a.
c, Digital polygon reconstruction of the specimen with most elements in b
placed in near-life position in right lateral view. Scale bar, 1 cm (a–c).
d, Reconstruction (not to scale) illustrating the preserved (in white) elements
of the cranium. Left (l) and right (r) sides are indicated. AOF, antorbital fenestra;
ect, ectopterygoid; EN, external nares; ITF, infratemporal fenestra; jpmx, jugal
process of the maxilla; ju, jugal; lc, lacrimal; mpmx, midline premaxilla; mx,
maxilla; na, nasal; pal, palatine; pmx, premaxilla; po, postorbital; pter,
pterygoid; qj, quadratojugal; sr, scleral ring; to, tooth.
enantiornithines (for example, Parapengornis)10. The caudodorsal
process is unique among avialans in morphology, being sub-triangular
in shape, dimpled and pneumatic (Extended Data Fig. 1d). The ventral
ramus is longer than either the caudodorsal or rostrodorsal process,
and is extensively excavated, as in pengornithids and bohaiornith-
ids2,19,27. The ventral ramus terminates as a caudally expanded boot
that sits just dorsal to the overlapping portions of the jugal and maxilla
(Fig. 1c).
The jugal is triradiate with a long, dorsoventrally restricted maxil-
lary process, a distinct postorbital process and an extended, bar-like
quadratojugal process (Extended Data Fig. 1e). In contrast to most
avialans21,28, the quadratojugal process is long and directed straight
caudally, forming the ventral border of the infratemporal fenestra.
The quadratojugal process is not bifurcated as in most non-avialan
theropods and some phylogenetically early branching birds such as
Sapeornis29. The right postorbital (Extended Data Fig. 1e) is represented
only by its ventral process, which is flat and tapering, and is unlike any
known among avialans or paravians in general30. At least 15 scleral ossi-
cles are present, enough to estimate the external diameter of the scleral
ring to be 16–18 mm (Fig. 1c).
The digitally reconstructed palate of UA 10015 (Extended Data Fig. 2)
reveals a substantial level of detail not typically observable in Mesozoic
Nature | Vol 588 | 10 December 2020 | 273
Article

Pisciv

hus
Longusunguis

is
orena

Shenqiornis

ohaiorn

hync
Fortu

rnis
*

Zhouorn

s
Bohaiorni
Sulcavis

ngo
ntiorn
Pte

aeor
Liny
ngua

ura
Du

Parab

cha

yx
ryg
nh
Ne

izoo
iorn

Arch
Go norn is

ter

vis
is
is
vis
uan yx
orn

Jian
uq

is
bip

Lo agop

So ngs usa
ron
Sch
is

rn
ue

is
gia

no
ter

r
Ve

Vo

is
Ho gic
Eo

ha
t

rn
Pa
s
Pe

co rnis

go
pe

is
ng

rn
ng

lin
Ra is
o
or

ng
Lo pa ni rn is
ng xa s a no orn
iro Y n is
str vis x i a rav
Lo
ng av Yi civo
ipt is Ornithuromorpha s
ery Pi is
Bo x rav
luo
ch Ite sus
n
Qili ia Ga avis
Falc ana sar
atak Ap is
ely Enantiornithes h yorn
Sape Icht is
ornis o r n
Confu
ciusorn Enali venus
is dui rnis ad
Bapto
Jinzhouo varneri
rnis Baptornis
Confuciusornis san
ctus Hesperornis
Changchenornis Parahesperornis
Vegavis
Eoconfuciusornis
is Anas
Jeholorn
ryx Gallus
e o p te
Archa

Fig. 2 | Mosaic evolution of the avialan facial skeleton as depicted among
select early branching forms. Phylogenetic analysis places Facatakely among
enantiornithine birds. The illustration of Xinghaiornis(*) is placed near its
approximate position in the phylogeny based on a previous publication15.
Illustrations are not to scale. Red, premaxilla; green, maxilla; yellow, nasal;
lavender, lacrimal; blue, dentary. Illustrations of Archaeopteryx, Ichthyornis,
Hesperornis and Gallus were modified from a previous publication16.
See Supplementary Information for additional details for included taxa and
phylogenetic analyses.

avialans. The palatine is triradiate, with a long, thin rostral process
that abuts the maxilla (Extended Data Fig. 2a). The palatine does not
contact the jugal and only modestly contacts the pterygoid, but shares
an elongate contact with the ectopterygoid. A dorsomedially directed
choanal process sweeps towards the midline to join its antimere. Only
the thin rostral processes of the pterygoids are preserved in UA 10015;
these processes are in close association with the palatines (Extended
Data Fig. 2a). The ectopterygoid, an element that is unknown in most
Cretaceous avialans24,31, is represented by a robust body and a thin,
elongate, uncinate process that contacts the jugal bar (Extended
Data Fig. 2a). The vomers are represented by two thin, dorsoventrally
restricted laminar plates that extend rostrally between the two maxillae
(Extended Data Fig. 2a, c). Thin sheets of bone are present just rostral
to the pterygoids, potentially representing the expanded caudal end
of the vomer, reminiscent of the condition in Gobipteryx24,31.
Mosaic evolution in the avian beak
Our phylogenetic analyses recover Falcatakely nested within Enantio-
rnithes (Fig. 2 and Extended Data Figs. 3, 4). The long, deep and narrow
rostrum of Falcatakely, dominated by an expanded maxilla, provides a
stark contrast to the facial region formed by the premaxilla and maxilla
in other enantiornithines and more-crownward non-neornithines. Even
among rostrally elongated ornithothoracine taxa such as Longipteryx,
Longirostravis and Dingavis, this morphology is achieved through a
concomitant reduction in premaxillary and maxillary height as bones
elongate along the rostrocaudal axis5,6,12,14,15.
Quantitative assessment of non-avialan and avialan (including Neor-
nithes) facial shape demonstrates the combination of a derived cranial
phenotype in Falcatakely (that is, a neornithine-like expanded rostrum)
formed by an underlying plesiomorphic paravian skeletal framework.
We used two-dimensional geometric morphometrics (Fig. 3) to com-
pare the maxillary and premaxillary shape in UA 10015 to that of a sam-
ple of fossil non-avialan theropods, as well as the crown birds Gallus
gallus (red junglefowl) and Nothoprocta pentlandii (Andean tinamou).
Principal component analysis reveals that species group together on the
basis of the ratio of the maxillary to premaxillary size (the first principal
component) and the ratio of the rostrocaudal length to dorsoventral
height of both elements (second principal component). Despite hav-
ing maxillary and premaxillary proportions that are similar to those of
non-avialan theropods (for example, paravians, oviraptorosaurs and
ornithomimosaurs), Falcatakely exhibits an overall rostrum phenotype
that is convergent on a number of neornithine groups.
The configuration of the individual skeletal elements in Falcatakely
is more similar to the non-avialans Microraptor and Zanabazar than to
ornithuromorphs (including neornithines) owing to the expanded max-
illa and relatively small premaxilla. Nonetheless, the three-dimensional
shape of the pre-orbital facial skeleton closely resembles that of
some extant birds (Extended Data Figs. 5, 6), as assessed using
three-dimensional geometric morphometrics to compare the shape
of the maxilla, premaxilla and nasal within a sample of 349 extant
birds32 (Supplementary Information). Principal component analy-
sis of the rostrum shape reveals that Falcatakely occupies a position
in whole-rostrum morphospace that is quantitatively similar to those

274 | Nature | Vol 588 | 10 December 2020

 0.8
PC2
max.
Citipati
0.6
Incisivosaurus
0.4
PC2 (30.2%)
Gallus
0.2
Sapeornis
Jeholornis
Tsaagan Nothoprocta
Velociraptor
0 Zhouornis Yanornis
Zanabazar Falcatakely
Bohaiornithid
Archaeopteryx
Ichthyornis
Microraptor Pengornis
Xinghaiornis
–0.2 PC2 Haplocheirus Longipteryx
min. Struthiomimus
Chanzuiornis
Rapaxavis
PC1
min.
–0.4
–0.2 0 0.2
PC1 (35.7%)
Fig. 3 | Geometric morphometric analyses of the facial shape of Falcatakely
among paravians. Plot of the first two principal components (PCs) of the
two-dimensional landmark analysis of maxillary (blue line segments) and
premaxillary (red line segments) morphology of select theropod taxa. The
configuration of maxilla and premaxilla in Falcatakely is more similar to that of
non-avialans in a two-dimensional analysis focused on fossil taxa, although the
overall three-dimensional rostrum phenotype occupies a morphospace
converged on by subsequent radiations of neornithine birds (Supplementary
Data). See Supplementary Information for analytical protocols.
of a number of unrelated neornithines, including members of the
Ramphastidae (toucans), Phaethonidae (tropicbirds), Columbidae
(pigeons and doves) and Tyrannidae (tyrant flycatchers) (an interactive
morphospace plot is included as Supplementary Data and at https://
doi.org/10.5061/dryad.mkkwh70wg).
The discovery of Falcatakely expands the realized cranial morphol-
ogy among known non-neornithine birds considerably. Analysis of its
three-dimensional anatomy shows that it is a stem bird that occupies a
previously unrealized position in rostrum morphospace and potentially
exploited an ecology that was not again seen until the diversification of
crown-group birds in the mid-Cenozoic era. A partial emancipation of
the palate from the facial skeleton (that is, loss of jugal contact with the
palatine) concurrent with heretofore unappreciated rostrum elabora-
tion suggests that these regions are functionally and developmentally
integrated16,31,33. Notably, a mosaic pattern of palatal release is found
among stem avialans, at least insofar as the functional demands of the
rostrum or beak in Falcatakely appears to have required reinforce-
ment of the connections to the rear of the face. These connections
are maintained through the ectopterygoid and with retention of the
robust postorbital linkage, despite the loss of the mid-face palatine
connection to the jugal bar. Such an arrangement was probably nec-
essary to stabilize the mid-portion of the cranium and the long, high
and extremely narrow rostrum. Although incomplete, the presence of
a robust postorbital further indicates a rigidly enforced caudal region
of the cranium7.
The maxilla-dominated facial skeleton of Falcatakely reveals two
important insights for the evolutionary history of birds. First, the ances-
tral developmental patterning of rostrum construction in basal avialans
has generated neornithine-like cranial phenotypes that have not been
recognized in the fossil record until now. Second, the developmental
reduction of the maxilla previously inferred for Ornithothoraces7,9
was not a fixed trait, at least among enantiornithine birds. Thus, con-
Non-avialan theropods
Avialae
solidation to a premaxilla-dominated rostrum, a hallmark of all living
Confuciusornithiformes birds, was not an evolutionary prerequisite for rostrum and, therefore,
Enantiornithes beak enlargement. More generally, this is consistent with a growing
Falcatakely
appreciation of the flexibility of the underlying developmental mecha-
Ornithuromorpha
Ornithurae nisms8,34,35 that may be responsible for the generation of convergent
Neornithes morphologies among distantly related forms. With Falcatakely, this
appreciation can now be extended to the deep-time avialan record.
The discovery of Falcatakely expands the ecomorphological potential
realized by enantiornithines and Mesozoic birds more generally3,18.
This new appreciation of avialan anatomy underscores the potential
for considerable variability in trophic ecology during the first great
diversification of the group during the Cretaceous period5,6,36,37.
Confuciusornis
Hesperornis
Gansus Online content
Any methods, additional references, Nature Research reporting sum-
maries, source data, extended data, supplementary information,
acknowledgements, peer review information; details of author con-
tributions and competing interests; and statements of data and code
PC1
max. availability are available at https://doi.org/10.1038/s41586-020-2945-x.
0.4
1. Xu, X. et al. An integrative approach to understanding bird origins. Science 346, 1253293
(2014).
2. Zhou, Z., Clarke, J. & Zhang, F. Insight into diversity, body size and morphological
evolution from the largest Early Cretaceous enantiornithine bird. J. Anat. 212, 565–577
(2008).
3. Brusatte, S. L., O’Connor, J. K. & Jarvis, E. D. The origin and diversification of birds. Curr.
Biol. 25, R888–R898 (2015).
4. O’Connor, J. K. in The Evolution of Feathers (eds Foth, C. & Rauhut, O. W. M.) 147–172
(Springer, 2020).
5. O’Connor, J. K. & Chiappe, L. M. A revision of enantiornithine (Aves: Ornithothoraces) skull
morphology. J. Syst. Palaeontol. 9, 135–157 (2011).
6. Huang, J. et al. A new ornithurine from the Early Cretaceous of China sheds light on the
evolution of early ecological and cranial diversity in birds. PeerJ 4, e1765 (2016).
7. Bhullar, B.-A. S. et al. How to make a bird skull: major transitions in the evolution of the
avian cranium, paedomorphosis, and the beak as a surrogate hand. Integr. Comp. Biol.
56, 389–403 (2016).
8. Young, N. M. et al. Embryonic bauplans and the developmental origins of facial diversity
and constraint. Development 141, 1059–1063 (2014).
9. Mayr, G. Comparative morphology of the avian maxillary bone (os maxillare) based on an
examination of macerated juvenile skeletons. Acta Zool. 101, 24–38 (2020).
10. Hu, H., O’Connor, J. K. & Zhou, Z. A new species of Pengornithidae (Aves: Enantiornithes)
from the Lower Cretaceous of China suggests a specialized scansorial habitat previously
unknown in early birds. PLoS ONE 10, e0126791 (2015).
11. Bailleul, A. M. et al. An Early Cretaceous enantiornithine (Aves) preserving an unlaid egg
and probable medullary bone. Nat. Commun. 10, 1275 (2019).
12. Hou, L., Chiappe, L. M., Zhang, F. & Chuong, C.-M. New Early Cretaceous fossil from China
documents a novel trophic specialization for Mesozoic birds. Naturwissenschaften 91,
22–25 (2004).
13. O’Connor, J. K. et al. Phylogenetic support for a specialized clade of Cretaceous
enantiornithine birds with information from a new species. J. Vertebr. Paleontol. 29,
188–204 (2009).
14. O’Connor, J. K., Chiappe, L. M., Gao, C. & Zhao, B. Anatomy of the Early Cretaceous
enantiornithine bird Rapaxavis pani. Acta Palaeontol. Pol. 56, 463–475 (2011).
15. O’Connor, J. K., Wang, M. & Hu, H. A new ornithuromorph (Aves) with an elongate rostrum
from the Jehol Biota, and the early evolution of rostralization in birds. J. Syst. Palaeontol.
14, 939–948 (2016).
16. Field, D. J. et al. Complete Ichthyornis skull illuminates mosaic assembly of the avian
head. Nature 557, 96–100 (2018).
17. Field, D. J., Benito, J., Chen, A., Jagt, J. W. M. & Ksepka, D. T. Late Cretaceous neornithine
from Europe illuminates the origins of crown birds. Nature 579, 397–401 (2020).
18. Chiappe, L. M. & Witmer, L. M. Mesozoic Birds: Above the Heads of Dinosaurs (Univ.
California Press, 2004).
19. Li, Z., Zhou, Z., Wang, M. & Clarke, J. A. A new specimen of large-bodied basal
enantiornithine Bohaiornis from the Early Cretaceous of China and the inference of
feeding ecology in Mesozoic birds. J. Paleontol. 88, 99–108 (2014).
20. Wang, M., Hu, H. & Li, Z. A new small enantiornithine bird from the Jehol Biota, with
implications for early evolution of avian skull morphology. J. Syst. Palaeontol. 14, 481–497
(2016).
21. Wang, M., O’Connor, J. K. & Zhou, Z. A new robust enantiornithine bird from the Lower
Cretaceous of China with scansorial adaptations. J. Vertebr. Paleontol. 34, 657–671
(2014).
22. Rogers, R. R., Hartman, J. H. & Krause, D. W. Stratigraphic analysis of Upper Cretaceous
rocks in the Mahajanga Basin, northwestern Madagascar: implications for ancient and
modern faunas. J. Geol. 108, 275–301 (2000).
23. Gauthier, J. A. & de Queiroz, K. in New Perspectives on the Origin and Early Evolution of
Birds: Proceedings of the International Symposium in Honor of John H. Ostrom (eds
Gauthier, J. & Gall, L. F.). 7–41 (Peabody Museum of Natural History, Yale Univ., 2001).
Nature | Vol 588 | 10 December 2020 | 275
Article
24. Chiappe, L. M., Norell, M. A. & Clark, J. M. A new skull of Gobipteryx minuta (Aves:
Enantiornithes) from the Cretaceous of the Gobi Desert. Am. Mus. Novit. 3346, 1–15
(2001).
25. Wang, M. & Zhou, Z. A new enantiornithine (Aves: Ornithothoraces) with completely
fused premaxillae from the Early Cretaceous of China. J. Syst. Palaeontol. 17, 1299–1312
(2019).
26. Hieronymus, T. L. & Witmer, L. M. Homology and evolution of avian compound
Rhamphothecae. Auk 127, 590–604 (2010).
27. Wang, M., Zhou, Z.-H., O’Connor, J. K. & Zelenkov, N. V. A new diverse enantiornithine
family (Bohaiornithidae fam. nov.) from the Lower Cretaceous of China with information
from two new species. Vert. Palasiat. 52, 31–76 (2014).
28. Wang, M. & Hu, H. A comparative morphological study of the jugal and quadratojugal in
early birds and their dinosaurian relatives. Anat. Rec. 300, 62–75 (2017).
29. Wang, Y. et al. A previously undescribed specimen reveals new information on the
dentition of Sapeornis chaoyangensis. Cretac. Res. 74, 1–10 (2017).
30. Hu, H., O’Connor, J. K., Wang, M., Wroe, S. & McDonald, P. G. New anatomical information
on the bohaiornithid Longusunguis and the presence of a plesiomorphic diapsid skull in
Enantiornithes. J. Syst. Palaeontol. 18, 1481–1495 (2020).
31. Hu, H. et al. Evolution of the vomer and its implications for cranial kinesis in Paraves.
Proc. Natl Acad. Sci. USA 116, 19571–19578 (2019).
276 | Nature | Vol 588 | 10 December 2020
32. Felice, R. N. & Goswami, A. Developmental origins of mosaic evolution in the avian
cranium. Proc. Natl Acad. Sci. USA 115, 555–560 (2018).
33. Bhullar, B.-A. S. et al. A molecular mechanism for the origin of a key evolutionary
innovation, the bird beak and palate, revealed by an integrative approach to major
transitions in vertebrate history. Evolution 69, 1665–1677 (2015).
34. Mallarino, R. et al. Closely related bird species demonstrate flexibility between beak
morphology and underlying developmental programs. Proc. Natl Acad. Sci. USA 109,
16222–16227 (2012).
35. Tokita, M., Yano, W., James, H. F. & Abzhanov, A. Cranial shape evolution in adaptive
radiations of birds: comparative morphometrics of Darwin’s finches and Hawaiian
honeycreepers. Phil. Trans. R. Soc. Lond. B 372, 20150481 (2017).
36. Bell, A. & Chiappe, L. M. Statistical approaches for inferring ecology in Mesozoic birds.
J. Syst. Palaeontol. 9, 119–133 (2011).
37. O’Connor, J. K. The trophic habits of early birds. Palaeogeogr. Palaeoclimatol. Palaeoecol.
513, 178–195 (2019).
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in
published maps and institutional affiliations.
© The Author(s), under exclusive licence to Springer Nature Limited 2020
Methods
Temporal and stratigraphic context
UA 10015 was recovered in 2010 at the locality MAD05-42 in the Beriv-
otra Study Area of the Mahajanga Basin Project. The bone-bearing
horizon lies within facies 2 of the Anembalemba Member in the Upper
Cretaceous (Maastrichtian) Maevarano Formation22 (Supplementary
Information). Many specimens—including UA 10015—that were recov-
ered from the Anembalemba Member were entombed by debris flows,
which often results in high-quality preservation with only minimal
displacement and taphonomic distortion during burial38,39.
Phylogenetic methods
Given the extremely derived condition in Falcatakely and the notable
amount of homoplasy among non-avialan paravians and basal avialans,
we used a two-tiered dataset approach in an effort to best constrain the
phylogenetic affinities of Falcatakely (Supplementary Information).
First, we used the densely sampled, coelurosaur-wide matrix from the
Theropod Working Group (TWiG)40,41 to broadly assess and confirm
the position of Falcatakely among paravians (Extended Data Fig. 3
and Supplementary Information). Next, we used a modified version
of a well-established Mesozoic avialan-focused (WEA) matrix25, along
with previously described modifications16, to further examine the
relationship of Falcatakely among avialans (Fig. 2 and Extended Data
Fig. 4). Bayesian inference trees were estimated for each dataset using
MrBayes v.3.242. The standard model (Markov k-state variable model)43
was specified with gamma-distributed rate variation44. A subset of char-
acters was set as ordered, following the previous use of the included
datasets. During the analysis, Markov chain Monte Carlo convergence
was assessed using the average standard deviation of split frequencies
and by examining the trace files in Tracer45. Convergence to stationar-
ity was assumed for split frequencies below 0.01 and effective sample
size values >200. All analyses were performed with two runs of four
chains each that were run for 10 million generations while sampling
parameters every 1,000 generations. The first 25% of samples were
discarded as burn-in. Results are summarized using a majority rule con-
sensus (MRC) tree46. MRC trees for both datasets depict Falcatakely as
a member of Enantiornithes. The TWiG dataset recovers Falcatakely as
the sister taxon to Pengornis, whereas the WEA matrix finds Falcatakely
in a large polytomy with other enantiornithines (Extended Data Figs. 3,
4). Given the denser avialan sampling in the WEA dataset, the phylo-
genetic results from this matrix are used here as the primary results.
Clade support was assessed using the estimated posterior probabilities
from the Bayesian inference trees. Morphological character support
was established for the MRC trees using the map and apo commands
in TNT47–49. Additional details for the phylogenetic results and clade
support are presented in the Supplementary Information.
To further investigate the robustness of our inferred trees, three sen-
sitivity analyses were performed examining the influence of cranial
versus postcranial data and of cranial-only character scorings for select
taxa (for example, Archaeopteryx and Sapeornis) on tree inference.
These analyses reveal no significant topological alterations relative to
the standard analysis described above, lending support to the primary
results in which Falcatakely is placed among enantiornithine birds.
Moreover, additional explicit hypothesis testing using Bayes factor com-
parisons was conducted with Falcatakely constrained to stemward posi-
tions (for example, with Falcatakely excluded from Pygostylia), which
resulted in suboptimal solutions. Details for these analyses and the
specifics of the results are provided in the Supplementary Information;
executable files for the sensitivity and alternative hypothesis testing
are available on DRYAD (https://doi.org/10.5061/dryad.mkkwh70wg).
Reporting summary
Further information on research design is available in the Nature
Research Reporting Summary linked to this paper.
Data availability
UA 10015 is catalogued into the collections at the Université
d’Antananarivo. Details regarding the development of the digital
files and the derivatives of these files (such as DICOM or PLY) used as
part of the study are included in the Supplementary Information and
archived on the MorphoSource website (https://www.morphosource.
org/Detail/ProjectDetail/Show/project_id/7894). Phylogenetic char-
acter information and parameters used in the analyses are provided
in the Supplementary Information. Executable files for phylogenetic
analyses, character–taxon matrices, an interactive three-dimensional
morphospace plot and interactive three-dimensional PDFs are
hosted on DRYAD (https://doi.org/10.5061/dryad.mkkwh70wg).
This published study, including the novel genus (urn:lsid:zoobank.
org:act:5BA26059-B428-4896-BFEA-2475419C61FC) and species
(urn:lsid:zoobank.org:act:69314771-F0D8-4C15-946C-524164385FB7)
along with the associated nomenclatural acts, have been registered
in ZooBank: urn:lsid:zoobank.org:pub:4595D69E-FE12-4DAD-B155-
89F084254F73.
38. Rogers, R. R. Fine-grained debris flows and extraordinary vertebrate burials in the Late
Cretaceous of Madagascar. Geology 33, 297–300 (2005).
39. Rogers, R. R., Krause, D. W., Curry Rogers, K., Rasoamiaramanana, A. H. & Rahantarisoa, L.
Paleoenvironment and paleoecology of Majungasaurus crenatissimus (Theropoda:
Abelisauridae) from the Late Cretaceous of Madagascar. J. Vertebr. Paleontol. 27, 21–31
(2007).
40. Brusatte, S. L., Lloyd, G. T., Wang, S. C. & Norell, M. A. Gradual assembly of avian body
plan culminated in rapid rates of evolution across the dinosaur–bird transition. Curr. Biol.
24, 2386–2392 (2014).
41. Turner, A. H., Makovicky, P. J. & Norell, M. A. A review of dromaeosaurid systematics and
paravian phylogeny. Bull. Am. Mus. Nat. Hist. 371, 1–206 (2012).
42. Ronquist, F. et al. MrBayes 3.2: efficient Bayesian phylogenetic inference and model
choice across a large model space. Syst. Biol. 61, 539–542 (2012).
43. Lewis, P. O. A likelihood approach to estimating phylogeny from discrete morphological
character data. Syst. Biol. 50, 913–925 (2001).
44. Clarke, J. A. & Middleton, K. M. Mosaicism, modules, and the evolution of birds: results
from a Bayesian approach to the study of morphological evolution using discrete
character data. Syst. Biol. 57, 185–201 (2008).
45. Rambaut, A., Suchard, M. A., Xie, D. & Drummond, A. J. Tracer v.1.6. http://beast.
community/tracer (2014).
46. O’Reilly, J. E. & Donoghue, P. C. J. The efficacy of consensus tree methods for
summarizing phylogenetic relationships from a posterior sample of trees estimated from
morphological data. Syst. Biol. 67, 354–362 (2018).
47. Goloboff, P. A., Farris, J. & Nixon, K. TNT, a free program for phylogenetic analysis.
Cladistics 24, 774–786 (2008).
48. Goloboff, P. A., Farris, J. S. & Nixon, K. C. TNT: tree analysis using new technology. version
1.1 (Willi Hennig Society Edition) http://www.lillo.org.ar/phylogeny/tnt/ (2008).
49. Goloboff, P. A. & Catalano, S. A. TNT version 1.5, including a full implementation of
phylogenetic morphometrics. Cladistics 32, 221–238 (2016).
Acknowledgements We thank the Université d’Antananarivo, the Mahajanga Basin Project field
teams and the villagers of the Berivotra Study Area for support; the ministries of Mines, Higher
Education and Culture of the Republic of Madagascar for permission to conduct field research;
the National Geographic Society (8597-09) and the US National Science Foundation
(EAR–0446488, EAR–1525915, EAR–1664432) for funding; and M. Witton for drafting the line
drawings used in Fig. 1 and Extended Data Figs. 1, 2. Collection of avian three-dimensional
morphometric data was funded by European Research Council grant no. STG-2014-637171
(to A. Goswami). Full acknowledgments are provided in the Supplementary Information.
Author contributions P.M.O., A.H.T. and J.R.G. designed the project; P.M.O., A.H.T., J.R.G., R.R.R.,
D.W.K. and L.J.R. conducted the fieldwork. J.R.G. performed the mechanical preparation of the
specimen; J.R.G. and P.M.O. conducted the digital preparation and interpretation of the
specimen using microcomputed tomography and carried out the rapid prototyping of
UA 10015; R.R.R. and L.J.R. provided geological data and taphonomic interpretation; P.M.O.,
A.H.T., J.R.G. and R.N.F. completed the laboratory work on and digital representation of the
fossil and provided input on descriptions and comparisons; A.H.T. and P.M.O. contributed to
the character coding and phylogenetic analysis; R.N.F. completed the morphometric analyses;
P.M.O., A.H.T. and J.R.G. developed the manuscript, with contributions and/or editing from all
authors.
Competing interests The authors declare no competing interests.
Additional information
Supplementary information is available for this paper at https://doi.org/10.1038/s41586-020-
2945-x.
Correspondence and requests for materials should be addressed to P.M.O.
Peer review information Nature thanks Bhart-Anjan Bhullar and Daniel Field for their
contribution to the peer review of this work.
Reprints and permissions information is available at http://www.nature.com/reprints.
Article

Extended Data Fig. 1 | See next page for caption.

Extended Data Fig. 1 | Rostrum of the Cretaceous enantiornithine bird
Falcatakely (UA 10015, holotype). a, Reconstruction (not to scale) illustrating
the preserved (in white) elements of the cranium. b, Digital polygon surface
reconstruction (from microcomputed tomography scans) of the right nasal in
rostrodorsal view (caudal to the top) highlighting the midline depression and
dimpled surface texture. c, Digital polygon surface reconstruction of the right
nasal in dorsal view illustrating the dimpled architecture on the frontal and
rostral portions, which extends laterally onto the lacrimal. d, Digital polygon
surface reconstruction of the right facial elements in right lateral view to
illustrate the shape and inter-element relationships of the nasal, maxilla and
lacrimal (note the surface texture of the right maxilla with neurovascular sulci
broadly expressed over the lateral surface, deep to the inferred keratinous
covering (that is, beak)). e, Digital polygon surface reconstruction of the lower
lateral face to highlight arrangement of the maxilla, lacrimal, jugal and
postorbital (all elements from the right side). f, Digital polygon surface
reconstruction of left maxilla and premaxilla articulation (rostral to the left).
AOF, antorbital fenestra; cdp, caudodorsal process of the lacrimal; cp, choanal
process of the palatine; ect, ectopterygoid; EN, external nares; ITF,
infratemporal fenestra; fpn, frontal process of the nasal; inb, internarial bar;
jpmx, jugal process of the maxilla; ju, jugal; lbo, lacrimal boot; lc, lacrimal; ld,
lacrimal dimpling; le, lacrimal excavation; lf, lacrimal foramen; mpmx, midline
premaxilla; mx, maxilla; mxpj, maxillary process of the jugal; na, nasal; nd,
nasal dimpling; nf, nasal fossa; nvs, neurovascular sulci; pal, palatine; pmpm,
premaxillary process of the maxilla; pmx, premaxilla; po, postorbital; qj,
quadratojugal; rdp, rostrodorsal process of the lacrimal; rpn, rostral process of
the nasal; tm, tomial margin; to, tooth; vr, ventral ramus of the lacrimal.
Article
Extended Data Fig. 2 | Palatal and lateral facial regions of the Cretaceous
enantiornithine bird Falcatakely (UA 10015, holotype). a, Digital polygon
surface reconstruction (from microcomputed tomography scans) of the palate
and lateral face in ventral view. b, Reconstructed outline drawing of Falcatakely
in palatal view (shaded regions are not preserved). c, Digital polygon surface
reconstruction of internal aspect of left facial skeleton (premaxilla, maxilla and
nasal) and palate in right lateral view. The left and right sides are indicated as (l)
and (r), respectively. The dashed line in c represents the approximate contour
of the caudal margin (that is, the ventral ramus of the lacrimal) of the antorbital
fenestra. Scale bar, 5 mm; the scale bar is representative for a and c; the
reconstruction in b is not to the same scale. AOF, antorbital fenestra; bs,
basisphenoid rostrum; cp, choanal process of the (right) palatine; ect,
ectopterygoid; EN, external nares; jpmx, jugal process of the maxilla; mpmx,
midline premaxilla; mx, maxilla; na, nasal; pal, palatine; pmx, premaxilla; pter,
pterygoid; to, tooth; up, uncinate process of the ectopterygoid; vm, vomers.
Extended Data Fig. 3 | Majority- rule tree of Falcatakely among coelurosaurians from the Bayesian analysis of the TWiG matrix. Clades outside of the Avialae
are collapsed for brevity. Posterior probabilities are placed above the nodes.
Article

Extended Data Fig. 4 | Majority -rule tree of Falcatakely among avialans from the Bayesian analysis of a modified matrix that was previously published.
A matrix modified from a previous study25 was used. Posterior probabilities are placed above the nodes.
Extended Data Fig. 5 | Geometric morphometric analysis of rostrum shape
in Falcatakely among avians. Plot of the first two principal components of the
three-dimensional landmark analysis of total rostrum shape of Falcatakely and
extant avian taxa. Whereas the unique configuration of the maxilla and
premaxilla in Falcatakely is more similar to those of non-avialan paravians
(Fig. 3), the overall three-dimensional rostrum phenotype occupies the
morphospace that is converged on by subsequent radiations of neornithine
birds (Supplementary Data). See Supplementary Information for analytical
protocols.
Article

Extended Data Fig. 6 | Landmarking procedure for three-dimensional geometric morphometric analysis in dorsal and lateral views. a, Dorsal view.
b, Lateral view. Red spheres represent anatomical (type I) landmarks; yellow spheres are sliding semi-landmarks.
 nature research | reporting summary
Corresponding author(s): Patrick M. O'Connor
Last updated by author(s): Aug 17, 2020

Reporting Summary
Nature Research wishes to improve the reproducibility of the work that we publish. This form provides structure for consistency and transparency
in reporting. For further information on Nature Research policies, see our Editorial Policies and the Editorial Policy Checklist.

Statistics
For all statistical analyses, confirm that the following items are present in the figure legend, table legend, main text, or Methods section.
n/a Confirmed
The exact sample size (n) for each experimental group/condition, given as a discrete number and unit of measurement
A statement on whether measurements were taken from distinct samples or whether the same sample was measured repeatedly
The statistical test(s) used AND whether they are one- or two-sided
Only common tests should be described solely by name; describe more complex techniques in the Methods section.

A description of all covariates tested

A description of any assumptions or corrections, such as tests of normality and adjustment for multiple comparisons
A full description of the statistical parameters including central tendency (e.g. means) or other basic estimates (e.g. regression coefficient)
AND variation (e.g. standard deviation) or associated estimates of uncertainty (e.g. confidence intervals)

For null hypothesis testing, the test statistic (e.g. F, t, r) with confidence intervals, effect sizes, degrees of freedom and P value noted
Give P values as exact values whenever suitable.

For Bayesian analysis, information on the choice of priors and Markov chain Monte Carlo settings
For hierarchical and complex designs, identification of the appropriate level for tests and full reporting of outcomes
Estimates of effect sizes (e.g. Cohen's d, Pearson's r), indicating how they were calculated
Our web collection on statistics for biologists contains articles on many of the points above.

Software and code

Policy information about availability of computer code
Data collection Avizo 7.1 (VSG), 9.2 (FEI/Thermo-Fisher Scientific), and Avizo Lite 2019 (ThermoScientific); ImageJ v1.48 (National Institutes of Health);
Checkpoint (Stratovan); Blender v2.79b.

Data analysis MrBayes v3.2; TNT 1.5; Tracer v1.6; Geomorph (R package), Adams and Otárola-Castillo, 2013; StereoMorph (R package); Avizo 7 (VSG), 9
(FEI/Thermo-Fisher Scientific), and Avizo Lite 2019 (ThermoScientific); Animation Producer in Avizo; Adobe Acrobat Pro DC (Continuous
Release) Version 2020, Adobe Premiere Pro (Creative Cloud edition).
For manuscripts utilizing custom algorithms or software that are central to the research but not yet described in published literature, software must be made available to editors and
reviewers. We strongly encourage code deposition in a community repository (e.g. GitHub). See the Nature Research guidelines for submitting code & software for further information.

Data
Policy information about availability of data
All manuscripts must include a data availability statement. This statement should provide the following information, where applicable:
April 2020

- Accession codes, unique identifiers, or web links for publicly available datasets
- A list of figures that have associated raw data
- A description of any restrictions on data availability

UA 10015 is cataloged into the collections at the Université d’Antananarivo. Details regarding digital file development and derivatives of files (e.g., DICOM, PLY) used
as part of the study are included in the Supplementary Information and archived on the MorphoSource website (https://www.morphosource.org/Detail/
ProjectDetail/Show/project_id/7894). Phylogenetic character information and parameters used in the analyses are provided in the Supplementary Information.
Executable files for phylogenetic analyses, interactive 3D morphospace plot, and interactive 3D PDFs are hosted on DRYAD: https://doi.org/10.5061/
dryad.mkkwh70wg. This published work, including the novel genus (urn:lsid:zoobank.org:act:5BA26059-B428-4896-BFEA-2475419C61FC) and species

1
(urn:lsid:zoobank.org:act:69314771-F0D8-4C15-946C-524164385FB7) along with the associated nomenclatural acts, have been registered in ZooBank:
urn:lsid:zoobank.org:pub:4595D69E-FE12-4DAD-B155-89F084254F73.

nature research | reporting summary

Field-specific reporting
Please select the one below that is the best fit for your research. If you are not sure, read the appropriate sections before making your selection.
Life sciences Behavioural & social sciences Ecological, evolutionary & environmental sciences
For a reference copy of the document with all sections, see nature.com/documents/nr-reporting-summary-flat.pdf

Ecological, evolutionary & environmental sciences study design

All studies must disclose on these points even when the disclosure is negative.
Study description Descriptive/comparative study of a new fossil bird (Falcatakely forsterae) from the Late Cretaceous of Madagascar.

Research sample This single cranium of Falcatakely (UA 10015) is the only known material of this taxon thus far discovered; it is known exclusively from
the Upper Cretaceous (Maastrichtian) of northwestern Madagascar.

Sampling strategy The study developed herein involves the description of a new taxon based on direct observation, light microscopy, and micro-
computed tomography of fossil represented by this single cranium. Digital preparation allowed for the a complete analysis and
reconstruction of individual elements of the cranium.

Data collection The holotype of Falcatakely forsterae (UA 10015) was collected from locality MAD05-42 by hand quarrying (ice pick, brush, rock
hammer), with subsequent emplacement in a plaster jacket prior to removal for laboratory processing. Mechanical and digital
preparation of the fossil was completed by J.R. Groenke, with interpretation of the anatomy (both of the fossil itself and digital
reconstructions/interpretations) by P.M. O'Connor, A.H. Turner, and J.R. Groenke. R.N. Felice led the morphometric analyses
included herein. Character scorings assessed by A.H. Turner and P.M. O'Connor, with phylogenetic analyses completed by A.H.
Turner.

Timing and spatial scale The specimen was originally collected during the 2010 calendar year, but only initially prepped and CT scanned (medical CT scanner
of the plaster jacket) that same year, yielding an ambiguous identification. J.R. Groenke (Ohio University) did additional mechanical
preparation in March 2017, immediately followed by a high-resolution microCT scan in April 2017. Intensive digital preparation then
ensued between April 2017 and January 2018, with subsequent, albeit intermittent, digital preparation, interpretation and
refinement of models through January 2019.

Data exclusions No data were excluded.

Reproducibility Not applicable; given that this paper focuses on a single specimen thus far known to humankind, it does not fall into the category for
being reproducible. However, the datasets assembled for this study are publicly available for future reanalyses by other workers.

Randomization Not Applicable.

Blinding Not applicable

Did the study involve field work? Yes No

Field work, collection and transport

Field conditions Fieldwork was conducted during the austral summer (i.e., the dry season) in 2010 in the Mahajanga Basin, near the village of
Berivotra, Madagascar.

Location The holotypic specimen was collected from the Upper Cretaceous Maevarano Formation, Mahajanga Basin, Madagascar.
Approximate coordinates: S 15 degrees, 54' 20.94", E 46 degrees, 35' 00.23"

Access & import/export The specimen was collected under a Collaborative Agreement with the University of Antananarivo and various ministries (Ministry of
Mines, Ministry of Higher Education) of the Madagascar government. Permits from the Ministry of Mines (Scientific Studies
Authorization No 005/2010) and the Ministry of Higher Education/University of Antananarivo (No 76 PAB/10, Supporting
documentation: - Scientific Authorization Studies No 007/2010, 005/2010, 006/2010, 009/2010) were used in support of field
research were issued on 17 June 2010 and 18 June 2010, respectively.
April 2020

Disturbance This study involved minimal disturbance to the environment, as the fossil-bearing layer was within 0.70 meters of the surface in the
locality.

Reporting for specific materials, systems and methods

We require information from authors about some types of materials, experimental systems and methods used in many studies. Here, indicate whether each material,
system or method listed is relevant to your study. If you are not sure if a list item applies to your research, read the appropriate section before selecting a response.

2
Materials & experimental systems Methods

nature research | reporting summary

n/a Involved in the study n/a Involved in the study
Antibodies ChIP-seq
Eukaryotic cell lines Flow cytometry
Palaeontology and archaeology MRI-based neuroimaging
Animals and other organisms
Human research participants
Clinical data
Dual use research of concern

Palaeontology and Archaeology

Specimen provenance
The holotype specimen (UA 10015) was recovered from locality MAD05-42 from the Upper Cretaceous Maevarano Formation,
Mahajanga Basin, Madagascar. Permits from the Ministry of Mines (Scientific Studies Authorization No 005/2010) and the Ministry of
Higher Education/University of Antananarivo (No 76 PAB/10, Supporting documentation: - Scientific Authorization Studies No
007/2010, 005/2010, 006/2010, 009/2010) were used in support of field research were issued on 17 June 2010 and 18 June 2010,
respectively.

Specimen deposition The holotype specimen of Falcatakely forsterae is reposited in the University of Antananarivo (UA), Madagascar with the collection
number UA 10015 .

Dating methods No new dates were obtained for this contribution; age constraint for the Maevarano Fm. is developed in Rogers et al. 2000.

Tick this box to confirm that the raw and calibrated dates are available in the paper or in Supplementary Information.

Ethics oversight Fossil collection and exportation were completed in compliance with permits issued by the Ministry of Mines (United Republic of
Madagascar) and through a Collaborative Agreement with the University of Antananarivo and various ministries (Ministry of Mines,
Ministry of Higher Education) of the Madagascar government.
Note that full information on the approval of the study protocol must also be provided in the manuscript.

April 2020

3
Article

Multiple wheat genomes reveal global

variation in modern breeding

https://doi.org/10.1038/s41586-020-2961-x Sean Walkowiak1,2,41, Liangliang Gao3,41, Cecile Monat4,41, Georg Haberer5,

Mulualem T. Kassa6, Jemima Brinton7, Ricardo H. Ramirez-Gonzalez7, Markus C. Kolodziej8,
Received: 3 April 2020
Emily Delorean3, Dinushika Thambugala9, Valentyna Klymiuk1, Brook Byrns1,
Accepted: 9 September 2020 Heidrun Gundlach5, Venkat Bandi10, Jorge Nunez Siri10, Kirby Nilsen1,11, Catharine Aquino12,
Axel Himmelbach4, Dario Copetti13,14, Tomohiro Ban15, Luca Venturini16, Michael Bevan7,
Published online: 25 November 2020
Bernardo Clavijo17, Dal-Hoe Koo3, Jennifer Ens1, Krystalee Wiebe1, Amidou N’Diaye1,
Open access Allen K. Fritz3, Carl Gutwin10, Anne Fiebig4, Christine Fosker17, Bin Xiao Fu2,
Gonzalo Garcia Accinelli17, Keith A. Gardner18, Nick Fradgley18, Juan Gutierrez-Gonzalez19,
Check for updates
Gwyneth Halstead-Nussloch13, Masaomi Hatakeyama12,13, Chu Shin Koh20, Jasline Deek21,
Alejandro C. Costamagna22, Pierre Fobert6, Darren Heavens17, Hiroyuki Kanamori23,
Kanako Kawaura15, Fuminori Kobayashi23, Ksenia Krasileva17, Tony Kuo24,25, Neil McKenzie7,
Kazuki Murata26, Yusuke Nabeka26, Timothy Paape13, Sudharsan Padmarasu4,
Lawrence Percival-Alwyn18, Sateesh Kagale6, Uwe Scholz4, Jun Sese25,27, Philomin Juliana28,
Ravi Singh28, Rie Shimizu-Inatsugi13, David Swarbreck17, James Cockram18, Hikmet Budak29,
Toshiaki Tameshige15, Tsuyoshi Tanaka23, Hiroyuki Tsuji15, Jonathan Wright17, Jianzhong Wu23,
Burkhard Steuernagel7, Ian Small30, Sylvie Cloutier31, Gabriel Keeble-Gagnère32,
Gary Muehlbauer19, Josquin Tibbets32, Shuhei Nasuda26, Joanna Melonek30, Pierre J. Hucl1,
Andrew G. Sharpe20, Matthew Clark16, Erik Legg33, Arvind Bharti33, Peter Langridge34,
Anthony Hall17, Cristobal Uauy7, Martin Mascher4,35, Simon G. Krattinger8,36,
Hirokazu Handa23,37, Kentaro K. Shimizu13,15, Assaf Distelfeld38, Ken Chalmers34,
Beat Keller8, Klaus F. X. Mayer5,39, Jesse Poland3, Nils Stein4,40, Curt A. McCartney9 ✉,
Manuel Spannagl5 ✉, Thomas Wicker8 ✉ & Curtis J. Pozniak1 ✉

Advances in genomics have expedited the improvement of several agriculturally

important crops but similar efforts in wheat (Triticum spp.) have been more
challenging. This is largely owing to the size and complexity of the wheat genome1,
and the lack of genome-assembly data for multiple wheat lines2,3. Here we generated
ten chromosome pseudomolecule and five scaffold assemblies of hexaploid wheat to
explore the genomic diversity among wheat lines from global breeding programs.
Comparative analysis revealed extensive structural rearrangements, introgressions
from wild relatives and differences in gene content resulting from complex breeding
histories aimed at improving adaptation to diverse environments, grain yield and
quality, and resistance to stresses4,5. We provide examples outlining the utility of these
genomes, including a detailed multi-genome-derived nucleotide-binding leucine-rich
repeat protein repertoire involved in disease resistance and the characterization of
Sm16, a gene associated with insect resistance. These genome assemblies will provide
a basis for functional gene discovery and breeding to deliver the next generation of
modern wheat cultivars.

Wheat is a staple food across all parts of the world and is one of the
most widely grown and consumed crops7. As the human population
continues to grow, wheat production must increase by more than 50%
over current levels by 2050 to meet demand7. Efforts to increase wheat
production may be aided by comprehensive genomic resources from
global breeding programs to identify within-species allelic diversity and
determine the best allele combinations to produce superior cultivars2,8.
Two species dominate current global wheat production: allotetra-
ploid (AABB) durum wheat (Triticum turgidum ssp. durum), which
is used to make couscous and pasta9, and allohexaploid (AABBDD)
bread wheat (Triticum aestivum), used for making bread and noodles.
A, B and D in these designations correspond to separate subgenomes
derived from three ancestral diploid species with similar but distinct
genome structure and gene content that diverged between 2.5 and
6 million years ago10. The large genome size (16 Gb for bread wheat),
high sequence similarity between subgenomes and abundance of
repetitive elements (about 85% of the genome) hampered early wheat
genome-assembly efforts3. However, chromosome-level assemblies
have recently become available for both tetraploid11,12 and hexaploid
wheat1,13. Although these genome assemblies are valuable resources,

*A list of affiliations appears at the end of the paper.

Nature | Vol 588 | 10 December 2020 | 277

Article
they do not fully capture within-species genomic variation that can be
used for crop improvement, and comparative genome data from mul-
tiple individuals is still needed to expedite bread wheat research and
breeding. Until now, comparative genomics of multiple bread wheat
lines have been limited to exome-capture sequencing4,5,14, low-coverage
sequencing2 and whole-genome scaffolded assemblies13,15–17. Here we
report multiple reference-quality genome assemblies and explore
genome variation that, owing to past breeder selection, differs greatly
between bread wheat lines. These genome assemblies usher a new era
for bread wheat and equip researchers and breeders with the tools
needed to improve bread wheat and meet future food demands.
Global variation in wheat genomes
To expand on the genome assembly of wheat for Chinese Spring1, we
generated ten reference-quality pseudomolecule assemblies (RQAs)
and five scaffold-level assemblies of hexaploid wheat (Supplemen-
tary Note 1, Supplementary Tables 1–3). For each RQA, we performed
de novo assembly of contigs (contig N50 > 48 kb) that were combined
into scaffolds (N50 > 10 Mb) spanning more than 14.2 Gb (Supplemen-
tary Note 1). The completeness of the genomes was supported by a
universal single-copy orthologue (BUSCO) analysis that identified more
than 97% of the expected gene content in each genome (Supplemen-
tary Note 1). More than 94% of the scaffolds were ordered, oriented
and curated using 10X Genomics linked reads and three-dimensional
chromosome conformation capture sequencing (Hi-C) to generate
21 pseudomolecules, as done previously for wheat1,12 and barley (Hor-
deum vulgare)18. The size and structure of the genomes were similar to
that of Chinese Spring, and we observed high collinearity between the
pseudomolecules (Extended Data Fig. 1). We also independently vali-
dated the scaffold placement and orientation in the pseudomolecule
assembly of CDC Landmark by Oxford Nanopore long-read sequencing
(Extended Data Fig. 2a, Supplementary Note 2). To complement the
RQAs, we generated scaffold-level assemblies of five additional bread
wheat lines (Supplementary Note 1). To determine the global context
of the 15 assemblies, we combined our data with existing datasets4,5,19
(Fig. 1a, Supplementary Table 4). The genetic relationships were in
agreement with those reported in previous studies4,5 and reflected
pedigree, geographical location and growth habit (that is, spring ver-
sus winter type). There was also a clear separation between the newly
assembled genomes and Chinese Spring, supporting that they capture
geographical and historical variation not represented in the Chinese
Spring assembly.
Polyploidy and CNV drive gene diversification
Single-nucleotide polymorphisms (SNPs), insertions or deletions
(indels), presence/absence variation (PAV) and gene copy number varia-
tion (CNV) influence agronomically important traits. This is particularly
true for polyploid species such as wheat, in which gene redundancy
can buffer the effect of genome variation17. To assess gene content, we
projected around 107,000 high-confidence gene models from Chinese
Spring1 onto the RQAs (Supplementary Note 3). The total number of
projected genes exhibited a narrow range, between 118,734 and 120,967
(Supplementary Table 5). We identified orthologous groups among
projected genes and used the alignment of the orthologous groups to
examine SNPs in coding sequences (Supplementary Note 3). The peak
positions of nucleotide diversity across the three subgenomes were
highly similar to those reported in previous studies20, supporting a
strong representation of breeding diversity within the RQAs (Extended
Data Fig. 3a, b). The correlation of synonymous nucleotide diversity π
(r = 0.11–0.29) and Tajima’s D (r = 0.02–0.06) between homeologues
was low (Supplementary Tables 6–8). This suggested that polyploidiza-
tion increased the number of targets of selection and contributed to
broad adaptation of bread wheat, as in wild polyploid plant species20–22.
278 | Nature | Vol 588 | 10 December 2020
Further investigation of orthologous groups indicated that 88.1% were
unambiguous (clusters containing at most one member in each cultivar)
(Extended Data Fig. 3c, Supplementary Table 5). Orthologous groups
comprising exactly one gene in each line (‘complete’) were the most
frequent (approximately 73.5% of genes per cultivar), suggesting strong
retention of orthologous genes within the ten RQAs. The residual genes
represented either singleton genes with no reciprocal best BLAST hits
or genes located in complex clusters in at least one cultivar. Roughly
12% of genes showed PAVs, and their clustering resulted in relationships
(Fig. 1b) that were consistent with SNP-based phylogenetic similarities
(Fig. 1a). In addition, approximately 26% of the projected genes were
found in tandem duplications, indicating that CNV is a strong contribu-
tor of genetic variation in wheat.
To provide an example of gene expansion on emerging breeding
targets, we performed a more detailed analysis of the restorer of fertil-
ity (Rf) gene families (Supplementary Note 4). Rf genes are involved in
restoring pollen fertility in hybrid breeding programs23, and we iden-
tified a previously undescribed clade within the mitochondrial tran-
scription termination factor (mTERF) family (Supplementary Table 9),
which has recently been implicated in fertility restoration in barley24. Of
note, this clade shows evolutionary patterns similar to those of Rf-like
pentatricopeptide repeat (PPR) proteins, representatives of which
are associated with Rf3, a major locus used in hybrid wheat breeding
programs (Extended Data Fig. 4). Although wheat is currently not a
hybrid crop, there is substantial interest in Rf genes and their potential
application in hybrid wheat production systems25. To our knowledge,
no Rf genes have been cloned in wheat and our analysis of Rf genes in
multiple RQAs and identification of an Rf clade in wheat is an impor-
tant step forward in tackling the challenges of hybrid wheat breeding.
The wheat NLR repertoire
To further exemplify the use of multi-genome comparisons for char-
acterizing agronomically relevant gene families, we examined gene
expansion in nucleotide-binding leucine-rich repeat (NLR) proteins,
which are major components of the innate immune system and are
often causal genes for disease resistance in plants26,27. We performed
de novo annotation of loci that contain conserved NLR motifs (NB-ARC–
leucine-rich repeat) and identified around 2,500 loci with NLR signa-
tures in each RQA (Supplementary Tables 10, 11). A redundancy analysis
showed that only 31–34% of the NLR signatures are shared across all
genomes, and the number of unique signatures ranged from 22 to 192
per wheat cultivar. We estimated the number of unique NLR signatures
that can be detected by incrementally adding more wheat genomes to
the dataset; this revealed that 90% of the NLR complement is reached
at between 8 (considering 95% sequence identity) and 11 wheat lines
(considering 100% protein sequence identity) (Fig. 1c). The total NLR
complement of all wheat lines consisted of 5,905 (98% identity) to 7,780
(100% identity) unique NLR signatures, highlighting the size and com-
plexity of the repertoire of receptors involved in disease resistance.
Transposon signatures identify introgressions
Transposable elements make up a large majority of the wheat genome
and have a critical role in genome structure and gene regulation. We
characterized the overall transposable element content (81.6%) and its
composition (69% long terminal-repeat retrotransposons (LTR) and
12.5% DNA transposons) in the RQAs (Supplementary Table 5). Across all
RQAs, we annotated 1.22 × 106 full length (fl)-LTRs, which clustered lines
into the same groups we observed from our analysis of PAV and SNPs
(Fig. 1a, b, Extended Data Fig. 3d). Generally, unique fl-LTRs (147,450)
were young (median of 0.9 million years) and were enriched in the
highly recombining, more distal chromosomal regions (Fig. 1d). By con-
trast, shared fl-LTRs were older (median of 1.3 million years) and were
more evenly distributed across the pericentric regions (Fig. 1d). The
 a Principal component 1 b Similarity in PAV
Norin 61 88 92
Robigus Julius
LongReach Lancer
SY Mattis
Mace
Jagger Chinese Spring
Claire
Norin 61
Principal component 2

ArinaLrFor Jagger
Chinese Julius
Spring Paragon
ArinaLrFor
Cadenza
SY Mattis
Weebill 1 LongReach Lancer
PI190962 (spelt wheat)
CDC Stanley
CDC Stanley
CDC Landmark
CDC Landmark
PI190962
(spelt wheat) Mace

c d LTR-retrotransposon density
Sequence
Low High
identity
8,000 2 Unique
100% 0
2 2
0
2 3
0
2
Unique NLRs

Insertion time (Myr)

98% 0

Number of lines
2
0 5
5,000 2
95% 6
0
2
0 7
2
0 8
2
0 9
2
0 10
2,000 2
0 11
0 5 10 15 0 25 50 75 100
Number of lines Chromosomal location (% length)

Fig. 1 | Patterns of variation in the wheat genome. a, Principal component
analysis of polymorphisms from exome-capture sequencing of about 1,200
lines (grey markers), 16 lines from whole-genome shotgun resequencing
(orange markers) and our new assemblies (black markers). Text colours reflect
different geographical locations and winter or spring growth. b, Dendrogram
of pairwise Jaccard similarities for gene PAV between all RQA assemblies.
c, Number of unique NLRs at different per cent identity cut-offs as the number
of genomes increases. Dashed vertical lines represent 90% of the NLR
complement. Markers indicate the mean values of all permutations of the order
of adding genomes. Whiskers show maximum and minimum values based on
one million random permutations. d, Chromosomal location versus insertion
age distribution of unique to (reading downward) increasingly shared syntenic
full-length LTR retrotransposons.

RLC-Angela fl-LTRs were the most abundant (21,000–27,000 full-length
copies per genome) and analysis of variant patterns identified several
chromosomal segments that contained numerous unique or rare ret-
rotransposon insertions (Extended Data Fig. 5), which, on the basis
of breeding history, we hypothesize to represent introgressions. For
example, the LongReach Lancer RQA revealed two unique regions, a
pericentric region on chromosome 2B and a segment on the end of
chromosome 3D (Fig. 2a, b), both of which affect chromosome length
(Extended Data Fig. 5). We used pedigree analysis to postulate the
source of the introgressions and performed whole-genome sequenc-
ing of multiple accessions of putative donors. LongReach Lancer carries
the stem rust resistance gene Sr36, derived from an introgression from
Triticum timopheevii, and the resistance genes Lr24 (leaf rust) and Sr24
(stem rust), derived from tall wheatgrass28,29 (Thinopyrum ponticum).
We generated whole-genome sequence reads from multiple T. ponticum
and T. timopheevii accessions (Supplementary Table 12) and alignment
to the LongReach Lancer RQA confirmed a T. ponticum introgression
spanning a region of approximately 60 Mb of chromosome 3D (Fig. 2a),
whereas T. timopheevii aligned to the majority (427 Mb) of chromo-
some 2B (Fig. 2b). Overall, we identified 341 chromosomal segments
larger than 20 Mb with unique or rare fl-LTR insertion patterns that
were present in only 1 to 4 of the RQA genomes, of which 273 insertion
patterns were uniquely associated with a single genome (Supplemen-
tary Tables 13–16). The majority of unique regions were in PI190962
(spelt wheat; Triticum aestivum ssp. spelta), which was expected, given
that it diverged from modern bread wheat several thousand years ago.
A similar strategy was used to confirm RLC-Angela variation at the
telomeric region of chromosome 2A in Jagger, Mace, SY Mattis and
CDC Stanley (Fig. 2c), which corresponds to the 2NvS introgression
from Aegilops ventricosa (Supplementary Note 5). This introgression
is a well-known source of resistance to wheat blast30, and contains the
Lr37–Yr17–Sr38 gene cluster, which provides resistance to several rust
diseases31. Sequencing of A. ventricosa accessions (Supplementary
Table 12) followed by comparison of chromosomes with the RQAs con-
firmed that Jagger, Mace, SY Mattis and CDC Stanley carry the 2NvS
introgression, which spans about 33 Mb on chromosome 2A (Fig. 2c,
Extended Data Fig. 6a). We annotated the coding genes within this
region and identified 535 high-confidence genes; more than 10% were
predicted to be associated with disease resistance, including genes that
encode putative NB-ARC and NLRs (Extended Data Fig. 6b, Supplemen-
tary Tables 17, 18). Furthermore, we used genotyping by sequencing to
detect the 2NvS segment in three wheat panels and discovered that its
frequency has been increasing in breeding germplasm and its pres-
ence is consistently associated with higher grain yield (Extended Data

Nature | Vol 588 | 10 December 2020 | 279

Article
a LongReach Lancer chromosome 3D e Chinese Spring (5B/7B non-carrier)
i 762 325 305 1
7BL 7BS
ii
iii Translocation breakpoint
iv
5BS 5BL
T. ponticum 1 166 222 737

b LongReach Lancer chromosome 2B

1 428 480 993
i 7BL 5BL
ii
iii 5BS 7BS
iv 1 157 174 488
ArinaLrFor (5B/7B carrier)
T. timopheevii

c Jagger chromosome 2A f SY Mattis cytology (5B/7B carrier)

i
ii 1B
7D 7B
iii 7A
7BL/5BL 3D
iv
2B
4D 2D
A. ventricosa
6B 5D
No. of lines Min. Max. 7BS/5BS
1 2 3 4
3A 3B
RLC_Angela Read depth 10 μm

d 500 g
SY Mattis Hi-C Norin 61 Hi-C
Julius chromosome 4D

400 (5B/7B carrier) (5B/7B non-carrier)

400
position (Mb)

Chromosome 7B (Mb)

300
350
Centromere
200 shift
300
100
250

0
200
0 100 200 300 400 500 100 120 140 160 180 200 100 120 140 160 180 200
Chinese Spring chromosome 4D Chromosome 5B (Mb)
position (Mb)

Fig. 2 | Introgressions and large-scale structural variation in wheat.
a–c, T. ponticum introgression on chromosome 3D in LongReach Lancer (a),
T. timopheevi introgression on chromosome 2B in LongReach Lancer (b) and
A. ventricosa introgression on chromosome 3D in Jagger (c). Track i, map of
polymorphic RLC-Angela retrotransposon insertions (legend at bottom); track
ii, density of projected gene annotations from Chinese Spring (blue bars,
scaled to maximum value); track iii, per cent identity to Chinese Spring based
on chromosome alignment (yellow; scale is 0–100%); track iv, read depth of
wheat wild relatives (blue–yellow heat map; legend at bottom). d, Dot plot
alignment showing chromosome-level collinearity (black) with relative density
of CENH3 ChIP–seq mapped to 100-kb bins for Chinese Spring (blue) and Julius
(red); the arrow indicates a centromere shift. e, Robertsonian translocation
between chromosomes 5B and 7B in ArinaLrFor. f, g, Cytology (f) and Hi-C (g)
confirm the 5B/7B translocation in SY Mattis (left) compared with the
non-carrier Norin 61 (right). In f, five independent cells were observed; the
translocation was confirmed independently ten times. Scale bar, 10 μm.

Fig. 6c, d, Supplementary Tables 19, 20). Of note, we identified about
60 genes belonging to the cytochrome P450 superfamily, which have
been implicated in abiotic and biotic stress tolerance32 and have been
functionally validated to influence grain yield in wheat33. Together,
these data indicate that the modern wheat gene pool contains many
chromosomal segments of diverse ancestral origins, which can be iden-
tified by their transposable-element signatures. We also confirmed the
wild-relative origins of three introgressions within the RQA assemblies—
a first step towards characterizing causal genes for breeding targets,
such as resistance to wheat blast and rust fungi.
Centromere dynamics
Centromeres are vital for cell division and chromosome pairing during
meiosis. In plants, functional centromeres are defined by the epige-
netic placement of the modified histone CENH334. We therefore used
CENH3 chromatin immunoprecipitation and sequencing (ChIP–seq)35
to determine the positions and sizes (about 7.5–9.6 Mb) of the cen-
tromeres for each RQA (Supplementary Tables 21, 22), which were
consistent with previous estimates for wheat1. Furthermore, all chro-
mosomes showed a single active site, implying that previous reports
of multiple active centromeres in Chinese Spring1 were artefacts of
misoriented scaffolds. However, we found examples in which the rela-
tive position of the centromere was shifted owing to several pericentric
inversions, including inversions on chromosomes 4B and 5B (Extended
Data Fig. 7a, b). We also observed one instance in which the centro-
meric position changed, but was not associated with a structural event.
Specifically, on chromosome 4D in Chinese Spring, the centromere is
shifted by around 25 Mb relative to the consensus position (Fig. 2d).
This shift was previously recognized by cytology but was hypothesized
to result from a pericentric inversion36. However, the high degree of
collinearity between genomes supports the hypothesis that Cen4D in

280 | Nature | Vol 588 | 10 December 2020

 a b CDC Landmark
Paragon
Robigus
Mace
Adult
Claire
CDC Stanley
Chinese Spring
Weebill 1
Norin 61
Cadenza
Julius
ArinaLrFor
Larvae
LongReach Lancer
SY Mattis
Jagger
0 Mb
5 Mb Sm1
Healthy
CDC Landmark
Paragon
Robigus
Mace
Claire
CDC Stanley
Chinese Spring
Weebill 1
Damaged
Norin 61
Cadenza
Julius
ArinaLrFor
LongReach Lancer
SY Mattis
Jagger
Fig. 3 | Cloning of the gene Sm1. a, The orange wheat blossom midge oviposits
eggs on wheat spikes and the larvae feed on developing wheat grains, resulting
in moderate to severe damage to mature kernels. b, Top, sections of
chromosome 2B of the same colour in the same position share haplotypes
(based on 5-Mb bins), with the exception of those in grey, which indicates a
line-specific haplotype. The position of Sm1 is indicated with respect to the
CDC Landmark assembly. Bottom, zoomed-in view of haplotype blocks (based
on 250-kb bins) from 5 to 25 Mb positions on chromosome 2B, surrounding
Sm1. CDC Landmark, Robigus and Paragon all carry the same haplotype
Chinese Spring has shifted to a non-homologous position; this shifting
of centromeres to non-homologous sites has also been reported in
maize37. By characterizing the centromere positions for these diverse
wheat lines, we provide strong evidence for changes in centromere
position caused by structural rearrangements and centromere shifts.
Large-scale structural variation between genomes
Structural variants are common in wheat38, and impact genome struc-
ture and gene content. We characterized large structural variants
using pairwise genome alignments (Extended Data Fig. 1), changes in
three-dimensional topology of chromosomes revealed by Hi-C confor-
mation capture directionality biases along the genome39,40 (Extended
Data Fig. 8, Supplementary Table 23), which were confirmed by Oxford
Nanopore long-read sequencing (Extended Data Fig. 2) and cytological
karyotyping (Extended Data Fig. 7c, Supplementary Table 24, Sup-
plementary Note 6). The most prominent event was a translocation
between chromosomes 5B and 7B, observed in ArinaLrFor, SY Mattis
(Fig. 2e–g) and Claire. Normally, chromosomes 5B and 7B are approxi-
mately 737 and 762 Mb long, respectively, and we estimated that the
recombined chromosomes are 488 Mb (5BS/7BS) and 993 Mb (7BL/5BL)
long, making 7BL/5BL the largest wheat chromosome (Extended Data
Fig. 9a). In ArinaLrFor and SY Mattis, the 7BL/5BL breakpoint resides
within an approximately 5-kb GAA microsatellite, which we were
able to span using polymerase chain reaction (PCR) (Extended Data
Fig. 9b, c). By contrast, the breakpoint on 5BS/7BS was less syntenic,
and we detected polymorphic fluorescence in situ hybridization signals
between ArinaLrFor and SY Mattis on the 5BS portion of the translo-
cated chromosome segment, suggesting that the regions adjacent to
the translocation events differ on 5BS/7BS (Supplementary Note 6).
To determine the stability of the translocation in breeding, we geno-
typed for the translocation event in a panel of 538 wheat lines that
c
15.7 Mb
17.0 Mb
Sm1
Chinese Spring
(Sm1 non-carrier)
15.2 Mb
15.7 Mb
CDC Landmark
(Sm1 carrier)
Haplotypes
1 Sm1 carrier
2 Sm1 non-carrier
800 Mb
3 Sm1 non-carrier
25 Mb
1 Sm1 carrier
G182R
W 98*
2 Sm1 non-carrier
3 Sm1 non-carrier
(that is, Waskada)
NB-ARC LRR S/T kinase MSP
Transmembrane Mutations Alternative haplotype
surrounding Sm1 (teal). c, Top, anchoring of the Sm1 fine map to the physical
maps of Chinese Spring and CDC Landmark and graphical genotypes of three
haplotypes critical to localizing the Sm1 candidate gene. Bottom, annotation of
the Sm1 candidate gene, which encodes NB-ARC and LRR motifs in addition to
the integrated serine/threonine (S/T) kinase and MSP domains. Two
independent ethyl-methanesulfonate-induced mutations (W98* and G182R)
result in loss of function and susceptibility to the orange wheat blossom midge
(light blue lines). An alternative haplotype was observed in the kinase region of
Waskada (black).
represent most of the UK wheat gene pool grown since the 1920s41.
The translocation occurred in 66% of the lines and was selectively neu-
tral (Supplementary Note 7). Notably, the Ph1 locus on chromosome
5B, which controls the pairing of homeologous chromosomes during
meiosis42, is near the translocation breakpoint, but remained highly
syntenic between translocation carriers and non-carriers. Genetic
mapping and analysis of short-read sequencing data indicated that
the 5B/7B translocated chromosomes recombine freely with 5B and 7B
chromosomes (Extended Data Fig. 9d), suggesting that chromosome
pairing is not affected by the translocation.
Haplotype-based gene mapping
To develop improved wheat cultivars, breeders shuffle allelic vari-
ants by making targeted crosses and exploiting the recombination
that occurs during meiosis. These alleles, however, are not inherited
independently, but rather as haplotype blocks that often extend
across multiple genes that are in genetic linkage43,44. We quantified
haplotype variation along chromosomes across the assemblies, and
developed visualization software to support its utility (Supplemen-
tary Note 8). We used these haplotypes to characterize a locus that
provides resistance to the orange wheat blossom midge (OWBM, Sito-
diplosis mosellana Géhin), one of the most damaging insect pests of
wheat, which is endemic in Europe, North America, west Asia and the
Far East. Upon hatching, the first-instar larvae feed on the developing
grains and damage the kernels (Fig. 3a). Sm1 is the only gene in wheat
known to provide resistance to OWBM6. CDC Landmark, Robigus and
Paragon are all resistant to the OWBM, and all three carry the same
7.3-Mb haplotype within the Sm1 locus on chromosome 2B (Fig. 3b).
To identify Sm1 gene candidates, we used high-resolution genetic
mapping and refined the locus to a 587-kb interval in the CDC Land-
mark RQA (Fig. 3c, Extended Data Fig. 10a, Supplementary Table 25).
Nature | Vol 588 | 10 December 2020 | 281
Article
Through extensive genotyping of diverse breeding lines, we found an
OWBM-susceptible line, Waskada, that displayed a resistant haplo-
type except near one gene, which we annotated in CDC Landmark to
encode a canonical NLR with kinase and major sperm protein (MSP)
integrated domains (Fig. 3c). Oxford Nanopore long-read sequenc-
ing further confirmed the structure of the gene in CDC Landmark
(Extended Data Fig. 10b). By contrast, the remaining assemblies (sus-
ceptible to OWBM) lacked the NB-ARC domain, but the kinase and MSP
domains remained intact (Fig. 3c). We sequenced the Waskada allele
and found it contains the NB-ARC domain, but an alternative haplotype
within the kinase domain (Fig. 3c, Extended Data Fig. 10c). This gene
is expressed in wheat kernels and seedlings of Sm1 carrier lines, and
the lack of cDNA amplification of the NB-ARC domain for non-carrier
lines further supported an alternative gene structure (Extended Data
Fig. 10c). We generated two knockout-mutant lines of this candidate
gene in the Sm1 carrier line Unity45, and both were consistently rated
as susceptible to OWBM (Supplementary Table 26). Sequencing of the
candidate gene in these two mutants revealed a single point mutation
in each line: a G>A mutation resulting in a Gly>Arg (G182R) amino acid
substitution in the NB-ARC domain, and a G>A mutation, resulting in
a stop codon (W98*) before the NB-ARC domain (Fig. 3c). The kinase
domain encoded by Sm1 belongs to the serine/threonine class46, similar
to those of Rpg5, which provides stem rust resistance47, and Tsn1, which
encodes sensitivity to the necrotrophic effector ToxA produced by Par-
astagonospora nodorum and Pyrenophora tritici-repentis48; however,
both Rpg5 and Tsn1 lack the MSP domain. To our knowledge, this is the
first report of an NB-ARC-LRR-kinase-MSP coding gene associated with
insect resistance. Additional research is needed to functionally validate
these domains and their putative role in OWBM resistance using tools
such as gene editing. Nevertheless, we developed a high-throughput
and low-cost competitive allele-specific PCR marker (KASP) that dis-
criminates between OWBM-susceptible and OWBM-resistant lines with
perfect accuracy (Extended Data Fig. 10d, Supplementary Table 27).
Our analyses, along with the haplotype and synteny viewers (https://
kiranbandi.github.io/10wheatgenomes/, http://10wheatgenomes.
plantinformatics.io/ and http://www.crop-haplotypes.com/), laid
the foundation for identifying haplotypes for Sm1. Haplotypes can
now be genotyped in breeding programs using single-marker or
high-throughput-sequencing-based approaches, which can integrate
desirable genes into improved cultivars more efficiently.
Discussion
We have built on the genome-sequence resources available for wheat
and related species to produce ten RQAs and five scaffolded assem-
blies that represent hexaploid wheat lines from different regions,
growth habits and breeding programs1,11,12,18,20,49. We have identified
and characterized SNPs, PAV, CNV, centromere shifts, large-scale
structural variants and introgressions from wild relatives of wheat
that can be used to identify and characterize important breeding
targets. This was complemented by a transposable-element-analysis
approach to identify candidate introgressions from wild relatives of
wheat, for which we provided high-quality assemblies of segments
already used in global breeding programs. Together, these RQAs
present an opportunity for breeders and researchers to perform
high-resolution manipulation of genomic segments and pave the
way to identifying genes responsible for in-demand traits, as we
demonstrated for resistance to the insect pest OWBM. Functional
gene studies will also be facilitated by comparative gene analyses,
as exemplified by our analyses of orthologous groups, Rf genes and
NLR immune receptors26. Finally, we highlight haplotype blocks,
which will facilitate marker development for applied breeding 43,50.
Equipped with multiple layers of data describing variation in wheat,
we now have powerful tools to increase the rate of wheat improve-
ment to meet future food demands.
282 | Nature | Vol 588 | 10 December 2020
Online content
Any methods, additional references, Nature Research reporting sum-
maries, source data, extended data, supplementary information,
acknowledgements, peer review information; details of author con-
tributions and competing interests; and statements of data and code
availability are available at https://doi.org/10.1038/s41586-020-2961-x.
1. The International Wheat Genome Sequencing Consortium. Shifting the limits in wheat
research and breeding using a fully annotated reference genome. Science 361, eaar7191
(2018).
2. Montenegro, J. D. et al. The pangenome of hexaploid bread wheat. Plant J. 90, 1007–1013
(2017).
3. International Wheat Genome Sequencing Consortium (IWGSC). A chromosome-based
draft sequence of the hexaploid bread wheat (Triticum aestivum) genome. Science 345,
1251788 (2014).
4. He, F. et al. Exome sequencing highlights the role of wild-relative introgression in shaping
the adaptive landscape of the wheat genome. Nat. Genet. 51, 896–904 (2019); correction
51, 1194 (2019).
5. Pont, C. et al. Tracing the ancestry of modern bread wheats. Nat. Genet. 51, 905–911
(2019).
6. Kassa, M. T. et al. A saturated SNP linkage map for the orange wheat blossom midge
resistance gene Sm1. Theor. Appl. Genet. 129, 1507–1517 (2016).
7. Tadesse, W. et al. Genetic gains in wheat breeding and its role in feeding the world. Crop
Breed. Genet. Genom. 1, e190005 (2019).
8. Zhao, Q. et al. Pan-genome analysis highlights the extent of genomic variation in
cultivated and wild rice. Nat. Genet. 50, 278–284 (2018).
9. Dubcovsky, J. & Dvorak, J. Genome plasticity a key factor in the success of polyploid
wheat under domestication. Science 316, 1862–1866 (2007).
10. Marcussen, T. et al. Ancient hybridizations among the ancestral genomes of bread wheat.
Science 345, 1250092 (2014).
11. Avni, R. et al. Wild emmer genome architecture and diversity elucidate wheat evolution
and domestication. Science 357, 93–97 (2017).
12. Maccaferri, M. et al. Durum wheat genome highlights past domestication signatures and
future improvement targets. Nat. Genet. 51, 885–895 (2019).
13. Zimin, A. V. et al. The first near-complete assembly of the hexaploid bread wheat
genome, Triticum aestivum. Gigascience 6, 1–7 (2017).
14. Winfield, M. O. et al. Targeted re-sequencing of the allohexaploid wheat exome. Plant
Biotechnol. J. 10, 733–742 (2012).
15. Arora, D., Gross, T. & Brueggeman, R. Allele characterization of genes required for
rpg4-mediated wheat stem rust resistance identifies Rpg5 as the R gene. Phytopathology
103, 1153–1161 (2013).
16. Adamski, N. M. et al. A roadmap for gene functional characterisation in crops with large
genomes: lessons from polyploid wheat. eLife 9, e55646 (2020).
17. Uauy, C. Wheat genomics comes of age. Curr. Opin. Plant Biol. 36, 142–148 (2017).
18. Mascher, M. et al. A chromosome conformation capture ordered sequence of the barley
genome. Nature 544, 427–433 (2017).
19. Edwards, D. et al. Bread matters: a national initiative to profile the genetic diversity of
Australian wheat. Plant Biotechnol. J. 10, 703–708 (2012).
20. Jordan, K. W. et al. A haplotype map of allohexaploid wheat reveals distinct patterns of
selection on homoeologous genomes. Genome Biol. 16, 48 (2015).
21. Paape, T. et al. Patterns of polymorphism and selection in the subgenomes of the
allopolyploid Arabidopsis kamchatica. Nat. Commun. 9, 3909 (2018).
22. Paape, T. et al. Conserved but attenuated parental gene expression in allopolyploids:
Constitutive zinc hyperaccumulation in the allotetraploid Arabidopsis kamchatica. Mol.
Biol. Evol. 33, 2781–2800 (2016).
23. Melonek, J., Stone, J. D. & Small, I. Evolutionary plasticity of restorer-of-fertility-like
proteins in rice. Sci. Rep. 6, 35152 (2016).
24. Bernhard, T., Koch, M., Snowdon, R. J., Friedt, W. & Wittkop, B. Undesired fertility
restoration in msm1 barley associates with two mTERF genes. Theor. Appl. Genet. 132,
1335–1350 (2019).
25. Whitford, R. et al. Hybrid breeding in wheat: technologies to improve hybrid wheat seed
production. J. Exp. Bot. 64, 5411–5428 (2013).
26. Keller, B., Wicker, T. & Krattinger, S. G. Advances in wheat and pathogen genomics:
Implications for disease control. Annu. Rev. Phytopathol. 56, 67–87 (2018).
27. Steuernagel, B. et al. Rapid cloning of disease-resistance genes in plants using
mutagenesis and sequence capture. Nat. Biotechnol. 34, 652–655 (2016).
28. Bariana, H. S. et al. Mapping of durable adult plant and seedling resistances to stripe rust
and stem rust diseases in wheat. Aust. J. Agric. Res. 52, 1247–1255 (2001).
29. Chemayek, B. et al. Tight repulsion linkage between Sr36 and Sr39 was revealed by
genetic, cytogenetic and molecular analyses. Theor. Appl. Genet. 130, 587–595
(2017).
30. Cruz, C. D. et al. The 2NS translocation from Aegilops ventricosa confers resistance to the
Triticum pathotype of Magnaporthe oryzae. Crop Sci. 56, 990–1000 (2016).
31. Helguera, M. et al. PCR assays for the Lr37-Yr17-Sr38 cluster of rust resistance genes and
their use to develop isogenic hard red spring wheat lines. Crop Sci. 43, 1839–1847 (2003).
32. Li, Y. & Wei, K. Comparative functional genomics analysis of cytochrome P450 gene
superfamily in wheat and maize. BMC Plant Biol. 20, 93 (2020).
33. Gunupuru, L. R. et al. A wheat cytochrome P450 enhances both resistance to
deoxynivalenol and grain yield. PLoS ONE 13, e0204992 (2018).
34. Li, B. et al. Wheat centromeric retrotransposons: the new ones take a major role in
centromeric structure. Plant J. 73, 952–965 (2013).
35. Gent, J. I., Wang, K., Jiang, J. & Dawe, R. K. Stable patterns of CENH3 occupancy through
maize lineages containing genetically similar centromeres. Genetics 200, 1105–1116 (2015).
36. Koo, D. H., Sehgal, S. K., Friebe, B. & Gill, B. S. Structure and stability of telocentric
chromosomes in wheat. PLoS ONE 10, e0137747 (2015).
37. Schneider, K. L., Xie, Z., Wolfgruber, T. K. & Presting, G. G. Inbreeding drives maize
centromere evolution. Proc. Natl Acad. Sci. USA 113, E987–E996 (2016).
38. Saxena, R. K., Edwards, D. & Varshney, R. K. Structural variations in plant genomes. Brief.
Funct. Genomics 13, 296–307 (2014).
39. Harewood, L. et al. Hi-C as a tool for precise detection and characterisation of
chromosomal rearrangements and copy number variation in human tumours. Genome
Biol. 18, 125 (2017).
40. Himmelbach, A. et al. Discovery of multi-megabase polymorphic inversions by
chromosome conformation capture sequencing in large-genome plant species. Plant J.
96, 1309–1316 (2018).
41. Fradgley, N. et al. A large-scale pedigree resource of wheat reveals evidence for
adaptation and selection by breeders. PLoS Biol. 17, e3000071 (2019).
42. Martín, A. C., Rey, M. D., Shaw, P. & Moore, G. Dual effect of the wheat Ph1 locus on
chromosome synapsis and crossover. Chromosoma 126, 669–680 (2017).
43. Bevan, M. W. et al. Genomic innovation for crop improvement. Nature 543, 346–354
(2017).
44. Luján Basile, S. M. et al. Haplotype block analysis of an Argentinean hexaploid wheat
collection and GWAS for yield components and adaptation. BMC Plant Biol. 19, 553 (2019).
45. Fox, S. L. et al. Unity hard red spring wheat. Can. J. Plant Sci. 90, 71–78 (2010).
46. Hanks, S. K., Quinn, A. M. & Hunter, T. The protein kinase family: conserved features and
deduced phylogeny of the catalytic domains. Science 241, 42–52 (1988).
47. Brueggeman, R. et al. The stem rust resistance gene Rpg5 encodes a protein with
nucleotide-binding-site, leucine-rich, and protein kinase domains. Proc. Natl Acad. Sci.
USA 105, 14970–14975 (2008).
48. Faris, J. D. et al. A unique wheat disease resistance-like gene governs effector-triggered
susceptibility to necrotrophic pathogens. Proc. Natl Acad. Sci. USA 107, 13544–13549
(2010).
49. Luo, M. C. et al. Genome sequence of the progenitor of the wheat D genome Aegilops
tauschii. Nature 551, 498–502 (2017).
50. Borrill, P., Harrington, S. A. & Uauy, C. Applying the latest advances in genomics and
phenomics for trait discovery in polyploid wheat. Plant J. 97, 56–72 (2019).
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in
published maps and institutional affiliations.
Open Access This article is licensed under a Creative Commons Attribution
4.0 International License, which permits use, sharing, adaptation, distribution
and reproduction in any medium or format, as long as you give appropriate
credit to the original author(s) and the source, provide a link to the Creative Commons license,
and indicate if changes were made. The images or other third party material in this article are
included in the article’s Creative Commons license, unless indicated otherwise in a credit line
to the material. If material is not included in the article’s Creative Commons license and your
intended use is not permitted by statutory regulation or exceeds the permitted use, you will
need to obtain permission directly from the copyright holder. To view a copy of this license,
visit http://creativecommons.org/licenses/by/4.0/.
© The Author(s) 2020
1
Crop Development Centre, University of Saskatchewan, Saskatoon, Saskatchewan,
Canada. 2Grain Research Laboratory, Canadian Grain Commission, Winnipeg, Manitoba,
Canada. 3Department of Plant Pathology, Kansas State University, Manhattan, KS, USA.
4
Leibniz Institute of Plant Genetics and Crop Plant Research (IPK) Gatersleben, Seeland,
Germany. 5Helmholtz Zentrum München—German Research Center for Environmental
Health, Neuherberg, Germany. 6Aquatic and Crop Resource Development, National
Research Council Canada, Saskatoon, Saskatchewan, Canada. 7John Innes Centre, Norwich
Research Park, Norwich, UK. 8Department of Plant and Microbial Biology, University of
Zurich, Zurich, Switzerland. 9Morden Research and Development Centre, Agriculture and
Agri-Food Canada, Morden, Manitoba, Canada. 10Department of Computer Science,
University of Saskatchewan, Saskatoon, Saskatchewan, Canada. 11Brandon Research and
Development Centre, Agriculture and Agri-Food Canada, Brandon, Manitoba, Canada.
12
Genomics/Transcriptomics group, Functional Genomics Center Zurich, Zurich,
Switzerland. 13Department of Evolutionary Biology and Environmental Studies, University of
Zurich, Zurich, Switzerland. 14Institute of Agricultural Sciences, ETHZ, Zurich, Switzerland.
15
Kihara Institute for Biological Research, Yokohama City University, Yokohama, Japan. 16Life
Sciences Department, Natural History Museum, London, UK. 17Earlham Institute, Norwich
Research Park, Norwich, UK. 18The John Bingham Laboratory, NIAB, Cambridge, UK.
19
Department of Agronomy and Plant Genetics, University of Minnesota, Saint Paul, MN,
USA. 20Global Institute for Food Security, University of Saskatchewan, Saskatoon,
Saskatchewan, Canada. 21School of Plant Sciences and Food Security, Tel Aviv University,
Ramat Aviv, Israel. 22Department of Entomology, University of Manitoba, Winnipeg,
Manitoba, Canada. 23Institute of Crop Science, NARO, Tsukuba, Japan. 24Centre for
Biodiversity Genomics, University of Guelph, Guelph, Ontario, Canada. 25National Institute
of Advanced Industrial Science and Technology (AIST), Tokyo, Japan. 26Laboratory of Plant
Genetics, Graduate School of Agriculture, Kyoto University, Kyoto, Japan. 27Humanome Lab,
Tokyo, Japan. 28Global Wheat Program, International Maize and Wheat Improvement Center
(CIMMYT), Texcoco, Mexico. 29Montana BioAg, Missoula, MT, USA. 30Australian Research
Council Centre of Excellence in Plant Energy Biology, School of Molecular Sciences,
University of Western Australia, Perth, Western Australia, Australia. 31Ottawa Research and
Development Centre, Agriculture and Agri-Food Canada, Ottawa, Ontario, Canada.
32
Agriculture Victoria, AgriBio, Centre for AgriBioscience, Bundoora, Victoria, Australia.
33
Syngenta, Durham, NC, USA. 34School of Agriculture, Food and Wine, University of
Adelaide, Adelaide, South Australia, Australia. 35German Centre for Integrative Biodiversity
Research (iDiv) Halle-Jena-Leipzig, Leipzig, Germany. 36Biological and Environmental
Science & Engineering Division, King Abdullah University of Science and Technology,
Thuwal, Saudi Arabia. 37Graduate School of Life and Environmental Sciences, Kyoto
Prefectural University, Kyoto, Japan. 38Institute of Evolution and Department of Evolutionary
and Environmental Biology, University of Haifa, Haifa, Israel. 39School of Life Sciences
Weihenstephan, Technical University of Munich, Freising, Germany. 40Center for Integrated
Breeding Research (CiBreed), Georg-August-University Göttingen, Göttingen, Germany.
41
These authors contributed equally: Sean Walkowiak, Liangliang Gao, Cecile Monat.
✉e-mail: curt.mccartney@canada.ca; manuel.spannagl@helmholtz-muenchen.de; wicker@
botinst.uzh.ch; curtis.pozniak@usask.ca
Nature | Vol 588 | 10 December 2020 | 283
Article
Methods
No statistical methods were used to predetermine sample size. The
field experiments were randomized, but the wheat lines sequenced
and assembled were not selected at random. The investigators were
not blinded to allocation during experiments and outcome assessment.
Assemblies and annotation
Genome assemblies. We assembled the genomes of 15 diverse wheat
lines using two approaches (Supplementary Table 1). The RQA approach
used the DeNovoMAGIC v.3.0 assembly pipeline, previously used for
the wild emmer wheat11, durum wheat12 and Chinese Spring RefSeqv1.0
assemblies. In brief, high-molecular-weight DNA was extracted from
wheat seedlings as described previously51. Illumina 450-bp paired-end
(PE), 800-bp PE and mate-pair (MP) libraries of three different sizes (3
kb, 6 kb and 9 kb) were generated. Sequencing was performed at the
University of Illinois Roy J. Carver Biotechnology Center. 10X Genom-
ics Chromium libraries were prepared and sequenced at the Genome
Canada Genome Innovation Centre using the manufacturers’ recom-
mendations to achieve a minimum of 30 × coverage. Hi-C libraries were
prepared using previously described methods40. Using the Illumina PE,
MP, 10X Genomics Chromium, and Hi-C, chromosome scale assemblies
were prepared as described previously18. For cultivars assembled to a
scaffold level, we used the W2RAP-contigger using k = 200 (Supple-
mentary Note 1). Two MP libraries (10 kb and 13 kb) were produced for
each line except Weebill 1, for which two additional MP libraries were
used. Mate pairs were processed, filtered and used to scaffold contigs
as described in the W2RAP pipeline (https://github.com/bioinfolog-
ics/w2rap). Scaffolds of less than 500 bp were removed from the final
assemblies. Additionally, we performed Oxford Nanopore sequenc-
ing of CDC Landmark using R9 flow cells and the GridION sequencing
technology (Supplementary Note 2).
Nucleotide diversity analysis. The variant call format data files from
two wheat exome-capture studies4,5 were retrieved, combined, and
filtered to retain hexaploid accessions and polymorphisms detected
in both studies. The 10X Genomics Chromium sequencing data for
each of the RQA lines were aligned to Chinese Spring RefSeqv1.0 using
the LongRanger v.2.1.6 software. Alignment files from the accessions
assembled here and 16 Bioplatforms Australia lines19 with alignments
obtained from the DAWN project52 were then used for variant calling by
GATK v.3.8 at the same genomic positions identified by exome-capture
sequencing. The variant files from the exome-capture studies, DAWN
project and 10+ Wheat Genomes lines were then merged and subjected
to principal component analysis (PCA) using the prcomp function in
R v.3.6.1.
Gene projections. We used the previously published high-confidence
gene models for Chinese Spring to assess the gene content in each
assembly. Representative coding sequences of each informant locus
were aligned to pseudomolecules of each line separately using BLAT53
v.3.5 with the ‘fine’ parameter and a maximal intron size of 70 kb. BLAT
matches seeded an additional alignment by exonerate54 in the genomic
neighbourhood encompassing 20 kb upstream and downstream of the
match position. Exonerate alignments required a minimal and maximal
intron sizes of 30 bp and 20 kb, respectively. A linear regression of
colocalized matches with complete alignments of the informant were
computed for 10,000 such pairs to derive a normalization function
and to render comparable scoring schemes for both methods. Sub-
sequently, we selected the top-scoring match for each mapping pair
as the locus for the gene projection. Projections were then filtered by
alignment coverage (Supplementary Note 3), the open reading frame
(ORF) contiguity, the observed mapping frequency of the informant,
coverage of start and stop codons, and the orthology or potential dis-
location of the match scaffold relative to its informant chromosome.
Identification of orthologous groups was analogous to the approach
used previously55. Reciprocal best BLAST hit (RBH) graphs were derived
from pairwise all-against-all BLASTn v2.8 transcript searches (minimal
e-value ≤ 1 × 10−30). Hits were assigned to homeologous groups on the
basis of gene models of Chinese Spring following a previously described
homeologue classification9. Multiple sequence alignments for the
population genetics analysis were performed using MUSCLE v.3.8 with
default parameters (Supplementary Note 3). Using the gene projec-
tions, we quantified average pairwise genetic diversity (π), polymor-
phism (Watterson’s θW), and Tajima’s D using compute and polydNdS
in the libsequence v.1.0.3-1 package56. We retained diversity estimates
for genes that were in all of the genomes and had ≤100 segregating
sites. PAV was determined from the orthologous groups limited to
one-to-one relations where there was no match in at least one genome.
Analysis of the Rf-like gene family. For Rf genes, the genome se-
quences were scanned for ORFs in six frame translations with the getorf
program of the EMBOSS v.6.6.0 package. ORFs longer than 89 codons
were searched for the presence of PPR motifs using hmmsearch from
the HMMER v.3.2.1 package (http://hmmer.org) and the hidden Markov
models defined previously. The PF02536 profile from the Pfam v32.0
database (http://pfam.xfam.org) was used to screen for ORFs carrying
mTERF motifs. Downstream processing of the hmmsearch results fol-
lowed the pipeline described previously57. ORFs with low hmmsearch
scores were removed from the analysis as they are unlikely to repre-
sent functional PPR proteins. Only genes encoding mTERF proteins
longer than 100 amino acids were included in the analysis. RFL-PPR
sequences were identified as described23. The phylogenetic analyses
were performed as described previously23. Conserved, non-PPR genes
delimiting the borders of analysed RFL clusters were identified in the
Chinese Spring RefSeqv1.0 reference genome and used to search for
syntenic regions in the remaining wheat accessions with BLAST v.2.8.
See Supplementary Note 4 for more details.
NLR repertoire. NLR signatures were annotated using
NLR-Annotator58,59 (https://github.com/steuernb/NLR-Annotator)
with the option -a. We estimated redundancy of NLR signatures between
genomes at different thresholds of identity: 95%, 98% and 100%. For the
165 amino acids in the consensus of all NB-ARC motifs, this translates to
8, 3 and 0 mismatches of a concatenated motif sequence. To calculate
the overall redundancy in all genomes, we counted the number of LR
signatures added to a non-redundant set by adding genomes iteratively.
This was done for 1 million random permutations.
Repeat annotation. Transposons were detected and classified by a
homology search against the REdat_9.7_Poaceae section of the PGSB
transposon library60 using vmatch (http://www.vmatch.de) with the fol-
lowing parameters: identity ≥70%, minimal hit length 75 bp, seedlength
12 bp (exact command line: -d -p -l 75 -identity 70 -seedlength 12 -exdrop
5). To remove overlapping annotations, the output was filtered for
redundant hits via a priority-based approach in which higher-scoring
matches where assigned first and lower-scoring hits at overlapping
positions were either shortened or removed if there was ≥90% overlap
with a priority hit or if <50 bp remained. Tandem repeats where iden-
tified with TandemRepeatFinder v.4.09 under default parameters61
and subjected to overlap removal as described above. Full-length LTR
retrotransposons were identified with LTRharvest (http://genometools.
org/documents/ltrharvest.pdf). All candidates were subsequently an-
notated for PfamA domains using HMMER v.3.0 and filtered to remove
false positives, non-canonical hybrids and gene-containing elements.
The inner domain order served as a criterion for the LTR retrotranspo-
son superfamily classification, either Gypsy (RLG: RT-RH-INT), Copia
(RLC: INT-RT-RH) or undetermined (RLX). The insertion age of fl-LTRs
was calculated from the divergence between the 5′ and 3′ long terminal
repeats, which are identical upon insertion. The genetic distance was
calculated with EMBOSS v.6.6.0 distmat (Kimura2-parameter correc-
tion) using a random mutation rate of 1.3 × 10−8.
Analysis of centromeric regions. For each line with a RQA, ChIP
was performed according to previous methods62 with slight modi-
fication using a wheat-specific CENH3 antibody36. An antigen with
the peptide sequence RTKHPAVRKTKALPKK, corresponding to the
N terminus of wheat CENH3, was used to produce an antibody using
the custom-antibody production facility provided by Thermo Fisher
Scientific. The customized antibody was purified and obtained as pel-
lets. The antibody pellet (0.396 mg) was dissolved in 2 ml PBS buffer,
pH 7.4, resulting in a working concentration of 198 ng μl−1. Nuclei were
isolated from 2-week-old seedlings, digested with micrococcal nuclease
and incubated overnight at 4 °C with 3 μg of antibody or rabbit serum
(control). Antibodies were captured using Dynabeads Protein G and
the chromatin eluted using 100 μl of 1% sodium dodecyl sulfate, 0.1
M NaHCO3 preheated to 65 °C. DNA isolation was then performed us-
ing ChIP DNA Clean & Concentrator Kit, and ChIP–seq libraries were
constructed using TruSeq ChIP Library Preparation Kit and sequenced
with a NovoSeq S4, which generated 150-bp paired-end reads.
For Chinese Spring, we used two datasets, SRR168679963 (dataset 1)
and the dataset generated in this study (dataset 2). Sequence reads were
de-multiplexed, trimmed and aligned to each of the respective RQAs
using HISAT2 v.2.1.064. Alignments were sorted, filtered for minimum
alignment quality of 30, counted in 100-kb bins using samtools v.1.10
and BEDtools v.2.29, and visualized in R v.3.6.1. To define the midpoint of
each centromere, we identified the highest density of CENH3 ChIP–seq
reads using a smoothing spline in R v.3.6.1 with smooth.spline function
(number of knots = 1,000) and identified the peak of the smooth spline
as the centre of the respective centromere for a given chromosome.
To compare centromeric positions of different genomes, the CENH3
ChIP–seq density was plotted along with MUMmer v.4.0 chromosome
alignments. To determine the overall size of wheat centromeres, we
considered each 100-kb bin with CENH3 ChIP–seq read density that
was greater than three times the background (genome average) level of
read density to be an active centromeric bin. The number of enriched
bins for each genome were counted and averaged to a total of 21 chro-
mosomes. This calculation included counting of unanchored bins.
Analysis of introgressions
Identification of full-length RLC-Angela retrotransposons. Retro-
transposon profiles were created for each genome using the RLC-Angela
family65 and consensus sequences obtained from the TREP database
(www.botinst.uzh.ch/en/research/genetics/thomasWicker/trep-db.
html). First, BLASTn was used to compare the ~1,700-bp LTR of
RLC-Angela to each genome. Matching elements and 500 bp of flank-
ing sequences were aligned to identify precise LTR borders as well as
different sub-families and/or sequences variants. We then used BLASTn
to compare the 18 consensus LTR sequences against each genome and
then screened for pairs of full-length LTRs that are found in the same
orientation within a window of 7.5–9.5 kb (RLC-Angela elements are ~8.7
kb long). These initial candidate full-length elements were screened for
the presence of RLC-Angela polyprotein sequences by BLASTx, as well
as for the typical 5-bp target-site duplications. We allowed a maximum
of two mismatches between the two target-site duplications. All identi-
fied full-length RLC-Angela copies were then aligned to a RLC-Angela
consensus sequence with the program Water from the EMBOSS v.6.6.0
package (www.ebi.ac.uk/Tools/emboss/). These alignments were used
to compile all nucleotide polymorphisms into a single file. The variant
call file was then used for PCA using the snpgdsPCA function in the R
package SNPrelate v.3.11.
Sequencing of the tertiary gene pool of wheat. Genomic DNA
(gDNA) was extracted and purified from young leaf tissue collected
from multiple accessions of T. timopheevii, A. ventricosa and T. ponticum
(Supplementary Table 12) following a standard CTAB–chloroform
extraction method. Yield and integrity were evaluated by fluorom-
etry (Qubit 2.0) and agarose gel electrophoresis. Paired-end libraries
were prepared following the Nextera DNA Flex protocol. In brief, 500
ng gDNA from each accession was fragmented and amplified with a
limited-cycle PCR. Each library was uniquely dual-indexed with a dis-
tinct 10-bp index code (IDT for Illumina Nextera DNA UD) for multiplex-
ing, and quantified by qPCR (Kapa Biosystems). Final average library
size was estimated on a Tapestation 2200. Libraries were normalized
and pooled for sequencing on an Illumina NovaSeq 6000 S4 to generate
~5× coverage per genotype. Sequencing data were de-multiplexed and
aligned to appropriate RQAs (Supplementary Table 12) in semi-perfect
mode using the BBMap v.38 short-read alignment software (https://
sourceforge.net/projects/bbmap/).
Structural variation
We karyotyped the lines using mitotic metaphase chromosomes
prepared by the conventional acetocarmine-squash method.
Non-denaturing fluorescence in situ hybridization (ND-FISH) of
three repetitive sequence probes, Oligo-pSc119.2-1, Oligo-pTa535
and Oligo-pTa713, was performed as described66,67 (Supplementary
Note 6). Chromosomes were counterstained with DAPI. Chromo-
some images were captured with an Olympus BX61 epifluorescence
microscope and a CCD camera DP80. Images were processed and
pseudocoloured with ImageJ v.1.51n in the Fiji package. For karyo-
typing, at least four chromosomes per accession were examined
and compared to the karyotype of Chinese Spring as described pre-
viously68. Hierarchal clustering of karyotype polymorphisms was
performed using the Ward method in R v.3.0.2, which was used to
estimate distance. Next, we applied Hi-C analysis for inversion calling
as described previously40. In brief, adapters were removed and reads
were mapped to Chinese Spring using minimap2 v.2.1069 as we have
done previously21. The raw Hi-C link counts were calculated in 1 Mb
non-overlapping sliding windows and then normalized as described
in our previous work40. Finally, the normalized Hi-C link matrix was
subjected to inversion calling using R.
We performed flow cytometry of wheat cultivars Arina and Forno
as previously described70, except that we used a FACSAria SORP flow
cytometer and cell sorter (Becton Dickinson). The 5B/7B translocation
breakpoints were identified by comparison of chromosomes 5B and 7B
from ArinaLrFor and Julius. Sequence collinearity between ArinaLrFor
and Julius was detected by BLASTn searches of 1,000-bp sequence
windows every 100 kb along the chromosomes. Once an interruption
of synteny was detected, sequence segments at the positions of syn-
teny loss were extracted and used for local alignments to determine
the precise breakpoint positions. PCR amplification of the 5BS/7BS
and 7BL/5BL translocation sites was performed using standard PCR
cycling conditions.
Characterization of haplotypes
Development of a wheat genome haplotype database. To iden-
tify haplotypes, pairwise chromosome alignments were performed
between the RQA using MUMmer v.4.0, which were combined
with pairwise nucleotide BLASTn analyses of the genes ± 2,000 bp
using custom scripts in R v.3.6.1 (https://github.com/Uauy-Lab/
pangenome-haplotypes)71 (Supplementary Note 8). The resultant
haplotypes were uploaded to an interactive viewer (http://www.
crop-haplotypes.com/). Pairwise BLASTn comparisons of the genes
were also used to identify structural variants, and were uploaded into
AccuSyn (https://accusyn.usask.ca/) and SynVisio (https://synvisio.
github.io/#/) to create a wheat-specific database (https://kiranbandi.
github.io/10wheatgenomes/). Pretzel (https://github.com/plantinfor-
matics/pretzel) was also used to visualize and compare the RQA and
the projected gene annotations (http://10wheatgenomes.plantinfor-
matics.io/).
Article
Characterization of Sm1. Sm1-linked markers6 were located in RQAs
using BLAST v.2.8.0. Two high-resolution mapping populations were
developed, 99B60-EJ2D/Thatcher and 99B60-EJ2G/Infinity. Progeny
heterozygous for crossover events near Sm1 were identified in the F2
generation, and the crossovers were fixed in the F3 generation. The
resulting F2-derived F3 families were analysed with KASP markers within
the Sm1 region and tested for resistance to OWBM in field nurseries
to identify markers associated with Sm1. Ethyl methanesulfonate was
used to develop knockout mutants in the Sm1 gene. Approximately
3,200 seeds of the Canadian spring wheat variety Unity (an Sm1 carrier)
were soaked in a 0.2% (v/v) aqueous ethyl methanesulfonate solution
for 22 h at 22 °C. The seed was then rinsed in distilled water and sown
in a field nursery. The M1 seed was grown to maturity and bulk har-
vested. Approximately 6,000 M2 seeds were space planted in two field
nurseries located in Brandon and Glenlea, Manitoba, Canada. Spikes
were collected on a per-plant basis at maturity and were classified as
resistant, susceptible or undamaged as done previously6,72. Putative
Sm1-knockout mutants were re-tested for OWBM resistance in indoor
cage tests73 in the M3 and M4 generations. M4-derived families were
tested for resistance to OWBM in field nurseries (randomized complete
block design, six environments, and eight replicates per environment).
Candidate genes were identified between Sm1 flanking markers on
the CDC Landmark assembly using the projected gene annotations and
FGENESH v.2.6 (http://www.softberry.com/), which were compared to
the projected genes of non-carriers. Both 5′ and 3′ rapid amplification
of cDNA ends (5′ and 3′ RACE) were used to verify the transcription
initiation and termination sites of the gene candidate, whose structure
was predicted by FGENESH v.2.6. In brief, RNA was extracted from the
leaves of Unity (Sm1 carrier) seedlings (using the Qiagen RNeasy kit),
RACE PCR performed (Invitrogen GeneRacer kit), and the PCR product
cloned (Invitrogen TOPO TA Cloning kit for sequencing) and sequenced
by Sanger sequencing. Prediction of the conserved domains was done
using the NCBI Conserved Domain Search tool (https://www.ncbi.
nlm.nih.gov/Structure/cdd/wrpsb.cgi) and PROSITE (release 2020_01;
https://prosite.expasy.org/). The LRR domain was defined on the basis
of the presence of 2–42 LRR motif repeats of 20–30 amino acids each.
LRR motifs were manually annotated74. Prediction of transmembrane
regions and orientation was performed using the program TMpred
NCBI Conserved Domain Search tool (https://embnet.vital-it.ch/soft-
ware/TMPRED_form.html).
To study the expression of Sm1, total RNA was extracted from four
biological replicates from four wheat genotypes (Unity, CDC Landmark,
Waskada and Thatcher) from two different tissues; seedling leaves and
developing kernels (five days post anthesis) using NucleoSpin RNA Plant
kit (Macherey-Nagel) according to the manufacturer’s instructions.
RNA was treated with RNase-free DNase (rDNase) (Macherey-Nagel)
and reversed transcribed into cDNA using SuperScript IV Reverse Tran-
scriptase kit (Invitrogen) according to the manufacturer’s instructions
and the NB-ARC domain amplified by PCR.
Reporting summary
Further information on research design is available in the Nature
Research Reporting Summary linked to this paper.
Data availability
All sequence reads assemblies have been deposited into the National
Center for Biotechnology Information sequence read archive (SRA)
(see Supplementary Table 1 for accession numbers). Sequence reads
for the RQAs, T. ponticum, A. ventricosa and T. timopheevii have been
deposited into the SRA (accession no. PRJNA544491) and ChIP–seq
short read-data used for centromere characterization is deposited
under accession no. PRJNA625537. All Hi-C data have been deposited in
the European Nucleotide Archive (Supplementary Table 1). The RQAs
are available for direct user download at https://wheat.ipk-gatersleben.
de/. All assemblies and projected annotations are available for com-
parative analysis at Ensembl Plants (https://plants.ensembl.org/index.
html). Comparative analysis viewers are also online for synteny (https://
kiranbandi.github.io/10wheatgenomes/, http://10wheatgenomes.
plantinformatics.io/) and haplotypes (http://www.crop-haplotypes.
com/). Seed stocks of the assembled lines are available at the UK Germ-
plasm Resources Unit (https://www.seedstor.ac.uk/).
Code availability
Code for custom genome visualizers have been deposited in the
public domain for haplotype viewer (https://github.com/Uauy-Lab/
pangenome-haplotypes), Pretzel (https://github.com/plantinformat-
ics/pretzel), AccuSyn (https://github.com/jorgenunezsiri/accusyn) and
SynVisio (https://github.com/kiranbandi/synvisio). Additional scripts
used for ChIP–seq analysis of the centromeres are provided at https://
github.com/wheatgenetics/centromere.
51. Dvorak, J., Mcguire, P. E. & Cassidy, B. Apparent sources of the A genomes of wheats
inferred from polymorphism in abundance and restriction fragment length of repeated
nucleotide-sequences. Genome 30, 680–689 (1988).
52. Watson-Haigh, N. S., Suchecki, R., Kalashyan, E., Garcia, M. & Baumann, U. DAWN: a
resource for yielding insights into the diversity among wheat genomes. BMC Genomics
19, 941 (2018).
53. Kent, W. J. BLAT—the BLAST-like alignment tool. Genome Res. 12, 656–664 (2002).
54. Slater, G. S. & Birney, E. Automated generation of heuristics for biological sequence
comparison. BMC Bioinformatics 6, 31 (2005).
55. Tatusov, R. L., Galperin, M. Y., Natale, D. A. & Koonin, E. V. The COG database: a tool for
genome-scale analysis of protein functions and evolution. Nucleic Acids Res. 28, 33–36
(2000).
56. Thornton, K. Libsequence: a C++ class library for evolutionary genetic analysis.
Bioinformatics 19, 2325–2327 (2003).
57. Cheng, S. et al. Redefining the structural motifs that determine RNA binding and RNA
editing by pentatricopeptide repeat proteins in land plants. Plant J. 85, 532–547 (2016).
58. Steuernagel, B. et al. Physical and transcriptional organisation of the bread wheat
intracellular immune receptor repertoire. Preprint at https://doi.org/10.1101/339424 (2018).
59. Steuernagel, B. et al. The NLR-Annotator tool enables annotation of the intracellular
immune receptor repertoire. Plant Physiol. 183, 468–482 (2020).
60. Spannagl, M. et al. PGSB PlantsDB: updates to the database framework for comparative
plant genome research. Nucleic Acids Res. 44 (D1), D1141–D1147 (2016).
61. Benson, G. Tandem repeats finder: a program to analyze DNA sequences. Nucleic Acids
Res. 27, 573–580 (1999).
62. Nagaki, K. et al. Chromatin immunoprecipitation reveals that the 180-bp satellite repeat
is the key functional DNA element of Arabidopsis thaliana centromeres. Genetics 163,
1221–1225 (2003).
63. Guo, X. et al. De novo centromere formation and centromeric sequence expansion in
wheat and its wide hybrids. PLoS Genet. 12, e1005997 (2016).
64. Kim, D., Paggi, J. M., Park, C., Bennett, C. & Salzberg, S. L. Graph-based genome
alignment and genotyping with HISAT2 and HISAT-genotype. Nat. Biotechnol. 37, 907–
915 (2019).
65. Wicker, T. et al. Impact of transposable elements on genome structure and evolution in
bread wheat. Genome Biol. 19, 103 (2018).
66. Tang, Z., Yang, Z. & Fu, S. Oligonucleotides replacing the roles of repetitive sequences
pAs1, pSc119.2, pTa-535, pTa71, CCS1, and pAWRC.1 for FISH analysis. J. Appl. Genet. 55,
313–318 (2014).
67. Zhao, L. et al. Cytological identification of an Aegilops variabilis chromosome carrying
stripe rust resistance in wheat. Breed. Sci. 66, 522–529 (2016).
68. Komuro, S., Endo, R., Shikata, K. & Kato, A. Genomic and chromosomal distribution
patterns of various repeated DNA sequences in wheat revealed by a fluorescence in situ
hybridization procedure. Genome 56, 131–137 (2013).
69. Li, H. Minimap2: pairwise alignment for nucleotide sequences. Bioinformatics 34, 3094–
3100 (2018).
70. Kubaláková, M., Vrána, J., Cíhalíková, J., Simková, H. & Doležel, J. Flow karyotyping and
chromosome sorting in bread wheat (Triticum aestivum L.). Theor. Appl. Genet. 104,
1362–1372 (2002).
71. Brinton, J. et al. A haplotype-led approach to increase the precision of wheat breeding.
Commun. Biol. https://doi.org/10.1038/s42003-020-01413-2 (2020).
72. Thomas, J. et al. Chromosome location and markers of Sm1: a gene of wheat that
conditions antibiotic resistance to orange wheat blossom midge. Mol. Breed. 15, 183–192
(2005).
73. Lamb, R. J. et al. Resistance to Sitodiplosis mosellana (Diptera: Cecidomyiidae) in spring
wheat (Gramineae). Can. Entomol. 132, 591–605 (2000).
74. la Cour, T. et al. Analysis and prediction of leucine-rich nuclear export signals. Protein
Eng. Des. Sel. 17, 527–536 (2004).
Acknowledgements We are grateful for funding from the Canadian Triticum Applied Genomics
research project (CTAG2) funded by Genome Canada, Genome Prairie, the Western Grains
Research Foundation, Government of Saskatchewan, Saskatchewan Wheat Development
Commission, Alberta Wheat Commission, Viterra and Manitoba Wheat and Barley Growers
Association. Funding was also provided by the Biotechnology and Biological Sciences Research
Council (BBSRC) via the projects Designing Future Wheat (BB/P016855/1), sLOLA (BB/
J003557/1) and MAGIC Pangenome (BB/P010741/1, BB/P010733/1 and BB/P010768/1), by AMED
NBRP (JP17km0210142), the German Federal Ministry of Education and Research (FKZ
031B0190, WHEATSeq, 2819103915 and 2819104015), German Network for Bioinformatics and
Infrastructure de.NBI (FKZ 031A536A, 031A536B), German Federal Ministry of Food and
Agriculture (BMEL FKZ 2819103915 WHEATSEQ), Israel Science Foundation (Grant 1137/17), JST
CREST (JPMJCR16O3), US National Science Foundation (1339389), Kansas Wheat Commission
and Kansas State University, MEXT KAKENHI, The Birth of New Plant Species (JP16H06469,
JP16H06464, JP16H06466 and JP16K21727), National Agriculture and Food Research
Organization (NARO) Vice President Fund, Swiss Federal Office of Agriculture (NAP-PGREL),
Agroscope, Delley Seeds and Plants, ETH Zurich Institute of Agricultural Sciences, Fenaco
Co-operative, IP-SUISSE, swisssem, JOWA, SGPV-FSPC, Swiss National Science Foundation
(31003A_182318 and CRSII5_183578), University of Zurich Research Priority Program Evolution in
Action, King Abdullah University of Science and Technology, Grains Research and Development
Corporation (GRDC), Australian Research Council (CE140100008) and Groupe Limagrain. We
are grateful for the computational support of the Functional Genomics Center Zurich, the
Molecular Plant Breeding Group—ETH Zurich, and the Global Institute of Food Security (GIFS),
Saskatoon. We acknowledge the contribution of the Australian Wheat Pathogens Consortium
(https://data.bioplatforms.com/organization/edit/bpa-wheat-cultivars) in the generation of data
used in this publication. The Initiative is supported by funding from Bioplatforms Australia
through the Australian Government National Collaborative Research Infrastructure Strategy
(NCRIS). We thank S. Wu for DNA preparations for assembly and ChIP–seq library preparations;
O. Francisco-Pabalan and J. Santos, T. Wisk and S. Wolfe for their provision of OWBM images;
M. Knauft, I. Walde, S. König, T. Münch, J. Bauernfeind and D. Schüler for their contribution to
Hi-C data generation and sequencing, DNA sequencing and IT administration and sequence
data management; J. Vrána for karyotyping the wheat cultivars Arina and Forno; and R. Regier
for project management, administration and support.
Author contributions Project establishment: K.C., A.D., A. Hall, B.K., S.G.K., E.L., P.L.,
K.F.X.M., J.P., C.J.P., K.K.S., M.S. and N.S. Project coordination: A. Hall, C.J.P. and N.S.
Genome assemblies were contributed as follows: CDC Stanley and CDC Landmark: P.J.H.,
C.J.P., A.G.S., B.B., C.S.K., A.N., K.N. and S.W.; Julius: K.F.X.M., N.S., M.M., C.M. and U.S.;
Jagger: G.M., J.P. and L.G.; ArinaLrFor: B.K., S.G.K. and M.C.K.; Mace and LongReach Lancer:
K.C., P.L., G.K.-G. and J.T.; Norin 61: K.K.S., H.H., S.N., J.S., K. Kawaura, H.T., T. Tameshige, T.B.,
D.C., M.H., R.S.-I., C.A., F.K., J.G.-G. and N.S.; SY Mattis: E.L. and A.B.; spelt (PI190962): A.D.,
C.J.P. and J.D.; Robigus, Claire, Paragon and Cadenza: M.B., M.C., B.C., C.F., N.F. and D.H.;
Weebill 1: M.C., B.C., J.C., K.A.G., L.P.-A. and L.V. Sequencing, assembly and analysis were
contributed by WRA2P computational assembly: A. Hall, B.C., G.G.A., K. Krasileva, N.M.,
D.S. and J. Wright; 10X Genomics: H.B., C.J.P., J.E., S.K. and K.W.; Hi-C and structural
analysis: M.M., N.S., A. Himmelbach, C.M., S.P. and L.G.; pseudomolecule assemblies: M.M.,
C.M. and N.S.; gene projections and TE analysis: K.F.X.M., M.S., H.G. and G.H.; diversity and
polymorphism analysis: K.K.S., E.D., T.P., G.H.-N., D.C., M.H., G.H., H.H., H.K., M.S., K.M., T.
Tameshige, T. Tanaka, J.S. and J. Wu; centromere diversity: J.P. and D.H.K.; 5B/7B
translocation: S.G.K., T.W., J.C. and M.C.K; 2NvS introgression: J.P., A.K.F., L.G., P.J., C.J.P., R.S.
and S.W.; TE-based introgressions: T.W., B.B., J.E., M.C.K., J.P., C.J.P., J.T. and S.W; cytological
karyotyping: S.N., K.M., Y.N., J.S. and T.K.; diversification of Rf genes: J.M. and I.S.; NLR
repertoire: S.G.K. and B.S.; Sm1 gene cloning: C.A.M., C.J.P., C.U., J.B., A.C.C., S.C., P.F.,
M.T.K., V.K., D.T. and K.W.; haplotype database: C.U., J.B. and R.H.R.-G.; visualization software:
C.G., V.B., G.K.-G., J.N.S., J.T. and J.M.; BLAST server: M.M., A.F. and U.S.; C.J.P and S.W.
drafted the manuscript with input from all authors. All co-authors contributed to and edited
the final version.
Competing interests The authors declare no competing interests.
Additional information
Supplementary information is available for this paper at https://doi.org/10.1038/s41586-020-
2961-x.
Correspondence and requests for materials should be addressed to C.A.M., M.S., T.W. and C.J.P.
Peer review information Nature thanks Victor Albert, Rudi Appels and the other, anonymous,
reviewer(s) for their contribution to the peer review of this work.
Reprints and permissions information is available at http://www.nature.com/reprints.
Article

Extended Data Fig. 1 | Chromosome-scale collinearity between the RQA. Lancer (red rectangles) and 5B/7B translocation in SY Mattis and ArinaLrFor
Genomes were aligned chromosome by chromosome using MUMmer and are (purple rectangles) are indicated.
represented as dot plots. The introgression on chromosome 2B of LongReach
Extended Data Fig. 2 | Evaluation of the CDC Landmark RQA using Oxford
Nanopore Long Reads. a, Scaffold-scaffold long read contact map showing
shared read IDs between scaffold ends along the ordered scaffolds in the CDC
Landmark pseudomolecules. The diagonal pattern indicates that adjacent
scaffolds share the same long reads and are therefore properly ordered and
oriented by Hi-C in the RQA. b, Characterization of inversion events on
chromosomes 2A, 3A, and 3D. The directionality biases estimated from
alignments of Hi-C data against Chinese Spring (left, top), and chromosome
alignment of the inversion events between CDC Landmark and Chinese Spring
RQAs (left, bottom) are shown. Long reads spanning the inversion events and
magnified views of the reads aligning to the left and right boundaries of the
inversions (right) are provided.
Article
Extended Data Fig. 3 | Diversity of genes and TEs. a, Average pairwise genetic
diversity of the homeologues (coding sequences only) of the A, B and D
subgenomes. The mode of the A, B and D subgenome is 0.00057, 0.00082, and
0.0002, respectively. b, Tajima’s D estimates of coding sequences for each
wheat subgenome. The lower and upper range of the boxplot hinges
correspond to the first and third quartiles (the 25th and 75th percentiles).
Boxplots show centre line, median; box limits, upper and lower quartiles;
whiskers, 1.5 × interquartile range. c, Total gene counts and orthologues for the
RQA. Genes in orthologous groups with exactly one gene for each line
(Complete; dark brown), genes contained in unambiguous orthologous groups
missing an orthologue for at least one line, that is, PAV (2-10 Lines; light brown),
and genes with ambiguous orthologues or CNV (Other; pink) are indicated. d,
Per cent of pairwise shared syntenic fl-LTRs between wheat lines.
Extended Data Fig. 4 | Evolutionary relationships among PPR and mTERF
gene sequences. a, The RFL clade is in blue and all remaining P-class PPRs are in
green. b, Clustered mTERF sequences are in blue and the remaining mTERFs are
shown in green. The scale bar represents number of substitutions per site. c,
Sequence inversions and copy number variation at the Rf3 locus on
chromosome 1B. RFL genes are shown as light pink triangles above the
chromosome scale. Conserved non-PPR genes used as syntenic anchors are
shown on the chromosome scale as coloured triangles. The total number (T)
and the number of putatively functional RFL genes with 10 or more PPR motifs
(F) are indicated on the right side of each panel.
Article
Extended Data Fig. 5 | Identification of alien introgressions from wheat
relatives. A feature of foreign chromosomal introgressions is that they contain
unique patterns of TE insertions. Shown are stretches of >20 Mb containing
multiple polymorphic RLC-Angela retrotransposons that are found only in one
or a few (≤4) of the sequenced lines. One representative chromosome for each
wheat subgenome is shown. Individual polymorphic retrotransposons are
indicated as coloured vertical lines. Colours correspond to the number of
cultivars a foreign segment is found in. Regions of particular interest are
indicated by black rectangles. These include the 2NvS alien introgression from
A. ventricosa at the end of chromosome 2A in Jagger, Mace, SY Mattis and CDC
Stanley, as well as introgression in the central region of chromosome 2B from
T. timopheevi in LongReach Lancer, and introgression at the end of
chromosome 3D from T. ponticum in LongReach Lancer.
Extended Data Fig. 6 | Detailed characterization of the 2NvS introgression
from A. ventricosa. a, Pairwise alignments of the first 50 Mb of chromosome
2A. The black arrow indicates a possible unique haplotype within spelt. b,
Orthologous genes between the 2NvS introgression from A. ventricosa in Jagger
and the genes on chromosomes 2A, 2B, and 2D in Chinese Spring. c, Frequency
of 2NvS introgression carriers in North American datasets from CIMMYT,
Kansas State, and the USDA Winter Wheat Regional Performance Nursery
(RPN) over time. d, Per cent yield difference in lines that carry the 2NvS
introgression. Two sided t-tests were performed to test for the significance of
the impact of the 2NvS introgression. **P < 0.01; ***P < 0.001.
Article
Extended Data Fig. 7 | Centromere positions and karyotype variation.
Functional centromere positions in the RQA have undergone structural and
positional rearrangement. Chromosome alignments showing collinearity
(black scaffolds in same orientation, grey scaffolds in opposite orientation)
with relative density of CENH3 ChIP–seq mapped to 100 kb genomic bins for
Chinese Spring (blue) and a representative genome of comparison (red) for
chromosome 4B of CDC Stanley (a), and chromosome 5B of Julius (b). c,
Detailed list and clustering of cytological features carried by each wheat line
(Supplementary Note 6). Features that are identical (dark grey) or have a gain
(black) or loss (light grey) relative to Chinese Spring are indicated.
Extended Data Fig. 8 | Hi-C validates inversions identified from pairwise Spring are shown. Boundaries of diagonal segments are indicative of inversions
chromosome alignments. Pairwise alignments of chromosome 6B from the and coincide with inversion boundaries identified from the chromosome
RQA and Chinese Spring are shown. Above each alignment dot plot, the alignments.
directionality biases estimated from alignments of Hi-C data against Chinese
Article
Extended Data Fig. 9 | Characterization of a translocation involving wheat
chromosomes 5B and 7B. a, Cytogenetic karyotypes of Forno (left) and Arina
(right), the parents of ArinaLrFor. Note that the large recombinant
chromosome 7B is represented by a distinct peak. b, Sequence of the
translocation breakpoint on chromosome 7B of ArinaLrFor. Note that the exact
breakpoint lies in a sequence gap (stretch of Ns). The bp positions are indicated
at the left. Forward PCR primers are shown in red and reverse primers in blue.
The overlap of the two reverse primers is shown in purple. The outer primer
pair was used for PCR, while the inner pair was used for a nested PCR. c, PCR
amplification of the fragment spanning the translocation breakpoint. The
nested PCR yielded a ~5 kb fragment that spanned the translocation breakpoint
and its identity was confirmed by sequencing. Both PCR and nested PCR were
performed in duplicate; both replicates of the nested PCR were sequenced
using the Sanger method. For gel source data, see Supplementary Fig. 1. d,
Mapping of Illumina reads from the cultivars Arina and Forno on to the
pseudomolecules of ArinaLrFor. Sequence derived from Forno is shown in blue,
while sequenced derived from Arina is in red. Note that chromosomes 5B and
7B are derived from both parents, indicating that these parental chromosomes
can recombine freely, despite the presence of a large 5B/7B translocation in
Arina.
Extended Data Fig. 10 | See next page for caption.
Article
Extended Data Fig. 10 | Confirmation of gene expression and gene
structure for Sm1. a, Critical recombinants from the 99B60-EJ2G/Infinity and
99B60-EJ2D/Thatcher populations used to fine map Sm1. The 99B60-EJ2G/
Infinity cross had 5,170 F2 plants, while 99B60-EJ2D/Thatcher cross had 5,264
F2 plants; only recombinant haplotypes between orange wheat blossom midge
resistant (R) and susceptible (S) genotypes are shown. b, Oxford Nanopore
long read confirmation of the Sm1 gene candidate in the CDC Landmark RQA
(left), and alternative haplotype in Chinese Spring (right). Vertical coloured
lines indicate sequence variants. c, Amplification of cDNA for the NB-ARC
domain of the Sm1 gene candidate (top) and actin control (bottom) derived
from RNA isolated from developing kernels (left) and wheat seedlings (right).
Unity and CDC Landmark are carriers of Sm1. Waskada carries an alternative
haplotype and does not carry Sm1 (see main text). Thatcher was used as a
susceptible parent for fine mapping of Sm1 and does not contain the associated
NB-ARC domain. The experiment was replicated on four independent
biological samples for each condition. d, Distribution of an Sm1 allele-specific
PCR marker in a diverse panel of >300 wheat lines.
 nature research | reporting summary
Corresponding author(s): Curtis Pozniak
Last updated by author(s): Aug 11, 2020

Reporting Summary
Nature Research wishes to improve the reproducibility of the work that we publish. This form provides structure for consistency and transparency
in reporting. For further information on Nature Research policies, see Authors & Referees and the Editorial Policy Checklist.

Statistics
For all statistical analyses, confirm that the following items are present in the figure legend, table legend, main text, or Methods section.
n/a Confirmed
The exact sample size (n) for each experimental group/condition, given as a discrete number and unit of measurement
A statement on whether measurements were taken from distinct samples or whether the same sample was measured repeatedly
The statistical test(s) used AND whether they are one- or two-sided
Only common tests should be described solely by name; describe more complex techniques in the Methods section.

A description of all covariates tested

A description of any assumptions or corrections, such as tests of normality and adjustment for multiple comparisons
A full description of the statistical parameters including central tendency (e.g. means) or other basic estimates (e.g. regression coefficient)
AND variation (e.g. standard deviation) or associated estimates of uncertainty (e.g. confidence intervals)

For null hypothesis testing, the test statistic (e.g. F, t, r) with confidence intervals, effect sizes, degrees of freedom and P value noted
Give P values as exact values whenever suitable.

For Bayesian analysis, information on the choice of priors and Markov chain Monte Carlo settings
For hierarchical and complex designs, identification of the appropriate level for tests and full reporting of outcomes
Estimates of effect sizes (e.g. Cohen's d, Pearson's r), indicating how they were calculated
Our web collection on statistics for biologists contains articles on many of the points above.

Software and code

Policy information about availability of computer code
Data collection No software was used to collect data for this study.

Data analysis A multitude of software and databases were used in this study, all of which have been listed, cited, or provided. These include:
DeNovoMAGIC v3.0, W2RAP (no versions, https://github.com/bioinfologics/w2rap), LongRanger v2.1.6, GATK v3.8, R v3.6.1 and v3.0.2,
BLAT v3.5, BLAST v2.8 , MUSCLE v3.8, libsequence v1.8.3, EMBOSS v6.6.0, HMMER 3.1b2, PFAM v32.0, NLR-Annotator (no version,
https://github.com/steuernb/NLR-Annotator), Vmatch v2.3.0, TandemRepeatFinder v4.07b, LTRharvest genometools-1.5.9, HMMER
v3.0, MUMmer v3.23 (haplotype database) and v4 (all other analyses), HISAT v2.1.0, SNPrelate v3.11, BBTools/BBMap v38, ImageJ
v1.51n, minimap2 v2.13, FGENESH v2.6, NCBI Conserved Domain Search tool (no version, https://www.ncbi.nlm.nih.gov/Structure/cdd/
wrpsb), PROSITE release 2020_01, TMpred v25, STAR v2.6.0b., AUGUSTUS v3.2.3., GMAP v2017-06-20, EvidenceModeler v1.1.1, AHRD
v1.6, MCScanX v2.0, samtools v1.10, BEDtools v2.29, and custom data scripts (https://github.com/Uauy-Lab/pangenome-haplotypes;
http://people.beocat.ksu.edu/~jpoland/centromeres/).
For manuscripts utilizing custom algorithms or software that are central to the research but not yet described in published literature, software must be made available to editors/reviewers.
We strongly encourage code deposition in a community repository (e.g. GitHub). See the Nature Research guidelines for submitting code & software for further information.

Data
October 2018

Policy information about availability of data

All manuscripts must include a data availability statement. This statement should provide the following information, where applicable:
- Accession codes, unique identifiers, or web links for publicly available datasets
- A list of figures that have associated raw data
- A description of any restrictions on data availability

All sequence reads have been deposited into the National Center for Biotechnology Information sequence read archive (SRA) (see Supplementary Table 1 for
accession numbers). Sequence reads for the RQAs, Th. ponticum, Ae. ventricosa and T. timopheevii have been deposited into the SRA (no. PRJNA544491) and ChIP-

1
seq short read-data used for centromere characterization is deposited as PRJNA625537. All Hi-C data has been deposited in the European Nucleotide Archive
(Supplementary Table 1). The RQAs and projected annotations are available for direct user download at https://wheat.ipk-gatersleben.de/. All RQA assemblies have

nature research | reporting summary

also been deposited at EBI with the following accession numbers: GCA_903993795; GCA_903993985; GCA_903993975; GCA_903994175; GCA_903994195;
GCA_904066035; GCA_903994155; GCA_903994165; GCA_903994185; GCA_903995565. These data will be syncrhonized across multiple platforms including NCBI
and at Ensembl Plants (https://plants.ensembl.org/index.html). Comparative analysis viewers are also online for synteny (https://
kiranbandi.github.io/10wheatgenomes/; http://10wheatgenomes.plantinformatics.io/) and haplotypes (http://www.crop-haplotypes.com/). Seed stocks of the
assembled lines are available at the UK Germplasm Resources Unit (https://www.seedstor.ac.uk/).

Field-specific reporting
Please select the one below that is the best fit for your research. If you are not sure, read the appropriate sections before making your selection.
Life sciences Behavioural & social sciences Ecological, evolutionary & environmental sciences
For a reference copy of the document with all sections, see nature.com/documents/nr-reporting-summary-flat.pdf

Life sciences study design

All studies must disclose on these points even when the disclosure is negative.
Sample size No statistical methods were used to establish sample size. The samples that were sequenced were selected to represent modern breeding
material from different continents that had known differences in pedigree and were known to carry different genes/traits/chromosomal
segments of interest.

Data exclusions All sequencing data generated was used in the genome assembly and analyses. Whenever possible, all data was included in the supporting
analyses. Data exclusion applies only to some of the subsequent supporting analysis, which was pre-established based on limitations in the
data. For example, we excluded the scaffolded assemblies from some analyses because the analyses required chromosome
pseudomolecules. We performed diversity analysis both with the spelt genome but also excluding the spelt genome because it is a different
species and is much more diverged and biased the results.

Replication In all analyses that support the genome assemblies, the number of replicates or iterations are indicated in materials and methods or
supplemental tables. In each case, all replications were successful and were used. The genome assemblies themselves were validated using
multiple methods (i.e. BUSCO, genetic maps, HiC, 10x Genomics, cytology, and comparions to Chinese Spring). The CDC Landmark assembly
was further validated using Oxford Nanopore long read sequencing. This helped validate the other approaches.

Randomization Randomization does not directly apply to the genome sequencing and assembly; however it applies to some of the supporting analyses. In
these cases, the group design and data seeding for computational analysis are described in the materials and methods and adhere to widely
accepted standards. For example, analysis of NLRs (Fig. 1c), 1 million random permutations were used. For the field experiments established
for phenotyping of Sm1, all samples were replicated and randomized using appropriate experimental designs.

Blinding Blinding does not apply to this study, as the study focuses on genome sequencing. This study focuses on plants genomics and the results of
the study are not impacted by the concealment of treatment, data, or groups.

Reporting for specific materials, systems and methods

We require information from authors about some types of materials, experimental systems and methods used in many studies. Here, indicate whether each material,
system or method listed is relevant to your study. If you are not sure if a list item applies to your research, read the appropriate section before selecting a response.

Materials & experimental systems Methods

n/a Involved in the study n/a Involved in the study
Antibodies ChIP-seq
Eukaryotic cell lines Flow cytometry
Palaeontology MRI-based neuroimaging
Animals and other organisms
Human research participants
Clinical data
October 2018

Antibodies
Antibodies used Chromatin immunoprecipitation (ChIP) was performed ausing wheat CenH3 antibody (Koo et al., 2015). A antigen with the
peptide sequence ‘RTKHPAVRKTKALPKK’ corresponding to the N-terminus of wheat CENH3 was used to produce antibody
utilizing the custom-antibody production facility provided by the Thermo Fisher Scientific, Illinois, USA (abs@thermofisher.com).
A 0.396 mg of the antibody pellet was dissolved in 2 ml of PBS buffer, pH 7.4 resulting in 198 ng/uL of the working concentration.

Validation In the manuscript, we validate the antibody according to a previous study of Chinese Spring (Koo et al., 2015) and achieved near

2
 identical results (Supplementary Table 12). Additional controls were used in the study where the antibody was substituted with
rabbit serum, which serves as nonspecific binding control in chromatin immunoprecipitation assay.

nature research | reporting summary

ChIP-seq
Data deposition
Confirm that both raw and final processed data have been deposited in a public database such as GEO.
Confirm that you have deposited or provided access to graph files (e.g. BED files) for the called peaks.

Data access links The data for the project has been deposited at NCBI: PRJNA625537 and analysis files are available for download: http://
May remain private before publication. people.beocat.ksu.edu/~jpoland/centromeres/

Files in database submission BED files, delta files (MUMmer), data analysis scripts

Genome browser session Data for visualization is available at http://people.beocat.ksu.edu/~jpoland/centromeres/

(e.g. UCSC)

Methodology
Replicates NA. Samples were obtained from 2-week-old seedlings.

Sequencing depth Paired-end reads were generated at varying levels of read depth, data was deposited at NCBI (PRJNA625537).

Antibodies Wheat CenH3 antibody - see: Koo DH, Sehgal SK, Friebe B, Gill BS (2015) Structure and stability of telocentric chromosomes
in wheat. PLoS One 10: e0137747.

Peak calling parameters Reads mapped per 100kb bin were counted for each sample using BEDtools and output as a bed file. Scripts for data
analysis are provided at http://people.beocat.ksu.edu/~jpoland/centromeres/. Unlike studies involving transcription factors,
CENH3 ChIP-seq provides clear distinct peaks that are ~100 fold greater than background.

Data quality SAM output files from HISAT2 were converted to BAM, sorted and filtered for minimum alignment quality of 30 using
SAMtools.

Software Reads for each sample were aligned to each of the respective genome assemblies using HISAT2.Reads mapped per 100kb
bin were counted for each sample using BEDtools and output as a bed file. Scripts for data analysis are provided at http://
people.beocat.ksu.edu/~jpoland/centromeres/.

October 2018

3
Article

The barley pan-genome reveals the hidden

legacy of mutation breeding

https://doi.org/10.1038/s41586-020-2947-8 Murukarthick Jayakodi1,20, Sudharsan Padmarasu1,20, Georg Haberer2,

Venkata Suresh Bonthala2, Heidrun Gundlach2, Cécile Monat1, Thomas Lux2, Nadia Kamal2,
Received: 3 April 2020
Daniel Lang2, Axel Himmelbach1, Jennifer Ens3, Xiao-Qi Zhang4, Tefera T. Angessa4,
Accepted: 9 September 2020 Gaofeng Zhou4,5, Cong Tan4, Camilla Hill4, Penghao Wang4, Miriam Schreiber6,
Lori B. Boston7, Christopher Plott7, Jerry Jenkins7, Yu Guo1, Anne Fiebig1, Hikmet Budak8,
Published online: 25 November 2020
Dongdong Xu9, Jing Zhang9, Chunchao Wang9, Jane Grimwood7, Jeremy Schmutz7,
Open access Ganggang Guo9, Guoping Zhang10, Keiichi Mochida11,12,13, Takashi Hirayama13, Kazuhiro Sato13,
Kenneth J. Chalmers14, Peter Langridge14, Robbie Waugh6,14,15, Curtis J. Pozniak3, Uwe Scholz1,
Check for updates
Klaus F. X. Mayer2,16, Manuel Spannagl2, Chengdao Li4,5,17 ✉, Martin Mascher1,18 ✉ & Nils Stein1,19 ✉

Genetic diversity is key to crop improvement. Owing to pervasive genomic structural

variation, a single reference genome assembly cannot capture the full complement of
sequence diversity of a crop species (known as the ‘pan-genome’1). Multiple
high-quality sequence assemblies are an indispensable component of a pan-genome
infrastructure. Barley (Hordeum vulgare L.) is an important cereal crop with a long
history of cultivation that is adapted to a wide range of agro-climatic conditions2. Here
we report the construction of chromosome-scale sequence assemblies for the
genotypes of 20 varieties of barley—comprising landraces, cultivars and a wild
barley—that were selected as representatives of global barley diversity. We catalogued
genomic presence/absence variants and explored the use of structural variants for
quantitative genetic analysis through whole-genome shotgun sequencing of
300 gene bank accessions. We discovered abundant large inversion polymorphisms
and analysed in detail two inversions that are frequently found in current elite barley
germplasm; one is probably the product of mutation breeding and the other is tightly
linked to a locus that is involved in the expansion of geographical range. This
first-generation barley pan-genome makes previously hidden genetic variation
accessible to genetic studies and breeding.

A staple food of ancient civilizations, today barley is used mainly for in gene content and copy number in the control of agronomic traits.
animal feed and malting. Barley is more adaptable to harsh environ- The concept of the pan-genome refers to a species-wide catalogue
mental conditions than its close relative wheat, and maintains an impor- of genic presence/absence variation (PAV)12, or more generally,
tant role in human nutrition in harsh climatic regions that include structural variation that affects (potentially non-coding) sequences
the Ethiopian and Tibetan highlands2. As in other crops, genomics of 50 or more base pairs (bp) in size. Although several methods of
has been a major driver of progress in barley genetics and breeding pan-genomic analysis that use short-read sequence data in the con-
in the past decade3. The first draft reference genome for barley4, and text of a single reference genome have been devised13, large and com-
its subsequent revisions5,6, have formed the basis for gene isolation7, plex genomes require multiple high-quality sequence assemblies to
compiling a single-nucleotide polymorphism (SNP) variation atlas capture and contextualize sequences that are absent in—or highly
for wild and domesticated germplasm8, and activating plant genetic diverged from—a single reference genotype14. Progress in sequenc-
resources9. At the same time, reduced-representation surveys of struc- ing and genome mapping technologies has only recently made pos-
tural variation10 and map-based cloning11 have implicated variation sible the fast and cost-effective assembly of tens of genotypes of
1
Leibniz Institute of Plant Genetics and Crop Plant Research (IPK) Gatersleben, Seeland, Germany. 2Plant Genome and Systems Biology (PGSB), Helmholtz Center Munich, German Research
Center for Environmental Health, Neuherberg, Germany. 3Department of Plant Sciences, University of Saskatchewan, Saskatoon, Saskatchewan, Canada. 4Western Barley Genetics
Alliance, State Agricultural Biotechnology Centre, College of Science, Health, Engineering and Education, Murdoch University, Murdoch, Western Australia, Australia. 5Agriculture and
Food, Department of Primary Industries and Regional Development, South Perth, Western Australia, Australia. 6The James Hutton Institute, Dundee, UK. 7HudsonAlpha, Institute for
Biotechnology, Huntsville, AL, USA. 8Montana BioAg Inc, Missoula, MT, USA. 9Institute of Crop Sciences, Chinese Academy of Agricultural Sciences (ICS-CAAS), Beijing, China. 10College of
Agriculture and Biotechnology, Zhejiang University, Hangzhou, China. 11Bioproductivity Informatics Research Team, RIKEN Center for Sustainable Resource Science, Yokohama, Japan.
12
Kihara Institute for Biological Research, Yokohama City University, Yokohama, Japan. 13Institute of Plant Science and Resources, Okayama University, Kurashiki, Japan. 14School of
Agriculture, Food and Wine, University of Adelaide, Glen Osmond, South Australia, Australia. 15School of Life Sciences, University of Dundee, Dundee, UK. 16School of Life Sciences
Weihenstephan, Technical University of Munich, Freising, Germany. 17Hubei Collaborative Innovation Centre for Grain Industry, Yangtze University, Jingzhou, China. 18German Centre for
Integrative Biodiversity Research (iDiv) Halle-Jena-Leipzig, Leipzig, Germany. 19Center for Integrated Breeding Research (CiBreed), Georg-August-University Göttingen, Göttingen, Germany.
20
These authors contributed equally: Murukarthick Jayakodi, Sudharsan Padmarasu. ✉e-mail: C.Li@murdoch.edu.au; mascher@ipk-gatersleben.de; stein@ipk-gatersleben.de

284 | Nature | Vol 588 | 10 December 2020


a Barke
0.02
0.01
PC4 (1.4%)
0 HOR 3365
–0.01
–0.02
–0.015 –0.005 0.005
PC3 (2%)
Fig. 1 | Chromosome-scale sequences of 20 representative barley
genotypes reveal large structural variants. a, We selected 20 barley
genotypes to represent the genetic diversity space, as revealed by PCA of
genotyping-by-sequencing data of 19,778 domesticated varieties of barley 9.
Principal component (PC)3 and PC4 are shown. The proportion of variance
explained by the principal components is indicated in the axis labels. Further
large-genome plant species, such as barley (haploid genome size of
5 Gb)15.
Twenty barley reference genomes
The starting point for pan-genomics in barley was the comprehensive
survey of species-wide diversity on the basis of the genome-wide geno-
typing of more than 22,000 barley accessions, mainly from the German
national gene bank9. To achieve a good representation of major barley
gene pools, we selected accessions that were located in the branches
of the first six principal components from the previously published
principal component analysis (PCA)9 (Fig. 1a, Extended Data Fig. 1),
reflecting the key determinants of population structure: geographical
origin, row type and annual growth habit. In addition to these gene pool
representatives, our panel included the reference cultivar Morex5, two
current or former elite malting varieties (RGT Planet and Hockett), two
founder lines of Chinese barley breeding (ZDM01467 and ZDM02064),
Golden Promise and Igri (two genotypes with high transformation
efficiency16,17), Barke (a successful German variety and the parent of
several mutant and mapping populations18,19) and one wild barley
(H. vulgare subsp. spontaneum (K. Koch) Thell.) genotype from Israel
(B1K-04-12, a desert ecotype collected at Ein Prat)20.
We constructed chromosome-scale sequence assemblies for
20 accessions (Extended Data Table 1). In brief, paired-end and
mate-pair Illumina short reads were assembled into scaffolds of
megabase (Mb)-scale contiguity (Extended Data Table 1). Scaffold
assembly was done with Minia21 and SOAPDenovo22 following the TRI-
TEX method6 (n = 16), DeNovoMagic from NRGene (n = 3) or W2rap23
(n = 1). We used 10X Genomics Chromium linked-reads and chromo-
some conformation capture (Hi-C) data to arrange scaffolds into chro-
mosomal pseudomolecules using the TRITEX pipeline6 (Extended Data
Table 1). A comparison of the short-read assembly of the Morex cultivar
to a long-read assembly of this genotype generated from PacBio long
reads showed high co-linearity at chromosomal scale, good concord-
ance in gene space representation and similar power to detect PAV
(Extended Data Fig. 2), indicating that short-read assemblies are ame-
nable to pan-genomic analyses in barley. Although the assemblies of
the 20 diverse accessions differed in contiguity and the extent of gap
sequence in the intergenic space, they had a similar representation of
reference gene models (Morex V2) and were highly co-linear to each
other at the whole-chromosome scale (Fig. 1b, Extended Data Fig. 3).
A similar proportion (about 80%) of the assembled sequence of each
genotype was composed of transposable elements, with an average of
113,200 intact full-length long-terminal repeat retro-elements (LTRs)
Akashinriki
b
Golden Promise
HOR 10350 600
HOR 13821
HOR 13942 500
Barke position (Mb)
HOR 21599
HOR 3081 400
HOR 7552 300
HOR 8148
HOR 9043 200
Hockett
Igri 100
Morex
OUN333 0
RGT Planet
ZDM01467 0 200 400 600
ZDM02064 Morex position (Mb)
principal components are shown in Extended Data Fig. 1a. b, Alignment of the
pseudomolecules of chromosome 2H of the Morex and Barke cultivars. The
inset zooms in on a 10-Mb inversion that is frequently found in germplasm from
northern Europe. Co-linearity plots for all assemblies and chromosomes are
shown in Extended Data Fig. 3a.
per assembly (Supplementary Table 1). However, we found pronounced
differences in the number of shared intact full-length LTR locations:
only 17 to 25% of full-length LTR locations present in the wild barley
B1K-04-12 were shared at 98% sequence identity and 98% alignment
coverage with any domesticated genotype (Extended Data Fig. 4). By
contrast, more closely related domesticated genotypes shared between
53% and 67% of their full-length LTRs, consistent with previous reports
of rapid sequence turn-over in the non-coding space in large-genome
plant species24,25.
De novo gene annotation using Illumina RNA sequencing and PacBio
Iso-Seq data (Supplementary Table 2) was performed for three geno-
types: Morex (which has previously been reported6), Barke and the Ethio-
pian landrace HOR 10350 (Extended Data Fig. 5). Gene models defined
on the basis of these three assemblies were consolidated and projected
onto the remaining 17 assemblies (Extended Data Fig. 5). Between 35,859
and 40,044 gene models were annotated by projection in each assem-
bly (Extended Data Table 1) with an average of 37,515 (s.d. = 896). The
number of gene models was about 20% higher in the projections than in
de novo annotations (Extended Data Fig. 5e), which indicates that some
of the models lack transcript support: possible explanations for the
discrepancy are highly tissue-specific expression or pseudogenization.
The clustering of orthologous gene models yielded 40,176 orthologous
groups. Of these, 21,992 occurred as a single copy in all 20 assemblies;
3,236 occurred in multiple copies in at least one of the 20 assemblies;
13,188 were absent from at least one assembly; and 1,760 were present
in only one assembly. On average, 14.7% of gene models annotated in
each assembly occurred in tandem arrays that comprised two or more
adjacent copies. These results point to abundant genic copy-number
variation between barley genotypes. Future transcriptomic studies will
ascertain the effect of structural variants on gene expression.
Pan-genome as a tool for genetics and breeding
High-quality genome assemblies are a resource for ascertaining and
providing context to structural variants, which can then be genotyped in
a wider set of germplasm using low-coverage or reduced-representation
sequence data. We used two complementary approaches to detect
structural variation: assembly comparison and clustering of single-copy
sequences to derive markers that can be scored in short-read data. We
used the Assemblytics26 software to discover PAV by pair-wise compari-
son of 19 chromosome-scale assemblies to the Morex reference. We
identified 1,586,262 PAVs, ranging in size from 50 to 999,568 bp, and
observed an enrichment for low-frequency variants (Extended Data
Fig. 6a, b). PAV density was higher in distal, gene-rich regions (Extended
Nature | Vol 588 | 10 December 2020 | 285
Article
a

Single-copy pan-genome size (Mb)

650

600

550

500

2 2
R 9
2

U 2
O 33

M 50

R 4
M 43

R 7
1

M 1

R x
as 8

i
ri

Ba t
G GT ke

Pr et

e
ik

t
H ore

is
O 4-1
H 159

O 55

O 36

H 206

H 146

O 08

82

Ak 814

Ig

ke

n
nr

r
9

ZD 03

ZD 90

om
de Pla
7
13

3
13

oc
hi
0

0
R
K-

H
R

R
O
O

O
B1

n
R
H

ol
b 20
–log10(P value)

Absent in Morex
15 Present in Morex
10
5
0
1H 2H 3H 4H 5H 6H 7H

c d 0
16,682 bp

Morex 7H –0.5

Normalized k-mer count

529.04 Mb 529.07 Mb
–1.0

–1.5
HOR 7552 7H
528.80 Mb –2.0

–2.5
Breakpoints Nud Deletion
–3.0

Hulled Naked

Fig. 2 | Single-copy pan-genome and use of PAVs in association mapping.
a, Cumulative size of single-copy regions in genome assemblies of 20 barley
genotypes. The genotypes were ordered according to the size of their unique
single-copy sequence. b, Genome-wide association scan for lemma adherence
on the basis of PAV markers. The black and red dots in the Manhattan plot
denote single-copy sequences that are present and absent in Morex,
respectively. c, The most highly associated PAV marker was a 16.7-kb region
that is deleted in the naked accession HOR 7552 and that contains the NUD
gene11. d, Allelic status of the NUD deletion in 196 domesticated varieties of
barley. Normalized single-copy k-mer counts within the 16.7-kb region are
shown for hulled (n = 160 genotypes) and naked varieties (n = 36 genotypes).

Data Fig. 6c), which are characterized by higher nucleotide diversity
and recombination rates8. A total of 5,446 out of 5,602 deletions longer
than 5 kilobases (kb) found in Barke relative to Morex were mapped
genetically in the 90 recombinant inbred lines of the Morex × Barke
population19 with highly concordant positions (Spearman correla-
tion = 0.99) (Extended Data Fig. 6d), which provides support for the
accuracy of the detected polymorphisms. At least one member of 18,562
(46%) groups of orthologous genes overlapped with structural vari-
ants discovered in the 20 sequence assemblies. As observed in other
plant species27, resistance-gene homologues containing NB-ARC and
protein kinase domains were frequently found among PAV genes (Sup-
plementary Table 3).
Structural variants cover non-genic regions composed of repetitive
sequence, making it hard to establish orthologous relationships or
the presence of specific alleles from short-read data only. To derive
quantitative estimates of the extent of pan-genomic variation and
as a tool for genetic analysis such as association scans, we focused
on single-copy regions extracted from each of the 20 assemblies and
clustered into a non-redundant set of sequences (hereafter referred to
as the ‘single-copy pan-genome’) (Extended Data Fig. 7a). The average
cumulative size of single-copy sequence in each accession was 478 Mb
(that is, 9.5% of the assembly genome). The total size of non-redundant
single-copy sequence was 638.6 Mb, represented by 1,472,508 clus-
ters with an N50 of 1,087 bp (Extended Data Fig. 7b). The single-copy
sequence shared among all 20 genotypes amounted to 402.5 Mb,
whereas 235.9 Mb were variable (that is, absent or present in higher
copy number in at least one assembly) (Fig. 2a). On average, each of the
20 genotypes contained 2.9 Mb of single-copy sequence not present in
any other assembly. As observed for transposable element divergence,
the wild barley B1K-04-12 had the highest amount of unique single-copy
sequence (Extended Data Table 1).
To test the suitability of the single-copy pan-genome for genetic analy-
sis in a wider diversity panel without high-quality genome sequences,
we collected whole-genome shotgun data (threefold coverage) for
200 domesticated and 100 wild varieties of barley (Supplementary
Table 4). The abundance of 160,716 single-copy clusters that overlap
structural variants was estimated by counting cluster-constituent
k-mers (k = 31) in sequence reads of the diversity panel. In addition, we
analysed genotyping-by-sequencing data of 19,778 gene bank accessions
of domesticated barley9 using the same approach. Abundance estimates
based on k-mers (hereafter referred to as ‘pan-genome markers’) showed
that loci detected as single-copy sequence in one genome assembly can
vary in copy number from zero to many in diverse germplasm (Extended
Data Fig. 7c). A PCA of pan-genome markers genotyped in whole-genome
shotgun and genotyping-by-sequencing data highlighted the same driv-
ers of global population structure as SNPs (Extended Data Fig. 7d–g). In
genome-wide association scans for morphological traits, pan-genome
markers revealed—with a good signal-to-noise ratio—peaks that are

286 | Nature | Vol 588 | 10 December 2020

 a d

RGT Planet

RGT Planet
600

RGT Planet position (Mb)

Morex

Morex
500
Valticky Diamant
400
141.5 Mb
300
200
100
0
0 100 200 300 400 500 600
Morex position (Mb)
b
140

M × B genetic position (cM)

R × H genetic position (cM)

100 120
80 100

60 80
60
40
40
20 20
0 0
0 100 200 300 400 500 600 0 100 200 300 400 500 600
7H position (Mb) 7H position (Mb)
c

M × B recombination (cM Mb–1)

R × H recombination (cM Mb–1)

10 10

8 8

6 6

4 4

2 2

0 0
0 100 200 300 400 500 600 0 100 200 300 400 500 600
7H position (Mb) 7H position (Mb)

Fig. 3 | Identification and characterization of a large inversion on
chromosome 7H. a, Alignment of the 7H pseudomolecules of the Morex and
RGT Planet cultivars. b, Alignment of physical and genetic positions mapped in
the RGT Planet × Hindmarsh (R × H) (left) and Morex × Barke (M × B) (right)
populations. Red shading marks the inverted region. c, We converted genetic
distances to recombination rates in the R × H (left) and M × B (right)
populations. A single marker per recombination block is shown. d, We
designed a PCR marker (Supplementary Figs. 1, 2a) to screen for the presence of
the inversion in gene bank accessions that represent the Valticky and Diamant
cultivars.

consistent with previous reports9 (Fig. 2b, Extended Data Fig. 8). Nota-
bly, the pan-genome marker that was most highly associated with lemma
adherence covered the NUDUM (NUD) gene11 (Fig. 2c). All varieties of
naked barley—in which lemmas can be easily separated from grains—are
thought to trace back to a single mutational event, deleting the entire
NUD sequence11. Another putative knockout allele of NUD (nud1.g) that
contains a likely disruptive SNP variant was recently found in Tibetan
barley28. All 36 naked accessions in our panel contained the known dele-
tion (Fig. 2d), indicating that broader sampling of barley diversity—with
a particular focus on centres of (morphological) diversity—is needed
to discover novel rare alleles by genomic analyses.
Compared to reference-free approaches for k-mer-based
genome-wide association scans such as AgRenSeq29, trait-associated
pan-genome markers are assigned with high precision to genomic
positions, and aligning sequence assemblies in their vicinity provides
immediate information about differences between haplotypes (Fig. 2c).
Furthermore, the reduction of marker number by implicit clustering
of k-mers into single-copy loci allows the use of standard mixed linear
models30,31 to correct for genomic relatedness.
A map of polymorphic inversions
Chromosome-scale sequence assemblies can reveal large-scale rear-
rangements that are challenging to detect with other methods. Large
inversions (more than 5 Mb in size) were prominent in the genome
alignments of our 20 assemblies (Fig. 1b, Extended Data Fig. 3a, c). Previ-
ous reports on segregating inversions in barley are anecdotal and have
focused on induced mutants32,33. To discover inversions in a broader set
of germplasm, we mined patterns of contact frequencies in Hi-C data of
a diversity panel mapped to a single reference genome34. Among 69 bar-
ley genotypes (67 domesticated and 2 wild accessions) (Supplementary
Table 5), Hi-C-based inversion scans revealed a total of 42 events that
ranged from 4 to 141 Mb in size (mean size of 23.9 Mb) (Extended Data
Fig. 9a). Most of these events occurred in the low-recombining peri-
centromeric regions of the barley chromosomes and segregated at low
frequency: 25 events were observed only once (Extended Data Fig. 9b, c,
Supplementary Table 6). We focus here on two notable examples: a
frequent event on chromosome 2H and an inversion in the distal region
of the long arm of chromosome 7H.
The inversion in chromosome 7H detected in the RGT Planet cultivar
was the largest event that segregated in our panel (141 Mb) (Fig. 3a).
In a biparental mapping population derived from a cross between
RGT Planet and the non-carrier cultivar Hindmarsh (Fig. 3b), this
event repressed recombination in an interval that spanned 49 cM in
the genetic map of the Morex × Barke population19, which is isogenic
for absence of the inversion (Fig. 3c, Supplementary Table 7). We also
observed a moderately distorted segregation (57% allele frequency,
χ2 = 4.88, P < 0.05) in favour of the Hindmarsh allele in this interval.

Nature | Vol 588 | 10 December 2020 | 287

Article
a b
0.010
0.005
PC2 (2.4%)
0 0.05
–0.005
–0.010
–0.015
–0.005 0 0.005 0.010
PC1 (2.5%)
c
1,415
458,374,957
Morex: 2H
458,368,537 458,373,542
24,516
444,665,384
Barke: 2H
444,635,865 444,640,868
432,644
Fig. 4 | Analysis of a frequent inversion on chromosome 2H. a, A PCA
showing the localization of inversion carriers in the diversity space of global
domesticated barley. The correspondence of PCA coordinates to correlates of
population structure is shown in Extended Data Fig. 1. Red dots denote carriers
of the inverted haplotype (n = 87) in a panel of 200 domesticated varieties of
barley. b, PCA for a diversity panel comprising 200 domesticated (red and
green dots) and 100 wild varieties of barley (blue dots). SNP markers detected
Recombination frequencies were increased in the flanking regions
of the inversion in the RGT Planet × Hindmarsh population relative to
Morex × Barke, which suggests a compensatory mechanism to maintain
an average number of one-to-two crossovers per chromosome in the
presence of large tracts of suppressed recombination35.
By focusing on the inversion breakpoints in the RGT Planet sequence
assembly, we designed a diagnostic PCR assay (Supplementary Fig. 2a,
b, d) to rapidly genotype the presence of the inversion in 1,406 acces-
sions (Supplementary Table 8). The inverted haplotype occurred at
low frequency (1.3%) in the whole panel, but was found in many lines
in the RGT Planet pedigree (Supplementary Fig. 3)—including com-
mercially successful barley cultivars of past decades, such as Triumph,
Quench and Sebastian. The earliest cultivar that carried the inversion
was Diamant. As one of the donors of the semi-dwarf growth habit,
Diamant was a highly influential founder line of modern barley breed-
ing and traces back to a mutant induced by gamma irradiation of the
Czech cultivar Valticky36. We genotyped several gene bank accessions
and germplasm samples of both Valticky and Diamant. Notably, none
of the Valticky samples carried the inversion, whereas it segregated
in the Diamant samples (Fig. 3d). Quantitative trait loci mapping for
yield-related traits in the RGT Planet × Hindmarsh population did not
show signals on chromosome 7H (Supplementary Fig. 2e, Supplemen-
tary Table 9), consistent with selective neutrality of the inversion. This
strongly suggests that mutation breeding in the 1960s has given rise to
a cryptic large inversion, which—unbeknownst to breeders—segregates
in elite varieties of barley.
The second inversion we focused on spanned 10 Mb in the interstitial
region of chromosome 2H (Fig. 1b) and was present in 26 out of 69 Hi-C
samples (Supplementary Table 8). Local PCA and haplotype analysis in
288 | Nature | Vol 588 | 10 December 2020
Inverted (domesticated)
Wild
Non-inverted (domesticated)
0.10
4%)
(6.4%)
PC2 (6
0
2H inversion –0.05
Other
0.015 –0.05 0 0.05 0.10
PC1 (9.6%)
46 HC genes 4
468739147 468,739,151 468,752,788
3 1 2
14,369
454,475,094 454,489,463 454,503,097
44 HC genes
448,638
HvCEN
HvCEN
in whole-genome shotgun data and located in the inverted regions were used.
c, Schematic of the inverted region. The HvCEN gene is closest to the
breakpoint that is distal in Morex (distance of 449 kb) and proximal in Barke
(distance of 433 kb) assemblies. A total of 46 and 44 high-confidence (HC)
genes were annotated in the Morex and Barke assemblies, respectively. The
yellow arrows (not drawn to scale) mark the positions of PCR primers to probe
for the presence of the inversion (Supplementary Fig. 2c).
our panel of 200 domesticated and 100 wild varieties of barley indicated
a single origin of the inverted haplotype (Fig. 4a, b, Supplementary
Fig. 2c). The inversion occurred only among domesticated barley of
Western geographical origin9, indicating that it arose or has risen to
high frequency after domestication. The inverted region contains
46 high-confidence genes in the Morex cultivar. The closest gene to
the inversion breakpoint—at 448 kb distance from the distal breakpoint
in the non-carrier Morex—was HvCENTRORADIALIS (HvCEN)37 (Fig. 4c).
Although induced mutants of HvCEN flower very early, natural variation
in HvCEN has previously been implicated in environmental adaptation
to northern European climates37. All of the inversion carriers we ana-
lysed had HvCEN haplotype III, which is associated with later flowering
in spring barley varieties from northern Europe37,38. Further research is
required to determine whether the inversion close to HvCEN has direct
functional consequences (for instance, by modulating HvCEN expres-
sion) or whether it hitchhiked along with a tightly linked causal variant.
Discussion
The digital representation of the pan-genome can expand the repertoire
of natural or induced sequence variation that is accessible to genetic
analyses and breeding. Our comparison of 20 chromosome-scale
sequence assemblies has revealed pervasive variation in genes and
non-coding regions. Focusing on single-copy sequences, we trans-
lated this variation into scorable markers that are amenable to popu-
lation genetic analysis and association scans. A notable finding was
the prevalence of large (more than 5 Mb in size) inversion polymor-
phisms in current elite germplasm. It is likely that the suppression
of genetic recombination in inversion heterozygotes has manifested
itself in hard-to-explain patterns of long-range linkage and segrega-
tion distortion between elite lines in breeding programmes. Our map
of inversion polymorphisms will provide breeders with a point of
reference to avoid—or interpret correctly—crosses between carriers
and non-carriers. We found abundant structural variation in 20 rep-
resentative barley genotypes, but individual events occurred at low
frequency (Extended Data Figs. 6, 9). This observation, combined with
the slow saturation of the single-copy pan-genome (Fig. 2a), moti-
vate the genomic analysis of more genotypes to expand the barley
pan-genome. The next phase of barley pan-genomics will focus on
an augmented panel of domesticated and wild germplasm, working
towards the long-term goal of high-quality genome sequences of all
barley plant genetic resources as part of a biodigital resource centre39,40.
Online content
Any methods, additional references, Nature Research reporting sum-
maries, source data, extended data, supplementary information,
acknowledgements, peer review information; details of author con-
tributions and competing interests; and statements of data and code
availability are available at https://doi.org/10.1038/s41586-020-2947-8.
1. Bayer, P. E., Golicz, A. A., Scheben, A., Batley, J. & Edwards, D. Plant pan-genomes are the
new reference. Nat. Plants 6, 914–920 (2020).
2. Dawson, I. K. et al. Barley: a translational model for adaptation to climate change. New
Phytol. 206, 913–931 (2015).
3. Stein, N. & Muehlbauer, G. J. The Barley Genome (Springer, 2018).
4. International Barley Genome Sequencing Consortium. A physical, genetic and functional
sequence assembly of the barley genome. Nature 491, 711–716 (2012).
5. Mascher, M. et al. A chromosome conformation capture ordered sequence of the barley
genome. Nature 544, 427–433 (2017).
6. Monat, C. et al. TRITEX: chromosome-scale sequence assembly of Triticeae genomes
with open-source tools. Genome Biol. 20, 284 (2019).
7. Mascher, M. et al. Mapping-by-sequencing accelerates forward genetics in barley.
Genome Biol. 15, R78 (2014).
8. Russell, J. et al. Exome sequencing of geographically diverse barley landraces and wild
relatives gives insights into environmental adaptation. Nat. Genet. 48, 1024–1030 (2016).
9. Milner, S. G. et al. Genebank genomics highlights the diversity of a global barley
collection. Nat. Genet. 51, 319–326 (2019).
10. Muñoz-Amatriaín, M. et al. Distribution, functional impact, and origin mechanisms of
copy number variation in the barley genome. Genome Biol. 14, R58 (2013).
11. Taketa, S. et al. Barley grain with adhering hulls is controlled by an ERF family
transcription factor gene regulating a lipid biosynthesis pathway. Proc. Natl Acad. Sci.
USA 105, 4062–4067 (2008).
12. Tettelin, H. et al. Genome analysis of multiple pathogenic isolates of Streptococcus
agalactiae: implications for the microbial “pan-genome”. Proc. Natl Acad. Sci. USA 102,
13950–13955 (2005).
13. Ho, S. S., Urban, A. E. & Mills, R. E. Structural variation in the sequencing era. Nat. Rev.
Genet. 21, 171–189 (2019).
14. Danilevicz, M. F., Tay Fernandez, C. G., Marsh, J. I., Bayer, P. E. & Edwards, D. Plant
pangenomics: approaches, applications and advancements. Curr. Opin. Plant Biol. 54,
18–25 (2020).
15. Monat, C., Schreiber, M., Stein, N. & Mascher, M. Prospects of pan-genomics in barley.
Theor. Appl. Genet. 132, 785–796 (2019).
16. Coronado, M.-J., Hensel, G., Broeders, S., Otto, I. & Kumlehn, J. Immature pollen-derived
doubled haploid formation in barley cv. Golden Promise as a tool for transgene
recombination. Acta Physiol. Plant. 27, 591–599 (2005).
17. Schreiber, M. et al. A genome assembly of the barley ‘transformation reference’ cultivar
Golden Promise. G3 10, 1823–1827 (2020).
18. Gottwald, S., Bauer, P., Komatsuda, T., Lundqvist, U. & Stein, N. TILLING in the two-rowed
barley cultivar ‘Barke’ reveals preferred sites of functional diversity in the gene HvHox1.
BMC Res. Notes 2, 258 (2009).
19. Mascher, M. et al. Anchoring and ordering NGS contig assemblies by population
sequencing (POPSEQ). Plant J. 76, 718–727 (2013).
20. Hübner, S. et al. Strong correlation of wild barley (Hordeum spontaneum) population
structure with temperature and precipitation variation. Mol. Ecol. 18, 1523–1536 (2009).
21. Chikhi, R., Limasset, A. & Medvedev, P. Compacting de Bruijn graphs from sequencing
data quickly and in low memory. Bioinformatics 32, i201–i208 (2016).
22. Luo, R. et al. SOAPdenovo2: an empirically improved memory-efficient short-read
de novo assembler. Gigascience 1, 18 (2012).
23. Clavijo, B. J. et al. An improved assembly and annotation of the allohexaploid wheat
genome identifies complete families of agronomic genes and provides genomic
evidence for chromosomal translocations. Genome Res. 27, 885–896 (2017).
24. Anderson, S. N. et al. Transposable elements contribute to dynamic genome content in
maize. Plant J. 100, 1052–1065 (2019).
25. Brunner, S., Fengler, K., Morgante, M., Tingey, S. & Rafalski, A. Evolution of DNA sequence
nonhomologies among maize inbreds. Plant Cell 17, 343–360 (2005).
26. Nattestad, M. & Schatz, M. C. Assemblytics: a web analytics tool for the detection of
variants from an assembly. Bioinformatics 32, 3021–3023 (2016).
27. Gordon, S. P. et al. Extensive gene content variation in the Brachypodium distachyon
pan-genome correlates with population structure. Nat. Commun. 8, 2184 (2017).
28. Yu, S. et al. A single nucleotide polymorphism of Nud converts the caryopsis type of
barley (Hordeum vulgare L.). Plant Mol. Biol. Report. 34, 242–248 (2016).
29. Arora, S. et al. Resistance gene cloning from a wild crop relative by sequence capture
and association genetics. Nat. Biotechnol. 37, 139–143 (2019).
30. Lipka, A. E. et al. GAPIT: genome association and prediction integrated tool.
Bioinformatics 28, 2397–2399 (2012).
31. Yu, J. et al. A unified mixed-model method for association mapping that accounts for
multiple levels of relatedness. Nat. Genet. 38, 203–208 (2006).
32. Ekberg, I. Cytogenetic studies of three paracentric inversions in barley. Hereditas 76, 1–30
(1974).
33. Ramage, R. & Suneson, C. Translocation-gene linkages on barley chromosome 7. Crop
Sci. 1, 319–320 (1961).
34. Himmelbach, A. et al. Discovery of multi-megabase polymorphic inversions by
chromosome conformation capture sequencing in large-genome plant species. Plant J.
96, 1309–1316 (2018).
35. Ederveen, A., Lai, Y., van Driel, M. A., Gerats, T. & Peters, J. L. Modulating crossover
positioning by introducing large structural changes in chromosomes. BMC Genomics 16,
89 (2015).
36. Bouma, J. & Ohnoutka, Z. Importance and Application of the Mutant ‘Diamant’ in Spring
Barley Breeding (IAEA, 1991).
37. Comadran, J. et al. Natural variation in a homolog of Antirrhinum CENTRORADIALIS
contributed to spring growth habit and environmental adaptation in cultivated barley.
Nat. Genet. 44, 1388–1392 (2012).
38. Bustos-Korts, D. et al. Exome sequences and multi-environment field trials elucidate the
genetic basis of adaptation in barley. Plant J. 99, 1172–1191 (2019).
39. Mascher, M. et al. Genebank genomics bridges the gap between the conservation of crop
diversity and plant breeding. Nat. Genet. 51, 1076–1081 (2019).
40. Khan, A. W. et al. Super-pangenome by integrating the wild side of a species for
accelerated crop improvement. Trends Plant Sci. 25, 148–158 (2020).
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in
published maps and institutional affiliations.
Open Access This article is licensed under a Creative Commons Attribution
4.0 International License, which permits use, sharing, adaptation, distribution
and reproduction in any medium or format, as long as you give appropriate
credit to the original author(s) and the source, provide a link to the Creative Commons license,
and indicate if changes were made. The images or other third party material in this article are
included in the article’s Creative Commons license, unless indicated otherwise in a credit line
to the material. If material is not included in the article’s Creative Commons license and your
intended use is not permitted by statutory regulation or exceeds the permitted use, you will
need to obtain permission directly from the copyright holder. To view a copy of this license,
visit http://creativecommons.org/licenses/by/4.0/.
© The Author(s) 2020
Nature | Vol 588 | 10 December 2020 | 289
Article
Methods
No statistical methods were used to predetermine sample size. The
experiments were not randomized, and investigators were not blinded
to allocation during experiments and outcome assessment.
Library preparation, sequencing data generation and genome
assembly of 20 diverse varieties of barley
High-molecular-weight DNA was extracted from one-week-old seed-
lings of 20 diverse barley accessions given in Supplementary Table 10,
using a previously described large-scale DNA extraction protocol41.
For the NRGene DeNovoMAGIC3.0 assemblies, 450-bp paired-end
(PE450) libraries of Morex, Barke, HOR 10350 and B1K-04-12 were pre-
pared at the Leibniz Institute of Plant Genetics and Crop Plant Research
(IPK) Gatersleben. The 450-bp paired-end libraries for other acces-
sions, 800-bp paired-end libraries and mate-pair libraries of three
sizes were prepared and sequenced at the University of Illinois Roy J.
Carver Biotechnology Center. The 10X Genomics Chromium libraries
were prepared at the University of Saskatchewan Wheat Molecular
Breeding Laboratoryand sequenced by Genome Quebec or prepared
and sequenced at the Roy J. Carver Biotechnology Center, using the
manufacturers’ recommendations. Published tethered chromosome
conformation data for Morex, Barke, HOR 10350 and B1K-04-12 (ref. 42)
was used for scaffolding the respective genome. For the other acces-
sions, in situ Hi-C libraries were prepared using a previously described
method43. Sequencing data generated from each of the libraries are
given in Supplementary Table 10. NRGene DeNovoMAGIC3.0 scaffold
assemblies were provided for Barke, HOR 10350 and B1K-04-12. The
10X Chromium, population sequencing (POPSEQ) and Hi-C data were
then used to prepare chromosome-scale assemblies using the TRITEX
pipeline6 (commit: 7041ff2). For the other assemblies, the TRITEX pipe-
line was also used for contig assembly and scaffolding with mate-pair
and 10X data (Extended Data Table 1). High-confidence gene models
annotated on the Morex V2 reference6 and full-length cDNA sequences44
were aligned to the assemblies to assess gene-space completeness with
the parameters of ≥90% query coverage and ≥97% (≥90% for full-length
cDNA) identity.
Tissue collection and RNA extraction
Plant material for the collection of tissues for RNA sequencing
(RNA-seq) and Iso-Seq was grown in the greenhouse at IPK Gatersle-
ben with day–night temperatures of 21 °C–18 °C. Embryonic tissue,
leaves, roots, internode, inflorescence (5 mm) and developing seeds (5
and 15 days after pollination) were collected as previously described4,
snap-frozen in liquid nitrogen and stored at −80 °C until RNA extrac-
tions were performed. RNA was extracted from the collected tissues
using a Trizol extraction protocol4 and purified using Qiagen RNeasy
miniprep columns as per the manufacturer’s instructions. RNA quality
was checked on Agilent RNA HS screen tape and RNA with RIN value
greater than 8 was used for RNA-seq and Iso-Seq library construction.
RNA-seq library preparation and data generation
RNA-seq libraries were prepared from purified RNA using the TruSeq
RNA sample preparation kit (Illumina) as per the manufacturer’s rec-
ommendation at IPK Gatersleben. Libraries were pooled at equimolar
concentrations, quantified by qPCR and paired-end-sequenced on an
Illumina HiSeq 2500 for 200 cycles. The data generated for each tissue
are given in Supplementary Table 2.
Iso-Seq data generation and analysis
Two libraries for each embryonic tissue RNA and pooled RNA from
seven tissues (described in ‘Tissue collection and RNA extraction’) were
prepared for Barke and HOR 10350 using the PacBio Iso-Seq protocol.
In brief, double-stranded RNA was synthesized using SMARTer PCR
cDNA synthesis kit (Clontech; cat. no. 634925). Two fractions of cDNA
with different size profiles were prepared by using differing ratios of
DNA to Ampure XP beads (Beckman Coulter, cat. no. A63882). Equi-
molar concentration of each fraction were pooled, and a minimum of
one microgram of double-stranded cDNA was used for Iso-Seq library
construction as per the PacBio library construction protocol. Two
additional libraries from pooled RNA tissues were prepared using cDNA
prepared from TeloPrime v.1.0 kit (Lexogen) following the manufac-
turer’s instructions. Libraries were quantified and sequenced on a
PacBio Sequel device at IPK Gatersleben. Data were analysed using
SMRTLink v.5.0 Isoseq v.1.0 pipeline or Isoseq3 pipeline (https://github.
com/PacificBiosciences/IsoSeq_SA3nUP/wiki/Tutorial:-Installing-
and-Running-Iso-Seq-3-using-Conda). The steps involved in Iso-Seq
data analysis were the generation of circular consensus sequences,
and then the classification of circular consensus sequence reads into
full-length non-chimeric reads and non-full length reads on the basis
of the presence of primer sequences and polyA sequences. Full-length
non-chimeric reads were then clustered on the basis of sequence simi-
larity to yield high- and low-quality isoforms. The data generated and
method of library preparation are given in Supplementary Table 2.
Gene projections and repeat annotation
Gene models for Morex, Barke and HOR 10350 were predicted using
transcriptome data (Supplementary Table 2) and protein homology
evidence, and derived by a previously described annotation pipeline5.
High-confidence gene models from these accessions were aligned to
pseudo-chromosomes of each accession separately using blat45. For
each genomic region identified by blat, additional alignments were
performed by exonerate46 in its genomic neighbourhood ranging
between 20 kb upstream and 20 kb downstream of the match posi-
tion. A series of quality criteria was applied to select high-confidence
gene models in each accession. The functional annotation for genes of
20 accessions was carried out using the AHRD pipeline v.3.3.3 (https://
github.com/groupschoof/AHRD). Orthologous gene groups between
the twenty accessions were predicted using OrthoFinder47 v.2.3.1 with
default parameters.
Repeat annotation
To obtain a consistent transposon annotation across all lines for trans-
posons and tandem repeats, the same methods were applied to all 20
barley lines. Transposons were detected and classified by homology
search against the REdat_9.7_Poaceae section of the PGSB transposon
library48. The program vmatch (http://www.vmatch.de, version 2.3.0)
was used for that purpose as a fast and efficient matching tool that is
well-suited for such large and highly repetitive genomes. Vmatch was
run with the following parameters: identity ≥ 70%, minimal hit length
75 bp, seed length 12 bp (exact command line: -d -p -l 75 -identity 70
-seedlength 12 -exdrop 5). To remove overlapping annotations, the
vmatch output was filtered for redundant hits via a priority-based
approach. Higher scoring matches were assigned first and lower scoring
hits at overlapping positions were either shortened or removed if they
were contained to ≥90% in the overlap or <50 bp of rest length remained.
The resulting transposon annotations are overlap-free, but disrupted
elements from nested insertions have not been defragmented into one
element. Still-intact full-length LTR retrotransposons were identified
with LTRharvest49, a program that scans the genome for LTR retrotrans-
poson specific structural hallmarks, such as long terminal repeats, RNA
cognate primer binding sites and target site duplications. LTRharvest
(included in genometools 1.5.9) was run with the following param-
eter settings: ‘overlaps best -seed 30 -minlenltr 100 -maxlenltr 2000
-mindistltr 3000 -maxdistltr 25000 -similar 85 -mintsd 4 -maxtsd 20
-motif tgca -motifmis 1 -vic 60 -xdrop 5 -mat 2 -mis -2 -ins -3 -del -3’. All
candidates were annotated for PfamA domains using hmmer3 (http://
hmmer.org, version 3.1b2) and filtered to remove false positives. The
inner domain order served as a criterion for the LTR-retrotransposon
superfamily classification into either Gypsy or Copia. In the cases of
insufficient domain information, the elements were assigned as still
undetermined.
Most of the transposons insert at random locations leading to novel
and usually unique sequence stretches at both borders around the
inserted element and the neighbouring original sequence. The de novo
detected full-length LTR set provides defined element borders, a pre-
requisite for the exact positioning of transposable element junctions.
We used 100-bp single transposable element junctions with 50 bp out-
side and 50 bp inside the element from both sides of the element and
merged them to 200 bp joined junctions per element. Junctions from
the reverse strand were reverse-complemented. The 200-bp joined
junctions from all 20 lines were clustered with vmatch dbcluster (http://
www.vmatch.de, version 2.3.0) at 98% identity and 98% mutual length
coverage (command-line parameters: 98 98 -e 2 -l 98 -d). About 97%
of the clusters belonged to the 1:1 type with a maximum of 1 mem-
ber per line and were used for the downstream analyses. Using the
above-described 200-bp joined junctions instead of full sequences
reduces the amount of data for clustering to 2%, from about 10 kb to
200 bp per full-length LTR element, thus allowing a sequence cluster-
ing of 2.2 million elements in the first place. By including sequence
information outside of the element, the repetitiveness of high-copy
transposable element families is removed and at the same time the
syntenic context is provided even for elements located on chrUn (that
is, not assigned to chromosomal pseudomolecules).
PAV detection and validation
Owing to higher sensitivity in detecting deletions over insertions, a
paired genome alignment strategy was used in which each assembly
was aligned to reference genome Morex reciprocally by treating Morex
as a query and reference using Minimap2 (v.2.17)50. From these two
alignments, insertion and deletions were called using Assemblytics
(v.1.2.1)26. Then, only deletions were selected in both alignments and
converted into PAVs with regard to Morex. In addition, a hard filter was
used to discard PAVs containing more than 5% gaps (Ns) and nested
PAVs. We used a previously described method51 to map deletions longer
than 5 kb in Barke relative to Morex using whole-genome shotgun data
for 90 Morex × Barke recombinant inbred lines19. Mosdepth (v.0.2.9)52
was used for determining read depth in genomic intervals.
k-mer-based genome-wide association
PAVs overlapping with single copy regions were identified by BedTools
(v.2.28.0)53. k-mer sequences with step size of 2 bp were retrieved
from single-copy regions residing within PAVs. The abundances of
the extracted k-mer sequences were counted in sequence reads using
BBDuk (BBMap_37.93) (https://sourceforge.net/projects/bbmap/).
k-mer counts were obtained for whole-genome shotgun data of
300 diverse varieties of barley generated in the present study and
previously published genotyping-by-sequencing data9. k-mer counts
were imported into R (v.3.5.1)54 and normalized for differences in read
depth between samples. The normalized k-mer counts were then used
for genome-wide association scans using GAPIT330 and PCA using stand-
ard R functions.
Construction of single-copy pan-genome
To identify single-copy regions in each genome, genomic regions cov-
ered by 31-mers occurring more than once were masked using BBDuk
(BBMap_37.93)55. Based on masking, single-copy regions in each assem-
bly were obtained in .bed format and subsequently related sequences
were retrieved using BEDTools (v2.28.0)53. Single-copy sequences from
all the assemblies were combined to perform an all-against-all blast
search. The blast results were filtered (>90% identity and minimum
80% alignment length) and then clustered using the igraph package56. A
representative from each cluster (the largest contained sequence) was
selected and used for estimating pan-genome size. Clusters shared by
all the 20 accessions are referred to as the core genome, and clusters
with sequences originating from 1 to 19 genotypes are considered as
the variable genome.
Hi-C library preparation, sequencing and inversion calling
In situ Hi-C libraries were prepared from one-week-old seedlings of
barley IPK core50 collection9 (Supplementary Table 5) based on a
previously described protocol43 Sequencing, Hi-C raw data process-
ing and inversion calling were performed as previously described34
using the MorexV2 reference genome sequence assembly6. The break-
point regions were identified by pairwise genome alignment using
Minimap2 (v.2.17)50 and PipMaker (http://pipmaker.bx.psu.edu/cgi-bin/
pipmaker?basic)57.
Resequencing, SNP calling and PCA
Raw reads (Supplementary Table 4) were trimmed with cutadapt
(v.1.15) and aligned to the MorexV2 genome assembly6 using Minimap2
(v.2.17)50. The alignments were sorted using Novosort (V3.06.05) (http://
www.novocraft.com). BCFtools (v.1.8)58 was used to call SNPs and short
insertions and deletions (indels). The resulting VCF file was converted
into Genomic Data Structure (GDS) format using SeqArray package59 in
R to obtain a SNP matrix. Finally, hard filtering was applied to remove
SNPs having more than 10% missing data and heterozygosity. Previ-
ously generated genotyping-by-sequencing data9 were aligned to the
MorexV2 reference and identified SNPs using a previously described
variant calling pipeline9. PCAs were performed using snpgdsPCA()
function of the package SNPrelate60.
RGT Planet × Hindmarsh mapping population
A cross was made between RGT Planet (maternal plant) and Hindmarsh
(pollen donor). In total, 38 F2 plants from the direct cross and 233 indi-
vidual heads from F3 seeds were progressed to the F6 generation by
single seed descent method. The F6 recombinant inbred lines (RIL)
(224 in total) were used for construction of a genetic linkage map.
Genomic DNA was extracted from the leaves of a single plant per RIL
using the cetyl-trimethyl-ammonium bromide method. DNA qual-
ity was assessed on 1% agarose gels and quantified using a NanoDrop
spectrophotometer (Thermo Scientific NanoDrop Products). DNA was
diluted into 50 ng/μl and placed in a 96-well plate for PCR. DArT-seq
genotyping-by-sequencing was performed using the DArT-seq platform
(DArT PL) according to the manufacturer’s protocol (https://www.
diversityarrays.com/). In brief, 100 μl of 50 ng μl−1 DNA was sent to
DArT PL, and genotyping-by-sequencing was performed using com-
plexity reduction followed by sequencing on a HiSeq Illumina plat-
form as previously described61 (Supplementary Table 9). Sequences
flanking polymorphisms detected by DArT-seq were aligned against
the MorexV2 genome assembly to determine their physical positions
(Supplementary Table 7).
Field experiments and phenotypic data
Field experiments were conducted at six sites: Gibson, Western Australia
(WA, −33.612176, 121.798438); Williams, Western Australia (−33.577668,
116.734934); Wongan Hills, Western Australia (−30.848953, 116.756461);
Merredin, Western Australia (−31.487009, 118.229668); South Perth,
Western Australia (−31.991186, 115.887944); and Shepperton, Victoria
(−36.487551, 145.388470). The distance between South Perth and Shep-
perton is over 3,300 km. The Merredin site is located inland and receives
little rainfall, whereas the Gibson site receives a high amount of rainfall:
the other sites are in between. The experimental design for field trial sites
was performed as previously described62. In brief, all regional field trials
(partially replicated design) were planted in a randomized complete
block design using plots of 1 by 5 m2, laid out in a row–column format
and the middle 3 m was harvested for grain yield. Field trials in South
Perth and Shepperton were conducted using double rows with a 40-cm
distance within and between rows, owing to space constraints. Seven
control varieties were used for spatial adjustment of the experimental
Article
data. Measurements were taken at each plot of each field experiment
in the study to determine flowering time (days to Zadoks stage (ZS)49),
plant height and grain yield. In brief, heading date was recorded as the
number of days from sowing to 50% awn emergence above the flag leaf
(ZS49), as a proxy for flowering time. Plant height was determined by
estimating the average height from the base to the tip of the head of all
plants in each plot. Grain yield (kg ha−1) was determined by destructively
harvesting all plant material from each plot to separate the grain, and
then determining grain mass. Grain yield data of the field experiments,
as well as plant height and heading data, were analysed using linear
mixed models in ASReml-R (https://www.vsni.co.uk/software/asreml-r/)
to determine best linear unbiased predictions or best linear unbiased
estimations for each trait for further analysis. Local best practices for
fertilization and disease control were adopted for each trial site.
Quantitative trait loci (QTL) mapping
Software MapQTL6 was used for the QTL analysis63. The genotypic
data, phenotypic data and genetic map were formatted and imported
to MapQTL6. Interval mapping was conducted for each trait, and then
the markers with a logarithm of odds (LOD) value of above 3.0 were
selected as cofactors. Multiple QTL model mapping was performed to
re-calculate the QTL. If the markers with the highest LOD value were
inconsistent with the cofactor markers, then the new markers were
selected as cofactors and re-calculated. The QTL results and charts
were exported from the software.
Long-read sequence assembly of the Morex cultivar
PacBio libraries were constructed using SMRTbell Template Prep Kit
1.0 and sized on a SAGE Blue Pippin instrument 20–80 kb. Sequenc-
ing was performed on Sequell II device at the HudsonAlpha Institute
using V.1.0 chemistry and 10-h movie time. Data were generated from
a total of five SMRT cells, yielding 604 Gb of raw sequence reads. A
total of 520.72 Gb of this set (104.15×) was used for assembly (Sup-
plementary Tables 11, 12). Previously published Illumina short-read
data (ERR3183748 and ERR31837496 (Supplementary Table 12)) was
used for polishing and error correction. Before use, Illumina fragment
reads were screened for phix contamination. Reads composed of >95%
simple sequences were removed. Illumina reads shorter than 50 bp
after trimming for adaptor and quality (q < 20) were removed. The final
read set consists of 605,178,701 reads, representing a total of 43.17×
of high-quality Illumina bases. The initial assembly was generated by
assembling 32,743,478 PacBio reads (104.15× sequence coverage) using
MECAT (v.1.1)64 and subsequently polished using Arrow65. This produced
an initial assembly of 1,577 scaffolds (1,577 contigs), with a contig N50
of 10.4 Mb, 987 scaffolds larger than 100 kb and a total genome size of
4,139.8 Mb (Supplementary Table 13).
A first round of breaking chimeric scaffolds was done using the POP-
SEQ genetic map19 to identify contigs bearing markers from distant
genomic regions. A total of 17 misjoins were identified and resolved.
Homozygous SNPs and indels were corrected in the release consensus
sequence using a subset of about 30× of the Illumina reads described
above in this section. Reads were aligned using BWA-MEM66. Homozy-
gous SNPs and indels were discovered with GATK’s UnifiedGenotyper
tool67. A total of 59 homozygous SNPs and 15,759 homozygous indels
were corrected. After these correction steps, the assembly contains
4,139.7 Mb of sequence, consisting of 1,594 contigs with a contig N50
of 10.2 Mb. A second round of chimaera breaking by inspecting Hi-C
contact matrices as described in the TRITEX pipeline6. Published Hi-C
data of the Morex cultivar was used5. Corrected contigs were arranged
into pseudomolecules using TRITEX.
Comparison of PacBio continuous long read (CLR) and TRITEX
assemblies of the Morex cultivar
Full-length cDNA sequences44 were aligned to the assemblies to assess
gene space completeness. Only alignments with query coverage ≥90%
and identity ≥90% were considered. Whole-genome assemblies were
done with Minimap2. Structural variant calling with Assemblytics
(v.1.2.1)26 (Morex TRITEX versus Morex CLR; Morex CLR versus Barke)
and extraction of single-copy regions were done as described in ‘PAV
detection and validation’.
Reporting summary
Further information on research design is available in the Nature
Research Reporting Summary linked to this paper.
Data availability
All raw sequence data collected in this study and sequence assemblies
have been deposited at the European Nucleotide Archive (ENA). Acces-
sion codes for raw data and assemblies are listed in Supplementary
Tables: Supplementary Table 14 (assemblies), Supplementary Table 10
(assembly raw data), Supplementary Table 4 (whole-genome shotgun
sequencing), Supplementary Table 5 (Hi-C) and Supplementary Table 9
(DArT-seq). Assemblies, annotations and analysis results were depos-
ited under a DOI in the PGP repository68 using the e!DAL submission
system69 and are accessible under the URL https://doi.org/10.5447/
ipk/2020/24. Assemblies and gene annotations can also be downloaded
from https://barley-pangenome.ipk-gatersleben.de. The Barley Pedi-
gree Catalogue is available at http://genbank.vurv.cz/barley/pedigree/.
Code availability
Source code is released in a public Bitbucket repository, at https://
bitbucket.org/ipk_dg_public/barley_pangenome/.
41. Dvorak, J., McGuire, P. E. & Cassidy, B. Apparent sources of the A genomes of wheats
inferred from polymorphism in abundance and restriction fragment length of repeated
nucleotide sequences. Genome 30, 680–689 (1988).
42. Himmelbach, A., Walde, I., Mascher, M. & Stein, N. Tethered chromosome conformation
capture sequencing in Triticeae: a valuable tool for genome assembly. Bio Protoc. 8,
e2955 (2018).
43. Padmarasu, S., Himmelbach, A., Mascher, M. & Stein, N. in Plant Long Non-Coding RNAs
(eds Chekanova, J. & Wang, H.-L.) 441–472 (Springer, 2019).
44. Matsumoto, T. et al. Comprehensive sequence analysis of 24,783 barley full-length cDNAs
derived from 12 clone libraries. Plant Physiol. 156, 20–28 (2011).
45. Kent, W. J. BLAT—the BLAST-like alignment tool. Genome Res. 12, 656–664 (2002).
46. Slater, G. S. C. & Birney, E. Automated generation of heuristics for biological sequence
comparison. BMC Bioinformatics 6, 31 (2005).
47. Emms, D. M. & Kelly, S. OrthoFinder: solving fundamental biases in whole genome
comparisons dramatically improves orthogroup inference accuracy. Genome Biol. 16,
157 (2015).
48. Spannagl, M. et al. PGSB PlantsDB: updates to the database framework for comparative
plant genome research. Nucleic Acids Res. 44, D1141–D1147 (2016).
49. Ellinghaus, D., Kurtz, S. & Willhoeft, U. LTRharvest, an efficient and flexible software for
de novo detection of LTR retrotransposons. BMC Bioinformatics 9, 18 (2008).
50. Li, H. Minimap2: pairwise alignment for nucleotide sequences. Bioinformatics 34,
3094–3100 (2018).
51. Gutierrez-Gonzalez, J. J., Mascher, M., Poland, J. & Muehlbauer, G. J. Dense
genotyping-by-sequencing linkage maps of two synthetic W7984×Opata reference
populations provide insights into wheat structural diversity. Sci. Rep. 9, 1793 (2019).
52. Pedersen, B. S. & Quinlan, A. R. Mosdepth: quick coverage calculation for genomes and
exomes. Bioinformatics 34, 867–868 (2018).
53. Quinlan, A. R. & Hall, I. M. BEDTools: a flexible suite of utilities for comparing genomic
features. Bioinformatics 26, 841–842 (2010).
54. R Core Team. R: A Language and Environment for Statistical Computing
http://www.R-project.org (R Foundation for Statistical Computing, 2013).
55. Bushnell, B. BBMap: A Fast, Accurate, Splice-aware Aligner (Lawrence Berkeley National
Laboratory, 2014).
56. Csardi, G. & Nepusz, T. The igraph software package for complex network research.
InterJournal Complex Syst. 1695, 1–9 (2006).
57. Schwartz, S. et al. PipMaker—a web server for aligning two genomic DNA sequences.
Genome Res. 10, 577–586 (2000).
58. Li, H. A statistical framework for SNP calling, mutation discovery, association mapping
and population genetical parameter estimation from sequencing data. Bioinformatics 27,
2987–2993 (2011).
59. Zheng, X. & Gogarten, S. SeqArray: big data management of genome-wide sequence
variants. R package version 1.10.6 https://github.com/zhengxwen/SeqArray (accessed
January 2017).
60. Zheng, X. et al. A high-performance computing toolset for relatedness and principal
component analysis of SNP data. Bioinformatics 28, 3326–3328 (2012).
61. Akbari, M. et al. Diversity arrays technology (DArT) for high-throughput profiling of the
hexaploid wheat genome. Theor. Appl. Genet. 113, 1409–1420 (2006).
62. Hill, C. B. et al. Hybridisation-based target enrichment of phenology genes to dissect
the genetic basis of yield and adaptation in barley. Plant Biotechnol. J. 17, 932–944
(2019).
63. Van Ooijen, J. MapQTL 5, Software for the Mapping of Quantitative Trait Loci in
Experimental Populations (Kyazma, 2004).
64. Xiao, C. L. et al. MECAT: fast mapping, error correction, and de novo assembly for
single-molecule sequencing reads. Nat. Methods 14, 1072–1074 (2017).
65. Chin, C. S. et al. Nonhybrid, finished microbial genome assemblies from long-read SMRT
sequencing data. Nat. Methods 10, 563–569 (2013).
66. Li, H. Aligning sequence reads, clone sequences and assembly contigs with BWA-MEM.
Preprint at https://arxiv.org/abs/1303.3997 (2013).
67. McKenna, A. et al. The Genome Analysis Toolkit: a MapReduce framework for analyzing
next-generation DNA sequencing data. Genome Res. 20, 1297–1303 (2010).
68. Arend, D. et al. PGP repository: a plant phenomics and genomics data publication
infrastructure. Database (Oxford) 2016, baw033 (2016).
69. Arend, D. et al. e!DAL—a framework to store, share and publish research data. BMC
Bioinformatics 15, 214 (2014).
Acknowledgements We thank M. Knauft, I. Walde and S. König for technical assistance;
D. Schüler for sequence data management; J. Bauernfeind, T. Münch and H. Miehe for IT
administration; D. Arend for help with data submission; M. Bayer for advice on
transcriptome analysis; and M. Herz for pedigree information. This research was supported
by grants from the German Federal Ministry of Education and Research to N.S., M.M., U.S.,
M.S. and K.F.X.M. (SHAPE, FKZ 031B0190), to U.S. and K.F.X.M. (de.NBI, FKZ 031A536) and to
N.S. (COBRA, FKZ 031A323A); the Australian Grain Research and Development Cooperation
(9176507) to C.L., K.C., P.L. and P.W.; JST CREST Japan (no. JPMJCR16O4 to K.M. and T.H.);
JST Mirai Program Japan (no. 18076896 to K.S.); the National Key R&D Program of China
(2018YFD1000701 and 2018YFD1000700) to D.X. and J.Z.; by funding from the China
Agriculture Research System (CARS-05) and the Agricultural Science and Technology
Innovation Program to C.W. and G.G. Support for 10X sequencing was provided by a
research grant from Genome Canada and Genome Prairie to C.P. and J.E.; and by the
Natural Science Foundation of China (31620103912) and the National Key R&D Program of
China (2018YFD1000706) to G.Z. We acknowledge support from the European Research
Council (ERC Shuffle, project identifier: 66918) to R.W.
Author contributions N.S. and M.M. designed the study. N.S. coordinated experiments and
sequencing. M.M. supervised sequence assembly. M. Spannagl and K.F.X.M. supervised
annotation. U.S. supervised data management and submission. S.P., A.H., J.E., D.X., L.B.B.
and J.G. performed sequencing experiments. M.J., C.M., Y.G., C.P., J.J. and J.S. performed
sequence assembly. M.J. performed structural variation and genome-wide association scan
analysis. A.F. submitted sequence data. G.H., T.L., H.G., V.S.B., N.K. and D.L. annotated and
analysed genes and transposable elements. S.P., M.J., X.-Q.Z., T.T.A., G. Zhou, C.T., C.H., P.W.,
M.M. and C.L. analysed polymorphic inversions. H.B., J.G., J.S., J.Z., C.W., G.G., G. Zhang,
K.M., T.H., K.S., K.J.C., P.L., C.J.P., C.L., M. Schreiber, R.W. and N.S. contributed sequence
data. M.J., S.P., C.L. and M.M. wrote the paper with input from all co-authors.
Competing interests The authors declare no competing interests.
Additional information
Supplementary information is available for this paper at https://doi.org/10.1038/s41586-020-
2947-8.
Correspondence and requests for materials should be addressed to C.L., M.M. or N.S.
Peer review information Nature thanks Victor Albert, Scott Jackson and the other, anonymous,
reviewer(s) for their contribution to the peer review of this work. Peer reviewer reports are
available.
Reprints and permissions information is available at http://www.nature.com/reprints.
Article

Extended Data Fig. 1 | Pan-genome selection in the global barley diversity pan-genome selection (first row), or according to geographic origin (second
space. PCA with genotyping-by-sequencing data of 19,778 varieties of row), row type (third row) or annual growth habit (fourth row). The proportion
domesticated barley sampled from the gene bank of the IPK9. The first six of variance explained by the principal components is indicated in the axis
principal components are shown. Samples are coloured to highlight the labels of the first row. The map was created with the R package mapdata54.
Extended Data Fig. 2 | Comparison between long-read and short-read
assemblies of the Morex cultivar. a, Co-linearity between Morex V2
(short-read) assembly and the Morex PacBio CLR assembly at the
pseudomolecule level. b, Summary statistics of the Morex PacBio CLR
assembly and Morex V2 assembly. c, Alignment of NUDUM locus (16 kb)
between Morex PacBio CLR and Morex V2. d, Structural variants between
Morex V2 and Morex PacBio CLR assemblies as detected and classified by
Assemblytics. e, PAVs between Barke and the Morex V2 and Morex CLR
assemblies.
Article

Extended Data Fig. 3 | Assessment of contiguity and completeness in annotation and full-length cDNAs (28,622 full-length cDNAs) in each assembly.
20 genome assemblies. a, Whole-genome alignments of assemblies of Alignments with less than 90% query coverage and 97% (less than 90% for
19 diverse barley accessions to the Morex V2 reference assembly. b, Alignment full-length cDNAs) identity were discarded. c, Whole-genome alignments show
summary of full-length coding sequences (32,878) from the MorexV2 some examples of large chromosomal inversions identified using Hi-C data.
Extended Data Fig. 4 | Pairwise shared syntenic full-length LTR locations. cultivars. The highest similarity is found between the Barke and RGT Planet
The wild variety B1K-04-12 is set apart as an outgroup, as it shares only 19–26% cultivars (67% shared full-length LTRs).
of its still-intact full-length LTR positions with the other landraces and
Article

Extended Data Fig. 5 | Gene projection and transposable element d, Summary of gene projections and transposable element annotation in
annotation. a, Schematic of the gene projection workflow. TE, transposable 20 accessions. e, Comparison between de novo annotations and gene
element. b, Pipeline for annotation and removing transposable elements. projections for three genotypes. Reported counts refer to
c, Steps to identify tandemly arrayed gene (TAG) clusters in each assembly. non-transposable-element genes.
Extended Data Fig. 6 | Summary of PAVs detected in pan-genome d, Co-linearity between physical position of PAVs detected between the Morex
assemblies. a, Size distribution of PAVs. b, Number of PAVs between and Barke cultivars, and mapped genetically in the POPSEQ population.
20 genome assemblies. c, Distribution of PAVs along the barley genome.
Article
Extended Data Fig. 7 | Analysis of the single-copy pan-genome. a, Pipeline
used to select single-copy k-mers in PAVs as markers for genome-wide
association scan analysis. b, Summary of single-copy sequence in 20 genome
assemblies and results of their clustering. c, Copy number of single-copy
sequences in a diversity panel comprising 200 domesticated and 100 wild
accessions. Frequency ranges from blue (low) to red (high). d–g, Comparison
of PCA on the basis of PAV and SNP variants in whole-genome shotgun data of
200 diverse accessions (d, e) and 19,778 varieties of domesticated barley 9 (f, g).
Top panels show PCA results from 160,716 PAVs; bottom panels show PCA
results from 779,503 of genotyping-by-sequencing SNPs. The accessions are
coloured according to geographical origin and row type (using the colour code
defined in Extended Data Fig. 1).
Extended Data Fig. 8 | PAV-based genome-wide association scans using
whole-genome shotgun and genotyping-by-sequencing data. a, Manhattan
plots of PAV-based genome-wide association scans for morphological traits,
including adherence of grain hull, row type, length of rachilla hairs and awn
roughness, using whole-genome shotgun data from 200 diverse varieties of
domesticated barley. b, PAV-based genome-wide association scan results for
these traits using genotyping-by-sequencing data from 1,000 diverse varieties
of domesticated barley collected from the gene bank of the IPK9. The 200
varieties of barley used for whole-genome shotgun sequencing are a subset of
the 1,000 genotyping-by-sequencing genotypes.
Article

Extended Data Fig. 9 | Characterization of large inversions in barley. (n = 90 genotypes). c, Number of inversions present as singletons or shared
a, Inversion size distribution. b, Recombination in inverted regions. between two or more accessions on each chromosome.
Recombination rate was determined in the Morex × Barke RIL population19
Extended Data Table 1 | Summary statistics of 20 pan-genome assemblies and annotation

#Chromosomes 1H to 7H.
§Non-transposable element models or transposable-element filtered.
 12345656762589 56 5317425
!"#$"%&'()*+, 3$71($"=C&($"C&4)=)"%#&-$
-&('!#&(#./&'()*+,>'%nonpnp
03&('0!&(4)5$$")%1 ' 22&  /
 ($2!6()!#'4$.$7$(/8()59()&(5!'.7$):;)$82!6$# ('4('84"$("4/&"#(&"!&"4/
$"!($"%:<8'()$"82&($""3&('0&4)!7$4$=>'()?08 &"#()@#$($&7A7$4/)497$(:

1(&($($4
<&77(&($($4&7&"&7/=4"8$2()&(()8775$"%$(2 &!"($"()8$%'7%"#=(&.77%"#=2&$"(B(=C()# 4($":
"D& "8$2#
q ;)B&4(&2!7  $E*F+8&4)B!$2"(&7%'!D4"#$($"=%$6"& &#$4("'2.&"#'"$(82&'2"(
q > (&(2"("5)()2&'2"(5(&9"82#$($"4(&2!7 5)()() &2 &2!75& 2&'#!&(#7/
q; ) (&($($4&7((*+'#>3G5)()()/&"H(5H$##
IFJKLMNOONFLPQRPRLRSNTJULVQLUQRMWXVQULRNJQJKLVKLFYOQZLUQRMWXVQLONWQLMNO[JQ\LPQMSFX]TQRLXFLPSQL^QPSNURLRQMPXNF_
q >#4$!($"8&7746&$&( ((#
q >#4$!($"8&"/& '2!($" 44($"='4)& ((8"2&7$(/&"#&#`'(2"(82'7($!742!&$"
q >8'77#4$!($"8() (&($($4&7!&&2($"47'#$"%4"(&7("#"4/*:%:2&"+().&$4($2&( *:%:% $"488$4$"(+
>3G6&$&($"*:%:(&"#&##6$&($"+& 4$&(#($2&( 8'"4(&$"(/*:%:4"8$#"4$"(6&7+
q <"'7 7)/!()$(($"%=()(((&($($4*:%:a=P=W+5$()4"8$#"4$"(6&7=884($E=#% 88#2&"#b6&7'"(#
cXdQLbLdYJTQRLYRLQ\YMPLdYJTQRLeSQFQdQWLRTXPYVJQ_
q <f&/$&"&"&7/$=$"82&($""()4)$48!$&"#C&964)&$"C"(&7 (($"%
q <)$&4)$4&7&"#42!7B#$%"=$#"($8$4&($"8()&!!!$&(7678((&"#8'77!($"%8'(42
q @($2&( 8884($E *:%:)"gU=A&"gW+=$"#$4&($"%)5()/54&74'7&(#
ITWLeQVLMNJJQMPXNFLNFLRPYPXRPXMRLhNWLVXNJNiXRPRLMNFPYXFRLYWPXMJQRLNFLOYFKLNhLPSQL[NXFPRLYVNdQ_
18(5&&"#4#
A7$4/$"82&($"&.'(&6&$7&.$7$(/842!'(4#
G&(&4774($" 3 8(5&5& '#8#&(&4774($":
G&(&&"&7/$ ;0r;@s& 2.7/!$!7$"*422$(,tpuv88n+=30k"G36C&%$46w:p=C$"$2&!n*6$"n:vt+=A$!2&9*7&npvvHpxHvnHpv+=
kC>A*7&npvxHptHpu+=1>C(7*6v:x+=36(*yw:pz:po+=f<(7*6v:x+=0*6w:o:v+=> 2.7/($4*6v:n:v+=rm!$!7$"
*1C0;-$"96o:prm6v:p+=.7&(*6woBv+=B"&(*6n:n:p+=>l0G*6w:w:w+={()<$"#*6n:w:v+=62&(4)*n:w:p+=%"2(7*v:o:|+=
ffG'9*ffC&!}wt:|w+=k>Ar;*6w+=$%&!)*6v:v:n+=13A07&(*6v:vp:n+=C&!~;-*6z+=C@>;*6v:v+=>5*6n:w:w+
<2&"'4$!('($7$E$"%4'(2&7%$()2 8(5&()&(&4"(&7(()&4).'("(/(#4$.#$"!'.7$)#7$(&('=8(5&2'(.2&#&6&$7&.7(#$(D6$5:
j ("%7/"4'&%4##!$($"$"&422'"$(/!$(/*:%:k$(l'.+:1()3&('0&4)%'$#7$" 8'.2$(($"%4#? 8(5&88'()$"82&($":
G&(&
A7$4/$"82&($"&.'(&6&$7&.$7$(/8#&(&
>772&"'4$!(2'($"47'#&#&(&&6&$7&.$7$(/(&(2"(:;)$ (&(2"()'7#!6$#()8775$"%$"82&($"=5)&!!7$4&.7,
H>44 $"4#='"$m'$#"($8$=5.7$"98!'.7$47/&6&$7&.7#&(&(
H>7$(88$%' ()&()&6& 4$&(#&5#&(&
H>#4$!($"8&"/($4($" "#&(&&6&$7&.$7$(/


>77&5 m'"4#&(&4774(#$"()$ ('#/&"# m'"4& 2.7$ )&6."#!$(#&(()@'!&"3'47($#>4)$6*@3>+:>44 $"4# 8&5#&(&

&"#& 2.7$ &7$(#$" '!!72"(&/(&.7,1'!!72"(&/;&.7vu*& 2.7$+=1'!!72"(&/;&.7vp*& 2.7/&5#&(&+=1'!!72"(&/;&.7u
*jk1 m'"4$"%+=1'!!72"(&/;&.7o*l$H+=1'!!72"(&/;&.7|*G>;Hm+:> 2.7$=&""(&($" &"#&"&7/$'7(5#!$(#'"#&#$%$(&7
.`4($#"($8$*G{r+$"()AkA!$(/zt'$"%()G>-'.2$$" /(2zx&"#&44 $.7'"#()€0-)((!,DD#B:#$:%Dvp:ouutD$!9DnpnpDnu:
> 2.7$ &"#%"&""(&($" 4&"&7.#5"7&##82)((!,DD.&7/H!&"%"2:$!9H%&(7.":#:;)f&7/A#$%&(&7%'$&6&$7&.7&()((!,DD
%".&"9:6'6:4ED.&7/D!#$%D:

0
<A7$&7#7H4((!)"4.$87$5(4  !  ($ " %

12345656762589 56 5317
)&($().(8$(8/'&4):r8/'&"('=&#()&!!!$&( 4($" .82&9$"%/'74($":
q -$8 4$"4 f)&6$'&7? 4$&74$"4 @47%$4&7=67'($"&/?"6$"2"(&74$"4
<&8"44!/8()#4'2"(5$()&774($"="&(':42D#4'2"(D"H!($"%H'22&/H87&(:!#8

->7$78('#$2'4$(#"$447"(")(!'#$"/#  $% "
(6"5)"()#$47'$
$"%&($6:

7425
1&2!7 $E >44 $" 8%"2& 2.7/54)"(
npv|3&('k"($4+:
(4
 6()#$6$(/*A>+!&4 8‚np=ppp#2($4&(#.&7/%"(/! *C$7"(&7:=
>44 $" 8m'"4$"%*jk1=l$H+54)".&# "(
" )4 (#8$"#. ./C$
/ 7"(&7:
3 &2!7 $E4&74'7&($"5& !82#:1&2!7 54)"822&`%2!7&2%'! 6$#"($"A " > .4&'876&"48
.&7/%"($4*f&9=k7#"A2$=r%$+
G&(&B47'$" 33#&(&5B47'##:
0!7$4&($" >!&($&77/!7$4&(##$%"5& 2!7/#$$"8 "$7#($&7:3
3!7$4&($"5& #"$$"%
" "2& 2.7/&"# m'"4$"%:
0&"#2$E&($" <&$"7##($&75#"$$"
" &"#2$E#.749#$%":@B!$2"(&7.6&($" 5#"5$()'(!#8$"#%'!$"%:{()82 8
2$E&($"5"(76&"(((
 )$ ('#/:
f7$"#$"% f7$"#(($"%5& "(#"& & $(5& "(76&"((!
 7&"(274'7&%"($4&"#%"2$459:

0jj!m'$$"8(2&$"($%8   !  4$8$42& (   $& 7 = / ( 2 & " #2

"82&'()&.'(2(/! 82&($&7=B!$2"(&7/(2 &"#2()# '#$
$"2&
( )  #
" "/('#$:l=$"#$4&(5)()&4)2&($&7=
/(22()#7$(#$
$76&"((/
 '('#/:r8/'&"('$
$8&7$($(2&!!7$ (/
 '&4)=&#()&!!!$&( 4($".8 74($"%&!":
C&($&7?B!$2"(&7/(2 C()#
"D& r"676#$
$"(
" ) ('#/ "D& r"676#$
$"(
" ) ('#/
q >"($.#$ q )rAHm
q @'9&/($44777$" q <754/(2(/
q A&7&"(7%/ q C0rH.&#"'$2&%$"%
q >"$2&7&"#()%&"$2
q l'2&"&4)!&($4$!&"(
q 7$"$4&7#&(&




Article

PIEZO2 in sensory neurons and urothelial

cells coordinates urination

https://doi.org/10.1038/s41586-020-2830-7 Kara L. Marshall1, Dimah Saade2, Nima Ghitani3, Adam M. Coombs1, Marcin Szczot3,
Jason Keller4,5, Tracy Ogata2, Ihab Daou1, Lisa T. Stowers4, Carsten G. Bönnemann2,
Received: 18 March 2020
Alexander T. Chesler2,3 ✉ & Ardem Patapoutian1 ✉
Accepted: 22 July 2020

Published online: 14 October 2020

Henry Miller stated that “to relieve a full bladder is one of the great human joys”.
Check for updates Urination is critically important in health and ailments of the lower urinary tract cause
high pathological burden. Although there have been advances in understanding the
central circuitry in the brain that facilitates urination1–3, there is a lack of in-depth
mechanistic insight into the process. In addition to central control, micturition
reflexes that govern urination are all initiated by peripheral mechanical stimuli such as
bladder stretch and urethral flow4. The mechanotransduction molecules and cell
types that function as the primary stretch and pressure detectors in the urinary tract
mostly remain unknown. Here we identify expression of the mechanosensitive ion
channel PIEZO2 in lower urinary tract tissues, where it is required for low-threshold
bladder-stretch sensing and urethral micturition reflexes. We show that PIEZO2 acts
as a sensor in both the bladder urothelium and innervating sensory neurons. Humans
and mice lacking functional PIEZO2 have impaired bladder control, and humans
lacking functional PIEZO2 report deficient bladder-filling sensation. This study
identifies PIEZO2 as a key mechanosensor in urinary function. These findings set the
foundation for future work to identify the interactions between urothelial cells and
sensory neurons that control urination.

The mechanotransduction channel necessary for urinary reflexes feeling the need to void and therefore followed a voiding schedule.
remains unknown. Several ion channels have been implicated in uri- A healthy frequency is defined as five to six voids per day. Despite the
nary tract function in vivo5–7, but none have been shown to be required lack of normal sensory feedback, all patients had achieved continence
for micturition reflexes. Moreover, it is not clear which cells are the at the time of evaluation except for one nine year old. However, many
primary sensors: umbrella cells of the innermost layer in the urothelium patients reported sudden urge incontinence, where any delay in voiding
have been proposed to be mechanosensory8,9, but the bladder is also resulted in urinary accidents. Two individuals reported occasional noc-
innervated by mechanically sensitive afferents from dorsal root ganglia turnal enuresis, and four had stress incontinence caused by laughter,
(DRG)1,10. PIEZO2 is the primary mechanosensor that mediates touch, cough and/or postural changes, with one case being severe enough
proprioception and mechanical allodynia in mice11–15. Loss-of-function to require treatment. Several patients had a sensation of incomplete
mutations in PIEZO2 also resulted in complete deficits in these senses voiding and an irregular urinary stream. Three adults described a sen-
in humans13,16. Furthermore, PIEZO2 mediates interoceptive processes sation of pelvic heaviness when their bladder was full, and all three
such as lung-stretch sensing and baroreception in mice17,18, but intero- independently reported voiding by leaning over or using their hands
ceptive deficits have not been studied in humans who are deficient in to apply pressure to their lower abdomen. Overall, these data suggest
PIEZO2. As urination is driven by mechanical interoceptive reflexes, we that PIEZO2 has a key functional role in human urination.
investigated whether PIEZO2 is important for urination. We next carried out studies in mice to understand where and how
To understand how PIEZO2 contributes to urination in humans, PIEZO2 functions in the urinary tract. To test whether Piezo2 is present in
PIEZO2-deficient individuals (n = 12; 5–43 years of age) answered ques- bladder sensory neurons, we used RNA fluorescent in situ hybridization
tionnaires designed to capture pathology and validated against healthy (FISH) in DRG tissue taken from three mice after injection of cholera
control individuals to screen for voiding and elimination dysfunc- toxin B–Alexa Fluor 488 (CTB), a neuronal tracer, into the bladder wall
tion19 (Fig. 1). We also assessed urological history, previous medical (Fig. 2a). Out of 92 bladder-innervating neurons labelled with CTB, 75
evaluations and non-invasive bladder ultrasound scans (Supplementary expressed Piezo2 transcript (81.5%). Piezo2 transcript was also detected
Table 1). All patients reported decreased voiding frequency, as low as in a subset of bladder urothelial cells expressing Krt20 (Fig. 2b), a marker
once or twice daily, regardless of hydration status. Notably, the major- of umbrella cells that line the bladder lumen and have been proposed
ity of individuals reported that they could spend an entire day without to contribute to detection of bladder filling8. Seventy-four per cent of
1
Howard Hughes Medical Institute, Department of Neuroscience, Dorris Neuroscience Center, The Scripps Research Institute, La Jolla, CA, USA. 2National Institute of Neurological Disorders
and Stroke, National Institutes of Health, Bethesda, MD, USA. 3National Center for Complementary and Integrative Health, National Institutes of Health, Bethesda, MD, USA. 4Department of
Neuroscience, Dorris Neuroscience Center, The Scripps Research Institute, La Jolla, CA, USA. 5Present address: Howard Hughes Medical Institute, Janelia Research Campus, Ashburn, VA, USA.
✉e-mail: alexander.chesler@nih.gov; ardem@scripps.edu

290 | Nature | Vol 588 | 10 December 2020

 Patient 1 2
Age 23 13
Sex F F
In a normal day I go to the
* 3–4
washroom to pee (no. of times)
I pee in my underwear during
the day (days per week)
When I pee in my
underwear, it is
I feel that I have to rush to
the washroom to pee
I hold my pee by crossing
my legs or sitting down
It hurts when I pee
I wet my bed at night
I wake up to pee at night
3–4
(nights per month)
When I pee, it
stops and starts
I have to push or wait for
my pee to start
Fig. 1 | Urinary dysfunction in individuals deficient in PIEZO2. Patient
numbers correspond to those in Supplementary Table 1. Grey indicates a
neutral answer or one not indicating pathology. Urinary frequency information
is scored differently to the other questions, and is colour-coded according
to the pathological score assigned to the answer in the questionnaire.
Unless otherwise noted, the colour code indicates the following: grey, never
umbrella cell nuclei were associated with Piezo2 transcript, and 12.6% of
cells showed high expression of Piezo2 (1 s.d. above the mean). The den-
sity of Piezo2-positive cells varied across the bladder. These results show
that Piezo2 is expressed in two distinct cell types in the lower urinary
tract and could function in detecting relevant mechanical stimuli. We
confirmed that urothelial cells express other mechanosensory proteins
(Extended Data Fig. 1), but the functional deficits in PIEZO2-deficient
humans focused our work on this protein.
We next used calcium imaging to determine whether bladder-stretch
responses in sensory neurons were dependent on Piezo2. Whether neu-
rons detect bladder stretch directly or downstream of urothelial cell
activation, we expect this stimulus to cause calcium influx. We injected
a viral vector carrying Cre recombinase into postnatal day 0 (P0) to P3
pups carrying the Cre-dependent calcium indicator GCaMP6f and a
conditional Piezo2-knockout allele, Piezo2cKO (Piezo2fl/fl-GCaMP6f+/+; con-
trols were GCaMP6f+/+)13. Thus, Piezo2 was deleted anywhere GCaMP6f
was expressed. We observed rapid, robust responses in control sacral
level 1 (S1) DRG neurons in response to manual, high-pressure blad-
der filling with saline (Fig. 2c), but these responses were markedly
attenuated in Piezo2cKO DRG cells (Fig. 2d). Notably, cells responding
to low-pressure stimuli were completely absent in Piezo2-knockout
DRG (Fig. 2e). Calcium traces for cells in wild-type DRG revealed graded
responses to pressure stimuli, with many cells responding to low and
high pressures, but Piezo2cKO cells were silent at low pressures. Piezo2cKO
DRG also had fewer cells responding to bladder stretch (Fig. 2f–m), but
normal numbers of cells responding to painful pinch (Extended Data
Fig. 1). This suggests that PIEZO2 is a key sensor of bladder stretch.
Mechanically-evoked micturition reflexes coordinate the bladder
detrusor muscle and the urethral sphincter muscles to mediate efficient
urinary control, and are critical for efficient voiding4. We hypothe-
sized that reflex responses rely on PIEZO2 function to provide feed-
back control over bladder pressure and urethra activity. We therefore
investigated micturition reflexes in mice lacking PIEZO2 in all caudal
tissues. These Hoxb8Cre;Piezo2fl/fl mice express Cre recombinase in
bladder-innervating DRG neurons14 and bladder urothelium (Extended
Data Fig. 2a). We used cystometry with urethral electromyography
to simultaneously monitor bladder pressure and sphincter activity.
With continual filling at 30 μl min−1, control mice initiated bladder
contractions at regular intervals, which are observed as pressure peaks
3 5 7 9 10 12
16 18 10 9 5 36
M M F F F F
1–2 1–2 1–2 3–4 3–4 5–6†
1 1
Damp Damp
(no pathology); blue, less than half of the time (pathology score of 1); yellow,
half of the time (pathology score of 2); orange, more than half of the time
(pathology score of 3); red, every day or every night if night time is indicated in
question (pathology score of 4). Asterisk indicates an unanswered question.
Dagger indicates that the individual answered twice per day during clinical
interview.
that correspond to micturition events (Fig. 3a, Extended Data Fig. 2b).
Piezo2-knockout mice displayed irregular micturition timing (Fig. 3b,
Extended Data Fig. 2c) and, on average, longer intervals between blad-
der contractions that resulted in urination (Fig. 3c, Extended Data
Fig. 2h). Therefore, PIEZO2-deficient mice are less sensitive to blad-
der filling, as it takes more volume to initiate bladder contractions.
Piezo2-knockout mice also displayed higher bladder pressures five
seconds before contraction peaks (Fig. 3d, Extended Data Fig. 2i). In
healthy animals, low pressures are maintained during bladder filling
because the detrusor muscle relaxes. The higher pressures that were
observed before contractions suggest that this relaxation process or
bladder compliance is impaired in Piezo2-knockout mice.
We next investigated individual micturition events to determine
whether the bladder pressure required to sustain micturition was
abnormal in these mice. We observed consistent pressure increases
within and among wild-type mice (Fig. 3e, Extended Data Fig. 2d), but
bladder pressure traces in Piezo2-knockout mice were highly variable
(Fig. 3f, Extended Data Fig. 2e). The knockout mice exhibited higher
peak bladder pressures (Fig. 3g), and required significantly more pres-
sure during contractions, suggesting that the detrusor muscle must
work harder to accomplish micturition (Fig. 3h, Extended Data Fig. 2j).
We also assessed whether sensory input via PIEZO2 is important for
urethral reflexes, which sustain efficient urination. During bladder
contractions in wild-type mice, there was coordinated engagement of
the urethra muscle (Fig. 3i). This reliable urethral activity was markedly
attenuated in Piezo2-knockout mice (Fig. 3j, k, Extended Data Fig. 2k).
Knockout urethra responses varied from silent or weak coordination to
inappropriately timed hyperactivity (Extended Data Fig. 2c). Hyperac-
tivity is a sign of detrusor–sphincter dyssynergia, a condition involving
uncoordinated communication between the muscle groups responsi-
ble for urination. This indicates that the urethra was not receiving the
appropriate sensory input to govern its activity during micturition.
Mice lacking PIEZO2 also had more variable, larger void volumes, as
longer periods between contractions allow more bladder filling (Fig. 3l).
Together, these data indicate that PIEZO2 sets stretch sensitivity in
the lower urinary tract and initiates appropriately timed reflexes that
contribute to efficient urination.
Next, we tested whether urination behaviour is altered in
Piezo2-knockout mice. We placed mice on filter paper for 4 h and
Nature | Vol 588 | 10 December 2020 | 291
Article
a b
CTB Piezo2 Merge
c Wild type
d
f g
Pressure
20 mmHg
h i
300 s
Calcium avg.
(all cells)
10% ΔF/F
j 20
Calcium
% ΔF/F
traces
3 cells
5 5
l 20
% ΔF/F
Calcium
peak
5 5
Fig. 2 | Piezo2 is expressed in the lower urinary tract, and sensory neurons
require PIEZO2 to detect low-pressure bladder filling. a, DRG neurons were
retrogradely labelled using CTB (cyan, left) and fluorescent in situ
hybridization (FISH) of DRGs with probes targeting Piezo2 (magenta, middle).
Arrowheads point to Piezo2-expressing bladder neurons. Scale bar, 50 μm. The
tracing experiment was repeated using three mice, n = 22–36 cells analysed per
mouse. b, FISH for Krt20 (green) and Piezo2 (magenta) in bladder. Arrowheads
point to Piezo2-expressing umbrella cells. Scale bar, 50 μm. FISH was
performed on three bladders, with two technical replications. Analysis was
performed on 80–117 nuclei per bladder. c, d, Image z-stack from GCaMP6f+/+
control mouse (c) and Piezo2 cKO mouse (d) S1 DRG during bladder filling.
e, Count of cells responding to low-pressure (black) and high-pressure (red)
imaged the resulting urination patterns with UV illumination. We used
only female mice to preclude territorial scent-marking behaviour.
Wild-type mice typically urinated in the corners and edges of the cage
in large spots (Fig. 3m). Piezo2-knockout mice had a variety of urination
patterns, and some displayed urine leaking (small spots) or large voids
towards the cage centre (Fig. 3n, o). This phenotype was not attributed
to the knockout mice spending more time in the middle of the cage
(Fig. 3p). Thus, knockout mice have abnormal urination behaviour,
including some apparent incontinence.
We next studied whether this observed urinary dysfunction in
Piezo2-knockout mice led to long-term consequences. Chronic uri-
nary tract dysfunction typically causes tissue remodelling as the
bladder wall grows thicker to compensate for inefficient voiding20,21.
This remodelling can eventually result in ‘decompensation’, which is
marked by a flaccid, ineffective bladder with sequelae of incomplete
voiding, vesicoureteral reflux and increased frequency of urinary tract
infections. Bladder-wall thickening was observed by haematoxylin and
eosin staining in Piezo2-deficient mice (Fig. 3q–s). The weight of freshly
excised bladders also revealed bladder-wall remodelling, as bladders
from the knockout mice were significantly heavier than those from
wild-type littermates (Fig. 3t, Extended Data Fig. 2m). Thus, impaired
292 | Nature | Vol 588 | 10 December 2020
DAPI Krt20 Piezo2 Merge
Knockout e DRG bladder-fill responses
High
No of. responding
30
(>20 mmHg)
20 Low
cells
(<20 mmHg)
10
0
WT KO
Pressure
20 mmHg
300 s
Calcium avg.
(all cells)
10% ΔF/F
k 20
Calcium
% ΔF/F
traces
m 20
% ΔF/F
Calcium
peak
bladder-filling stimuli in wild-type (WT) (n = 3 mice) and Piezo2 cKO (KO) (n = 4
mice) DRGs. f, g, Example pressure trace from wild-type (f) and Piezo2 cKO (g)
DRG. Stimuli were interleaved during recording, but are shown sorted low to
high, hence the discontinuous line. Data below these graphs are sorted
together with the respective pressure peaks. h, i, Average per cent change in
calcium fluorescence for all responding cells during the pressure peaks shown
in f, g, respectively. j, k, Calcium traces for individual wild-type cells that
responded to pressure stimuli in f (n = 17) ( j) and for Piezo2 cKO cells responding
to pressure stimuli shown in g (n = 6) (k). Each cell’s responses are shown on the
same horizontal line. Cells are sorted by cumulative response to the four
lowest-pressure stimuli. l, m, Maximum calcium response for the
corresponding cells in j, k, 1 s after pressure peak.
urinary reflexes in Piezo2-knockout mice lead to detrusor hypertrophy,
an indicator of chronic voiding dysfunction.
We next investigated the cell types in which PIEZO2 was required.
We tested whether PIEZO2 deficiency in urothelial cells changed the
pressure threshold that is required to initiate micturition. We used the
Upk2-cre allele to abrogate PIEZO2 activity in urothelial cells (Extended
Data Fig. 3a), which have been proposed to act as stretch sensors and
communicate to underlying neurons using ATP7,9,22–24. We found that
Upk2-cre;Piezo2fl/fl knockout mice displayed similar phenotypes to the
Hoxb8-cre;Piezo2fl/fl knockout mice, with higher bladder stretch thresh-
olds, increased bladder pressure during micturition and attenuated
urethral reflexes (Fig. 4a–h). In combination with expression data from
FISH (Fig. 2b), these data indicate that PIEZO2 acts in umbrella cells to
help set bladder-stretch sensitivity and initiate appropriate micturition
reflexes. These results confirm the proposed role for umbrella cells
as mechanosensory cells that participate in initiating micturition8.
We observed similar phenotypes in mice that lacked PIEZO2
only in sensory neurons (Fig. 4i–p). Deleting PIEZO2 in all sensory
neurons is lethal, so we used Scn10a-cre mice25 (Scn10a encodes the
voltage-gated sodium channel Nav1.8) to delete PIEZO2 in the Aδ- and
c-fibre subsets, which are the primary sensory neuron types described
 a Wild type
b Knockout
c d
Contraction intervals Pre-contraction pressure
600 **** 25 ****

Interpeak interval (s)

20

Pressure (mmHg)
400
15

10
20 mmHg 200
5

0 0
100 s
WT KO WT KO
e Wild type f Knockout g h
25
Bladder pressure (mmHg)

Peak bladder pressures Bladder contractions

Contraction AUC (mmHg s)

40 **** 20 ****

Peak pressures (mmHg)

20

15 30 15

10 20 10

5 10 5

0
0 0
–20 –10 0 10 20 –20 –10 0 10 20
Time from peak (s) Time from peak (s) WT KO WT KO

i j k Urethra contractions l Void volume

15
2 × 106 **** 150 *
Urethra activity (μV)

Sum activity (μV)

10

Volume (μl)
100
1 × 106
5 50

0 0 0
–20 –10 0 10 20 –20 –10 0 10 20 WT KO WT KO
Time from peak (s) Time from peak (s)
m o q s t
Filter paper urination
Bladder muscle Bladder weight
***
Urine in centre (%)

80 * ***
Wild type

60 600 50
40 Muscle thickness (μm)
40
20

Weight (mg)
400
0 30
n p WT KO r
Time in centre (%)

40 20
Knockout

200
30
20 10
10
0 0
0 WT KO WT KO
WT KO

Fig. 3 | PIEZO2 is required for efficient micturition reflexes. a, b, Example
pressure traces from three female wild-type mice (a) and three female
Hoxb8-cre;Piezo2fl/fl mice (KO) (b) during continuous bladder filling.
c, d, Bladder-contraction intervals (P < 0.0001) (c) and bladder pressure five
seconds before contraction peaks (P < 0.0001) (d). e, f, Heat maps showing
bladder contractions in wild-type (e) and Hoxb8-cre;Piezo2fl/fl (f) female mice.
Each row represents bladder pressure during a single micturition event, with
peaks aligned at 0. Arrowheads mark where data from one animal end and data
from another begin. g, h, Peak bladder pressures (P < 0.0001) (g) and area under
the curve (AUC) for bladder contractions (P < 0.0001) (h). i, j, Heat maps showing
urethra activity in wild-type (i) and Hoxb8-cre;Piezo2fl/fl (j) female mice from
e, f, with rows corresponding to bladder-contraction events in e, f. k, Urethra
activity during micturition (P < 0.0001). l, Void volume measurements (P = 0.03).
Student’s t-tests with Welch’s correction. In c–l, n = 6 (wild-type) and n = 5
(Hoxb8-cre;Piezo2fl/fl) female mice; n = 10–29 bladder contractions analysed per
mouse. m, n, Urination patterns of five wild-type (m) and five Hoxb8-cre;Piezo2fl/fl (n)
mice. o, Quantification of urine in the middle 50% of the cage (P = 0.0001). n = 11
female mice per group. p, Wild-type and Hoxb8-cre;Piezo2fl/fl mice spend similar
amounts of time in the cage centre. q, r, Haematoxylin and eosin staining from
wild-type (q) and Hoxb8-cre;Piezo2fl/fl (r) bladder sections, from 6- to 7-month-old
littermates. Scale bars, 100 μm. The muscle layer is marked with vertical lines.
s, t, Bladder muscle wall thickness (s; n = 5, P = 0.016) and total bladder weight
(t; n = 9 (wild type) and 8 (Hoxb8-cre;Piezo2fl/fl), P = 0.0002). In o, s, t, Mann–
Whitney test; otherwise, two-sided Student’s t-tests with Welch’s correction.
Data are mean ± s.d.

in the bladder10,26. This mouse line does not induce recombination in mice, but not in Scn10a-cre;Piezo2fl/fl mice. Neuronal PIEZO2-knockout
urothelial cells (Extended Data Fig. 3b–e). Sensory-neuron-specific mice do require more bladder pressure for micturition and have highly
Piezo2-knockout mice displayed longer intervals between contrac- attenuated urethral reflex responses (Fig. 4o, p). These data implicate
tions (Fig. 4i), but the pressure before contractions was not different PIEZO2 in mediating neuronal stretch responses that are critical for
from that in wild-type mice, as it was in urothelial-specific- and full downstream urethral reflexes. Of note, mice with Piezo2 knockout in
caudal-knockout mice (Fig. 4j). This implies that mechanosensory individual tissues did not display the marked bladder remodelling that
stimuli activate PIEZO2 in umbrella cells to initiate bladder relaxation was observed in full caudal-knockout mice (Fig. 3s, t, Extended Data
during filling (Fig. 4b) and that neuronal mechanosensing is dispensa- Fig. 3f, g), suggesting that urothelial or neuronal PIEZO2 alone could
ble for this process. Alternatively, it is possible that bladders become still contribute to urinary function. These results indicate that there is a
fibrotic and less compliant in Upk2-cre;Piezo2fl/fl and Hoxb8-cre;Piezo2fl/fl two-part signalling mechanism involving PIEZO2 in umbrella cells and

Nature | Vol 588 | 10 December 2020 | 293

Article
Upk2-cre;Piezo2fl/fl Scn10a-cre;Piezo2fl/fl

a Contraction intervals b Pre-contraction pressure i Contraction intervals j Pre-contraction pressure

600 **** 30 *** ** 30 NS

Pressure pre-peak (mmHg)

800

Pressure pre-peak (mmHg)

Interpeak interval (s)

Interpeak interval (s) 500

400 20 20

250
200 10 10

0 0 0 0
WT KO WT KO WT KO WT KO
c Wild type d Knockout k Wild type l Knockout
Bladder pressure (mmHg)

Bladder pressure (mmHg)

25 25
20 20

15 15

10 10
5 5
0 0
–20 –10 0 10 20 –20 –10 0 10 20 –20 –10 0 10 20 –20 –10 0 10 20
Time from peak (s) Time from peak (s) Time from peak (s) Time from peak (s)
e f m n
15 15
Urethra activity (μV)

Urethra activity (μV)

10 10

5 5

0 0
–20 –10 0 10 20 –20 –10 0 10 20 –20 –10 0 10 20–20 –10 0 10 20
Time from peak (s) Time from peak (s) Time from peak (s) Time from peak (s)
g Bladder contractions h Urethra contractions o Bladder contractions p Urethra contractions
**** * ** ****
15 20
Contraction AUC (mmHg)
Contraction AUC (mmHg)

2 × 106 2 × 106
Sum activity (μV)

Sum activity (μV)

15
10
1 × 106 10 1 × 106
5
5
0 0
0 0
WT KO WT KO WT KO WT KO
fl/fl
Fig. 4 | PIEZO2 functions in both bladder urothelium and sensory neurons. wild-type and Scn10a-cre;Piezo2 (KO) mice. n = 3 wild type and 3
a–h, Cystometry data from wild-type and Upk2-cre;Piezo2fl/fl (KO) mice. n = 5 Scn10a-cre;Piezo2 fl/fl female mice; 11–24 contractions per mouse. Cartoon in the
wild type and 4 Upk2-cre;Piezo2 fl/fl female mice; 18–49 bladder contractions top right depicts the lower urinary tract, with Piezo2 KO tissue in red.
analysed per mouse. Cartoon in the top right depicts the lower urinary tract, i, j, Intervals between bladder contractions (P = 0.002) (i) and bladder
with Piezo2 KO tissue in red. a, b, Intervals between bladder-contraction voids pressures five seconds before peak contraction ( j). k, Bladder pressure events
(P < 0.0001) (a) and bladder pressures five seconds before peak contraction during continuous filling cystometry in wild-type (k) and Scn10a-cre;Piezo2fl/fl (l)
(P = 0.001) (b). c, d, Bladder pressure events during continuous filling mice. m, n, Urethra activity recorded during the bladder contraction events
cystometry in wild-type (c) and Upk2-cre;Piezo2fl/fl (d) mice. e, f, Urethra activity shown in k, l. o, p, Bladder pressure during micturition events (P = 0.004) (o)
recorded during the bladder contraction events shown in c, d. g, h, Bladder and urethral reflex responses during micturition (P < 0.0001) (p). Data are
pressure during micturition events (P < 0.0001) (g) and urethral reflex mean ± s.d. Two-sided Student’s t-test with Welch’s correction. *P ≤ 0.05,
responses during micturition (P = 0.03) (h). i–p, Cystometry data from **P ≤ 0.01, ***P ≤ 0.001 and ****P ≤ 0.0001.

sensory neurons that set bladder sensitivity and promote micturition For example, the mechanotransduction ion channels TMEM63B and
reflexes. Further investigations are required to address how these cell PIEZO1 are widely expressed in the urothelium, and PIEZO1 partially
types communicate. mediates urothelial stretch responses in vitro27.
We have used evidence from mice and humans to identify the mecha- Our results suggest a two-part model of mechanosensory signalling
notransduction channel PIEZO2 as a critical mediator of urinary tract in the urinary tract, which is reminiscent of epithelial cell–neuronal
function. Absence of Piezo2 in mice does not result in urinary tract sensory machinery in the skin (Merkel cell–neurite complexes), lung
paralysis and death, and PIEZO2-deficient humans are still able to uri- (neuroepithelial bodies) and intestine (enterochromaffin cells)15,17,28,29.
nate. This indicates that there are mechanotransduction proteins other Our results also implicate umbrella cells in mediating bladder relaxa-
than PIEZO2 in the urothelium and lower urinary tract sensory neurons. tion during filling, perhaps by signalling to bladder muscle and/or

294 | Nature | Vol 588 | 10 December 2020

through stretch-induced cellular changes30. Future studies will address
how urothelial cells and sensory neurons cooperate to control urinary
function.
Online content
Any methods, additional references, Nature Research reporting sum-
maries, source data, extended data, supplementary information,
acknowledgements, peer review information; details of author con-
tributions and competing interests; and statements of data and code
availability are available at https://doi.org/10.1038/s41586-020-2830-7.
1. de Groat, W. C. & Yoshimura, N. Afferent nerve regulation of bladder function in health
and disease. Handb. Exp. Pharmacol. 194, 91–138 (2009).
2. Keller, J. A. et al. Voluntary urination control by brainstem neurons that relax the urethral
sphincter. Nat. Neurosci. 21, 1229–1238 (2018).
3. Hou, X. H. et al. Central control circuit for context-dependent micturition. Cell 167, 73–86
(2016).
4. Garry, R. C., Roberts, T. D. & Todd, J. K. Reflex responses of the external urethral sphincter
of the cat to filling of the bladder. J. Physiol. 139, 13–14 (1957).
5. Cockayne, D. A. et al. Urinary bladder hyporeflexia and reduced pain-related behaviour in
P2X3-deficient mice. Nature 407, 1011–1015 (2000).
6. Andersson, K. E., Gratzke, C. & Hedlund, P. The role of the transient receptor potential
(TRP) superfamily of cation-selective channels in the management of the overactive
bladder. BJU Int. 106, 1114–1127 (2010).
7. Mochizuki, T. et al. The TRPV4 cation channel mediates stretch-evoked Ca2+ influx and
ATP release in primary urothelial cell cultures. J. Biol. Chem. 284, 21257–21264 (2009).
8. Merrill, L., Gonzalez, E. J., Girard, B. M. & Vizzard, M. A. Receptors, channels, and
signalling in the urothelial sensory system in the bladder. Nat. Rev. Urol. 13, 193–204
(2016).
9. Apodaca, G., Balestreire, E. & Birder, L. A. The uroepithelial-associated sensory web.
Kidney Int. 72, 1057–1064 (2007).
10. Zagorodnyuk, V. P., Brookes, S. J., Spencer, N. J. & Gregory, S. Mechanotransduction and
chemosensitivity of two major classes of bladder afferents with endings in the vicinity to
the urothelium. J. Physiol. 587, 3523–3538 (2009).
11. Murthy, S. E. et al. The mechanosensitive ion channel Piezo2 mediates sensitivity to
mechanical pain in mice. Sci. Transl. Med. 10, (2018).
12. Ranade, S. S. et al. Piezo2 is the major transducer of mechanical forces for touch
sensation in mice. Nature 516, 121–125 (2014).
13. Szczot, M. et al. PIEZO2 mediates injury-induced tactile pain in mice and humans.
Sci. Transl. Med. 10, (2018).
14. Woo, S. H. et al. Piezo2 is the principal mechanotransduction channel for proprioception.
Nat. Neurosci. 18, 1756–1762 (2015).
15. Woo, S. H. et al. Piezo2 is required for Merkel-cell mechanotransduction. Nature 509,
622–626 (2014).
16. Chesler, A. T. et al. The role of PIEZO2 in human mechanosensation. N. Engl. J. Med. 375,
1355–1364 (2016).
17. Nonomura, K. et al. Piezo2 senses airway stretch and mediates lung inflation-induced
apnoea. Nature 541, 176–181 (2017).
18. Zeng, W. Z. et al. PIEZOs mediate neuronal sensing of blood pressure and the
baroreceptor reflex. Science 362, 464–467 (2018).
19. Afshar, K., Mirbagheri, A., Scott, H. & MacNeily, A. E. Development of a symptom score for
dysfunctional elimination syndrome. J. Urol. 182, 1939–1944 (2009).
20. Ehrhardt, A. et al. Urinary retention, incontinence, and dysregulation of muscarinic
receptors in male mice lacking Mras. PLoS ONE 10, e0141493 (2015).
21. Flum, A. S. et al. Testosterone modifies alterations to detrusor muscle after partial
bladder outlet obstruction in juvenile mice. Front Pediatr. 5, 132 (2017).
22. Takezawa, K. et al. Authentic role of ATP signaling in micturition reflex. Sci. Rep. 6, 19585
(2016).
23. Takezawa, K., Kondo, M., Nonomura, N. & Shimada, S. Urothelial ATP signaling: what is its
role in bladder sensation? Neurourol. Urodyn. 36, 966–972 (2017).
24. Ferguson, D. R., Kennedy, I. & Burton, T. J. ATP is released from rabbit urinary bladder
epithelial cells by hydrostatic pressure changes–a possible sensory mechanism?
J. Physiol. 505, 503–511 (1997).
25. Agarwal, N., Offermanns, S. & Kuner, R. Conditional gene deletion in primary nociceptive
neurons of trigeminal ganglia and dorsal root ganglia. Genesis 38, 122–129 (2004).
26. Sengupta, J. N. & Gebhart, G. F. Mechanosensitive properties of pelvic nerve afferent
fibers innervating the urinary bladder of the rat. J. Neurophysiol. 72, 2420–2430 (1994).
27. Miyamoto, T. et al. Functional role for Piezo1 in stretch-evoked Ca2+ influx and ATP release
in urothelial cell cultures. J. Biol. Chem. 289, 16565–16575 (2014).
28. Maksimovic, S. et al. Epidermal Merkel cells are mechanosensory cells that tune
mammalian touch receptors. Nature 509, 617–621 (2014).
29. Alcaino, C. et al. A population of gut epithelial enterochromaffin cells is
mechanosensitive and requires Piezo2 to convert force into serotonin release. Proc. Natl
Acad. Sci. USA 115, E7632–E7641 (2018).
30. Wang, E. C. et al. ATP and purinergic receptor-dependent membrane traffic in bladder
umbrella cells. J. Clin. Invest. 115, 2412–2422 (2005).
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in
published maps and institutional affiliations.
© The Author(s), under exclusive licence to Springer Nature Limited 2020
Nature | Vol 588 | 10 December 2020 | 295
Article
Methods
All experiments were performed within the protocols and guidelines
approved by the Institutional Animal Care and Use Committees of The
Scripps Research Institute in compliance with regulatory standards
established by the Association for Assessment and Accreditation of
Laboratory Animal Care International (AAALAC).
Statistics
Unless otherwise noted in the legends, groups were compared using
two-tailed Student’s t-test with Welch’s correction, as groups did not
have equal variances. For comparisons of groups with an n less than 15,
we used the non-parametric Mann–Whitney test to assess differences,
as we could not assess distributions. These tests were indicated in the
figure legends. No statistical test was used to pre-determine sample
size. Instead, sample size was determined by animal availability and
previous studies in the field, which found these sample sizes sufficient
to detect deficits.
Study design
We established exclusion criteria before collecting cystometry data:
data from the first 30 min of cystometry recording was not used because
bladder muscle activity has often not stabilized. Moreover, animals
that displayed bladder leaking during recording were excluded from
analysis, as leaking indicated a flawed seal and thus inaccurate filling
responses. To verify the reproducibility of experimental findings, we
restricted the time of day that cystometry recordings were done (Zeit-
geber 8–14) and we performed every experiment in a cohort of male
and female mice to compare to their wild-type littermates. The order of
recordings for different genotypes was randomized. The experimenter
was blind to genotype when possible. Hoxb8-cre+;Piezo2fl/fl knockout
mice have obvious motor impairments, so the experimenter was never
blind to genotype for these groups.
Mice
Mice were kept in standard housing with 12:12 h light:dark cycle set
with lights on from 06:00–18:00, with room temperature kept around
22 °C, with humidity between 30–80% (not controlled). Adult male and
female mice were used as indicated in the text. Age-matched knockout
and wild-type littermates were tested at the same age in each cohort, but
ages tested ranged from 5–8 months for Hoxb8-cre;Piezofl/fl, 6–12 months
for Scn10a-cre and 4–6 months for Upk2-cre). The Hoxb8-cre;Piezofl/fl l
mouse line has been previously described 11. GCaMP6f+/+ mice
(B6;129S-Gt(ROSA)26Sortm95.1(CAG-GCaMP6f )Hze/J, Jackson Laboratory:
Ai95, 024105) were bred to Piezo2fl/fl mice as described previously13,31.
Piezo2fl/fl mice were mated with Scn10a-cre mice25 or Upk2-cre mice
(B6(129)-Tg(Upk2-cre)1Rkl/WghJ, Jackson Laboratory: 029281) to cre-
ate sensory-neuron-specific and urothelial-specific Piezo2-knockout
animals, respectively. Each of these Cre lines was also crossed with
Ai9 mice (B6.Cg-Gt(ROSA)26Sortm9(CAG-tdTomato)Hze/J, Jackson Laboratory:
07909) to assess Cre expression. Genotyping was performed using
guidelines from Jackson Laboratory.
Retrograde labelling of sensory neurons
Mice were anesthetized with isoflurane. Their lower abdomen was
shaved and hair was removed using Nair (Fisher Scientific: NC0132811),
and sterilized with ethanol and iodine, cleaned and a midline laparot-
omy was performed to expose the bladder. Three to five injections of
1–2 μl of cholera toxin B–Alexa 488 (Fisher Scientific: C22841) were
made in the bladder wall using a Hamilton syringe. Care was taken to
avoid puncturing through to the bladder lumen. The abdominal wall and
skin were sutured separately, and mice were given subcutaneous
flunixin (0.1 ml per g body weight) for post-operative care. We waited
three to five days before taking tissue to allow the dye to reach the
cell soma.
Fluorescent in situ hybridization
Bladder and DRG tissues were removed immediately from eutha-
nized animals, and flash frozen in liquid nitrogen. The protocol for
RNAscope Multiplex Fluorescent Reagent Kit V2 (ACDBio: 323100) was
followed according to instructions for fresh frozen tissue. A gentler
protease (III) was applied for only 25 min to lessen CTB-fluorophore
degradation in DRG tissue and delicate bladder tissue. Probes for
Piezo2 (ACDBio: 400191 or 439971), Krt20 (ACDBio: 402301), Piezo1
(ACDBio:) or Tmem63b (ACDBio: 431531) were applied to detect tran-
script. Quantification of was performed in ImageJ using regions of
interest to define quantification area and then by measuring mean pixel
intensity. Only DRG cells with nuclei visible were quantified to prevent
double-counting of the same cells in different sections. Multinucleated
umbrella cell borders could not always be defined, so quantification
was done per nucleus, with a set region of interest (ROI) size used for all
nuclei (roughly twice the diameter of the nucleus). ROIs were centred
on each nucleus and used to measure mean pixel intensity. Control area
intensity was measured from tissue background to define an intensity
cut-off for negative cells.
DRG calcium imaging
Viral strategy is previously described13. Animals were anesthetized
with isoflurane and the bladder was exposed (described above). The
bladder apex was opened and phlanged PE20 tubing was filled with
saline and inserted as a catheter (Stoelting: 51154) into the bladder, and
tied off using silk suture (Fisher Scientific: NC9140103). A syringe and
lines filled with saline that were connected to the catheter was used to
fill the bladder and test for leaking. The abdomen was sewn shut with
the catheter extending out. The animal was flipped to prone position,
and the vertebral column was exposed. A microdrill was used to open a
window in the bone above the first sacral DRG. Epifluorescence imaging
was performed using an upright microscope (FVMPE-RS, Olympus)
equipped with a 4×, 0.28–numerical aperture air objective. Illumination
was provided with a 130-W halogen light source (U-HGLGPS, Olympus),
using a standard green excitation/emission filter cube. Images were
acquired using an ORCA-Flash 4.0 CMOS camera (Hamamatsu) at a
frame rate of 5 Hz using MetaMorph (Molecular Devices. Analysis was
previously described13.
Cystometry and electromyography
Male and female mice were anesthetized by isoflurane (5% induction,
1–2% maintenance, Kent Scientific SomnoSuite) and the bladder was
catheterized and connected to saline lines (described above). Tungsten
electrodes were inserted directly into the urethral muscle (A-M systems:
795500). Saline lines were connected to a pressure sensor (Biopac:
RX104A-MRI) which connected via pressure transducer (TSD104A)
to an MP160 Biopac system amplifier (DA100C). Electromyography
electrodes were connected to a differential amplifier (EMG100C: gain
1,000, sample rate 10 kHz, low-pass filter 5 kHz, 60 Hz notch filter and
100 Hz high-pass filter). Bladder was continuously filled at 20 μl min−1
using a syringe pump until regular urination cycles began. Data was
not collected until the mouse had been stably cycling (30–40 min after
beginning of recording), at which point filling rate was increased to
30 μl min−1. Data were logged with Acqknowledge software (v.4.4.2)
and processed in MATLAB (v.2018b).
Behavioural assays
Female mice were placed in normal home cages that were bottom-
less, set on filter paper (Fisher Scientific: 05-714-4) and left in a dark-
ened room for 4 h. Mice did not have access to water to prevent the
water leaks from disturbing urine marks. Paper was imaged using a
widefield camera (Logitech C930e) while illuminated by UV light.
Images were thresholded and converted to B&W binary in ImageJ
(v.2.0.0-rc-49/1.51d), and total number of black pixels was counted.
A region of interest corresponding to the middle 50% of the image area
was used to count the number of black pixels in the middle of the cage.
Bladder histology
Bladders used for histology were the same bladders used for the bladder
weight measurements. Wild-type and knockout littermate male blad-
ders were collected, opened and blotted on kimwipes before weighing.
After recording their freshly excised weight, they were fixed in 10%
Neutral Buffered Formalin for 24 h, and then stored in 70% ethanol.
Bladders were paraffin embedded, and central cross-sections were
used for haematoxylin and eosin staining. Processing and staining was
performed by Scripps Histology Core services. Muscle layer thickness
was measured in ImageJ at 7–11 different places along the muscle wall
per animal perpendicular to the muscle surface, and an average thick-
ness value is shown per mouse.
Clinical assessment
Twelve patients with PIEZO2 loss-of-function mutations from 11 families
(n = 4 males and 8 females, ranging in age from 4 to 43) were evalu-
ated at the National Institutes of Health (NIH) under research proto-
col approved by the Institutional Review Boards of National Institute
of Neurological Disorders and Stroke (NINDS, protocol 12-N-0095)
between April 2015 and May 2020. Written informed consent and/or
assent (for minor patients) was obtained from each participant in the
study. All of the patients had biparentally inherited bi-allelic homozy-
gous or compound heterozygous nonsense variants that are expected
to result in a ‘null’ status for protein expression. Patients with PIEZO2
loss-of-function either found us or were referred to our group through
our network of international collaborators. Genotype information
can be found in Supplementary Table 1, along with past treatments
and diagnoses. One patient, P10, carried a nonsense and a deleterious
splice site variant in compound heterozygosity. Also as stated above,
all patients presented with a profound congenital ubiquitous lack of
proprioception, vibration, and specific loss of touch discrimination
on glabrous skin. Detailed history, clinical evaluation and testing were
conducted including an in-depth review of urinary function, urologi-
cal history, review of previous evaluations and non-invasive blad-
der ultrasound. Patients were recruited from all over the world and
their age ranged between 5 to 43 years (see table). Four adult patients
(3 females and 1 male) provided their own history. None of the patients
were taking any medications that could affect urinary function at the
time of the questionnaire. Parents assisted with information gathering
from their children.
Reporting summary
Further information on research design is available in the Nature
Research Reporting Summary linked to this paper.
Data availability
The raw data that support the findings of this study are available from
the corresponding authors upon reasonable request.
Code availability
Code for calcium imaging analysis is previously published13. MATLAB
(v.2018b) code used for cystometry analysis is available at https://
github.com/PatapoutianLab/cystometry.
31. Chen, T. W. et al. Ultrasensitive fluorescent proteins for imaging neuronal activity. Nature
499, 295–300 (2013).
Acknowledgements We thank D. Barucha-Goebel, A. R. Foley and S. Donkervoort for help with
the clinical assessments; G. Averion, C. Jones and K. Brooks for experimental assistance; S. Ma
and S. Simpson-Dworschak for early work on the project; E. Lacefield for helpful discussions;
and the Scripps Histology Core for sample preparation. This work was supported by the
Howard Hughes Medical Institute; the NIH grants R35 NS105067 to A.P., F32 DK121494 to
K.L.M. and R01 NS108439 to L.S.; and the NIH Intramural Research Program funding from the
National Center for Complementary and Integrative Health (A.T.C.) and from the National
Institute of Neurological Disorders and Stroke (A.T.C. and C.G.B.).
Author contributions K.L.M. designed and performed all mouse cystometry, behavioural
experiments and tissue histology, analysed data and, together with A.P., wrote the manuscript.
D.S., T.O., C.G.B. and A.T.C. designed and performed the human clinical assessments. Calcium
imaging and analysis was performed by N.G., K.L.M. and M.S. Retrograde labelling and FISH
experiments were performed by K.L.M., A.M.C. and I.D. J.K. and L.T.S. contributed analytical
tools for data analysis, technical support and conceptual project design. C.G.B, A.T.C. and A.P.
contributed to project design and supervision. All authors discussed results and contributed
to manuscript editing.
Competing interests The authors declare no competing interests.
Additional information
Supplementary information is available for this paper at https://doi.org/10.1038/s41586-020-
2830-7.
Correspondence and requests for materials should be addressed to A.T.C. or A.P.
Peer review information Nature thanks Eric Honoré, Jon Levine and Mark Nelson for their
contribution to the peer review of this work.
Reprints and permissions information is available at http://www.nature.com/reprints.
Article
Extended Data Fig. 1 | The bladder urothelium expresses multiple
mechanosensitive proteins, and PIEZO2 is not required for sensory neuron
pinch responses. a, FISH in bladder tissue with probes against Krt20 (green)
and Piezo1 (white). DAPI in blue. b, FISH in bladder tissue with probes against
Krt20 (green) and Tmem63b (white). DAPI in blue. c, Z-projection of the
standard deviation of responses from genital pinch in WT and d, Piezo2 cKO DRG.
e, Quantification of peak responses during pinch shown as percent of baseline
(each data point is one cell). n = 3 DRGs, 40 cells for WT, 4 DRGs and 69 cells for
Piezo2 cKO DRGs.
Extended Data Fig. 2 | See next page for caption.
Article
Extended Data Fig. 2 | PIEZO2 is required for efficient micturition reflexes
in male mice. a, Hoxb8-cre;Ai9 bladder tissue, fixed, frozen and mounted to
show tdTomato (red) throughout the tissue, labelled with DAPI (blue).
Scale is 100 μm. Expression was evaluated in two mice. b, Example pressure
and urethra activity traces from three wild-type males and c, three Hoxb8-
cre;Piezo2 fl/fl knockout male littermates. d, Heat map of individual bladder
contraction events in wild-type and e, knockout male mice, with corresponding
urethra activity below in f and g respectively. h, Bladder contraction intervals
for males. i, Bladder pressures five seconds before peak contraction for males.
Note: 1,200 s was the length of one recording. These dots represent recording
periods in which the animal had no successful urination events. j, Total bladder
pressure for males and k, sum of urethra activity during bladder contractions.
n = 6 males per group. P < 0.0001 for graphs in h, i, j and k, two-sided Student’s
t-test with Welch’s correction. l, Body weights from a subset of mice whose
bladder weights are shown in Fig. 2t, and m, bladder weights from animals in l,
shown as a percentage of body weight. Red horizontal lines indicate means,
vertical red bars indicate +/− standard deviation (shown where possible).
 a Upk2-cre;Ai9 b Scn10a-cre;Ai9

200 µm

c d e

50 µm

f Upk2-cre bladder weight g Scn10a-cre bladder weight

50 50

40 40
Weight (mg)

Weight (mg)

30 30

20 20

10 10

0 0
WT KO WT KO

Extended Data Fig. 3 | Upk2- and Scn10a-cre expression and bladder

weights. a, Upk2-cre;Ai9 bladder tissue fixed, frozen and mounted to show
tdTomato (red) throughout the urothelium, labelled with DAPI (blue).
Expression was evaluated in two mice. b, Scn10a-cre;Ai9 bladder tissue fixed,
frozen and mounted to show tdTomato (red) is not present. Expression was
evaluated in two mice. Thin cryosections made neuronal endings difficult to
visualize. Scale: 200 μm, applies to a and b. c, Scn10a-cre;Ai9 DRG tissue
showing tdTomato (red) in the majority of neurons, and d, a cell backlabelled
with CTB-Alexa 488 injected into bladder. e, Merge of c and d, DAPI in blue.
9/9 backlabelled bladder cells analysed from two mice were tdTomato positive.
f, Quantification of freshly excised bladder weights from four Upk2-cre;Piezo2fl/fl
knockout and wild-type littermates. Age-matched littermates were 10–11
months old, which could account for greater variability. g, Bladder weights
from age-matched Scn10a-cre;Piezo2 fl/fl knockout mice and wild-type
littermates, 7–8 months old. Red lines indicate mean values.
 nature research | reporting summary
Corresponding author(s): Ardem Patapoutian and Alexander Chesler
Last updated by author(s): Jul 12, 2020

Reporting Summary
Nature Research wishes to improve the reproducibility of the work that we publish. This form provides structure for consistency and transparency
in reporting. For further information on Nature Research policies, see our Editorial Policies and the Editorial Policy Checklist.

Statistics
For all statistical analyses, confirm that the following items are present in the figure legend, table legend, main text, or Methods section.
n/a Confirmed
The exact sample size (n) for each experimental group/condition, given as a discrete number and unit of measurement
A statement on whether measurements were taken from distinct samples or whether the same sample was measured repeatedly
The statistical test(s) used AND whether they are one- or two-sided
Only common tests should be described solely by name; describe more complex techniques in the Methods section.

A description of all covariates tested

A description of any assumptions or corrections, such as tests of normality and adjustment for multiple comparisons
A full description of the statistical parameters including central tendency (e.g. means) or other basic estimates (e.g. regression coefficient)
AND variation (e.g. standard deviation) or associated estimates of uncertainty (e.g. confidence intervals)

For null hypothesis testing, the test statistic (e.g. F, t, r) with confidence intervals, effect sizes, degrees of freedom and P value noted
Give P values as exact values whenever suitable.

For Bayesian analysis, information on the choice of priors and Markov chain Monte Carlo settings
For hierarchical and complex designs, identification of the appropriate level for tests and full reporting of outcomes
Estimates of effect sizes (e.g. Cohen's d, Pearson's r), indicating how they were calculated
Our web collection on statistics for biologists contains articles on many of the points above.

Software and code

Policy information about availability of computer code
Data collection Acqknowledge software was used for data collection (version 4.4.2)

Data analysis ImageJ (version 2.0.0-rc-49/1.51d), and Acqknowledge software (version 4.4.2) were used, and versions added to the supplementary
information. Code availability statement in manuscript: "Code for calcium imaging analysis is previously published13. Matlab (R2018b) code
was used for cystometry analysis and is available at: https://github.com/PatapoutianLab/cystometry."
For manuscripts utilizing custom algorithms or software that are central to the research but not yet described in published literature, software must be made available to editors and
reviewers. We strongly encourage code deposition in a community repository (e.g. GitHub). See the Nature Research guidelines for submitting code & software for further information.

Data
Policy information about availability of data
All manuscripts must include a data availability statement. This statement should provide the following information, where applicable:
- Accession codes, unique identifiers, or web links for publicly available datasets
April 2020

- A list of figures that have associated raw data

- A description of any restrictions on data availability

The raw data that support the findings of this study are available from the corresponding author upon reasonable request.

1
 nature research | reporting summary
Field-specific reporting
Please select the one below that is the best fit for your research. If you are not sure, read the appropriate sections before making your selection.
Life sciences Behavioural & social sciences Ecological, evolutionary & environmental sciences
For a reference copy of the document with all sections, see nature.com/documents/nr-reporting-summary-flat.pdf

Life sciences study design

All studies must disclose on these points even when the disclosure is negative.
Sample size No statistical test was used to pre-determine sample size. Instead, sample size was determined by animal availability and previous studies in
the field, which found these sample sized sufficient to detect deficits during cystometry (Keller et al., 2018 PMID: 30104734) and histological
differences in remodeling and behavior (Everaerts and Zhen, 2010, PMID: 20956320)

Data exclusions We established exclusion criteria prior to collecting cystometry data: data from the first 30 minutes of cystometry recording was not used
because bladder muscle activity has often not stabilized. Moreover, animals that displayed bladder leaking during recording were excluded
from analysis, as leaking indicated a flawed seal and thus inaccurate filling responses.

Replication FISH experiments were independently replicated 2-3 times with the same results. Cystometry recordings were performed in independent
male and female cohorts, and results were replicated. To verify the reproducibility of experimental findings, we restricted the time of day that
cystometry recordings were done (Zeitgeber 8-14) and we performed every experiment in a cohort of male and female mice to compare to
their wildtype littermates.

Randomization The order of recordings for different genotypes was randomized. Beyond this, assigning animals to experimental groups is not relevant to this
study, as the groups are defined by genotype. Animals of different sexes were analyzed independently to remove this covariate.

Blinding The experimenter was blind to genotype when possible for all experiments. HoxB8Cre+;Piezo2f/f knockout mice have obvious motor
impairments, so it was impossible to keep the experimenter blind for these groups.

Reporting for specific materials, systems and methods

We require information from authors about some types of materials, experimental systems and methods used in many studies. Here, indicate whether each material,
system or method listed is relevant to your study. If you are not sure if a list item applies to your research, read the appropriate section before selecting a response.

Materials & experimental systems Methods

n/a Involved in the study n/a Involved in the study
Antibodies ChIP-seq
Eukaryotic cell lines Flow cytometry
Palaeontology and archaeology MRI-based neuroimaging
Animals and other organisms
Human research participants
Clinical data
Dual use research of concern

Animals and other organisms

Policy information about studies involving animals; ARRIVE guidelines recommended for reporting animal research
Laboratory animals Mice were kept in standard housing with 12 h light/dark cycle set with lights on from 6 AM to 6 PM, with room temperature kept
around 72 degrees Fahrenheit, with humidity between 30-80 % (not controlled). Adult male and female mice were used as indicated
in the text. Age-matched knockout and wildtype littermates were tested at the same age in each cohort, but ages tested ranged from
5-12 months). The HoxB8Cre;Piezof/f mouse line has been previously described11. GCaMP6f+/+ mice (B6;129S-
Gt(ROSA)26Sortm95.1(CAG-GCaMP6f)Hze/J, Jackson Laboratory: Ai95, #024105) were bred to Piezo2f/f mice as described
previously13,35 . Piezo2f/f mice were mated with SNSCre mice29 or UPK IICre mice (B6(129)-Tg(Upk2-cre)1Rkl/WghJ , Jackson
April 2020

Laboratory: #029281) to create sensory-neuron specific and urothelial specific Piezo2 knockout animals, respectively. Each of these
Cre lines was also crossed with Ai9 mice (B6.Cg-Gt(ROSA)26Sortm9(CAG-tdTomato)Hze/J, Jackson Laboratory: # 07909) to assess Cre
expression.

Wild animals No wild animals were used in this study

Field-collected samples No field collected samples were used in this study.

2
 Ethics oversight All experiments were performed within the protocols and guidelines approved by the Institutional Animal Care and Use Committees

nature research | reporting summary

of The Scripps Research Institute in compliance with regulatory standards established by the Association for Assessment and
Accreditation of Laboratory Animal Care International (AAALAC).
Note that full information on the approval of the study protocol must also be provided in the manuscript.

Human research participants

Policy information about studies involving human research participants
Population characteristics Twelve patients with PIEZO2 loss-of-function mutations from 11 families (N=4 males and 8 females, ranging in age from 4 to
43) were evaluated at the National Institutes of Health (NIH) under research protocol approved by the Institutional Review
Boards of National Institute of Neurological Disorders and Stroke (NINDS, protocol 12-N-0095) between April of 2015 and
May of 2020. Written informed consent and/or assent (for minor patients) was obtained from each participant in the study.
Genotype information can be found in Extended Data Table 1, along with past treatments and diagnoses.

Recruitment Patients were recruited on the basis of their biparentally inherited bi-allelic homozygous or compound heterozygous
nonsense variant mutations in the Piezo2 gene. Patients with PIEZO2 loss of function either found us, or were referred to our
group through our network of international collaborators. The nature of this group means that we are only analyzing patients
without functional Piezo2, which is the goal of the study.

Ethics oversight Research protocol approved by the Institutional Review Boards of National Institute of Neurological Disorders and Stroke
(NINDS, protocol 12-N-0095)
Note that full information on the approval of the study protocol must also be provided in the manuscript.

Clinical data
Policy information about clinical studies
All manuscripts should comply with the ICMJE guidelines for publication of clinical research and a completed CONSORT checklist must be included with all submissions.

Clinical trial registration This is not a clinical trial, but approval was through: NINDS, protocol 12-N-0095

Study protocol NINDS, protocol 12-N-0095

Data collection Data was collected at the NIH between April of 2015 and May of 2020.

Outcomes No outcomes measured.

April 2020

3
Article

Chemico-genetic discovery of astrocytic

control of inhibition in vivo

https://doi.org/10.1038/s41586-020-2926-0 Tetsuya Takano1 ✉, John T. Wallace1, Katherine T. Baldwin1, Alicia M. Purkey1, Akiyoshi Uezu1,
Jamie L. Courtland2, Erik J. Soderblom1,3, Tomomi Shimogori4, Patricia F. Maness5,6,
Received: 23 January 2020
Cagla Eroglu1,2 ✉ & Scott H. Soderling1,2 ✉
Accepted: 19 August 2020

Published online: 11 November 2020

Perisynaptic astrocytic processes are an integral part of central nervous system
Check for updates synapses1,2; however, the molecular mechanisms that govern astrocyte–synapse
adhesions and how astrocyte contacts control synapse formation and function are
largely unknown. Here we use an in vivo chemico-genetic approach that applies a
cell-surface fragment complementation strategy, Split-TurboID, and identify a
proteome that is enriched at astrocyte–neuron junctions in vivo, which includes
neuronal cell adhesion molecule (NRCAM). We find that NRCAM is expressed in
cortical astrocytes, localizes to perisynaptic contacts and is required to restrict
neuropil infiltration by astrocytic processes. Furthermore, we show that astrocytic
NRCAM interacts transcellularly with neuronal NRCAM coupled to gephyrin at
inhibitory postsynapses. Depletion of astrocytic NRCAM reduces numbers of
inhibitory synapses without altering glutamatergic synaptic density. Moreover, loss
of astrocytic NRCAM markedly decreases inhibitory synaptic function, with minor
effects on excitation. Thus, our results present a proteomic framework for how
astrocytes interface with neurons and reveal how astrocytes control GABAergic
synapse formation and function.

The majority of central nervous system (CNS) synapses are ensheathed (Extended Data Fig. 1a), the biotinylation activity of Split 1-TurboID
by tiny astrocytic processes1,2. These astrocytic contacts are an integral was higher than that of Split 2-TurboID (Extended Data Fig. 1b, lane 8).
functional compartment of the tripartite synapse, which is defined as We therefore used Split 1-TurboID for the remainder of this study, and
the combination of pre- and postsynaptic neuronal, and perisynaptic named the two molecules N-TurboID and C-TurboID. We also used a
astrocytic, processes3. At the synapse, astrocytes control basal synaptic GRAPHIC-tagged full-length TurboID construct, TurboID-surface, to
transmission, neuromodulation, ionic balance and neurotransmitter biotinylate astrocyte surface proteins (Extended Data Fig. 1a, bottom).
clearance4–8. Furthermore, astrocyte and synapse development are In astrocyte–neuron co-cultures, astrocytes expressing TurboID-
interdependent processes that are regulated by dynamic bidirectional surface under the control of the GfaABC1D promoter17 (Extended
intercellular communication via secreted factors and cell adhesion Data Fig. 1c) exhibited biotinylation activity along their mem-
molecules9–12. Historically, however, gaining molecular insights branes (Extended Data Fig. 1d). Moreover, the reconstituted activ-
into perisynaptic astrocyte–neuron signalling has been hampered ity of Split-TurboID was found only at contact sites between neurons
owing to the lack of biochemical methods for isolating this astrocytic and astrocytes (Extended Data Fig. 1c), but not when either of the
compartment. halves were expressed alone (Extended Data Fig. 1d). To investi-
To identify proteins at the extracellular clefts between astro- gate whether TurboID-surface or Split-TurboID biotinylates tripar-
cytes and neurons, we developed a chemico-genetic in vivo BioID tite synapses in these cultures, astrocytes were co-transduced with
(iBioID) approach, based on reconstituting the enzymatic activity of GfaABC1D-mCherry-CAAX to mark astrocyte membranes, and synapses
a proximity-biotinylating enzyme, TurboID13, at astrocyte–neuron were labelled by immunostaining with pre- and postsynaptic makers
junctions (Fig. 1a). Recent studies have shown that split biotinyla- (excitatory, VGLUT1 and HOMER 1; inhibitory, VGAT and gephyrin).
tion constructs could recover enzymatic activity when they were in Both constructs mediated biotinylation that overlapped with astro-
close proximity in the cell cytoplasm14,15. In this study, we used our cytic membranes and closely associated with excitatory and inhibi-
glycosylphosphatidylinositol-anchored reconstitution-activated tory synaptic markers (Extended Data Fig. 2a–d), demonstrating the
proteins highlight intercellular connections (GRAPHIC) strategy16 to functional reconstitution of TurboID transcellularly at perisynaptic
direct N- and C-terminal TurboID fragments to the extracellular sur- astrocyte–neuron junctions in vitro.
face of neurons and astrocytes (Fig. 1a, Extended Data Fig. 1a). Among To test their activity in vivo, the constructs were introduced into
the two Split-TurboID construct pairs that we tested in HEK 293T cells mouse brain astrocytes and/or neurons via retro-orbital injections
1
The Department of Cell Biology, Duke University Medical School, Durham, NC, USA. 2Department of Neurobiology, Duke University Medical School, Durham, NC, USA. 3Duke Proteomics and
Metabolomics Shared Resource and Duke Center for Genomic and Computational Biology, Duke University Medical School, Durham, NC, USA. 4Molecular Mechanisms of Brain Development,
Center for Brain Science (CBS), RIKEN, Saitama, Japan. 5Department of Biochemistry, University of North Carolina School of Medicine, Chapel Hill, NC, USA. 6Department of Biophysics,
University of North Carolina School of Medicine, Chapel Hill, NC, USA. ✉e-mail: tetsuya.takano@keio.jp; cagla.eroglu@duke.edu; scott.soderling@duke.edu

296 | Nature | Vol 588 | 10 December 2020

 a b Cell-type-specific AAVs
AAV PHP.eB-hSynI-N-TurboID
Astrocyte Presynapse
ITR hSynI N-TurboID GPI WPRE pA ITR
Biotinylation
Split-TurboID AAV PHP.eB-GfaABC1D-C-TurboID TurboID-HA

ITR GfaABC1D C-TurboID GPI WPRE pA ITR

AAV injection Biotin injection

Postsynapse
Analysis
P21 P42 P48
hSynI-eGFP GfaABC1D-mCherry–CAAX
c (neuron) (astrocyte) Streptavidin Merge

Control

TurboID-surface

Split-TurboID 5 μm

d e f

Streptavidin colocalization
TurboID-surface Split-TurboID TurboID-surface Split-TurboID 100 P < 0.001

(% puncta per 100 μm2)

PSD95 Gephyrin
VGLUT1 VGAT 80 P < 0.05
Streptavidin Streptavidin 60 P < 0.001
VGLUT1
40 PSD95
STED
STED

P < 0.001 VGAT

20 Gephyrin
Other
0

ce

ID
bo
rfa
1 μm 1 μm

ur
su

-T
D-

lit
oI

Fig. 1 | Identification of the astrocyte–neuron synaptic cleft proteome
using in vivo Split-TurboID. a, Schematic of the Split-surface iBioID
approach. b, Outline of Split-TurboID method using cell-type-specific AAVs.
ITR, inverted terminal repeats; hSyn1, human synapsin 1 promoter; GPI,
glycosylphosphatidylinositol; WPRE, woodchuck hepatitis virus post-
transcriptional regulatory element; pA, polyadenylation. c, Confocal images of
cortical expression of Split-TurboID or TurboID-surface coexpressed with
Sp
rb
Tu
neuronal eGFP and astrocyte mCherry–CAAX. d, e, Three-colour STED images
showing biotinylated proteins adjacent to excitatory synaptic markers PSD95
and VGLUT1 (d), and inhibitory synaptic markers gephyrin and VGAT (e). f, The
ratio of biotinylated proteins that colocalized with VGLUT1, PSD95, VGAT or
gephyrin. n = 15 cells per each condition from 3 mice. n = 3 biological repeats.
Student’s paired t-test, comparing TurboID-surface and Split-TurboID. Data are
mean ± s.e.m.

of adeno-associated viruses (AAVs)18 at postnatal day (P)21 (Fig. 1b,
Extended Data Fig. 3a–c) and the mice were given subcutaneous biotin
injections starting at P42 for 7 days (Fig. 1b)19. Biotinylated proteins were
detected by immunoblotting and immunohistochemistry (Extended
Data Fig. 4a–d) both for astrocyte-specific TurboID-surface and recon-
stituted Split-TurboID constructs. However, when Split-TurboID frag-
ments were expressed alone, no biotinylation was observed (Extended
Data Fig. 4a–c). These results show that the TurboID-surface and
Split-TurboID constructs generate extracellular biotinylation in vivo.
To confirm that biotinylated proteins localize to neuron–astro-
cyte contacts, we labelled neurons with eGFP (using AAV PHP.
eB-hSynI-eGFP) and astrocyte membranes with mCherry–CAAX
(using AAV PHP.eB-GfaBC1D-mCherry-CAAX) and co-injected either
astrocyte-specific TurboID-surface or Split-TurboID-expressing
viruses. In both conditions, biotinylated proteins were located at the
contacts between astrocytic and neuronal processes (Fig. 1c). Using
super-resolution stimulated emission depletion (STED) microscopy,
we found that biotinylated proteins surround excitatory and inhibi-
tory synapses (Fig. 1d, e). More than 50% of TurboID-surface-induced
biotinylation and more than 90% of Split-TurboID-induced biotinylation
was closely associated with synaptic markers (Fig. 1f). The densities of
synapses were not affected by either labelling approach (Extended Data
Fig. 4e, f). Together, these results show that the TurboID-surface and
Split-TurboID constructs effectively biotin-label perisynaptic contacts
between astrocytes and neurons in vivo.
Perisynaptic cleft proteome discovery
To identify the tripartite synaptic proteins, proteins biotinylated by
Split-TurboID or astrocyte-specific TurboID-surface constructs were
purified and analysed by quantitative high-resolution liquid chro-
matography–tandem mass spectrometry (LC–MS) (Fig. 2a). When
combined, the Split-TurboID and astrocyte-specific TurboID-surface
datasets identified 776,376 peptides corresponding to 3,171 distinct
proteins (Extended Data Fig. 4g). After three independent experiments
and following removal of known contaminants19, 173 and 178 proteins
were found to be significantly enriched (1.5. fold) in Split-TurboID and
astrocyte-specific TurboID-surface fractions, respectively, compared
with soluble TurboID control (Extended Data Fig. 4g–i, Supplementary
Tables 1, 2). This enrichment approach is stringent, and thus may not
identify all astrocytic proteins that are present at perisynaptic pro-
cesses, as it selects only those that are overrepresented at synapses
compared with other compartments.
A total of 118 proteins were common between the two datasets, yield-
ing a high-confidence tri-partite synapse proteome (Fig. 2b, Extended
Data Fig. 4g–i, Supplementary Table 3). This list includes known tripar-
tite synapse proteins such as neuroligin-3 and neurexin-19, calcium chan-
nel auxiliary subunits that also regulate glutamate receptor trafficking
(CACNA2D3, CACNG2 and CACNG3), excitatory synaptic proteins such
as AMPA receptors (GRIA2 and GRIA3), and inhibitory synaptic pro-
teins such as type A γ-aminobutyric acid (GABAA) receptors (GABRA1,
GABRA4, GABRB2 and GABRG2) (Fig. 2b). By cross-referencing our pro-
teomics data set with cell-type-specific gene-expression databases20,21,
we found that messenger RNA for 33 of these proteins were enriched in
astrocytes (RNA-sequencing expression ratio >1.0, diamonds in Fig. 2b),
76 were enriched in neurons (circles in Fig. 2b) and 5 proteins had equal
or unknown distribution (Fig. 2b). Bioinformatics analysis showed that
our high-confidence tripartite proteome contained known synaptic
cleft proteins (29 proteins, 25%), cell adhesion proteins (18 proteins,
15%), channels (18 proteins, 15%), G-protein-coupled receptors and

Nature | Vol 588 | 10 December 2020 | 297

Article
Lysate
a LC–MS/MS

Intensity
Streptavidin
beads m/z
AAV injection Biotin injection Mouse cortex Biotinylated proteins Data acquisition
Split-TurboID dissection purification
TurboID-surface
Syp
Previous synaptic cleft proteomics Cell adhesion molecules
b Thy1
Gjc3 Slc30a3

Tspan2 Cacng3
Serpina1a
Atp8a1 Ngly1 Sv2b
Ctsd Hapln4
Gria3 Hba-a1 Vdac3 Gstm7
Vdac2
Atp9a Tpt1
Cpe Ptk2b
Tenm1 Cst3 Cntnap2 Prxl2a
Syt7 Pdia3
Ppt1 Prdx4 Lynx1
Slc17a6 Gria2 Nrcam Cdipt
Nlgn3 Pten
Alb Cacng2
Nrxn1 Sacm1l
Apmap Clic4 Serpinb1a
Sord
C3 Gabrg2
Tm9sf2 Pzp Ppia
Scamp3 Channels and VGCCs Receptors
Ostc Tfrc Tenm2 Olfm1
Lrp1
C1qbp Adck1 Split- Hapln1 C1qb Hsp90b1
Cntfr TurboID
Split-TurboID and TurboID-surface

Lgi1 Lrba Gabra1 Hexb

Mal2 Slc8a2
Cntnap1 Lamp1 Sv2a Serpinb6a
Synj2bp
Btbd17 Ptprs Vcan
Tgfbrap1 Erlin2
Syngr1 Adam22 Lman2 Adam11
Lrpap1
Tmem106b
Prrt3 Lgi3 Slc2a3 Nomo1 Gabrb2
Car2 Tnr
Mesd Wdr1
Igsf8
Pcyox1l Prnp Itpr1 Cacna2d3
Tmed10 Tas2r116
Gabra4 Abcg2
Adcyap1r1
Eno2 Secreted and extracellular matrix Neurological disorders
Lsamp Lgalsl
Atp1b1 Slc5a7 Tenm4
Mag Slc17a7 Vdac1
Hcn2
Ntm Slc27a4 Slc8a1
Mbp Cpt1c
Prkacb
Excitatory synaptic proteins
Inhibitory synaptic proteins
Both excitatory and inhibitory synaptic proteins
Synaptic proteins without specificity information
Synaptic orphans

Fig. 2 | The astrocyte–neuron synaptic cleft proteome. a, Outline of enriched in neurons (mRNA expression in astrocyte/neuron <1) are in circles,
proteomic approach. b, Left, overlapping high-confidence proteins shared those for which gene expression is enriched in astrocytes (mRNA expression in
between the Split-TurboID and TurboID-surface enriched fractions. Right, astrocyte/neuron ≥1.0) are in diamonds. Edges are shaded according to the
clustergram topology of proteins in selected functional categories. Node titles type of interaction (grey, iBioID; black, previously reported protein–protein
show the corresponding gene symbols and node size represents log 2 fold interactions).
enrichment over negative control. Proteins for which gene expression is

associated proteins (4 proteins, 3%), other receptors and associated (measured by neuropil infiltration volume (NIV)) (Extended Data
proteins (16 proteins, 14%), secreted or extracellular matrix compo- Fig. 5h–k). By contrast, the deletion of NRCAM significantly increased
nents (34 proteins, 29%), and proteins encoded by genes implicated NIV (Extended Data Fig. 5j, k), indicating that NRCAM is a negative
in disorders, including autism spectrum disorder and schizophrenia regulator of astrocytic elaboration into the neuropil. Thus, we focused
(34 proteins, 29%) (Fig. 2b). on NRCAM for further analysis.
Adhesions between astrocytes and neurons have critical roles in
orchestrating the concurrent development of synapses and morpho-
genesis of astrocytes9,22. To identify regulators of this process, we NRCAM regulates astrocyte morphogenesis
selected teneurin-2 (TENM2), teneurin-4 (TENM4) and NRCAM as can- To confirm that endogenous NRCAM is labelled by Split-TurboID in vivo,
didate bridging molecules between astrocytes and neurons. To deplete we used STED imaging, which showed that NRCAM colocalizes with
target proteins in astrocytes, we used a CRISPR-based approach. We biotinylated proteins in vivo (Extended Data Fig. 6a). NRCAM has previ-
confirmed depletion of astrocytic NRCAM using this approach by ously been identified at contacts between axons and myelinating glia24,25
quantitative western blot analysis (Extended Data Fig. 5a). NRCAM and has been studied as a neuronal protein regulating dendritic spine
single guide RNA (sgRNA) in combination with astrocyte-specific pruning26,27 but not, to our knowledge, in astrocytes. Cell-type-specific
Cas9 significantly diminished the level of NRCAM protein in mixed transcriptome analysis shows that levels of mRNA encoding NRCAM
neuron–astrocyte cultures; this could be rescued by re-expression of are higher in astrocytes than in neurons or oligodendrocytes20,21. We
sgRNA-resistant human NRCAM in astrocytes (Extended Data Fig. 5b, c). confirmed NRCAM protein expression in cultured astrocytes by western
Next, we used this astrocyte-specific CRISPR-based approach in vivo to blot (Extended Data Fig. 6b). Next, we analysed NRCAM localization
rapidly gain preliminary data on candidate proteins23 (Extended Data in astrocytes in vivo by STED microscopy, observing that endogenous
Fig. 5d, e). We retro-orbitally injected AAVs containing sgRNA for each NRCAM puncta colocalized with astrocytic membranes (Extended
candidate gene together with Cre recombinase under the control of an Data Fig. 6c, d).
astrocyte-specific promoter (AAV PHP.eB-U6-sgRNA-GfaABC1D-Cre) NRCAM is known to function in part through a homophilic transcel-
into conditional Cas9 knock-in (KI) mice. Astrocyte-specific Cre lular interaction28. In agreement, when we injected neuron-specific and
expression was confirmed in vivo using a tdTomato Cre-reporter astrocyte-specific Nrcam-expressing viruses into P21 mice (Extended
line (Extended Data Fig. 5f, g). We used either a negative control virus Data Fig. 6e), we observed colocalization of sparsely expressed astro-
(AAV-empty sgRNA-GfaABC1D-Cre) or sgRNA virus against each cytic haemagglutinin-tagged NRCAM (NRCAM–HA) with neuronal
target gene along with astrocyte-specific mCherry–CAAX to quantify NRCAM–V5 (Extended Data Fig. 6f) by STED imaging at P42.
astrocyte morphology. NRCAM is also expressed during early postnatal development26,27.
Compared with controls, loss of TENM4 but not TENM2 in P42 Deletion of NRCAM from astrocytes during the first two weeks of devel-
mouse cortical astrocytes significantly decreased astrocyte territory opment significantly increased astrocytic territory size and enhanced
volume and the infiltration of fine astrocyte processes into the neuropil NIV when compared with controls (Extended Data Fig. 7a–g). These

298 | Nature | Vol 588 | 10 December 2020

 mCherry–CAAX
a c Excitatory synapse (astrocyte process)
AAV-GfaABC1D-NRCAM-HA (astrocytic NRCAM) Inhibitory synapse

VGLUT1
NRCAM
VGAT

Analysis
e

Distance from VGLUT1 (nm)

200
P21 P42 NS
Ezrin NRCAM 150
b d /mCherry–CAAX/VGLUT1 /mCherry–CAAX/VGLUT1
100
Astrocytic NRCAM–HA
PSD95
50
VGLUT1
0

rin

AM
Ez

RC
0.5 μm

N
f 200

Distance from VGAT (nm)

STED

STED
5 μm 1 μm Ezrin NRCAM NS
/mCherry–CAAX/VGAT /mCherry–CAAX/VGAT 150
Astrocytic NRCAM–HA
Gephyrin
VGAT 100

50

rin

AM
5 μm 1 μm 0.5 μm

Ez

RC
N
g Excitatory mCherry–CAAX h i j
synapse (astrocyte process) Control NRCAM sgRNA
Inhibitory synapse 200 200
NS

VGLUT1 distance (nm)

PSD95 distance (nm)

NS

Astrocyte process–

Astrocyte process–
VGLUT1 150 150
NRCAM
STED

100 100
PSD95 VGAT
50 50
Gephyrin 0.5 μm
0 0
mCherry–CAAX (astrocyte process)

A
A
l

sg l
tro

AM ntro

RN
RN
± NRCAM /PSD95/VGLUT1

on

sg

o
C

C
AM

RC
RC

N
N
NRCAM sgRNA NRCAM sgRNA
k Control NRCAM sgRNA + hNRCAM + neuroNRCAM sgRNA neuroNRCAM sgRNA
STED

1.0 μm

mCherry–CAAX (astrocyte process)/Gephyrin/VGAT

l P < 0.01 m P < 0.01

400 P < 0.01 300
gephyrin distance (nm)
VGAT distance (nm)

P < 0.01
Astrocyte process–

Astrocyte process–

300
200
NS NS
200
100
100

0 0
oN AM s M

oN AM s M
ne NR CA NR NA

ne NR CA NR NA
ur N + h sgR l

AM gR NA

AM gRNNA

A
A M tro

ur N + sgR l
A

A M ro
ur C M CA

ur C M CA

RN
sg A

sg A
RN
RN A on

RN A nt
RC s gR
N

RC s gR
sg RC Co
sg RC C

o R h
o R
N

N
AM

AM
ne

ne
RC

RC
+

+
N

Fig. 3 | NRCAM controls astrocyte-neuron contacts in vivo. a, Schematic of
the visualization of astrocytic NRCAM in vivo. b, Three-colour STED images
demonstrating that astrocytic NRCAM are adjacent to excitatory synapses or
inhibitory synapses. c, Schematic of astrocytic NRCAM distribution assay
in vivo. d, Three-colour STED images showing mCherry–CAAX-positive
NRCAM or ezrin adjacent to excitatory presynapses and inhibitory
presynapses. e, f, Quantification of average distance between astrocytic
NRCAM and VGLUT1 or VGAT (n = 30 puncta per each condition from 3 brains).
g, Schematic of in vivo astrocytic process–neuronal synapses contact assay.
h–m, Three-colour STED images of an astrocytic process following deletion of
astrocyte NRCAM adjacent to excitatory synapses (h) or inhibitory
synapses (k). hNRCAM, human NRCAM; neuroNRCAM sgRNA, deletion of
neuronal NRCAM. i, j, l, m, Quantification of average distance between
astrocytic process and excitatory synapses (i, j) or inhibitory synapses (l, m)
(n = 30 puncta per condition from 3 brains). n = 3 biological repeats. In
e, f, i, j, Student’s paired t-test. In l, m, one-way analysis of variance (ANOVA)
with Dunnett’s multiple comparison. Data are mean ± s.e.m. NS, not significant.
Arrows in b, d, h, highlight examples of adjacent synaptic fluorescent signals in
the images.

phenotypes were rescued by coexpression of sgRNA-resistant NRCAM–
HA in astrocytes (Extended Data Fig. 7b–g). NRCAM is a type I mem-
brane protein with a modular extracellular domain architecture that
is composed of repeated immunoglobulin and fibronectin domains
(Extended Data Fig. 7b). To determine whether extracellular interac-
tions of NRCAM are required for astrocytic morphogenesis in vivo,
we created two deletion mutants of human NRCAM: NRCAM(ΔIG),
comprising residues 620–1193 and lacking the immunoglobulin
domain; and NRCAM(ΔECD), comprising residues 1030–1193 and
lacking both immunoglobulin and fibronectin domains (Extended Data
Fig. 7b, c). Neither mutant rescued the morphology of NRCAM-deleted
astrocytes (Extended Data Fig. 7d–g), indicating that the extracel-
lular interactions via immunoglobulin domains of NRCAM are nec-
essary for maintaining the wild-type morphology. To test whether

Nature | Vol 588 | 10 December 2020 | 299

Article
Input
a (brain lysate) IP c d NS
P < 0.0001 P < 0.0001
eGFP HA GABA A R Merge 5

AM

AM

GABAA R signal (AU)

RC

RC

Control
an G

G
4

N
Ig

Ig
ti-

ti-
(kDa)

trl

trl
an
3

C
150
NRCAM 10 μm
100 2
100
Gephyrin

NRCAM–HA
75 1
100
PSD95 75 0

sg A
RC NR ol
M

A
150

N
tr
AM CA

RN
hy sgR
100

on
NRP2

C
neuroNRCAM

rin
NRCAM–HA
b HEK293T co-culture

ep
sgRNA

oN

G
ur
+
Inhibitory synapse NRCAM

ne
VGAT
NRCAM f

NRCAM–HA
P < 0.005

gephyrin
GABAA R

sgRNA

Gephyrin–VGAT colocalization
+
P < 0.001
HEK 293T cell 25

(puncta per 100 μm2)

Gephyrin NS
20
± NRCAM
± Gephyrin 15
e NRCAM sgRNA NRCAM sgRNA
Control NRCAM sgRNA + hNRCAM + neuroNRCAM sgRNA neuroNRCAM sgRNA 10
mCherry-CAAX/Gephyrin/VGAT

oN M sg M
l

ne NR CA NR A

A
RN AM tro

AM RN A
o R h N
ur CA M CA

RN
RC sg RN

sg A
on

ur N + gR
C

A s
sg RC
10 μm

AM N

ne
RC

+
N
2 μm
P < 0.01

mIPSC amplitude (pA)

15
g CRISPR–Cas9-mediated Recording h i
astrocytic NRCAM deletion Control 10
I

II/III 5

NRCAM sgRNA
0
IV

A
sg rol
RN
t
AM on
P42 Cas9 knock-in mice V

C
AAV-NRCAM sgRNA + Cre

RC
N
j k l m n Fast Slow
(< 2.8 ms) (> 2.8 ms)
4 P < 0.001
mIPSC frequency (Hz)

Cumulative frequency

1.5 1.5 1.5 20

Cumulative frequency

Control Control mIPSC amplitude (pA)

Cumulative frequency

Control P < 0.01 P = 0.08

NRCAM sgRNA NRCAM sgRNA NRCAM sgRNA
3 15
1.0 1.0 1.0
2 10
0.5 0.5 0.5
1 5
0 0 0 0 0
0 5 10 15 20 0 1 2 3 4 0 1 2 3 4 5
sg rol
A

A
l

sg l
RN

AM tro

AM ntro
RN

RN
t

mIPSC amplitude (pA) mIPSC inter-event

AM on

mIPSC rise time (ms)

on

sg
C

RC Co

interval (s)
C
RC

RC
N

Fig. 4 | Astrocytic NRCAM controls inhibitory synaptic organization and indicate examples of colocalizing synaptic markers. f, Average number of
function. a, Co-immunoprecipitation (IP) from cortical lysates of NRCAM with inhibitory synaptic colocalized puncta within astrocyte territories from
gephyrin, PSD95 and NRP2. b, Schematic of co-culture assay to identify effects cells as in e. n = 15 cells per each condition from 3 mice. g, Schematic of
of non-neuronal NRCAM–HA on inhibitory synaptic specializations. c, Images electrophysiology experiments in L2/3 pyramidal neurons of V1 cortex.
of NRCAM–HA coexpressed with eGFP in HEK 293T cells co-cultured with h, mIPSC traces from L2/3 pyramidal neurons following astrocyte treatment
neurons depleted of NRCAM or gephyrin. d, Mean integrated intensity of with control (empty sgRNA) or NRCAM sgRNA. i–m, mIPSC amplitude (i, j),
GABA A receptor in contact with transfected HEK 293T cells counted from frequency (k, l) and rise time (m) (n = 20 cells per condition from 4 mice).
cells as in c (number of cells: n = 418 control, n = 416 NRCAM–HA, n = 297 n, mIPSC amplitudes sorted by fast and slow rise times. In d, f, one-way
neuroNRCAM sgRNA + NRCAM, n = 356 gephyrin sgRNA + NRCAM). e, Images ANOVA with Dunnett’s multiple comparison; n = 3–6 biological repeats. In
of inhibitory synapses among NRCAM-deficient astrocytes. High i, k, n, Student’s paired t-test. Data are mean ± s.e.m.
magnification images (bottom) correspond to outlined areas (above), arrows

the transcellular homophilic binding between astrocytic and neu-

ronal NRCAMs is required for astrocyte morphogenesis, we targeted Astrocyte NRCAM modulates GABA synapses
neuronal NRCAM using AAV-NrCAM sgRNA-hSynI-Cre. Depletion Previous studies have shown that proper astrocyte morphogenesis is
of NRCAM from only neurons or from both astrocytes and neurons required for synaptic development, mediated through direct synaptic
enhanced astrocytic territory (at P14 but not at P42) and NIV (at both contact9. To determine whether astrocytic NRCAM is also important
P14 and P42) to a similar degree to astrocyte-specific NRCAM dele- for astrocyte–synapse contacts, we used STED microscopy to ana-
tion (Extended Data Fig. 7d–k). Together, these results indicate that lyse astrocyte-expressed NRCAM–HA with respect to excitatory and
homophilic binding between neuronal and astrocytic NRCAM restricts inhibitory synapses (Fig. 3a). We found that astrocytic NRCAM–HA
growth of astrocyte processes into the neuropil. This function of astro- closely associated with both synapse types (Fig. 3b). Then, to deter-
cytic NRCAM might be similar to its known role in promoting retraction mine whether endogenous astrocytic NRCAM is localized at tripartite
of dendritic spines via semaphorin–plexin signalling in neurons26. synaptic sites in vivo, we measured the distance from mCherry–
Notably, SEMA7A and PLXNA4 were also detected in our proteomic CAAX-positive NRCAM puncta to excitatory (VGLUT1+) and inhibitory
analysis (Fig. 2b). (VGAT+) presynapses (Fig. 3c) and compared this to localization of ezrin,

300 | Nature | Vol 588 | 10 December 2020

a protein known to be localized to perisynaptic astrocyte processes.
The distances of NRCAM and ezrin puncta were a similar distance from
presynapses29,30 (Fig. 3d–f), demonstrating that astrocytic NRCAM is
localized at astrocyte–synapse contacts in vivo.
To determine the effect of NRCAM loss on astrocyte–neuron contacts
in vivo, we measured the distance between mCherry–CAAX-labelled
astrocytic process and excitatory or inhibitory synapses (Fig. 3g). Dele-
tion of astrocytic NRCAM did not alter the distance between astrocytic
processes and excitatory pre- or postsynapses (Fig. 3h–j). However,
the distance of astrocytic processes from inhibitory pre- and postsyn-
apses was significantly increased (Fig. 3k–m). Furthermore, simultane-
ous deletion of both astrocytic and neuronal NRCAM, or of neuronal
NRCAM alone, similarly disrupted contacts between astrocytes and
inhibitory synapses (Fig. 3k–m).
The impairment of astrocyte–inhibitory synapse contacts due to
loss of astrocytic NRCAM was rescued by expression of human NRCAM
(Fig. 3k–m). This effect appeared to be directly related to NRCAM deple-
tion as the levels of other proteins implicated in astrocyte–neuron
interactions were unaffected (Extended Data Fig. 8a, b). Together, these
results strongly support a model in which homophilic NRCAM inter-
actions between astrocytes and neurons mediate adhesions between
astrocytes and inhibitory synapses.
In previous studies, we identified neuronal NRCAM in the pro-
teome of GABAergic postsynapses using iBioID with the inhibi-
tory synapse organizer gephyrin as the bait19. Indeed, NRCAM
co-immunoprecipitated with GFP–gephyrin when coexpressed in
HEK 293T cells (Extended Data Fig. 8c), and endogenous NRCAM
co-immunoprecipitated with gephyrin from brain lysate (Fig. 4a). The
positive controls PSD95 and neuropilin-2 (NRP2)26 were also detected in
these co-immunoprecipitations, whereas negative control IgG did not
precipitate gephyrin or positive-control proteins (Fig. 4a). These results
indicate that NRCAM forms a complex with the neuronal GABAergic
synaptic scaffolding protein gephyrin, and thus it may have a critical
role in inhibitory synapse development in vivo.
To test whether NRCAM functions as an organizer for inhibitory syn-
aptic specializations, we used an in vitro HEK 293T–neuron co-culture
assay31–33. In this assay, consistent with previous studies, expression
of NL2 in HEK 293T cells induced ectopic formation of excitatory
(VGLUT1+) and inhibitory (VGAT+) presynapses32,33 (Extended Data
Fig. 8e–l). Similarly, expression of presynaptic neurexin-1β (NRX1β),
induced excitatory (HOMER1+) and inhibitory (GABAA receptor-positive)
postsynapses around the HEK 293T cells31,33 (Extended Data Fig. 8e–l).
When NRCAM, NRCAM(ΔIG) or NRCAM(ΔECD) were expressed in HEK
293T cells co-cultured with neurons31–33 (Extended Data Fig. 8b–l), the
expression of NRCAM, but not the mutant NRCAMs, induced ectopic
formation of inhibitory pre- and postsynaptic contacts (Extended Data
Fig. 8e–h). NRCAM did not recruit excitatory synaptic specializations
onto HEK 293T cells (Extended Data Fig. 8i–l). Of note, when NRCAM
or gephyrin were deleted from neurons using specific sgRNAs (Fig. 4b),
the ability of NRCAM-expressing HEK 293T cells to promote clustering
of inhibitory post-synapses was abolished (Fig. 4c, d). Together, these
data indicate that transcellular homophilic NRCAM interactions control
the organization of inhibitory synaptic specializations via neuronal
gephyrin.
NRCAM controls inhibition in vivo
Next, we examined the requirement of astrocytic NRCAM for excitatory
or inhibitory synaptic structure and function in the mouse visual cortex.
When we quantified the intracortical synapses of layer 2/3 neurons that
are abundant in layer 1, we found that deletion of astrocytic NRCAM
did not alter excitatory synapse number (Extended Data Fig. 9a, b). By
contrast, deletion of astrocytic NRCAM significantly decreased inhibi-
tory synapses in layer 2/3 of the mouse visual cortex (Fig. 4e, f). The
effect of astrocytic NRCAM deletion on inhibitory synapse number was
rescued by the expression of human NRCAM (Fig. 4e, f). The deletion
of NRCAM from neurons alone or from both neurons and astrocytes
significantly decreased inhibitory synapse numbers (Fig. 4e, f). This
result further supports the idea that NRCAM bridges astrocytes and
neurons via homophilic interactions to control inhibitory synapses.
To determine the functional consequences of deleting NRCAM from
astrocytes, we performed whole-cell patch-clamp recordings of min-
iature excitatory postsynaptic currents (mEPSCs) and inhibitory post-
synaptic currents (mIPSCs) of pyramidal neurons in layer 2/3 (Fig. 4g).
The amplitude of mEPSCs was slightly decreased by the deletion of
astrocytic NRCAM, but the frequency was not altered (Extended Data
Fig. 9c–g). By contrast, both amplitude and frequency of mIPSCs were
significantly decreased following NRCAM deletion compared with of
controls (Fig. 4h–l). Inhibitory synapses that develop into pyramidal
neurons are established by a heterologous population of interneurons,
targeting either perisomatic or distal dendritic regions34. Owing to
their juxtaposition to the recording electrode, somatic mIPSC events
have much steeper rise kinetics than distal dendritic events and thus
can distinguish between these two populations35,36. Notably, we saw an
increase in the rise time of mIPSCs (Fig. 4m), which when separated by
fast and slow events (fast being less than 2.8 ms and slow being over
2.8 ms)36, showed a significant decrease of mIPSC amplitudes for the
fast (somatic) (Fig. 4n) compared to slower (dendritic) rise time events.
Thus, astrocytic NRCAM is probably important for proper somatic
inhibitory synaptic development and function in vivo. It will be interest-
ing to analyse how these effects modulate GABAergic networks, such
as during visual cortical critical periods, in future studies.
Dissection of the in vivo chemical-affinity codes that organize
the wiring of the brain in a cell-type-specific manner from tissue has
remained a considerable challenge. In this study, we have developed an
in vivo BioID approach for discovery of extracellular cell–cell contact
proteomes (Extended Data Fig. 10, top). Our Split-TurboID approach
differs from analogous methods in two ways: it can specify labelling
of junctions between two genetically defined cell types, and it can be
applied in vivo. Previously, synaptic cleft proteomics studies have been
performed in vitro with split horseradish peroxidase-conjugated with
neurexin and neuroligin37,38 or with the synaptic adhesion molecule
SynCAM139. Both approaches have been highly successful, identifying
excitatory and inhibitory synaptic cleft proteins in cultured cortical
neurons. HRP-based labelling has the advantage of labelling synaptic
clefts on a minute timescale in cultured cells or ex vivo37,38,40. However,
a concern with this method is that the labelling requires H2O2, which
is cytotoxic and difficult to use in living brain tissue while maintain-
ing complex multicellular interactions of the neuropil. We designed
TurboID-surface and Split-TurboID to overcome this issue. A different
version of split-TurboID has been described recently for intracellular
labelling between endoplasmic reticulum and mitochondria41. It will be
interesting to test how this version performs when displayed extracel-
lularly between cell types.
Astrocytes have been proposed to control inhibitory synapse forma-
tion via secreted proteins42,43; however, the presence of adhesion-based
mechanisms through which astrocyte contacts control inhibitory
synaptogenesis remain largely unknown. In this study we show that
astrocytic and neuronal NRCAMs bridge these two cell types to foster
inhibitory postsynaptic specializations via gephyrin. We propose that
these postsynaptic specializations then recruit presynaptic neuronal
partners, to direct the formation of tripartite inhibitory synapses
(Extended Data Fig. 10, bottom). Loss of perisynaptic NRCAM interac-
tions results in significant deficits of GABAergic transmission, with
slight reductions in the amplitudes of glutamatergic responses. These
reduced glutamatergic responses may be due to a well-documented
homeostatic response to reduced inhibition44,45, given the lack of
effects of NRCAM on excitatory synapse formation in co-culture and
depletion assays. Thus, our proteomic analysis reveals both a mecha-
nism for how astrocytes modulate inhibitory synapses, and a protein
Nature | Vol 588 | 10 December 2020 | 301
Article
map to provide a basis for future studies of astrocyte–neuron signal-
ling at synapses.
Online content
Any methods, additional references, Nature Research reporting sum-
maries, source data, extended data, supplementary information,
acknowledgements, peer review information; details of author con-
tributions and competing interests; and statements of data and code
availability are available at https://doi.org/10.1038/s41586-020-2926-0.
1. Yu, X., Nagai, J. & Khakh, B. S. Improved tools to study astrocytes. Nat. Rev. Neurosci. 21,
121–138 (2020).
2. Lanjakornsiripan, D. et al. Layer-specific morphological and molecular differences in
neocortical astrocytes and their dependence on neuronal layers. Nat. Commun. 9, 1623
(2018).
3. Araque, A., Parpura, V., Sanzgiri, R. P. & Haydon, P. G. Tripartite synapses: glia, the
unacknowledged partner. Trends Neurosci. 22, 208–215 (1999).
4. Khakh, B. S. & Sofroniew, M. V. Diversity of astrocyte functions and phenotypes in neural
circuits. Nat. Neurosci. 18, 942–952 (2015).
5. Ma, Z., Stork, T., Bergles, D. E. & Freeman, M. R. Neuromodulators signal through
astrocytes to alter neural circuit activity and behaviour. Nature 539, 428–432 (2016).
6. Papouin, T., Dunphy, J., Tolman, M., Foley, J. C. & Haydon, P. G. Astrocytic control of
synaptic function. Phil. Trans. R. Soc. Lond. B 372, 20160154 (2017).
7. Panatier, A. et al. Astrocytes are endogenous regulators of basal transmission at central
synapses. Cell 146, 785–798 (2011).
8. Araque, A. et al. Gliotransmitters travel in time and space. Neuron 81, 728–739 (2014).
9. Stogsdill, J. A. et al. Astrocytic neuroligins control astrocyte morphogenesis and
synaptogenesis. Nature 551, 192–197 (2017).
10. Stork, T., Sheehan, A., Tasdemir-Yilmaz, O. E. & Freeman, M. R. Neuron–glia interactions
through the Heartless FGF receptor signaling pathway mediate morphogenesis of
Drosophila astrocytes. Neuron 83, 388–403 (2014).
11. Sloan, S. A. & Barres, B. A. Mechanisms of astrocyte development and their
contributions to neurodevelopmental disorders. Curr. Opin. Neurobiol. 27, 75–81
(2014).
12. Allen, N. J. & Lyons, D. A. Glia as architects of central nervous system formation and
function. Science 362, 181–185 (2018).
13. Branon, T. C. et al. Efficient proximity labeling in living cells and organisms with TurboID.
Nat. Biotechnol. 36, 880–887 (2018).
14. Schopp, I. M. et al. Split-BioID a conditional proteomics approach to monitor the
composition of spatiotemporally defined protein complexes. Nat. Commun. 8, 15690
(2017).
15. De Munter, S. et al. Split-BioID: a proximity biotinylation assay for dimerization-dependent
protein interactions. FEBS Lett. 591, 415–424 (2017).
16. Kinoshita, N. et al. Genetically encoded fluorescent indicator GRAPHIC delineates
intercellular connections. iScience 15, 28–38 (2019).
17. Lee, Y., Messing, A., Su, M. & Brenner, M. GFAP promoter elements required for
region-specific and astrocyte-specific expression. Glia 56, 481–493 (2008).
18. Chan, K. Y. et al. Engineered AAVs for efficient noninvasive gene delivery to the central
and peripheral nervous systems. Nat. Neurosci. 20, 1172–1179 (2017).
19. Uezu, A. et al. Identification of an elaborate complex mediating postsynaptic inhibition.
Science 353, 1123–1129 (2016).
20. Zhang, Y. et al. An RNA-sequencing transcriptome and splicing database of glia,
neurons, and vascular cells of the cerebral cortex. J. Neurosci. 34, 11929–11947
(2014).
21. Zhang, Y. et al. Purification and characterization of progenitor and mature human
astrocytes reveals transcriptional and functional differences with mouse. Neuron 89,
37–53 (2016).
302 | Nature | Vol 588 | 10 December 2020
22. Sakers, K. & Eroglu, C. Control of neural development and function by glial neuroligins.
Curr. Opin. Neurobiol. 57, 163–170 (2019).
23. Incontro, S., Asensio, C. S., Edwards, R. H. & Nicoll, R. A. Efficient, complete deletion of
synaptic proteins using CRISPR. Neuron 83, 1051–1057 (2014).
24. Custer, A. W. et al. The role of the ankyrin-binding protein NrCAM in node of Ranvier
formation. J. Neurosci. 23, 10032–10039 (2003).
25. Feinberg, K. et al. A glial signal consisting of gliomedin and NrCAM clusters axonal Na+
channels during the formation of nodes of Ranvier. Neuron 65, 490–502 (2010).
26. Demyanenko, G. P. et al. Neural cell adhesion molecule NrCAM regulates Semaphorin
3F-induced dendritic spine remodeling. J. Neurosci. 34, 11274–11287 (2014).
27. Mohan, V. et al. Temporal regulation of dendritic spines through NrCAM-Semaphorin3F
receptor signaling in developing cortical pyramidal neurons. Cereb. Cortex 29, 963–977
(2019).
28. Mauro, V. P., Krushel, L. A., Cunningham, B. A. & Edelman, G. M. Homophilic and
heterophilic binding activities of Nr-CAM, a nervous system cell adhesion molecule.
J. Cell Biol. 119, 191–202 (1992).
29. Derouiche, A., Anlauf, E., Aumann, G., Mühlstädt, B. & Lavialle, M. Anatomical aspects of
glia-synapse interaction: the perisynaptic glial sheath consists of a specialized astrocyte
compartment. J. Physiol. Paris 96, 177–182 (2002).
30. Lavialle, M. et al. Structural plasticity of perisynaptic astrocyte processes involves ezrin
and metabotropic glutamate receptors. Proc. Natl Acad. Sci. USA 108, 12915–12919 (2011).
31. Scheiffele, P., Fan, J., Choih, J., Fetter, R. & Serafini, T. Neuroligin expressed in
nonneuronal cells triggers presynaptic development in contacting axons. Cell 101,
657–669 (2000).
32. Graf, E. R., Zhang, X., Jin, S. X., Linhoff, M. W. & Craig, A. M. Neurexins induce
differentiation of GABA and glutamate postsynaptic specializations via neuroligins. Cell
119, 1013–1026 (2004).
33. Chih, B., Gollan, L. & Scheiffele, P. Alternative splicing controls selective trans-synaptic
interactions of the neuroligin–neurexin complex. Neuron 51, 171–178 (2006).
34. Tremblay, R., Lee, S. & Rudy, B. GABAergic interneurons in the neocortex: from cellular
properties to circuits. Neuron 91, 260–292 (2016).
35. Miles, R., Tóth, K., Gulyás, A. I., Hájos, N. & Freund, T. F. Differences between somatic and
dendritic inhibition in the hippocampus. Neuron 16, 815–823 (1996).
36. Wierenga, C. J. & Wadman, W. J. Miniature inhibitory postsynaptic currents in CA1
pyramidal neurons after kindling epileptogenesis. J. Neurophysiol. 82, 1352–1362 (1999).
37. Martell, J. D. et al. A split horseradish peroxidase for the detection of intercellular
protein-protein interactions and sensitive visualization of synapses. Nat. Biotechnol. 34,
774–780 (2016).
38. Loh, K. H. et al. Proteomic analysis of unbounded cellular compartments: synaptic clefts.
Cell 166, 1295–1307 (2016).
39. Cijsouw, T. et al. Mapping the proteome of the synaptic cleft through proximity labeling
reveals new cleft proteins. Proteomes 6, E48 (2018).
40. Li, J. et al. Cell-surface proteomic profiling in the fly brain uncovers wiring regulators. Cell
180, 373–386 (2020).
41. Cho, K. F. et al. Split-TurboID enables contact-dependent proximity labeling in cells. Proc.
Natl Acad. Sci. USA 117, 12143–12154 (2020).
42. Elmariah, S. B., Oh, E. J., Hughes, E. G. & Balice-Gordon, R. J. Astrocytes regulate
inhibitory synapse formation via Trk-mediated modulation of postsynaptic GABAA
receptors. J. Neurosci. 25, 3638–3650 (2005).
43. Hughes, E. G., Elmariah, S. B. & Balice-Gordon, R. J. Astrocyte secreted proteins
selectively increase hippocampal GABAergic axon length, branching, and
synaptogenesis. Mol. Cell. Neurosci. 43, 136–145 (2010).
44. Turrigiano, G. G., Leslie, K. R., Desai, N. S., Rutherford, L. C. & Nelson, S. B.
Activity-dependent scaling of quantal amplitude in neocortical neurons. Nature 391,
892–896 (1998).
45. O’Brien, R. J. et al. Activity-dependent modulation of synaptic AMPA receptor
accumulation. Neuron 21, 1067–1078 (1998).
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in
published maps and institutional affiliations.
© The Author(s), under exclusive licence to Springer Nature Limited 2020
Methods
No statistical methods were used to predetermine sample size. The
experiments were not randomized. The investigators were blinded to
allocation during experiments and outcome assessment.
Animals
All mice were housed (2–5 mice per cage) at the Division of Laboratory
Animal Resources facilities at Duke University. All procedures were con-
ducted with a protocol approved by the Duke University Institutional
Animal Care and Use Committee in accordance with US National Insti-
tutes of Health guidelines. All mice were kept under typical day:night
conditions of 12 h cycles. CD1 (022, Charles River), Cas9 (028239) and
Ai14 (007914) mice were purchased from Jackson laboratory. Both
males and females were used, ages ranged from P0 to P42.
Plasmid construction
pZac2.1-GfaABC1D-Lck-GCaMP6f was a gift from B. Khakh (UCLA)
(Addgene plasmid #52924). pcDNA3-V5-TurboID-NES was a gift
from A. Ting (Stanford) (Addgene plasmid #107169). GRAPHIC was
obtained as previously described16. TurboID was subcloned into
pZac2.1-GfaABC1D vector. The split sites of Split 1-TurboID and
Split 2-TurboID were at the 256/257 and 140/141 amino acid posi-
tion, respectively. The N-TurboID and C-TurboID fragments were
subcloned into AAV-hSynI and pZac2.1-GfaABC1D vector, respec-
tively. AAV-hSynI-EGFP and pZac2.1-GfaABC1D-mCherry-CAAX were
previously described9,19. GfaABC1D was amplified and subcloned
into AAV-U6-sg-Cre vectors. The sgRNA sequences used were as
follows: Tenm2, 5′- ATCTGGAATAATGGATGTAAAGG-3′; Tenm4,
5′- GCCAGAGGCCATGGACGTGAAGG-3′; Nrcam, 5′- GTGCCAGA
TGATCAGCGCGCTGG-3′. The gephyrin sgRNA was obtained as previ-
ously described19. The cDNA encoding human NRCAM (Gene ID4897,
Dharmacon) was amplified and subcloned into AAV-Ef1α, AAV-hSynI,
pZac2.1-GfaABC1D vector. The fragments encoding human NRCAM
mutants (hNrCAM-ΔIg and hNrCAM-ΔECD) were subcloned into
AAV-Ef1α and pZac2.1-GfaABC1D vector. pCAG-HA-Nrxn1beta AS4(-)
and pNICE-NL2(-) were gifts from P. Scheiffele (University of Basel)
(Addgene plasmid #59409 and #15246, respectively). pEGFP-gephyrin
and pcDNA-PSD95-GFP were previously described19. All constructs
were confirmed by DNA sequencing. All primers are shown in Sup-
plementary Table 4.
Antibodies
The following antibodies were used: monoclonal anti-V5 (Thermo­
Fisher, R960-25, immunoblot (IB) 1:1,000, immunofluorescence (IF)
1:500, immunohistochemistry (IHC) 1:500), rat anti-HA (Sigma,
12158167001, IB 1:1,000, IF 1:500, IHC 1:200), mouse anti-HA (Biole-
gend, MMS-101P, IB 1:1,000), chicken anti-GFP (Abcam, ab13970, IB
1:1,000, IF 1:1,000, IHC 1:1,000), rabbit anti-mCherry (Abcam, ab167453,
IF 1:500, IHC 1:500), rabbit anti-PSD95 (Life Techonologies, 51-6900,
IHC 1:200), mouse anti-PSD95 (ThermoFisher, 7E3, IB 1:1,000), guinea
pig anti-VGLUT1 (Synaptic Systems, 135-304, IF 1:1,000, IHC 1:1,000),
rabbit anti-gephyrin (Synaptic Systems, 147-002, IF 1:1,000, IHC 1:500),
mouse anti-gephyrin (Synaptic Systems, 147-011, IB 1:1000, IF 1:300),
guinea pig anti-VGAT (Synaptic System, 131-004, IF 1:1,000, IHC 1:500),
rabbit anti-NL2 (Synaptic System, 129-202, IB 1:500), rabbit anti-NRCAM
(Abcam, ab24344, IB 1:1,000, IHC 1:200), rabbit anti-HOMER1 (Synaptic
Systems, 160002, IF 1:2,000), rabbit anti-GABA-A receptor β2 (Synaptic
Systems, 224-803, IF 1:1,000), goat anti-neuropilin-2 (R & D Systems,
AF567, IB 1:500), rat anti-tdTomato (Kerafast, EST203, IHC 1:1,000),
rat anti-tubulin (Santa Cruz, sc-53029, IB 1:1,000), rabbit anti-Ezrin
(Cell Signaling, 3142, IHC 1:200), rabbit anti-EAAT2 (GLT1) (Alamone,
AGC-022, IB 1:1,000), rabbit anti-KIR4.1 (Alamone, APC-035, IB 1:500),
rabbit anti-NL3 (Novus, NBP1-90080, IB 1:500), Alexa Fluor 488 Goat
anti-Mouse (ThermoFisher, A32723), Alexa Fluor 488 Goat anti-Rabbit
(ThermoFisher, A-11034), Alexa Fluor 488 Goat anti-Guinea pig (Ther-
moFisher, A11073), Alexa Fluor 488 Goat anti-Chicken (ThermoFisher,
A-11006), Oregon Green 488 Goat anti-Rabbit (ThermoFisher, O-11038),
Alexa Fluor 555 Goat anti-Rabbit (ThermoFisher, A21428), Alexa Fluor
568 Goat anti-Rat (ThermoFisher, A-11077), Alexa Fluor 594 Streptavidin
(ThermoFisher, S11227), Alexa Fluor 647 Donkey anti-rabbit (Ther-
moFisher, A31573), Alexa Fluor 647 Goat anti-Chicken (ThermoFisher,
A-21449), Alexa Fluor 647 Donkey anti-Guinea pig ( Jackson ImmunoRe-
search, 706-605-148), Alexa Fluor 647 Streptavidin (ThermoFisher,
S21374), Atto647N anti-Mouse (Sigma, 50185), Atto647N anti-rabbit
(Sigma, 40839), Donkey anti-Goat IRDye 800CW (LI-COR, 926-32214),
Goat anti-rat IRDye 800CW (LI-COR, 925-32219) and Goat anti-Mouse
IRDye 680RD (LI-COR, 925-6818).
AAV production
AAVs were produced as previously described19,46. In brief, HEK 293T cells
(obtained from ATCC, CRL-11268; short tandem repeat confirmed and
mycoplasma negative) were transfected with pAd-DELTA F6, serotype
plasmid AAV PHP.eB and AAV plasmid. After 72 h, the cells were lysed
in 15 mM NaCl, 5 mM Tris-HCl, pH 8.5, and incubated with 50 U ml−1
benzonase for 30 min at 37 °C. The cell lysate was then centrifuged
at 4,500 rpm for 30 min at 4 °C, and the supernatant containing AAV
was added to the top of an iodixanol gradient (15%, 25%, 40% and 60%
iodixanol solution, top to bottom) and centrifuged using a Beckman
Ti-70 rotor, spun at 67,000 rpm for 1 h. The viral solution extracted
from the virus layer (between the 40% and 60% iodixanol layers) with
a 24-gauge needle and 5-ml syringe, and concentrated with a 100-kDa
filter. Viral titres were measured by quantitative PCR using a linearized
genome plasmid as a standard47. For small-scale AAV supernatant, HEK
293T cells were transfected pAd-DELTA F6, serotype plasmid AAV PHP.
eB or AAV2/1 and AAV plasmid. After 72 h, the AAV-containing super-
natant medium was collected and filtered with a 0.45-μm cellulose
acetate Spin-X centrifuge tube filter (Costar 8162).
Primary neuronal, astrocytic and HEK 293T cell cultures
Cortical neurons and astrocytes were prepared from P1 mouse
pups. These cells were seeded on coverslips or dishes coated with
poly-l-lysine (Sigma) and cultured in neurobasal medium A (Invitrogen)
supplemented with B-27 (Invitrogen) and 1 mM GlutaMAX (Invitrogen).
Mouse cortical astrocytes were prepared as previously described9.
P0-3 mouse cortices were microdissected and papain digested fol-
lowed by trituration in low and high ovomucoid solutions. Cells were
passed through a 20-μm mesh filter, resuspended in astrocyte growth
medium (AGM; DMEM (Gibco 11960), 10% FBS, 10 μM, hydrocortisone,
100 U ml−1 penicillin/streptomycin, 2 mM l-glutamine, 5 μg ml−1 Insulin,
1 mM sodium pyruvate, 5 μg ml−1 N-acetyl-l-cysteine) and 30 million
cells were plated on 75-mm2 flasks (non-ventilated cap) coated with
poly-d-lysine. Flasks containing cells were incubated at 37 °C in 10%
CO2. On day in vitro (DIV) 3, AGM was removed and replaced with DPBS.
Flasks were then shaken vigorously by hand for 10–15 s until only the
adherent monolayer of astroglia remained. DPBS was then replaced
with fresh AGM. On DIV 4, the medium was supplemented with AraC
protein for 3 days to eliminate fast dividing cells, and astrocytes were
treated with AAVs. On DIV 7, astrocytes were passaged into 6-well dishes
(400,000 cells per well) and half the medium was replaced every 2–3
days. On DIV 14, astrocytes were collected for immunoblotting analy-
sis. HEK 293T (obtained from ATCC, CRL-11268; short tandem repeat
confirmed and mycoplasma negative) cells were maintained in DMEM
(Gibco) supplemented with 10% FBS (Gibco) and 100 U ml−1 penicillin/
streptomycin. Cell lines were incubated at 37 °C in 5% CO2. Cells were
regularly passaged every three days.
Immunostaining and imaging analysis
Cultured neurons and astrocytes were infected with small-scale AAVs at
DIV 14. After 3 days, these cells were treated with 500 μM biotin for 6 h.
Article
Neurons and astrocytes were fixed at indicated time points in 4% PFA,
4% sucrose for 20 min at room temperature. They were permeabilized
with 0.1% Triton-X 100 and 10% normal goat serum (NGS) for 30 min
at room temperature. Samples were then incubated for overnight at
4 °C with primary antibodies followed by Alexa Fluor 488-, Alexa Fluor
555 or Alexa 647-conjugated secondary antibodies diluted in PBS con-
taining 0.01% Triton X-100 and 10% NGS for 2 h at room temperature.
The neuron and HEK 293T cells mixed-culture assay was performed
as previously described32,33. In brief, HEK 293T cells were transfected
using Lipofectamine 2000 according to the manufacturer’s instruc-
tions. After 20 h, transfected HEK 293T cells were seeded on cultured
neurons at DIV 14. Fluorescence images were acquired with Zeiss Imager
M2 upright microscope equipped with an Apotome module, Zeiss 710,
Zeiss 780 or Zeiss 880 confocal microscopes using the Zen Software
or a stimulated emission depletion (STED) super resolution micro-
scope (TCS SP8 STED, Lecia Microsystems) using the Leica Application
Suite (LAS) software. The individual acquiring the images was always
blinded to the experiment. Images were quantified and post-processed
using FIJI.
Immunohistochemistry and imaging analysis
Immunohistochemistry was performed as previously described48,49. In
brief, brains were fixed in 4% PFA, 4% sucrose, and coronally or sagittally
sectioned with a cryostat (Leica Microsystems) at a thickness of 40 μm
or 100 μm. The slices were incubated with primary antibodies diluted
in PBS containing 0.1% Triton X-100 and 10% NGS at 4 °C for 2 days fol-
lowed by Alexa Fluor 488- or Alexa Fluor 555- or Alexa 647-conjugated
secondary antibodies diluted in PBS containing 0.1% Triton X-100 and
10% NGS for 2 h at room temperature. The nuclei were visualized by
staining with DAPI.
Astrocyte morphology was analysed as previously described9. For
the astrocyte territory volume analysis, entire astrocytes expressing
mCherry–CAAX in 100 μm-thick floating sections were imaged using
a 63× objective with 1× optical zoom images on the Zeiss 780 upright
confocal microscope (Zen Software) and processed with Imaris soft-
ware. The fluorescence signal from each astrocyte was reconstructed
using the surface tool. The intersecting nodes of the surface render
(vertices) were identified using the Matlab extension ‘Visualize Surface
Spots’. The Matlab Xtension ‘Convex hull’ identified the most terminal
vertices (outside edges of the 3D surface render) and created an addi-
tional surface render to connect these terminal vertices by the shortest
distance possible. Thus, a surface render of the outer rim (that is, terri-
tory) of each astrocyte was formed. The volume of each territory was
measured in Imaris and recorded. Astrocyte territory sizes between
experimental conditions were statistically analysed using one-way
ANOVA followed by Fisher’s least-squares difference (LSD) post hoc
test when necessary. The individual analysing the images was always
blinded to the experimental conditions. For the NIV analysis, astrocytes
expressing mCherry–CAAX were imaged by 63× plus 2× optical zoom
high magnification on Zeiss 710 confocal microscope. The images were
uploaded into Imaris Bitplane software for 3D reconstructions. We
chose at least three regions of interest (ROIs) measuring 200 pixels ×
200 pixels × 20 pixels from each astrocyte that were devoid of the soma
and large branches. ROIs were reconstructed using the surface tool in
Imaris. NIV was calculated in Imaris and statistically analysed using a
one-way ANOVA followed by a Fisher’s LSD post hoc test. Images were
analysed blinded to the experimental conditions.
Analysis of synaptic number
Synaptic number was analysed as previously described9. In brief, P42
control and experimental tissue sections were stained with an antibody
against mCherry, biotinylated proteins and the following antibodies
against pre- and postsynaptic protein pairs: VGLUT1 and PSD95 (makers
of excitatory synapses) and VGAT and gephyrin (markers of inhibitory
synapses). Five-micrometre-thick Z-stacks of 15 optical sections of
astrocytes expressing mCherry–CAAX were imaged by 63× plus 1×
optical zoom high magnification on Zeiss 780 confocal microscope
(Zen Software). Synapse number quantification by colocalization takes
advantage of the fact that pre- and postsynaptic proteins appear colo-
calized at synaptic junctions due to their close proximity. Each Z-stack
was converted into 5 maximum projection images by condensing three
consecutive optical sections using ImageJ. The number of colocal-
ized synaptic puncta of excitatory intracortical (VGLUT1–PSD95), and
inhibitory (VGAT–gephyrin) were obtained using the ImageJ plugin
Puncta Analyzer50 (B. Wark, available upon request from cagla.eroglu@
dm.duke.edu). For each image, colocalized synaptic puncta were quan-
tified within astrocytes from ROIs of 100 μm2 area that were focused
away from regions with neuronal cell bodies (areas lacking synaptic
puncta). Statistical analysis of the synaptic staining was performed with
a one-way ANOVA followed by a post hoc Fisher’s LSD test when neces-
sary. Images were analysed blinded to the experimental conditions.
Analysis of synaptic distance
Synaptic distance was analysed with super-resolution imaging as pre-
viously described51. In brief, P42 control and experimental tissue sec-
tions were stained with an antibody against mCherry, NRCAM, Ezrin
and synaptic makers (VGLUT1, PSD95, VGAT and gephyrin). Optical
sections of astrocytes expressing mCherry–CAAX were imaged by
93× plus 5× optical zoom high magnification on a STED microscope
(TCS SP8 STED, Lecia Microsystems). The distance was measured as
the distance between the peak positions of the two distributions of
localization points using the Leica Application Suite (LAS) software.
Statistical analysis was performed with a Student’s t-test or one-way
ANOVA followed by a post hoc Fisher’s LSD test when necessary. Images
were analysed blinded to the experimental conditions.
In vivo TurboID protein purification
In vivo TurboID experiments were performed as previously
described19,46, with some modifications. Each AAV-TurboID probe
virus was retro-orbitally injected into CD1 juvenile mouse brain (P21).
Three weeks after viral injection, biotin was subcutaneously injected
at 24 mg kg−1 for 7 consecutive days to increase the biotinylation effi-
ciency. For each TurboID probe, 4–10 mice were used for biotinylated
protein purification. Each purification was performed independently
at least three times. Each cortex was lysed in 50 mM Tris/HCl, pH 7.5;
150 mM NaCl; 1 mM EDTA; protease inhibitor mixture (cOmplete Mini
EDTA-free, Roche); and phosphatase inhibitor mixture (PhosSTOP,
Roche). The lysed samples were added to an equal volume of 50 mM
Tris-HCl, pH 7.5, 150 mM NaCl, 1 mM EDTA, 0.4% SDS, 2% TritonX-100,
2% deocycholate, protease inhibitor mixture and phosphatase inhibi-
tor mixture, and then sonicated and centrifuged at 15,000g for 10 min.
Supernatant was further ultracentrifuged at 100,000g for 30 min at
4 °C (Beckman TLA-100 ultracentrifuge, TLA-55 rotor). SDS was added
to the cleared supernatant to a final concentration of 1% and heated at
45 °C for 45 min. The sample was cooled on ice and incubated with Pierce
High Capacity NeutrAvidin Agarose (ThermoFisher) at 4 °C overnight.
Beads were washed twice with 2% SDS; twice with 1% TritonX-100, 1%
deoxycholate, 25 mM LiCl; twice with 1 M NaCl and 5 times with 50 mM
ammonium bicarbonate. Biotinylated proteins were eluted in 125 mM
Tris-HCl, pH6.8, 4% SDS, 0.2% β-mercaptoethanol, 20% glycerol and
3 mM biotin at 60 °C for 15 min.
Quantitative LC–MS/MS analysis
Samples were spiked with either a total of 120 or 240 fmol of casein and
reduced with 10 mM dithiolthreitol for 30 min at 80 °C and alkylated
with 20 mM iodoacetamide for 45 min at room temperature, then sup-
plemented with a final concentration of 1.2% phosphoric acid and 328 μl
of S-Trap (Protifi) binding buffer (90% methanol, 100 mM triethylam-
monium bicarbonate (TEAB)). Proteins were trapped on the S-Trap,
digested using 20 ng μl−1 sequencing grade trypsin (Promega) for
1 h at 47 °C, and eluted using 50 mM TEAB, followed by 0.2% formic acid
(FA), and lastly using 50% acetonitrile, 0.2% FA. All samples were then
lyophilized to dryness and resuspended in 12 μl 1% trifluoroacetic acid,
2% acetonitrile containing 12.5 fmol μl−1 yeast alcohol dehydrogenase.
From each sample, 3 μl was removed to create a QC pool sample which
was run periodically throughout the acquisition period.
Quantitative LC–MS/MS was performed on 2 μl of each sample,
using a nanoAcquity UPLC system (Waters Corp) coupled to a Thermo
Orbitrap Fusion Lumos high resolution accurate mass tandem mass
spectrometer (Thermo) via a nanoelectrospray ionization source. In
brief, the sample was first trapped on a Symmetry C18 20 mm × 180 μm
trapping column (5 μl min−1 at 99.9/0.1 v/v water/acetonitrile), after
which the analytical separation was performed using a 1.8 μm Acquity
HSS T3 C18 75 μm × 250 mm column (Waters) with a 90-min linear gradi-
ent of 5–30% acetonitrile with 0.1% formic acid at a flow rate of 400 nl
min−1 with a column temperature of 55 °C. Data collection on the Fusion
Lumos mass spectrometer was performed in a data-dependent acquisi-
tion (DDA) mode of acquisition with r = 120,000 (at m/z 200) full MS
scan from m/z 375 to 1,500 with a target AGC value of 2 × 105 ions. MS/MS
scans were acquired at rapid scan rate in the linear ion trap with an
AGC target of 5 × 103 ions and a max injection time of 100 ms. The total
cycle time for MS and MS/MS scans was 2 s. A 20 s dynamic exclusion
was employed to increase depth of coverage.
Following data collection, data were imported into Proteome
Discoverer 2.2 (Thermo Scientific), and individual LC–MS .raw files
were aligned on the basis of the accurate mass and retention time of
detected ions (‘features’) using the Minora Feature Detector algorithm
in Proteome Discoverer. Relative peptide abundance was calculated
based on peak intensities following integration of selected ion chro-
matograms of the aligned features across all runs. The MS/MS data was
searched against a SwissProt Mus musculus database (downloaded in
Apr 2018) containing an equal number of reversed-sequence ‘decoys’
for false discovery rate determination. Mascot Distiller and Mascot
Server (v.2.5, Matrix Sciences) were used to produce fragment ion
spectra and to perform the database searches using full trypsin
enzyme rules with 5 ppm precursor and 0.8 Da product ion match
tolerances. Database search parameters included fixed modifica-
tion on cysteine (carbamidomethyl) and variable modifications on
methionine (oxidation) and asparagine and glutamine (deamida-
tion). Peptide Validator and Protein FDR Validator nodes in Proteome
Discoverer were used to annotate the data at a maximum 1% protein
false discovery rate.
Split-TurboID protein network
Split-TuboID and TurboID-surface protein networks were performed
as previously described19,46 with modifications. Network figures were
created using Cytoscape (v.3.7), with nodes corresponding to the
gene name (multiple isoforms of proteins were collapsed into one
node based on gene nomenclature) for proteins identified in the prot-
eomic analysis. The known protein–protein interaction networks were
provided by Strings, HitPredict, HPRD, BioGrid and APID database. A
non-redundant list of protein–protein interactions was assembled
using the MGI database, GeneCard and UniProt (Supplementary
Tables 5–6). In all networks, node size is proportional to fold enrich-
ment over soluble TurboID alone. However, the bait (Split-TurboID and
TurboID-surface) node sizes were set manually. Clustergrams were
created by manual inspection on the basis of Uniprot and GeneCard
database annotation as previously described19,46.
Immunoprecipitation and immunoblotting
HEK 293T cells were transfected with Lipofectamine 2000 according
to the manufacturer’s instructions. After 20 h, transfected HEK 293T
cells were lysed in 25 mM HEPES, pH 7.5, 150 mM NaCl, 1 mM EDTA, 1%
NP-40, protease inhibitor mixture and phosphatase inhibitor mixture.
The cell lysate, which was obtained by centrifugation at 15,000g for
15 min at 4 °C, was incubated with GFP-Trap Agarose beads (Chromotek)
at 4 °C overnight. For the protein expression assay from cultured astro-
cytes, the cell was lysed in 25 mM Tris, pH 7.4, 150 mM NaCl, 1 mM CaCl2,
1 mM MgCl2, 0.5% NP-40 and protease inhibitor mixture. The lysed
samples were centrifuged at 15,000g for 5 min at 4 °C. For the endog-
enous NRCAM binding assay, juvenile mouse cortex (P42) was lysed
in 25 mM HEPES, pH 7.5; 150 mM NaCl; 1 mM EDTA; 1% NP-40; protease
inhibitor mixture; and phosphatase inhibitor mixture. The lysed sam-
ples were centrifuged at 15,000g for 10 min at 4 °C. Supernatant was
pre-cleared with at 4 °C for 30 min with Protein G Sepharose beads (Mil-
lipore). NRCAM was immunoprecipitated with anti-NRCAM antibody
followed by Protein G Sepharose beads overnight at 4 °C. SDS–PAGE
and immunoblotting were performed as previously described48,49.
The data were obtained with Odyssey Software v.4. Full gel images are
shown in Supplementary Fig. 1.
Electrophysiological analysis
For whole-cell patch clamp recordings, P42–48 mice were decapi-
tated under deep isoflurane anaesthesia. Brains were removed and
300-μm sagittal slices were prepared in ice cold, oxygenated cut-
ting solution containing (in mM) 85 NaCl, 3 KCl, 1.3 MgSO4·7H2O,
1.25 NaH2PO4·H2O, 26 NaHCO3, 25 dextrose, 2.5 CaCl2, and 75 sucrose
at ~320 mOsm l−1, with a vibratome (Leica VT 1000S). Slices were
recovered for 30 min at 31.5 °C in 95% O2, 5% CO2 bubbled artificial
cerebrospinal fluid (ACSF) (containing (in mM) 124 NaCl, 3 KCl,
1.3 MgSO4·7H2O, 1.25 NaH2PO4·H2O, 26 NaHCO3, 10 dextrose, 2.5 CaCl2
at ~310 mOsm l−1) and then at room temperature for at least 1 h. Slices
were superfused with oxygenated ACSF at room temperature. To
isolate mIPSCs, 50 μm D-APV, 10 μm NBQX and 0.5 μm TTX was added
to ACSF. To isolate mEPSCs, 0.5 μm picrotoxin and 0.5 μm tetrodo-
toxin (TTX) was added to ACSF. V1 cells were visually identified under
Zeiss Axio Examiner.D1 microscope with 20× dipping objective and
IR-1000 camera (DAGE-MTI) using an IR bandpass filter. Cortical
cells in layer 2/3 were patched using glass pipettes (4–7 MΩ resist-
ance) made from borosilicate glass capillaries (Sutter Instrument)
using a P-97 puller (Sutter Instrument). Pipettes were filled with
internal solution containing: (in mM) 135 CsMeSO3, 8 NaCl, 10 HEPES,
0.3 EGTA, 10 Na2-phosphocreatine, 4 MgATP, 0.3 Na2GTP, 5 TEA-Cl,
5 QX-314 at ~290 mOsm l−1. Miniature post-synaptic currents were
measured at −70 mV. Series resistance was monitored throughout
all recordings and only recordings that remained stable over the
recording period (≤30 MΩ resistance and <20% change in resist-
ance) were included. Data were recorded using a Multiclamp 700B
amplified (Molecular Devices), digitized at 50 kHz using a Digidata
1550 digitizer (Molecular Devices), and low-pass filtered at 1kHz. All
data were acquired using pClamp software and analysed in Clampfit
(Molecular Devices) including only events larger than 5 pA. Events
were initially identified using a custom-made template and manu-
ally assessed for inclusion with the template search function. Rise
time was defined as the time from 10–90% of the peak. All chemicals
were purchased from Sigma-Aldrich or Tocris. Experiments were
performed blinded to the condition.
Statistical analysis
Data are expressed as the mean ± s.e.m. Statistical analyses were
performed with GraphPad Prism version 6 (GraphPad Software).
We compared independent sample means using t-tests and one-way
ANOVAs as appropriate. Statistically significant F values detected in
the ANOVAs were followed by alpha-adjusted post hoc tests (Tukey’s
honestly significant difference). We confirmed necessary paramet-
ric test assumptions using the Shapiro–Wilk test (normality) and
Levene’s test (error variance homogeneity). P < 0.001, P < 0.01 and
P < 0.05 were considered to indicate statistical significance. Sam-
ple size for each experiment is indicated in the figure legend for
each experiment. Sample sizes were determined based on previous
Article
experience for each experiment to yield high power to detect spe-
cific effects. No statistical methods were used to predetermine
sample size. All results of the statistical analysis are shown in Sup-
plementary Table 7.
Reporting summary
Further information on research design is available in the Nature
Research Reporting Summary linked to this paper.
Data availability
Proteomics data are available in the MassIVE database under accession
MSV000085821. The data that support the findings of this study are
available from the corresponding author upon reasonable request.
46. Spence, E. F. et al. In vivo proximity proteomics of nascent synapses reveals a novel
regulator of cytoskeleton-mediated synaptic maturation. Nat. Commun. 10, 386
(2019).
47. Shin, J. H., Yue, Y. & Duan, D. Recombinant adeno-associated viral vector production and
purification. Methods Mol. Biol. 798, 267–284 (2012).
48. Takano, T. et al. LMTK1 regulates dendritic formation by regulating movement of
Rab11A-positive endosomes. Mol. Biol. Cell 25, 1755–1768 (2014).
49. Takano, T. et al. Discovery of long-range inhibitory signaling to ensure single axon
formation. Nat. Commun. 8, 33 (2017).
50. Ippolito, D. M, Eroglu, C. Quantifying synapses: an immunocytochemistry-based assay to
quantify synapse number. J. Vis. Exp. 16, 2270 (2010).
51. Dani, A., Huang, B., Bergan, J., Dulac, C. & Zhuang, X. Superresolution imaging of
chemical synapses in the brain. Neuron 68, 843–856 (2010).
Acknowledgements We thank H. Katsura for modifying the promoter of the surface and split
TurboID plasmids for HEK 293T cell expression, and B. Duncan for technical support. This work
was supported by Brain initiative RO1DA047258 from NIH (S.H.S. and C.E.), R01MH113280 from
NIH (P.F.M.), Kahn Neurotechnology Award (S.H.S. and C.E.), a Grant-in-Aid for JSPS Fellows
(PD) 20153173 from the Japan Society for the Promotion of Science (T.T.), The Uehara memorial
Foundation (T.T.), and National Institute of Mental Health Fellowship F30MH117851 (J.L.C).
Author contributions T.T., C.E. and S.H.S. designed the study. T.T., J.T.W., A.P., C.E. and S.H.S.
wrote the manuscript. T.T., J.T.W., A.U. and E.J.S. performed in vivo BioID-proteomics analysis.
T.T., J.T.W., J.L.C., T.S. and P.F.M. produced the constructs. T.T., J.T.W. and K.T.B. performed
imaging analysis and the morphological analysis of the astrocytes. A.P. performed
electrophysiological analysis. T.T. and K.T.B. performed the biological experiments. All authors
discussed the results and commented on the manuscript text.
Competing interests The authors declare no competing interests.
Additional information
Supplementary information is available for this paper at https://doi.org/10.1038/s41586-020-
2926-0.
Correspondence and requests for materials should be addressed to T.T., C.E. or S.H.S.
Peer review information Nature thanks Thomas Biederer, Peter Scheiffele and the other,
anonymous, reviewer(s) for their contribution to the peer review of this work.
Reprints and permissions information is available at http://www.nature.com/reprints.
Extended Data Fig. 1 | The reconstituted activity of Split-TurboID in
neurons and astrocytes in vitro. a, Schematics of constructs tested.
b, Immunoblot analysis of construct expression and biotinylation activity.
c, Schematic of neuron-astrocyte mixed-culture assay for Split-TurboID with
cell-type-specific AAVs in vitro. d, Cultured neurons and astrocytes were
infected with AAV1/2-GfaABC1D-TurboID-HA-surface, AAV1/2-hSynI-
V5-N-TurboID and/or AAV1/2-GfaABC1D-C-TurboID-HA. Representative images
of neuron and astrocyte at DIV14 after the treatment of 500 μM biotin for 6h are
shown. n = 3 biological repeats.
Article

Extended Data Fig. 2 | Split-TurboID maps excitatory and inhibitory HOMER1 (b), inhibitory presynaptic marker VGAT (c), and postsynaptic marker
perisynaptic proteins. a–d, Representative images demonstrating that gephyrin (d). Astrocytes were visualized with GfaABC1D-mCherry-CAAX. n = 3
proteins biotinylated by astrocytic TurboID-surface or Split-TurboID (cyan) are biological repeats.
adjacent to excitatory presynaptic marker VGLUT1 (a), postsynaptic marker
Extended Data Fig. 3 | Brain-wide transduction of astrocytes and neurons. expression throughout the cortex and other structures. c, Representative
a, Schematic of AAV PHP.eB viruses for neuronal-EGFP or astrocyte-mCherry- image from cortex, hippocampus or cerebellum showing high coverage of
CAAX and retro-orbital injection. b, Sagittal section of mouse brain showing neuronal and astrocytic expression.
Article

Extended Data Fig. 4 | See next page for caption.

Extended Data Fig. 4 | Mapping and identification of tripartite synaptic
cleft proteins by Split-TurboID in vivo. a, Biotinylation activity of Split-
TurboID in vivo. Lysates of mouse brain infected with cell-type-specific
TurboID-surface-HA, V5-N-TurboID and/or C-TurboID-HA. Brain lysates were
analysed by immunoblotting with anti-Streptavidin, anti-V5, anti-HA and anti-
Tubulin antibodies. b, The graph indicates the ratio of botinylation activity
in vivo (n = 4 brains per each condition). c, d, The biotinylation of Split-TurboID
in mouse cortex. e, f, Quantification of average number of excitatory or
inhibitory synaptic colocalized puncta in layer 2/3 of the visual cortex. n = 15
slices per each condition from 3 mice. g, Chart summarizing proteomic data set
identified by mass spectrometry and filters used to identify top candidates.
h, Venn diagram comparing proteome list of Split-TurboID and TurboID-
surface. i, Scale-free network of Split-TurboID (green) and TurboID-surface
(blue) identified proteins. High-confidence proteins enriched in both Split-
TurboID and TurboID-surface fractions are shown in red. Neuronal enriched
proteins (RNA-seq expression ratio <1) and astrocyte enriched proteins (RNA-
seq expression ratio≧1.0) are represented as circle or diamond, respectively.
At least n = 4 biological repeats. One-way ANOVA (Dunnett’s multiple
comparison, P < 0.0001, 0.001). Data are means ± s.e.m.
Article

Extended Data Fig. 5 | See next page for caption.

Extended Data Fig. 5 | The validation of candidate proteins with CRISPR-
based astrocytic candidate gene depletion strategy. a, Schematic of
CRISPR-based deletion of astrocytic NrCAM in vitro. b, Immunoblots showing
loss of NrCAM with sgRNA. AAV1/2-U6-empty sgRNA or AAV1/2-U6-NrCAM
sgRNA was co-infected with AAV1/2- GfaABC1D-Cas9 to cultured neurons and
astrocytes at DIV14. The cells were subjected to immunoblot analysis with an
anti-NrCAM antibody. Tubulin was used as a loading control. c, The bar graph
indicates the expression level of NrCAM from 3 independent experiments.
d, Schematic of CRISPR-based deletion strategy of candidate gene.
e, Experimental timeline of AAV-mediated CRISPR-based astrocytic gene
deletion strategy in Flex-TdTomato mice. f, AAV PHP.eB-U6-NrCAM sgRNA was
co-infected with AAV PHP.eB-GfaABC1D-Cas9 in Flex-TdTomato mice at P21.
Coronal sections were prepared and immunostained with an anti-TdTomato
antibody. g, A High-magnification image is shown. h, Images of Tenm2-, Tenm4-
or NrCAM-deleted astrocytes (cyan) and their territories (red outlines) in visual
cortexes of juvenile mice. i, Average territory volumes at P42 of Tenm2-,
Tenm4- or NrCAM-deleted astrocytes. Between 20-25 cells per condition
from 3 mice. j, Images of Tenm2-, Tenm4- or NrCAM-deleted astrocytes (cyan)
and their NIV reconstructions (orange) in visual cortexes of juvenile mice.
k, Average NIV at P42 of Tenm2-, Tenm4- or NrCAM-deleted astrocytes. 51 cells
per each condition from 3 mice. n = 3 biological repeats. One-way ANOVA
(Dunnett’s multiple comparison, P < 0.0001, 0.01). Data are means ± s.e.m.
Article
Extended Data Fig. 6 | NrCAM is a novel tripartite synaptic protein. a, A high
magnification STED image showing that endogenous NrCAM was enriched at
biotinylated proteins in vivo. b, Immunoblot analysis of endogenous NrCAM,
astrocyte marker GFAP, neuronal marker b-Tubulin III or loading control
α-Tubulin from mouse brain or purified astrocyte lysate. c, Schematic of the
visualization of astrocytic membrane and endogenous NrCAM in vivo. d, STED
images demonstrating the localization of endogenous NrCAM in vivo. Coronal
sections were immunostained with anti-NrCAM antibody (cyan). High
magnification image was shown (right panel). e, Schematic of the visualization
of both astrocytic and neuronal NrCAM in vivo. f, STED images demonstrating
that the colocalization of astrocytic NrCAM with neuronal NrCAM in vivo.
Coronal sections were prepared and co-immunostained with an anti-V5 (cyan)
and anti-HA (magenta) antibody. A high-magnification image is shown in the
right. n = 3 biological repeats. Data represent means ± s.e.m.
Extended Data Fig. 7 | The role of astrocytic NrCAM in astrocytic
morphogenesis in vivo. a, Schematic of CRISPR-based NrCAM deletion
in vivo. b, Schematic of hNrCAM domains and fragments. SP, signal peptide;
IG, immunoglobulin; FN, fibronectin; TMD, transmembrane domain; ECD,
extracellular domains. c, Immunoblots showing the expression of each NrCAM
fragments in HEK293T cells. d, f, h, j, Images of astrocytes following deletion of
astrocyte NrCAM alone (NrCAM sgRNA), with coexpression with indicated
constructs of sgRNA-resistant human NrCAM, neuronal NrCAM deletion
(neuroNrCAM sgRNA), or following neuronal NrCAM deletion alone. Images at
indicated ages represent. e, i, Analysis of astrocyte territory, 15–29 cells per
each condition from 3 mice; g, k, Analysis of neuropil infiltration volume.
50–51 cells per each condition from 3 mice. n = 3 biological repeats. One-way
ANOVA (Dunnett’s multiple comparison, P < 0.0001). Data represent
means ± s.e.m.
Article

Extended Data Fig. 8 | See next page for caption.

Extended Data Fig. 8 | NrCAM controls inhibitory synaptic specializations
through binding the gephyrin. a, Immunoblot analysis of endogenous
NrCAM, astrocyte marker GFAP, Neuroligin 2, Neuroligin 3, Kir4.1 or EAAT2
(GLT1) from purified astrocyte lysate. b, The bar graph indicates the expression
level. c, The interaction of NrCAM with PSD95 and gephyrin in HEK293T cells.
Cell lysates coexpressing NrCAM-HA with GFP, PSD95-GFP or GFP-gephyrin
were incubated with anti-GFP-bound beads. Immunoprecipitated (right) or
total (left) NrCAM, GFP, PSD95-GFP or GFP-gephyrin were detected by
immunoblotting with anti-HA and anti-GFP antibodies. d, Schematic of
HEK293T/neuronal mixed-cultured assay in vitro. e–h, Images of in vitro
inhibitory synapse formation assays. The graph shows average of the total
integrated intensity of VGAT (Cont = 258, NL2 = 222, NrCAM = 242, NrCAM-
ΔIG = 288, NrCAM-ΔECD = 303 cells) or GABAA receptor (Cont = 313,
NRX1β4(-) = 310, NrCAM = 300, NrCAM-ΔIG = 278, NrCAM-ΔECD = 278 cells)
clusters that contact transfected HEK293T cells. i–l, Images of in vitro
excitatory synapse formation assay. The graph shows average of the total
integrated intensity of VGLUT1 (Cont = 259, NL2 = 306, NrCAM = 286, NrCAM-
ΔIG = 321, NrCAM-ΔECD = 196 cells) or HOMER1 (Cont = 471, NRX1β4(-) = 214,
NrCAM = 247, NrCAM-ΔIG = 387, NrCAM-ΔECD = 251 cells) clusters that contact
transfected HEK293T cells. n = 3 biological repeats. One-way ANOVA (Dunnett’s
multiple comparison, P < 0.0001). Data are means ± s.e.m.
Article
Extended Data Fig. 9 | The effect of NrCAM on excitatory synapse
formation and function in vivo. a, Images of postsynapse PSD95 and
presynapse VGLUT1 within NrCAM-deletion astrocytes in L1 of the visual
cortex. High magnification images (bottom) correspond to boxes (above).
b, Quantification of average number of excitatory synaptic colocalized puncta
within astrocyte territories. n = 15 cells per each condition from 3 mice.
c, mEPSC traces from L2/3 pyramidal neurons following astrocyte control
empty sgRNA or NrCAM sgRNA expression. d–g, Quantification of mEPSC
amplitude (d, e, Cont = 16, NrCAM sgRNA = 14 cells from 4 mice) and frequency
(f, g, Cont = 14, NrCAM sgRNA = 17 cells from each of 4 mice). At least n = 3
biological repeats. Student’s t-test (paired, P < 0.05). Data represent
means ± s.e.m.
Extended Data Fig. 10 | In vivo chemogenetics method, Split-TurboID,
reveals a novel astrocytic cell adhesion molecule, NrCAM, that controls
inhibitory synaptic organization. Development of in vivo chemo-affinity
codes, Split-TurboID, and a working model of astrocytic NrCAM influencing
inhibitory synaptic function. Split-TurboID can map the molecular composition
of such intercellular contacts, even within the highly complex structure of the
tripartite synapse in vivo. Mapping this interface, we discovered a new
molecular mechanism by which astrocytes influence inhibitory synapses
within the tripartite synaptic cleft via NrCAM. NrCAM is expressed in cortical
astrocytes where it interacts with neuronal NrCAM that is coupled to gephyrin
at inhibitory postsynapses. Loss of astrocytic NrCAM dramatically alters
inhibitory synaptic organization and function in vivo.
 nature research | reporting summary
Scott H. Soderling, Cagla Eroglu and Tetsuya
Corresponding author(s): Takano
Last updated by author(s): Jul 24, 2020

Reporting Summary
Nature Research wishes to improve the reproducibility of the work that we publish. This form provides structure for consistency and transparency
in reporting. For further information on Nature Research policies, see our Editorial Policies and the Editorial Policy Checklist.

Statistics
For all statistical analyses, confirm that the following items are present in the figure legend, table legend, main text, or Methods section.
n/a Confirmed
The exact sample size (n) for each experimental group/condition, given as a discrete number and unit of measurement
A statement on whether measurements were taken from distinct samples or whether the same sample was measured repeatedly
The statistical test(s) used AND whether they are one- or two-sided
Only common tests should be described solely by name; describe more complex techniques in the Methods section.

A description of all covariates tested

A description of any assumptions or corrections, such as tests of normality and adjustment for multiple comparisons
A full description of the statistical parameters including central tendency (e.g. means) or other basic estimates (e.g. regression coefficient)
AND variation (e.g. standard deviation) or associated estimates of uncertainty (e.g. confidence intervals)

For null hypothesis testing, the test statistic (e.g. F, t, r) with confidence intervals, effect sizes, degrees of freedom and P value noted
Give P values as exact values whenever suitable.

For Bayesian analysis, information on the choice of priors and Markov chain Monte Carlo settings
For hierarchical and complex designs, identification of the appropriate level for tests and full reporting of outcomes
Estimates of effect sizes (e.g. Cohen's d, Pearson's r), indicating how they were calculated
Our web collection on statistics for biologists contains articles on many of the points above.

Software and code

Policy information about availability of computer code
Data collection Zen Software (Zen black 2.3, Zen black 2.1 SP3 FP1), Leica Application Suite (LAS) software(v3.5.5), Odyssey Software(v4), Proteome
Discoverer 2.2 (Thermo Scientific Inc.), GraphPad Prism (v8) and pClamp (v10) were used for data collection.

Data analysis Imaris software (v8.2.1.), Leica Application Suite (LAS) software (v3.5.5), ImageJ (v10.2), and Mascot Distiller and Mascot Server (v 2.5,
Matrix Sciences) were used for data analysis. Minora Feature Detection alogrithm is part of the Protein Discover Package version2.2.
For manuscripts utilizing custom algorithms or software that are central to the research but not yet described in published literature, software must be made available to editors and
reviewers. We strongly encourage code deposition in a community repository (e.g. GitHub). See the Nature Research guidelines for submitting code & software for further information.

Data
Policy information about availability of data
All manuscripts must include a data availability statement. This statement should provide the following information, where applicable:
- Accession codes, unique identifiers, or web links for publicly available datasets
April 2020

- A list of figures that have associated raw data

- A description of any restrictions on data availability

The data that support the findings of this study are available from the corresponding author upon reasonable request.

1
 nature research | reporting summary
Field-specific reporting
Please select the one below that is the best fit for your research. If you are not sure, read the appropriate sections before making your selection.
Life sciences Behavioural & social sciences Ecological, evolutionary & environmental sciences
For a reference copy of the document with all sections, see nature.com/documents/nr-reporting-summary-flat.pdf

Life sciences study design

All studies must disclose on these points even when the disclosure is negative.
Sample size Sample sizes were determined based on previous experience (Stogsdill et al., Nature, 2017) for each experiment to yield high power to detect
specific effects. No statistical
methods were used to predetermine sample size.

Data exclusions No data were excluded from the analyses.

Replication All attempts at replication (3-5 times) were successful.

Randomization Allocation was random.

Blinding Images data collection and statistical analyses were analyzed blinded to the experimental conditions.

Reporting for specific materials, systems and methods

We require information from authors about some types of materials, experimental systems and methods used in many studies. Here, indicate whether each material,
system or method listed is relevant to your study. If you are not sure if a list item applies to your research, read the appropriate section before selecting a response.

Materials & experimental systems Methods

n/a Involved in the study n/a Involved in the study
Antibodies ChIP-seq
Eukaryotic cell lines Flow cytometry
Palaeontology and archaeology MRI-based neuroimaging
Animals and other organisms
Human research participants
Clinical data
Dual use research of concern

Antibodies
Antibodies used The following antibodies were used: monoclonal anti-V5 (ThermoFisher, R960-25, IB 1:1000, IF 1:500, IHC 1:500), rat anti-HA (Sigma,
12158167001, IB 1:1000, IF 1:500, IHC 1:200), mouse anti-HA (Biolegend, MMS-101P, IB 1:1000), chicken anti-GFP (Abcam, ab13970,
IB 1:1000, IF 1:1000, IHC 1:1000), rabbit anti-mCherry (Abcam, ab167453, IF 1:500, IHC 1:500), rabbit anti-PSD95 (Life Techonologies,
51-6900, IHC 1:200), mouse anti-PSD-95 (ThermoFisher, 7E3, IB 1:1000), guinea pig anti-VGLUT1 (Synaptic Systems, 135-304, IF
1:1000, IHC 1:1000), rabbit anti-gephyrin (Synaptic Systems, 147-002, IF 1:1000, IHC 1:500), mouse anti-gephyrin (Synaptic Systems,
147-011, IB 1:1000, IF 1:300), guinea pig anti-VGAT (Synaptic System, 131-004, IF 1:1000, IHC 1:500), rabbit anti-NL2 (Synaptic
System, 129-202, IB 1:500), rabbit anti-NrCAM (Abcam, ab24344, IB 1:1000, IHC 1:200), rabbit anti-Homer1 (Synaptic Systems,
160002, IF 1:2000), rabbit anti-GABA-A receptor 2 (Synaptic Systems, 224-803, IF 1:1000), goat anti-Neuropilin-2 (R & D Systems,
AF567, IB 1:500), rat anti-tdTomato (Kerafast, EST203, IHC 1:1000), rat anti-Tubulin (Santa Cruz, sc-53029, IB 1:1000), rabbit anti-
Ezrin (Cell Signaling, #3142, IHC 1:200), rabbit anti-EAAT2 (GLT1) (Alamone, AGC-022, IB 1:1000), rabbit anti-Kir4.1 (Alamone,
APC-035, IB 1:500), rabbit anti-NL3 (Novus, NBP1-90080, IB 1:500), Alexa Fluor 488 Goat anti-Mouse (ThermoFisher, A32723), Alexa
Fluor 488 Goat anti-Rabbit (ThermoFisher, A-11034), Alexa Fluor 488 Goat anti-Guinea pig (ThermoFisher, A11073), Alexa Fluor 488
Goat anti-Chicken (ThermoFisher, A-11006), Oregon Green 488 Goat anti-Rabbit (ThermoFisher, O-11038), Alexa Fluor 555 Goat anti-
Rabbit (ThermoFisher, A21428), Alexa Fluor 568 Goat anti-Rat (ThermoFisher, A-11077), Alexa Fluor 594 Streptavidin (ThermoFisher,
April 2020

S11227), Alexa Fluor 647 Donkey anti-rabbit (ThermoFisher, A31573), Alexa Fluor 647 Goat anti-Chicken (ThermoFisher, A-21449),
Alexa Fluor 647 Donkey anti-Guinea pig (Jackson ImmunoResearch, 706-605-148), Alexa Fluor 647 Streptavidin (ThermoFisher,
S21374), Atto647N anti-Mouse (Sigma, 50185), Atto647N anti-rabbit (Sigma, 40839), Donkey anti-Goat IRDye 800CW (LI-COR,
926-32214), Goat anti-rat IRDye 800CW (LI-COR, 925-32219), Goat anti-Mouse IRDye 680RD (LI-COR, 925-6818).

Validation 1 monoclonal anti-V5 ThermoFisher R960-25 ELISA, Immunocytochemistry, Immunofluorescence, Immunoprecipitation, Western
Blot Vender (IB, IF)
2 rat anti-HA Sigma 12158167001 ELISA, Immunocytochemistry, Immunofluorescence, Immunoprecipitation, Western Blot
"Hougbing Liu et al., 2014. J AM Heart Assoc 20; 3(3) (IB)

2
 Fimiani et al., 2015. Nucleic Acids Res 18;43(16) (IB, IF)
Stogsdill et al., 2017. Nature "

nature research | reporting summary

3 mouse anti-HA Biolegend MMS-101P western blot (WB), immunocytochemistry (ICC), immunoprecipitation (IP), and flow
cytometry (FC). "Kim JY, et al. 2003. J Neurosci. 23:5561. (IP, WB)
Helliwell SB, et al. 2001. J Cell Biol. 153:649. (WB)
Bennett BD, et al. 2000. J Biol Chem. 275:37712. (IF, IP, WB)
Royer Y, et al. 2005. J. Biol. Chem. 29:27251. (FC)"
4 chicken anti-GFP Abcam ab13970 IHC-P, WB, IHC - Wholemount, IHC-FrFl, ICC/IF, IHC-Fr, IHC-FoFr Vender (IB, IF, IHC)
5 rabbit anti-mCherry Abcam ab167453 WB, ICC/IF, IHC-P "Stogsdill et al., 2017. Nature 551, 192-197 (IF, IHC)
Vender (IB, IF)"
6 rabbit anti-PSD95 Life Techonologies 51-6900 human mouse, rat ELISA, ICC, IF, IP, WB "Vender (IB, IF, IHC)
Stogsdill et al., 2017. Nature 551, 192-197 (IF, IHC)"
7 mouse anti-PSD-95 ThermoFisher MA1-046 human, mouse, rat, xenopus Flow, ICC, IF, IHC, IP, WB Vender (IF, IB)
8 guinea pig anti-VGLUT1 Synaptic Systems 135-304 rat, mouse, human, cow WB, IP, ICC, IHC, EM, FACS Vender (IB, IF, IHC)
9 rabbit anti-gephyrin Synaptic Systems 147-008 human, rat, mouse, pig, goldfish, zebrafish WB, IP, ICC, IHC Vender (ICC, IHC)
10 mouse anti-gephyrin Synaptic Systems 147-011 human, rat, mouse, pig, goldfish, zebrafish, chicken WB, IP, ICC, IHC, EM
"Davenport EC, Szulc BR, Drew J, Taylor J, Morgan T, Higgs NF, López-Doménech G, Kittler JT
Cell reports (2019) 268: 2037-2051.e6. 147 011 (WB, ICC, IHC)
Vender (ICC, IHC)"
11 guinea pig anti-VGAT Synaptic System 131-004 rat, mouse, zebrafish, ape WB, IP, ICC, IHC, EM Vender (WB, ICC, IHC)
12 rabbit anti-NL2 Synaptic System 129-202 human, rat, mouse, monkey, ape, cow WB, IP, ICC, IHC Stogsdill et al., 2017. Nature 551,
192-197 (IB)
13 rabbit anti-NrCAM Abcam ab24344 mouse, rat, human IHC-Fr, IHC-P, ICC/IF, WB, IP, IHC-FoFr Demynanenko et al., J Neurosci
34:1127-87 (IB, IHC)
14 rabbit anti-Homer1 Synaptic Systems 160002 human, rat, mouse WB, IP, ICC, IHC Vender (WB, ICC, IHC)
15 rabbit anti-GABA-A receptor b2 Synaptic Systems 224-803 rat, mouse WB, IP, ICC, IHC Vender (WB, ICC, IHC)
16 goat anti-Neuropilin-2 R & D Systems AF567 human, rat, mouse WB, IP, ICC, IHC, FACS Demynanenko et al., J Neurosci 34:1127-87
(IB)
17 rat anti-tdTomato Kerafast EST203 WB, ELISA, IF, IHC, IP "Stogsdill et al., 2017. Nature (IHC)
Vender (IF)"
18 rat anti-Tubulin Santa Cruz sc-53029 mouse, human, rat WB, IP, ICC, IHC, EM, FACS Vender (IB, IF, IHC)
19 rabbit anti-Ezrin Cell Signaling #3142 human, mouse, rat, monkey, bovine WB Vender (WB)
20 rabbit anti-EAAT2 (GLT1) Alamone AGC-022 human, mouse, rat ICC, IF, IHC, LCI, WB Vender (IB, IF, IHC)
21 rabbit anti-Kir4.1 Alamone APC-035 human, mouse, rat ICC, IF, IHC, IP, WB Vender (IB)
22 rabbit anti-NL3 Novus NBP1-90080 human, mouse, rat WB, IHC "Stogsdill et al., 2017. Nature 551, 192-197 (IB)
Vender (IB)"

Eukaryotic cell lines

Policy information about cell lines
Cell line source(s) HEK293T cell line was obtained from the Duke Cell Culture Facility.

Authentication The cell line was validated by STR testing.

Mycoplasma contamination The cell lines were tested for mycoplasma contamination and were negative.

Commonly misidentified lines No commonly misidentified cell lines were used in the study.
(See ICLAC register)

Animals and other organisms

Policy information about studies involving animals; ARRIVE guidelines recommended for reporting animal research
Laboratory animals P1~P42, male and female CD1 (022, Charles River), Cas9 (028239,JAX) and Ai14 (007914,JAX) mice. All animals were housed at 72F
+/-2 degrees at 30-70% humidity.

Wild animals This study did not involve wild animals.

Field-collected samples This study did not involve samples collected from the field.

Ethics oversight The Duke University Institutional Animal Care and Use Committee provided ethical approval and guidance.

Note that full information on the approval of the study protocol must also be provided in the manuscript.
April 2020

3
Article

The gut microbiota is associated with

immune cell dynamics in humans

https://doi.org/10.1038/s41586-020-2971-8 Jonas Schluter1,2 ✉, Jonathan U. Peled3,4, Bradford P. Taylor2, Kate A. Markey3,4,

Melody Smith3,4, Ying Taur5, Rene Niehus6, Anna Staffas7, Anqi Dai3, Emily Fontana5,
Received: 3 May 2019
Luigi A. Amoretti5, Roberta J. Wright5, Sejal Morjaria5, Maly Fenelus8, Melissa S. Pessin8,
Accepted: 30 September 2020 Nelson J. Chao9, Meagan Lew9, Lauren Bohannon9, Amy Bush9, Anthony D. Sung9,
Tobias M. Hohl5, Miguel-Angel Perales3,4, Marcel R. M. van den Brink3,4 & Joao B. Xavier2 ✉
Published online: 25 November 2020

Check for updates

The gut microbiota influences development1–3 and homeostasis4–7 of the mammalian
immune system, and is associated with human inflammatory8 and immune diseases9,10
as well as responses to immunotherapy11–14. Nevertheless, our understanding of how
gut bacteria modulate the immune system remains limited, particularly in humans,
where the difficulty of direct experimentation makes inference challenging. Here we
study hundreds of hospitalized—and closely monitored—patients with cancer
receiving haematopoietic cell transplantation as they recover from chemotherapy
and stem-cell engraftment. This aggressive treatment causes large shifts in both
circulatory immune cell and microbiota populations, enabling the relationships
between the two to be studied simultaneously. Analysis of observed daily changes in
circulating neutrophil, lymphocyte and monocyte counts and more than 10,000
longitudinal microbiota samples revealed consistent associations between gut
bacteria and immune cell dynamics. High-resolution clinical metadata and Bayesian
inference allowed us to compare the effects of bacterial genera in relation to those of
immunomodulatory medications, revealing a considerable influence of the gut
microbiota—together and over time—on systemic immune cell dynamics. Our analysis
establishes and quantifies the link between the gut microbiota and the human
immune system, with implications for microbiota-driven modulation of immunity.

The human gut microbiota is considered a major modulator of the
immune system during development3 and in health and disease8,9. For
example, preterm infants have distinct microbiome compositions
and distinct developmental trajectories of peripheral immune cell
populations3. In adults, the success of immunotherapies that rely on
peripheral immune cells, such as checkpoint inhibitor treatments,
has been associated with the composition of the microbiome11–13,15.
There is an increasing interest in using the microbiome to modulate the
immune system and augment treatments7,16, including the growing field
of chimeric antigen receptor T cell therapy17. However, our understand-
ing of how the microbiota influences the dynamics of immune cells
in humans and how this compares to deliberate immunomodulatory
interventions remains limited owing to a lack of feasible experiments
in human subjects.
To overcome this limitation, we investigated whether the gut micro-
biota could influence day-by-day changes in peripheral immune cell
counts. We collected a vast dataset of immune-reconstitution dynam-
ics after allogeneic haematopoietic cell transplantation (HCT) from
individuals treated at Memorial Sloan Kettering Cancer Centre (MSK)
between 2003 and 2019 (Fig. 1a, Supplementary Table 1). The condition-
ing regimen of radiation and chemotherapy administered to patients
before HCT is the most severe perturbation to the immune system
deliberately performed in humans: this offers a unique opportunity to
investigate links between the gut microbiota and immune dynamics
directly in humans.
Conditioning depletes white blood cell (WBC) counts, leading to
neutropenia (less than 500 neutrophils per μl blood) until transplanted
stem cells begin to release granulocytes from the bone marrow, initi-
ating immune reconstitution (Fig. 1a–c). HCT also damages the gut
microbiota18 and reduces its biodiversity (Fig. 1d–i), a collateral effect
associated with increased mortality in patients undergoing HCT19.
Immune and microbiome reconstitution vary considerably between
patients and treatment types (Fig. 1, Extended Data Fig. 1a), enabling
analyses of associations between microbiome and immune system, and
their comparison with immunomodulators such as granulocyte-colony
stimulating factor (GCSF).
To detect a directional and causal link between the microbiota and cir-
culatory WBCs, we first used data from a randomized trial of autologous

Institute for Computational Medicine, NYU Langone Health, New York, NY, USA. 2Computational and Systems Biology Program, Sloan Kettering Institute, Memorial Sloan Kettering Cancer
1

Center, New York, NY, USA. 3Adult Bone Marrow Transplantation Service, Department of Medicine, Memorial Sloan Kettering Cancer Center, New York, NY, USA. 4Weill Cornell Medical
College, New York, NY, USA. 5Infectious Disease Service, Department of Medicine, and Immunology Program, Sloan Kettering Institute, New York, NY, USA. 6Harvard University, T. H. Chan
School of Public Health, Boston, MA, USA. 7Sahlgrenska Cancer Center, Department of Microbiology and Immunology, Institute of Biomedicine, University of Gothenburg, Gothenburg,
Sweden. 8Department of Laboratory Medicine, Memorial Sloan Kettering Cancer Center, New York, NY, USA. 9Division of Hematologic Malignancies and Cellular Therapy, Duke University
School of Medicine, Durham, NC, USA. ✉e-mail: jonas.schluter@nyulangone.org; xavierj@mskcc.org

Nature | Vol 588 | 10 December 2020 | 303

Article
a I II III HCT phase b Neutrophil
Patient 1 (PBSC) c Neutrophil
Patient 2 (cord)
engraftment engraftment
20 Mean (n = 2,335)

(×1,000 μl–1)
15 Neutrophil counts
GCSF GCSF
10
5
0
3

(×1,000 μl–1)
2 Lymphocyte counts

0
(×1,000 μl–1)
3 Monocyte counts

1.5

0
0 20 40 56 0 20 40 56 0 20 40 56
Time after HCT (d) Time after HCT (d) Time after HCT (d)
d e f
Mean (n = 1,294)
Diversity

15
10
5

g 1
Mean (n = 1,294) h i
abundance
Relative

0.5 0.5 0.5

0 0 0
0 20 40 56 0 20 40 56 0 20 40 56
Time after HCT (d) Time after HCT (d) Time after HCT (d)
Akkermansiaceae Enterobacteriaceae Lactobacillaceae Actinomycetaceae
Lachnospiraceae Bacteroidaceae Enterococcaceae Peptostreptococcacea Streptococcaceae
Enterococcaceae Bifidobacteriaceae Erysipelotrichaceae Ruminococcaceae Veillonellaceae
Ruminococcaceae
Clostridiaceae 1 Lachnospiraceae Staphylococcaceae Other families

Fig. 1 | Immune reconstitution and microbiome dynamics after HCT.
a–c, Major phases of HCT: immunoablation during conditioning before HCT on
day 0 (I) is followed by post-HCT neutropenia (II) and reconstitution (III). Daily
mean counts (shaded area indicates s.d.) of neutrophils, lymphocytes and
monocytes from individuals receiving transplants between 2003 and 2019 (a),
compared with those from individuals (b, c) representative of the recovery
trajectories for different stem-cell graft sources. Patient 1 received a PBSC
graft and patient 2 received umbilical-cord blood. Line with circles shows data
from the patient; solid line and shaded region show mean ± s.d. for all patients
receiving transplants from the same source. In a, coloured bars on the left
indicate the range of cell counts in healthy individuals. d–f, Loss of microbial
diversity during HCT, measured by 16S rRNA gene sequencing of faecal
samples, supporting previous smaller studies23,24. In d, the line shows daily
mean across patients, shaded area shows s.d. e, f, Data from individual patients.
g–i, Relative abundance of commensal bacteria families. g, Mean (± s.d.)
relative abundance across all patients. h, i, Relative abundance in individual
patients.

faecal microbiota transplantation (auto-FMT)—a controlled micro-
biota manipulation experiment performed directly on our patients20
(Extended Data Fig. 2a). To investigate whether auto-FMT affected WBC
reconstitution, we compared the neutrophil, lymphocyte and mono-
cyte counts after neutrophil engraftment in 24 individuals (engraftment
defined as 3 consecutive days with over 500 neutrophils per μl). FMTs
were conducted at variable dates relative to neutrophil engraftment
(Fig. 2a, Supplementary Table 2). Overall, we observed higher counts
of each WBC type in individuals who received an auto-FMT during the
first 100 days after neutrophil engraftment (P < 0.001, Fig. 2b, c; total
WBCs, Extended Data Fig. 2b–g).
The higher WBCs in individuals receiving auto-FMT could result from
the successful reconstitution of a complex microbiota20 and associated
metabolic capabilities21, or they could be a systemic response to a severe
therapy that introduced billions of intestinal organisms at once via an
enema (no enema was administered to controls20). Moreover, chance
differences in extrinsic factors such as different immunomodulator
medications may have affected this result owing to the small cohort
size. Nonetheless, the results supported the notion that the microbiota
can modulate the peripheral immune system. High lymphocyte counts
during immune reconstitution have been associated with improved
clinical outcomes22, and 3-year survival was positively associated
with higher mean levels of WBCs during the 100 days after neutro-
phil engraftment in the individuals receiving HCT (hazard ratio = 0.91,
P = 0.04). Identifying the taxa that modulate immune dynamics could
therefore open new ways to improve immune reconstitution—critical
for clinical outcomes.
To investigate the links between the gut microbiota and the dynam-
ics of WBC recovery, we turned to our large observational cohort of
patients receiving HCT. Homeostasis of circulatory WBC counts is a
complex, dynamic process: WBCs are formed and released into the
blood de novo by differentiation of haematopoietic progenitor cells
from the bone marrow, and can be mobilized from thymus and lymph
nodes (lymphocytes), and spleen, liver and lungs (neutrophils); WBCs
can also migrate from the blood to other tissues when needed23. To iden-
tify modulators of these dynamic processes, we developed a two-stage
approach analysing the changes of WBC counts between two days
(Fig. 3a). Stage 1 served as a clinical- and metadata-feature-selection
stage using blood and medication data of 1,096 patients without avail-
able microbiome information (Extended Data Fig. 1b shows data inclu-
sion). Stage 2 was performed on data from an independent cohort of 841
different patients from whom concurrent microbiome samples were
available to detect associations between microbiome and peripheral
immune cell dynamics.
In stage 1, we analysed the changes in neutrophils, lymphocytes
and monocytes during recovery from more than 20,000 pairs of
post-engraftment blood samples separated by a single day (Fig. 3b).
A cross-validated feature-selection approach detected medications
and HCT parameters associated with WBC dynamics (Extended Data
Fig. 3a–c, Supplementary Table 3). In stage 2, we sought to identify the
additional contribution of the gut microbiome. We performed Bayes-
ian inferences using data from different sets of patients with available
microbiome samples (Supplementary Table 4). Stage 1 had identified—
as expected—that the sources of stem-cell grafts are associated with
immune reconstitution kinetics (for example, umbilical-cord blood
is associated with slower kinetics than peripheral blood24 (peripheral
blood stem cells (PBSCs)), and we therefore stratified our patients
by graft source in stage 2. The dynamic systems model of stage 2
thus included bacterial genera as predictors of daily changes in WBC
counts, in addition to the medications selected in stage 1, clinical fea-
tures (conditioning intensity, age and sex), and the current state of the
blood in the form of counts of neutrophils, lymphocytes, monocytes,

304 | Nature | Vol 588 | 10 December 2020

 a 10
Neutrophils (×1,000 μl–1)
1.5
Lymphocytes (×1,000 μl–1)
0.05 0.05
Control
FMT-treated
0 100 0 100
Days relative to neutrophil engraftment
b Neutrophils
10 10
FMT
Control
μl–1)
(×1,000 μl–1)
5 5
(×1,000
0.05 0.05
0 5 10 0 5
Weeks relative to FMT randomization date
c HPDI95 (+1%, +7%) per log10 difference) were associated with increased
Neutrophils
Epost-FMT
Eintercept
0 2.5
(×1,000 μl–1)
Fig. 2 | Neutrophil, lymphocyte and monocyte counts increased in
FMT-treated individuals in the weeks following treatment. a, Absolute
counts of neutrophils, lymphocytes and monocytes in 10 control and 14
FMT-treated individuals after neutrophil engraftment. Vertical lines indicate
randomization dates. b, Weekly mean cell counts aligned to the randomization
date. Line shows weekly mean, shaded region shows 95% CI. c, Coefficient
estimates from linear mixed-effects models of neutrophils, lymphocytes and
monocytes over time indicate an increase of each WBC type induced by
auto-FMT (corresponding coefficient, βFMT: P = 4 × 10 −11 (neutrophils),
P = 2 × 10 −10 (lymphocytes) and P = 2 × 10 −16, (monocytes); full regression results
are presented in the Supplementary Information). In b, c, N = 24 subjects,
n = 921 blood samples.
eosinophils and platelets (Fig. 3a). The dataset comprised 841 individu-
als, but approximately 60% of the stool samples paired with a daily
change in WBC counts were taken before neutrophil engraftment,
that is, when WBC counts were zero. Nevertheless, more than 2,000
post-engraftment observations of WBC changes during immune recon-
stitution provided a large sample for analysis of dynamics (Fig. 3b).
Stage 2 focused on data from the largest (Fig. 3c) cohort: PBSC graft
recipients. We withheld the other cohorts (bone marrow (BM); T cell
depleted ex vivo by CD34+ selection grafts (TCD); and umbilical cord
(cord)) for validation scoring, and included data from patients treated
at Duke University for additional validation (Supplementary Table 5).
Notably, as a verification of our approach, we detected associa-
tions between immunomodulator administrations and consequent
immune cell dynamics that were consistent with known biological
mechanisms (Fig. 3c, Extended Data Fig. 4a–f). The strongest asso-
ciation across all predictors is the well-known neutrophil-increasing
effect of GCSF25; GCSF administration—used to accelerate recovery
from chemotherapy-induced neutropenia25—was associated with a
+140% increase in the rate of neutrophil changes from one day to the
next (95% highest posterior density interval (HPDI95) (+114%, +170%)).
This finding was observed in all MSK (v-score = 3, Fig. 3d) and Duke
validation datasets (Extended Data Fig. 4g–j). Furthermore, we found
3
Monocytes (×1,000 μl–1)
a GCSF-associated increase of +43% (HPDI95 (+30%, +58%), v-score = 3)
0.05
in monocyte rates and a smaller increase in lymphocyte rates (+16%,
HPDI95 (+5%, +27%), v-score = 3). Neutrophil and lymphocyte rates
decreased following antihistamine or immunosuppressive medication
(cetirizine, −18%, HPDI95 (−35%, +5%); mycophenolate mofetil, −8%,
HPDI95 (−15%, +1%)). Finally, less intensive chemotherapeutic condi-
tioning regimens (non-ablative and reduced intensity) were associated
with increased lymphocyte and monocyte rates (Extended Data Fig. 4c).
Beyond the mechanistically plausible associations of medications,
our analysis detected associations between the current count of WBCs
and their rates of change: negative associations of neutrophil and lym-
phocyte counts with the rates of monocytes, and negative associa-
tions of platelet and lymphocyte counts with the rates of lymphocytes
and neutrophils (Fig. 3e). Conversely, we found positive associations
between monocytes and the rates of each of the investigated WBC
0 100
subsets. These associations, derived from daily counts of WBCs, could
reflect a complex network underlying the regulation of blood immune
Lymphocytes Monocytes cell composition23. More importantly, the associations quantified for
10
these potential homeostatic feedbacks and medications provided a
benchmark against which we could compare gut microbial taxa.
(×1,000 μl–1)
5 We identified bacterial genera that consistently associated with WBC
dynamics (Fig. 3f). Higher relative abundances of Faecalibacterium
0.05 (+8%, HPDI95 (+1%, +14%) per unit log10 difference), Ruminococcus
10 0 5 10
2 (+5%, HPDI95 (0%, +10%) per log10 difference) and Akkermansia (+4%,
Lymphocytes Monocytes
neutrophil rates, whereas Rothia (−3%, HPDI95 (−7%, 0%) per log10 dif-
ference) and Clostridium sensu stricto 1 (−3%, HPDI95 (−6%, 0%) per log10
difference) associated with reduced neutrophil rates. These results
0 0.38 0 0.5
(×1,000 μl–1) (×1,000 μl–1)
were validated in univariate analyses of the Duke cohort (Extended
Data Fig. 4g–i). We also used total bacterial abundances as predictors
instead of relative abundances; this confirmed Faecalibacterium as
most strongly associated with neutrophil dynamics (Extended Data
Fig. 5). Ruminococcus 2 (+5%, HPDI95 (+1%, +9%) per log10 difference)
and Staphylococcus (+4%, HPDI95 (+1%, +6%) per log10 difference) were
positively associated with lymphocyte rates. Faecalibacterium and
Ruminococcus 2 also associated with increases in monocyte rates; this
association was validated in other cohorts (v-score 3 and 1, respec-
tively), but there was higher uncertainty of the association estimate
(HPDI50 > 0). Clostridium sensu strictu 1 (−3%, HPDI95 (−5%, −1%) per
log10 difference) associated with decreased rates of monocytes. The
associations we identified—and validated in other cohorts—between
gut microbial taxa and daily changes in WBCs support the idea that
haematopoiesis and mobilization respond to the composition of the
gut microbiome, influencing systemic immunity26.
Intestinal bacteria may affect circulatory WBC counts by influencing
either their sources in the bone marrow (or their cytokine profiles27 and
proliferation rates in the blood), their sinks in different organs, or both.
The immune system in turn can interact with the microbiota and modu-
late its composition, for example, via immunoglobulin A, as studied in
mice28–30. To investigate the effect of the immune system on bacterial
populations, we used an analogous approach to the stage 1 analysis.
Dynamics of WBCs could be estimated from changes in absolute cell
counts, and to obtain the necessary absolute bacterial counts, we meas-
ured total bacterial 16S rRNA gene copies per gram of stool for a subset
of our samples (3,995 samples from 481 subjects) to jointly infer the
bidirectional association network between microbiota and the periph-
eral immune system dynamics. All of our subjects received antibiotics
at some point during their treatment18, and their strong influence on
microbiota dynamics were the dominant effects that survived feature
selection (Extended Data Fig. 6). However, relaxing the regulariza-
tion strength (Methods) revealed several bidirectional relationships
between WBCs and gut bacterial dynamics (Extended Data Fig. 7). Of
note, we detected a negative association of absolute [Ruminococcus]
gnavus group abundance with lymphocytes rates, confirming our main
result based on relative bacterial abundances (Fig. 3f). In the reverse
Nature | Vol 588 | 10 December 2020 | 305
Article
a Medications and treatments b Δt = 1 day, WBC dynamic data c Fig. 3 | Bayesian inference reveals associations between the microbiota and
Host state Neutrophils
With microbiota dynamics of circulatory WBC counts. a, Cartoon of the model: observed
Lymphocytes
Monocytes
changes in WBC counts between two consecutive days are associated with the
1,000 Validation
daily change current state of the host in the form of blood cell counts in circulation,

samples
No. of
log(Wt+1) – log(Wt ) 500

PC 2 (EV: 23%)
immunomodulatory medications, clinical metadata and the state of the
W: 0
neutrophils 300 microbiome. b, Visualization of the WBC dynamics data. Scatter plot of the

recipients
lymphocytes

No. of
monocytes 150 principal components (PC) of observed daily changes of neutrophils,
0 lymphocytes and monocytes without (grey; n = 20,751 (after neutrophil

PBSC

BM
TCD

cord
Microbiome engraftment), 81,253 (total)) and with (orange; n = 2,615 (after neutrophil
PC 1 (EV: 48%) Graft type
engraftment), 6,297 (total)) available concurrent microbiota samples. EV,
d V-score V-score V-score e V-score V-score V-score
explained variance. c, Recipients of PBSCs (N = 312) provided most paired
tes ** ** **
cy blood dynamics and microbiota samples (n = 995). Datasets from recipients of
GCSF *** *** *** no ils
Mo oph stem cells from TCD, bone marrow (BM) and umbilical cord (cord) grafts were
MM sin ils **
* * Eo oph used for validation. d–f, Bayesian inference results from PBSC graft recipient
utr tes ** * **
Cetirizine Ne ocy
h
mp ele
ts * data. d, Posterior coefficient distributions of associations between treatments
Ly Plat
–1 0 1 –1 0 1 –1 0 1 –0.3 0.0 0.5 –0.3 0 0.25 –0.3 0 0.25 (colour shows more than 95% posterior density (PD) of coefficient estimates
Effect on: ΔNeutrophils ΔLymphocytes ΔMonocytes ΔNeutrophils ΔLymphocytes ΔMonocytes
Posteriors
greater than zero (red) or less than zero (blue)). MM, mycophenolate mofetil.
e, WBC counts. f, Relative abundances of microbial genera and daily changes
50%

V-score V-score V-score

Mean f (Δ) in neutrophils, lymphocytes and monocytes. The v-score is the number of
95% HPDI
>95% prob.<0
Faecalibacterium validation cohorts confirming associations; it is set to zero if invalidated by
>95% prob.>0
Ruminococcus 2 validation cohorts (additional coefficients in Extended Data Fig. 4a–c). Unid.,
Akkermansia unidentified. g, One hundred microbiota samples with highest (left) or lowest
3 (right) relative abundance of Faecalibacterium, Ruminococcus 2 and
V-score

Unid. Lactobacillales 635

2
1
0 Veillonella Akkermansia. h, Simulation of neutrophil dynamics in the presence of GCSF
Bacteroides and microbiota compositions sampled from those with high (blue) or low (red)
[Clostridium] innocuum group relative abundance of Faecalibacterium, Ruminococcus 2 and Akkermansia as
Staphylococcus
shown in g. Lines show medians of 1,000 simulations and shaded regions show
Parabacteroides
the interquartile range of simulated trajectories. i, Time until neutrophil
[R.] gnavus group
counts first reach a density of 2,000 cells per μl in equivalent simulations
without GCSF.
Clostridium sensu stricto 1

Rothia

Faecalitalea
Ruminococcus 2 and Akkermansia that we associated with increased
–0.1
Effect on:
0.0 0.1
ΔNeutrophils
–0.1 0.0 0.1
ΔLymphocytes
–0.1 0.0
ΔMonocytes
0.1
WBC rates were also among those best reconstituted by auto-FMT20,
g h i potentially explaining the higher counts of neutrophils, monocytes
Faecalibacterium
Ruminococcus 2 25 + GCSF
4
– GCSF and lymphocytes in auto-FMT-treated individuals.
Simulated neutrophil count

Akkermansia
0.6 20 Our analyses show that the gut microbiome is associated with
Probability (%)
Relative abundance

immune cell dynamics in humans. The inferred associations should

(×1,000 μl–1)

15
100 highest 100 lowest
10
0.1
5 effect on other sources and sinks. Unlike the plausible role of obligate
0 0
Sample index Sample index
direction, we saw a positive association of lymphocyte counts with [R.]
gnavus group growth rates. Ruminococcus gnavus is associated with
inflammatory bowel diseases31 and autoimmune disorders10; our analy-
sis suggests that it may drive high neutrophil-to-lymphocyte ratios that
are broadly characteristic of poor disease outcomes in inflammatory
bowel diseases32 and other conditions33,34.
Several of the bacterial taxa positively associated with WBC rates
were obligate anaerobes, some of which produce cell-wall molecules1,35
and short-chain fatty acids36 that modulate immune responses and
granulopoiesis37. Ruminococcus 2, for example, contains keystone
species that release nutrients from complex dietary starch38, and such
nutritional support from the microbiota improved haematopoietic
reconstitution in mice21. To identify a similar association in our patients,
we estimated the microbiota reconstitution potency of each sample
(Methods). Shotgun metagenomic sequences from 124 of our sam-
ples revealed that samples with positive microbiota potency were
enriched in cholate degradation, vitamin B1 synthesis and butanoate
formation pathways (Extended Data Fig. 8). In line with evolutionary
theory39, our results suggest that such broadly available microbial
traits may be co-opted by the host as part of the homeostatic interplay
between immune system and microbiota. The genera Faecalibacterium,
2 be interpreted as net effects, since they do not, for example, distin-
>15 d guish the effect of the microbiota on de novo haematopoiesis from its
anaerobe fermenters in augmenting haematopoiesis via nutritional
0 1 2 3 0 5 10 15
Simulated days Time to 2,000 support21, the positive association detected between Staphylococcus
neutrophils per μl (d)
and lymphocyte dynamics could instead result from reduced extrava-
sation of T cells from circulation into the gut epithelium40, especially
since high abundances of Staphylococcus are associated with low gut
microbiota diversity (P < 0.001, Extended Data Fig. 9a), which indicates
a depleted microbiota.
Nevertheless, our approach enables us to leverage the chronology
of events and assess ‘mathematical causality’41. Owing to the observa-
tional nature of these data, there are risks of confounding, for exam-
ple, from undetected infections or dietary components, which could
explain some of the associations, but the close temporal correspond-
ence41 between microbiota and WBC dynamics reduces the number of
plausible confounders. Notwithstanding potential confounders, our
results suggest candidate bacterial taxa that might improve immune
reconstitution, and focused follow-up studies are required to evaluate
their immunomodulatory efficacy. Members of Faecalibacterium,
Ruminococcus12 in one study, and Akkermansia11 in another have been
associated with better responses to anti–PD-1 immunotherapy, which
suggested a disagreement regarding involved taxa42. Our results, how-
ever, identified Faecalibacterium, Ruminococcus 2 and Akkermansia as
the taxa with the strongest associations with immune cell dynamics,
agreeing with the findings of both these previous studies that these
taxa are associated with human immune modulation. Furthermore,
our work enables us to directly compare their inferred effect sizes with
the effects of immunomodulatory drugs. These genera are common in

306 | Nature | Vol 588 | 10 December 2020

healthy people43, but their abundance can fall below the detection limit
in patients after HCT18. Realistic ranges of 3–5 orders of magnitude in
bacterial relative abundances (Fig. 3g, Extended Data Fig. 9b, c) could
yield effect sizes similar to the homeostatic feedbacks inferred between
WBCs and immunomodulatory medications (for example, a change in
Ruminococcus 2 from below the detection limit to 1% relative abundance
was associated with a +67% change in neutrophils and a +63% change
in lymphocytes). The effect sizes of gut bacteria may initially appear
small relative to those of immunomodulatory drugs, but their effect
could be considerable, as they refer to changes in exponential rates of
WBCs and would therefore accumulate while those bacteria remain
abundant. To demonstrate this accumulation over time, we simulated
WBC dynamics using our posterior coefficient distributions (Meth-
ods). We simulated 1,000 time series for microbiota compositions
chosen from the 100 samples highest or lowest in Faecalibacterium,
Ruminococcus 2 and Akkermansia (Fig. 3g), in the presence (Fig. 3h)
or absence (Fig. 3i) of GCSF administration. Simulations predict that
microbiota enriched in these genera accelerate immune reconstitu-
tion, and reduce the time until neutrophils reach a level of more than
2,000 μl−1 in the absence of GCSF by 2.4 days, from a predicted 6.8
days (95% confidence interval (CI) (6.5, 7)) to 4.4 days (95% CI (4.3, 4.5))
days. Gut bacteria, together and over time, could therefore influence
steady-state immune homeostasis considerably, even in individuals
with less severely injured microbiomes.
In sum, our work links the human gut microbiota to the dynam-
ics of the immune system via peripheral WBC dynamics. Our analy-
sis uses WBCs counted directly from human subjects, which is a
coarse-grained clinical analysis conducted at large scale, but it lacks
details such as the subtypes of lymphocytes and other immune cells.
Because our study is in humans, it fills an important gap at a critical
time for microbiome research, when the clinical relevance of animal
models of microbiome-immune interaction has been questioned44.
By studying a large number of patients over time, we could infer and
quantify the association between gut bacteria and systemic immune
cell dynamics, and our results help to consolidate previous apparently
contradictory findings11,12,42. Our demonstration that the microbiota
influences systemic immunity in humans opens the door towards an
exploration of potential microbiota-targeted interventions to improve
immunotherapy and treatments for immune-mediated and inflamma-
tory diseases8,10–12.
Online content
Any methods, additional references, Nature Research reporting sum-
maries, source data, extended data, supplementary information,
acknowledgements, peer review information; details of author con-
tributions and competing interests; and statements of data and code
availability are available at https://doi.org/10.1038/s41586-020-2971-8.
1. Mazmanian, S. K., Liu, C. H., Tzianabos, A. O. & Kasper, D. L. An immunomodulatory
molecule of symbiotic bacteria directs maturation of the host immune system. Cell 122,
107–118 (2005).
2. Gomez de Agüero, M. et al. The maternal microbiota drives early postnatal innate
immune development. Science 351, 1296–1302 (2016).
3. Olin, A. et al. Stereotypic immune system development in newborn children. Cell 174,
1277–1292 (2018).
4. Tan, T. G. et al. Identifying species of symbiont bacteria from the human gut that, alone,
can induce intestinal Th17 cells in mice. Proc. Natl Acad. Sci. USA 113, E8141–E8150 (2016).
5. Deshmukh, H. S. et al. The microbiota regulates neutrophil homeostasis and host
resistance to Escherichia coli K1 sepsis in neonatal mice. Nat. Med. 20, 524–530 (2014).
6. Ivanov, I. I. et al. Specific microbiota direct the differentiation of IL-17-producing T-helper
cells in the mucosa of the small intestine. Cell Host Microbe 4, 337–349 (2008).
7. Geva-Zatorsky, N. et al. Mining the human gut microbiota for immunomodulatory
organisms. Cell 168, 928–943 (2017).
8. Lloyd-Price, J. et al. Multi-omics of the gut microbial ecosystem in inflammatory bowel
diseases. Nature 569, 655–662 (2019).
9. Markey, K. A. et al. The microbe-derived short-chain fatty acids butyrate and propionate
are associated with protection from chronic GVHD. Blood 136, 130–136 (2020).
10. Azzouz, D. et al. Lupus nephritis is linked to disease-activity associated expansions and
immunity to a gut commensal. Ann. Rheum. Dis. 78, 947–956 (2019).
11. Routy, B. et al. Gut microbiome influences efficacy of PD-1-based immunotherapy against
epithelial tumors. Science 359, 91–97 (2018).
12. Gopalakrishnan, V. et al. Gut microbiome modulates response to anti-PD-1
immunotherapy in melanoma patients. Science 359, 97–103 (2018).
13. Vétizou, M. et al. Anticancer immunotherapy by CTLA-4 blockade relies on the gut
microbiota. Science 350, 1079–1084 (2015).
14. Matson, V. et al. The commensal microbiome is associated with anti-PD-1 efficacy in
metastatic melanoma patients. Science 359, 104–108 (2018).
15. Tanoue, T. et al. A defined commensal consortium elicits CD8 T cells and anti-cancer
immunity. Nature 565, 600–605 (2019).
16. Brandi, G. & Frega, G. Microbiota: overview and implication in immunotherapy-based
cancer treatments. Int. J. Mol. Sci. 20, 2699 (2019).
17. Xin Yu, J., Hubbard-Lucey, V. M. & Tang, J. The global pipeline of cell therapies for cancer.
Nat. Rev. Drug Discov. 18, 821–822 (2019).
18. Morjaria, S. et al. Antibiotic-induced shifts in fecal microbiota density and composition
during hematopoietic stem cell transplantation. Infect. Immun. 87, e00206-19 (2019).
19. Peled, J. U. et al. Microbiota as predictor of mortality in allogeneic hematopoietic-cell
transplantation. N. Engl. J. Med. 382, 822–834 (2020).
20. Taur, Y. et al. Reconstitution of the gut microbiota of antibiotic-treated patients by
autologous fecal microbiota transplant. Sci. Transl. Med. 10, eaap9489 (2018).
21. Staffas, A. et al. Nutritional support from the intestinal microbiota improves
hematopoietic reconstitution after bone marrow transplantation in mice. Cell Host
Microbe 23, 447–457. (2018).
22. Savani, B. N. et al. Absolute lymphocyte count on day 30 is a surrogate for robust
hematopoietic recovery and strongly predicts outcome after T cell-depleted allogeneic
stem cell transplantation. Biol. Blood Marrow Transplant. 13, 1216–1223 (2007).
23. Scheiermann, C., Frenette, P. S. & Hidalgo, A. Regulation of leucocyte homeostasis in the
circulation. Cardiovasc. Res. 107, 340–351 (2015).
24. Thompson, P. A. et al. Umbilical cord blood graft engineering: challenges and
opportunities. Bone Marrow Transplant. 50 (Suppl 2), S55–S62 (2015).
25. Gabrilove, J. L. et al. Effect of granulocyte colony-stimulating factor on neutropenia and
associated morbidity due to chemotherapy for transitional-cell carcinoma of the
urothelium. N. Engl. J. Med. 318, 1414–1422 (1988).
26. Belkaid, Y. & Hand, T. W. Role of the microbiota in immunity and inflammation. Cell 157,
121–141 (2014).
27. Schirmer, M. et al. Linking the human gut microbiome to inflammatory cytokine
production capacity. Cell 167, 1125–1136 (2016).
28. McLoughlin, K., Schluter, J., Rakoff-Nahoum, S., Smith, A. L. & Foster, K. R. Host selection
of microbiota via differential adhesion. Cell Host Microbe 19, 550–559 (2016).
29. Hooper, L. V., Littman, D. R. & Macpherson, A. J. Interactions between the microbiota and
the immune system. Science 336, 1268–1273 (2012).
30. Palm, N. W. et al. Immunoglobulin A coating identifies colitogenic bacteria in
inflammatory bowel disease. Cell 158, 1000–1010 (2014).
31. Henke, M. T. et al. Ruminococcus gnavus, a member of the human gut microbiome
associated with Crohn’s disease, produces an inflammatory polysaccharide. Proc. Natl
Acad. Sci. USA 116, 12672–12677 (2019).
32. Okba, A. M. et al. Neutrophil/lymphocyte ratio and lymphocyte/monocyte ratio in
ulcerative colitis as non-invasive biomarkers of disease activity and severity. Auto Immun.
Highlights 10, 4 (2019).
33. Choi, S.-J. et al. High neutrophil-to-lymphocyte ratio predicts short survival duration in
amyotrophic lateral sclerosis. Sci. Rep. 10, 428 (2020).
34. Gao, Y. et al. Neutrophil/lymphocyte ratio is a more sensitive systemic inflammatory
response biomarker than platelet/lymphocyte ratio in the prognosis evaluation of
unresectable pancreatic cancer. Oncotarget 8, 88835–88844 (2017).
35. Hergott, C. B. et al. Peptidoglycan from the gut microbiota governs the lifespan of
circulating phagocytes at homeostasis. Blood 127, 2460–2471 (2016).
36. Smith, P. M. et al. The microbial metabolites, short-chain fatty acids, regulate colonic Treg
cell homeostasis. Science 341, 569–573 (2013).
37. Balmer, M. L. et al. Microbiota-derived compounds drive steady-state granulopoiesis via
MyD88/TICAM signaling. J. Immunol. 193, 5273–5283 (2014).
38. Ze, X., Duncan, S. H., Louis, P. & Flint, H. J. Ruminococcus bromii is a keystone species for
the degradation of resistant starch in the human colon. ISME J. 6, 1535–1543 (2012).
39. Foster, K. R., Schluter, J., Coyte, K. Z. & Rakoff-Nahoum, S. The evolution of the host
microbiome as an ecosystem on a leash. Nature 548, 43–51 (2017).
40. Fu, Y.-Y. et al. T cell recruitment to the intestinal stem cell compartment drives
immune-mediated intestinal damage after allogeneic transplantation. Immunity 51,
90–103 (2019).
41. Gerber, G. K. The dynamic microbiome. FEBS Lett. 588, 4131–4139 (2014).
42. Jobin, C. Precision medicine using microbiota. Science 359, 32–34 (2018).
43. The Integrative HMP (iHMP) Research Network Consortium. The integrative human
microbiome project. Nature 569, 641–648 (2019).
44. Walter, J., Armet, A. M., Finlay, B. B. & Shanahan, F. Establishing or exaggerating causality
for the gut microbiome: lessons from human microbiota-associated rodents. Cell 180,
221–232 (2020).
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in
published maps and institutional affiliations.
© The Author(s), under exclusive licence to Springer Nature Limited 2020
Nature | Vol 588 | 10 December 2020 | 307
Article
Methods
No statistical methods were used to predetermine sample size. The
experiments were not randomized, except for the auto-FMT trial as
explained in NCT02269150. The investigators were not blinded to allo-
cation during experiments and outcome assessment.
Ethics approval and informed consent
The participants in the auto-FMT trial (NCT02269150) provided written
informed consent to participate in the trial (#14-025). Participants in
the observational cohorts at both MSK and at Duke provided written
informed consent for the use of their faecal specimens and clinical
data. The use and analysis of these specimens for the work herein was
approved by Institutional Research Boards at both institutions: MSK
(#16-834) and Duke (PRO0006268 and Pro00050975).
Complete blood count collection and characterization
Absolute WBC count data were obtained from routine complete blood
counts ordered by clinicians during normal clinical practice, used to
obtain informative diagnostic and monitoring information. Blood
samples received in the clinical haematology laboratory were analysed
using Sysmex XN automated haematology analysers (Sysmex) and,
when needed based on specific flags and parameters as per MSKCC
standard operating procedures, were validated manually using the
Sysmex DI-60 Slide Processing System or CellaVision DM9600 Auto-
mated Digital Morphology System (Sysmex).
16S rRNA gene amplification and multiparallel sequencing
For each sample, duplicate 50-μl PCRs were performed, each containing
50 ng purified DNA, 0.2 mM deoxynucleotide triphosphates, 1.5 mM
MgCl2, 2.5 U Platinum Taq DNA polymerase, 2.5 μl of 10× PCR buffer,
and 0.5 μM of each primer designed to amplify the V4–V5 region:
563F (5′-nnnnnnnnNNNNNNNNNNNNAYTGGGYDTAAAGNG-3′) and
926R (5′-nnnnnnnnNNNNNNNNNNNNCCGTCAATTYHTTTRAGT-3′
). A unique 12-base Golay barcode (Ns) precedes the primers for sam-
ple identification45, and one to eight additional nucleotides (n) were
placed in front of the barcode to offset the sequencing of the primers.
Cycling conditions were 94 °C for 3 min, followed by 27 cycles of 94 °C
for 50 s, 51 °C for 30 s, and 72 °C for 1 min. For the final elongation step,
72 °C for 5 min was used. Replicate PCRs were pooled, and amplicons
were purified using the QIAquick PCR Purification Kit (Qiagen). PCR
products were quantified and pooled at equimolar amounts before
Illumina barcodes and adaptors were ligated, using the Illumina TruSeq
Sample Preparation protocol. The completed library was sequenced
on an Illumina MiSeq platform following the Illumina recommended
procedures with a paired-end 250 × 250-bp kit
Sequence analysis
The 16S (V4-V5) paired-end reads were merged and demultiplexed.
Amplicon sequence variants (ASVs) were identified using the divisive
amplicon denoising algorithm (DADA2) pipeline including filtering
and trimming of the reads46. Reads were trimmed to the first 180 bp
or the first point with a quality score Q < 2. Reads were removed if they
contained ambiguous nucleotides (N) or if two or more errors were
expected based on the quality of the trimmed read. We assigned tax-
onomy to ASVs using an octamer-based classifier trained by IDTaxa47
using the SILVA database48.
Quantification of total microbiota density per gram of stool and
estimation of total genus abundances
qPCR was performed on DNA extracted from 1 g wet weight of a stool sam-
ple using DyNAmo SYBR Green qPCR kit (Finnzymes) and 0.2 μM of the
universal bacterial primer 8F (5′-AGAGTTTGATCCTGGCTCAG-3′) and the
broad-range bacterial primer 338R (5′-TGCTGCCTCCCGTAGGAGT-3′).
Standard curves were prepared by serial dilution of the PCR blunt vector
(Invitrogen) containing 1 copy of the 16 s rRNA gene. Cycling conditions
were 95 °C for 10 min followed by 40 cycles of 95 °C for 30 s, 52 °C for
30 s and 72 °C for 1 min. We used the measurements of total 16S rRNA
gene counts per gram of stool to multiply the relative abundances of
taxa obtained from 16S amplicon sequencing to obtain the estimate
of their total abundance per gram of stool (supplementary informa-
tion). Of note, this does not account for 16S copy-number variation
between taxa, but the observed dynamic ranges in total abundances
of taxa in our dataset span up to nine orders of magnitude, exceeding
the potential inaccuracies due to copy-number variation.
Diversity calculations
Microbiome alpha-diversity was measured by the inverse Simpson
(IS) index of a sample. It was calculated by ISi = N 1 2 , where p is the
∑ j =1 pij
relative abundance of the jth ASV out of N total ASVs in sample i.
Linear mixed-effects model of WBC counts
To study the effect of auto-FMT on WBCs, we investigated the WBC
counts of 24 subjects enrolled in this trial from the day of neutrophil
engraftment until 100 days after engraftment. FMT was performed on
different days relative to neutrophil engraftment. Thus, we performed
an analogous analysis to that conducted in the original publication
that demonstrated how FMT re-established a diverse microbiome in
the post-FMT period20. To determine whether WBC counts differed
after FMT, we used a linear mixed-effects model of WBC counts, y,
modelled as a function of the FMT treatment as well as patient- and
time-point-specific random effects. We included random intercept
terms for each day i and each patient j, and a fixed-effects term for
the post-FMT period with associated coefficient ‘armpost’, using the
indicator variable FMT, which has a value of 1 when a patient was from
the FMT-treated arm of the trial and day was greater than or equal to
the day of the FMT procedure. We conducted independent analyses
for neutrophil, lymphocyte and monocyte counts. This resulted in the
following model of a cell count, y, for patient j on day i:
yij = β0 + armpost × FMTij + dayi + patientj + εij , i = 0, …, D, j = 1, …, P
2 2
with prior distributions dayi ~ N(0, σ day), and patientj ~ N(0, σ patient),
independent error εij ~ N(0, σ 2) and fixed intercept β0, for the D days
after neutrophils engraftment and P individuals, (D = 100, P = 24). For
convenience of those interested in reanalysing our data, the part of
our data concerning the auto-FMT analysis is available in tidy format
(Supplementary Information), and the analysis code conducted in
the R programming language is available as an exported notebook
(fmt_effect_on_wbc.pdf) on github: https://github.com/jsevo/
wbcdynamics_microbiome/ 49. We conducted an additional analysis
with ‘day’ as a continuous predictor, which did not change our conclu-
sions (Supplementary Information).
Dynamic systems analyses
We analysed factors associated with the observed changes of absolute
counts of neutrophils, lymphocytes and monocytes between two days.
In the following we describe how chronology of events and biological
samples were encoded, and the models used to infer a role of medica-
tions, clinical parameters and the microbiome on dynamics of WBCs.
To reveal factors that associate with day-to-day changes in WBC
counts, we started from a first-order differential equation of WBC
(W) dynamics:
d(W )  P 
dt
= W gr +
 ∑ βj Xj 
 j =1 
Where gr represents the intercept, that is, the base line rate of change
during immune reconstitution, and βj are the to-be-estimated
coefficients of the P predictors Xj, j ∈ P of the WBC dynamics. This
equation was then linearized to
P
d(ln W )
dt
= gr + ∑ βj Xj
j =1
And we parameterized the corresponding discrete difference
equation:
P
Δ ln(W )
Δt
= gr + ∑ βj Xj
j =1
where Δln(W) is the log-difference between single days of neutrophils,
lymphocytes or monocytes counts, and Δt = 1 for all intervals. Predictors
include the counts of neutrophils, lymphocytes, monocytes, eosino-
phils and platelets during an interval (homeostatic feedbacks), immu-
nomodulatory medication and clinical observations such as a blood
stream infection and the onset of graft versus host disease, HCT param-
eters such as graft types and conditioning regimens, and, additionally,
the microbiota composition in stage 2 of our analysis (Supplementary
Information for data exclusion and additional details on interval defini-
tions). Importantly, by parameterizing a dynamic equation and analys-
ing rates of change, our coefficient estimates have an immediate causal
interpretation within our modelling framework (that is, a βj >0 implies
that higher levels of the corresponding Xj increases the rate of change
of WBC type, W). To differentiate such results from other associations,
they have been described by the term ‘mathematical causality’41.
Stage 1 analysis
This includes feature selection: identifying medications and clinical
observations associated with WBC dynamics from patients without
microbiome data. Stage 1 uses data of patients without any available
microbiome samples and the following model of WBC changes, y:
p
y = gr + ∑ βjX j,
j=1
with intercept, gr. The predictors, X, include dummy variables for the
HCT graft type, patients’ age on the date of HCT, sex, 13 most frequently
observed positive blood cultures with remaining other blood stream
infections grouped into a separate category ‘other infections’, an indi-
cator for the onset of graft versus host disease, administrations of 55
different, most common immunomodulatory medications and platelet
transfusion events, and HCT conditioning intensity regimens as well
as the log-transformed geometric mean counts of neutrophils, lym-
phocytes, monocytes, eosinophils and platelets during the respective
interval. We used elastic net regression50 for feature selection using the
sklearn package for the Python programming language51. For elastic net
regression with 50% L1 penalty, predictors were scaled between zero
and 1, and we used tenfold cross validation (that is, leaving out 10% of
patients at each cross-validation step) to choose the regularization
strength, λ, solving for
 N  p 
2
p p
 1 1 1
argmingr , β  ∑ yi − gr − ∑ xijβj  + λ ∑ |βj | + 2 λ ∑ β j2
 2N i =1  j=1 
2 j =1 j =1
 
Stage 1 yielded a sparse coefficient matrix of predictors used to design
the model in stage 2.
Stage 2
Stage 2 of the analysis comprises expanded analysis on patients
with microbiome data. To identify associations between microbiota
and WBC dynamics, we conducted an analogous, Bayesian regression
using the package PyMC3 for the Python programming language52.
Stage 1 identified important differences between transplant types,
and we therefore stratified our data into four cohorts according to
the source of the stem-cell graft. Using data independently from each
cohort, we applied ‘no u-turn’ sampling53 to produce 10,000 posterior
samples from 5 independent MCMC chains that parameterized the
model:
y ~ N(μ, σ 2)
Pˆ
μ = gr + ∑ xj βj
j =1
with uninformative prior distributions
gr ~ N(mean = 0, standard deviation = 100)
βj ~ N(mean = 0, standard deviation = 100)
σ ~ half Cauchy(beta = 2)
where y is the observed daily change of a focal WBC type as in stage 1
with normal distributed mean μ, and σ, the model uncertainty with a
thick-tailed half Cauchy prior (importantly, our posterior estimates
do not depend on this choice as we obtain the same results with an
inverse gamma prior, Extended Data Fig. 10b). μ was a function of the
baseline growth rate gr, and predictors P̂: medications with non-zero
coefficients in stage 1, the WBC counts, patient age and sex, and HCT
conditioning intensities; additionally, P̂ now included the
log-abundances of microbial genera as measured by 16S sequencing
from DNA in the stool collected on the second day of a daily interval
(see supplementary information for details). We considered taxa that
were among the 100 most abundant, or had reached maximum relative
abundances of at least 10%, and selected those who were non-zero in
more than 75% of our samples. WBC counts and microbiota data present
during a daily interval were log-transformed, and zeros were filled with
half of the minimum observed non-zero counts (that is, 0.5 × 103 and
2 × 10−6, respectively). We focused on the largest cohort (PBSC) and used
the independent inference results from TCD, bone marrow and cord
cohorts for validation.
Validation score
Coefficients learned from the PBSC cohort were assigned a valida-
tion score based on the results obtained from the other three MSK
patient cohorts. Our requirements for validation were conservative;
we required evidence from our validation datasets as well as absence
of counter evidence. For regression results from each of the validation
graft type cohorts, that is, TCD, bone marrow and cord, we checked if a
coefficient had more than 75% probability (50%HPDI) to have the same
sign as the mean of the PBSC coefficient posterior for a given predictor.
If so, this was considered evidence of validation, and we summed the
evidence over the three validation sets (that is, maximum score of 3, 1
from each of TCD, bone marrow and cord cohorts). Conversely, if we
 found more than 75% probability among any of the validation datasets

that a given predictor had the opposite sign as the posterior mean
 calculated from PBSC data, this was considered counter evidence and
the validation score was always set to zero.
Analysis of WBC dynamics with absolute bacterial abundances
as predictors instead of relative abundances
We conducted an ordinary least-squares regression using the statsmod-
els package in the Python programming language of the same model
as in the main Bayesian analysis using total bacterial abundances as
predictors. This was only possible on a subset of 389 neutrophil, 331
lymphocyte and 376 monocyte rate observations from PBSC patients.
Article
Forwards simulation of predicted immune system
reconstitution kinetics
To assess the impact of the estimated microbiota coefficients on immune
system dynamics, we conducted 1,000 simulations of the system of 3 dif-
ferential equations describing the dynamics of neutrophils, lymphocytes
and monocytes. We ran 1,000 simulations four times: in the presence and
absence of GCSF, each with microbiota compositions enriched or depleted
in Faecalibacterium, Ruminococcus 2 and Akkermansia. To identify these
compositions, we ranked the observed microbiota compositions by these
taxa, and chose randomly either from the top or bottom 100. The coef-
ficients for WBC interactions, interactions with the microbiota and the
effect of GCSF were sampled from our posterior coefficient distributions.
Using these coefficients sampled at the start of the simulation, and using
50 cells μl−1 of neutrophils, lymphocytes and monocytes as initial values,
we simulated these differential equations forwards in time using the odeint
function of the scipy package for the Python programming language.
Validation on data from Duke University
We analysed 9,603 blood samples with 25,581 associated administra-
tions of immunomodulatory medications, and 741 microbiota sam-
ples from Duke as an orthogonal dataset to validate our findings. The
temporal resolution of this data was much lower, and after filtering for
samples from the relevant post-neutrophil engraftment period, and by
requiring daily intervals, 83 valid, complete data points were available.
Using these data, we correlated daily blood cell changes individually
in univariate, or jointly in a partial least squares regression, with those
predictors that achieved more than 95% probability density in the posi-
tive or negative domain in the PBSC data regression. For each of these
predictors, we present the sign of slopes and Bonferroni-corrected P
values from individual linear regressions.
Joint analysis of the effect of antibiotics and WBC counts on
the microbiota and the microbiota and immunomodulatory
medications on WBC counts
Analogous to stage 1, we performed cross-validated, regularized linear
regressions (ElasticNet) using the scikit-learn package for the Python
programming language to jointly estimate the association network
between microbiota and circulatory WBCs. For this, we constructed a
block matrix X of predictor matrices Xi that include the absolute bacte-
rial abundances, drug data (antibiotics for bacterial dynamics and
immune modulators for WBC dynamics), as well as the counts of WBCs
and a separate intercept term per block. Each block X ln l , pl , with nl obser-
vations and pl predictors (l = 0,...,k), on the diagonal of X corresponds
to the indices of the observed daily log-changes of one of the 41 bacte-
rial genera considered in our main analysis or the log changes in neu-
trophil, lymphocyte and monocyte counts from PBSC patients
contained in Y (in total we calculated 15,833 rates from 256 patients).
Our regression problem can thus be written as:
X n00, p0 ⋯ 0n 0, pk
argminβ (Y − Xβ) where X = ⋮ ⋱ ⋮
0n k , p0 ⋯ X nk k , pk
with k = 44, that is, 41 bacterial genera and 3 WBC types, the to-be estimated
coefficient vector β and 0 the zero matrix. This system is underdetermined
and we therefore chose the same approach as in stage 1, elastic net regres-
sion, for feature selection. Predictors were scaled between zero and 1, and
we used threefold cross validation, leaving out one-third of the patients
at each iteration to identify a global regularization strength λ, solving for
  
2 
1 1 1  classification of microbiome sequences. Microbiome 6, 140 (2018).
argminβ  ∑ yi − ∑ xij βj  +
2
λ ∑ |βj | + 2
λ ∑ β 2j 
2 i =1  j =1  j =1 j =1 
 
where η is the total number of observed daily log changes in genera and
WBCs, and ρ the total number of predictors. This yielded a strongly
regularizing λs, and thus few predictors. To characterize potential bidi-
rectional relationships between WBC counts and the gut microbiota,
we iteratively reduced the regularization strength until the strongest
interaction between microbiota and WBC dynamics, that is, Faecali-
bacterium with neutrophil dynamics, was detected. We than re-ran the
regression with this reduced regularization strength λr.
Shotgun sequencing
Sequencing of 124 post-neutrophil engraftment was conducted on
the Illumina HiSeq platform. For details and the processing of the
FASTQ files, see supplementary information. We used the HUMAnN2
pipeline54 with default settings for functional profiling of our sam-
ples, with the UniRef90 data base and ChocoPhlAn for alignment,
and we renormalized our samples by library depth to copies per mil-
lion. We used MetaCyc to obtain stratified and unstratified pathway
abundances.
Statistical analysis of shotgun data
We calculated the predicted microbiota potency score for each sample
and separately for neutrophils, lymphocytes and monocytes, by multi-
plying the abundances of taxa in each of the 124 samples with the cor-
responding posterior coefficients obtained from the PBSC inference.
To distinguish the sets of metabolic functions that separate samples
with positive and negative predicted potencies, we converted the path-
way abundances into presence and absences profiles. We performed
a linear discriminant analysis between positive and negative potency
samples with a least squares solver and automatic shrinkage using the
Ledoit–Wolf lemma using the sklearn package for the Python program-
ming language51. To assess differences in the presence or absence of
pathways between samples with positive and negative potency, we
used Fisher’s exact test for each pathway.
Reporting summary
Further information on research design is available in the Nature
Research Reporting Summary linked to this paper.
Data availability
All data supporting the findings of this study are available within the
paper and its Supplementary Information files. The data used in our
study are organized in Excel-compatible comma-separated value files
as Supplementary Tables (data-tables.zip). All sequencing data have
been made available publicly, and the NCBI SRA accession numbers are
listed in the Supplementary Tables. Metadata and processed sequenc-
ing data are made available on a public repository via Figshare: meta
data, https://doi.org/10.6084/m9.figshare.12016986.v4; samples,
https://doi.org/10.6084/m9.figshare.12016983.v4; 16S counts, https://
doi.org/10.6084/m9.figshare.12016989.v3; and 16S taxonomy, https://
doi.org/10.6084/m9.figshare.12016992.v1.
Code availability
All of the steps of the analyses that were performed in this study
are described in detail to allow reproduction of the results. Rele-
vant analysis code is available publicly at https://github.com/jsevo/
wbcdynamics_microbiome.
45. Caporaso, J. G. et al. Ultra-high-throughput microbial community analysis on the Illumina
HiSeq and MiSeq platforms. ISME J. 6, 1621–1624 (2012).
46. Callahan, B. J. et al. DADA2: high-resolution sample inference from Illumina amplicon
data. Nat. Methods 13, 581–583 (2016).
47. Murali, A., Bhargava, A. & Wright, E. S. IDTAXA: a novel approach for accurate taxonomic
48. Quast, C. et al. The SILVA ribosomal RNA gene database project: improved data
processing and web-based tools. Nucleic Acids Res. 41, D590–D596 (2013).
49. Pinheiro, J. C., Bates, D. M., DebRoy, S. S. & Sarkar, D. nlme: Linear and Nonlinear Mixed
Effects Models. R package version 3.1-150 (2013).
50. Tibshirani, R. Regression shrinkage and selection via the lasso. J. R. Stat. Soc. B 58,
267–288 (1996).
51. Pedregosa, F. et al. Scikit-learn: machine learning in Python. J. Mach. Learn. Res. 12, 2825
(2011).
52. Salvatier, J., Wiecki, T. V. & Fonnesbeck, C. Probabilistic programming in Python using
PyMC3. PeerJ Comput. Sci. 2, e55 (2016).
53. Hoffman, M. D. & Gelman, A. The No-U-turn sampler: adaptively setting path lengths in
Hamiltonian Monte Carlo. J. Mach. Learn. Res. 15, 1593–1623 (2014).
54. Franzosa, E. A. et al. Species-level functional profiling of metagenomes and
metatranscriptomes. Nat. Methods 15, 962–968 (2018).
Acknowledgements We thank M. Lipsitch, S. B. Andersen, K. R. Foster, J. K. Sia, E. G. Pamer,
K. Coyte, S. Mitschka and the members of the Xavier lab for helpful discussion and comments
on the manuscript. This work was supported by the National Institutes of Health (NIH) grants
U01 AI124275, R01 AI137269 and U54 CA209975 to JBX, by the MSKCC Cancer Center Core
Grant P30 CA008748, the Parker Institute for Cancer Immunotherapy at Memorial Sloan
Kettering Cancer Center, the Sawiris Foundation, the Society of Memorial Sloan Kettering
Cancer Center, MSKCC Cancer Systems Immunology Pilot Grant and Empire Clinical Research
Investigator Program. M.S. received funding from the Burroughs Wellcome Fund Postdoctoral
Enrichment Program, the Damon Runyon Physician-Scientist Award, and the Robert Wood
Johnson Foundation. T.M.H. is investigator in the Pathogenesis of Infectious Diseases from the
Burroughs Wellcome Fund, and funded via an award from Geoffrey Beene Foundation, and NIH
RO1 AI093808. The funders had no role in study design, data collection and analysis, decision
to publish or preparation of the manuscript.
Author contributions J.S. and J.B.X. wrote the manuscript. J.S. and J.B.X. designed the
analyses with expert help from R.N. J.U.P. and Y.T. contributed to the clinical data preparation,
B.P.T. provided the 16S data-processing pipelines, K.A.M., M.S., A.S., S.M., M.F., M.S.P., T.M.H.,
M.-A.P. and M.R.M.v.d.B. provided clinical context and helped with variable selection, N.J.C.,
M.L., L.B., A.B. and A.D.S. provided clinical and other data from Duke, A.D. provided the
shotgun processing pipelines. E.F., L.A.A. and R.J.W. processed patients’ stool samples,
including for 16S sequencing, shotgun metagenomics and qPCR quantification of total 16S
rRNA gene. All authors contributed to the writing and interpretation of the results.
Competing interests M.R.M.v.d.B. and J.U.P. received financial support from Seres
Therapeutics. M.-A.P. has received honoraria from AbbVie, Bellicum, Bristol-Myers Squibb,
Incyte, Merck, Novartis, Nektar Therapeutics, and Takeda, research support for clinical trials
from Incyte, Kite (Gilead) and Miltenyi Biotec, and serves on data and safety monitoring boards
for Servier and Medigene and scientific advisory boards for MolMed and NexImmune.
Additional information
Supplementary information is available for this paper at https://doi.org/10.1038/s41586-020-
2971-8.
Correspondence and requests for materials should be addressed to J.S. or J.B.X.
Peer review information Nature thanks Henrik Nielsen and the other, anonymous, reviewer(s)
for their contribution to the peer review of this work.
Reprints and permissions information is available at http://www.nature.com/reprints.
Article

Extended Data Fig. 1 | Blood cell counts over time. a, WBC counts and platelet counts per graft source over the first 100 days post HCT per day relative to HCT
from N = 2,235 adult patients (detailed demographics in supplementary Table 1); lines: mean, shaded: ± standard deviations. b, Data exclusion diagram.
Extended Data Fig. 2 | FMT increases WBC counts. a, HCT patient who
received an autologous faecal microbiota transplant (auto-FMT, dashed red
line) that restored commensal microbial families and ecological diversity in
the gut microbiota, with concurrent cell counts of peripheral neutrophils,
lymphocytes and monocytes and immunomodulatory drug administrations.
b, Total WBC counts in 24 enrolled patients (10 control, 14 treated)
post-neutrophil engraftment; vertical lines indicate randomization dates.
c, Weekly mean WBC counts aligned to the randomization date (FMT-treated:
red, control: black). Line: mean per week, shaded region: 95% CI. d, Coefficient
estimates (mean vs. mean + FMT effect) from linear mixed effects models of
total WBC counts over time indicate an auto-FMT-induced increase of WBCs
(βFMT: P = 7 × 10 −14). e–g, Respectively: neutrophil, lymphocyte and monocyte
count trajectories of 24 FMT trial patients. Thin lines: raw data (blue:
post-FMT); thick black: mean per day, thick blue: mean+post-FMT coefficient.
Means and confidence intervals (shaded region) without (black) and after FMT
(blue), as well as the coefficient estimate for FMT treatment and its P value from
a linear mixed effects model relating cell counts over time to the FMT
treatment (Methods).
Article

Extended Data Fig. 3 | Results of the feature selection stage 1 regression. gr: intercept; TCD: T cell depleted graft (ex-vivo) by CD34+ selection; PBSC:
a–c, Stage 1 regression on neutrophil, lymphocyte, and monocyte dynamics, peripheral blood stem cells; BM: bone marrow; cord: umbilical cord blood;
respectively, on patients without microbiome data. Coefficients from NONABL: Nonmyeloablative; REDUCE: reduced-intensity conditioning
tenfold cross-validated elastic net regression daily changes in neutrophils. regimen; F: female; N: patients, n: samples (daily changes in neutrophils).
Extended Data Fig. 4 | Additional coefficients, posterior convergence
evaluation and validation. a–c, Additional posterior coefficient estimates of
medications, additional genera and HCT metadata from the Bayesian stage 2
regression, see also Fig. 3. REDUCE: reduced-intensity conditioning regimen;
NONABL: non-myeloablative conditioning regimen. F: female. d–f, posterior
sampling convergence. Histograms of the ranked posterior draws from the
model of neutrophil, lymphocyte and monocyte dynamics, respectively, in
PBSC patients (ranked over all chains), plotted separately for each chain show
no substantial differences between chains. g–i, Predictors of WBC dynamics
using data from patients treated at Duke. Heatmaps indicate the slope
coefficients from individual univariate regressions of microbiome and clinical
predictors with changes in neutrophils, lymphocytes and monocyte, and for
comparison the corresponding coefficients signs from the Bayesian multiple
linear regressions in stage 2 of the analysis of WBC dynamics in MSK patients
(Fig. 3). Pvalues were adjusted for multiple hypothesis testing using Bonferroni
correction: ***P < 0.001, **P < 0.01, *P < 0.05; P > 0.05: n.s. Sign of coefficients
from MSK PBSC patients for comparison. j, Equivalent validation analysis from
patients treated at Duke using partial least squares regression of microbiome
and clinical predictors identified in stage 2 of our analysis on daily changes in
neutrophils, lymphocytes and monocyte.
Article
Extended Data Fig. 5 | Validation using absolute instead of relative
abundance bacterial genus data. a–d, Validation analysis of the main model
using absolute bacterial abundances as predictors instead of relative
abundances in Fig. 3. Results show inferred coefficients and P values from
multiple linear regressions. One regression per analysed WBC type dynamics,
that is, neutrophil, lymphocyte and monocyte daily log-changes, was
conducted, and coefficients for medications (a), WBC feedbacks (b)
metadata (c) and total genus abundances (d) are shown. This was only possible
for only a subset of the data used in the main analysis for which we obtained
absolute bacterial abundance estimates (Methods), n: samples, N: patients.
Extended Data Fig. 6 | Jointly inferred association network between WBC and bacterial genus dynamics. Strong regularization yields few non-zero
coefficients and antibiotics dominate the dynamics.
Article

Extended Data Fig. 7 | Jointly inferred association network between WBC for example, between lymphocytes and [Ruminococcus] gnavus group
and bacterial genus dynamics with reduced regularization. Reducing (highlighter green boxes, and cartoon).
regularization strength (Methods) indicates potential bidirectional feedbacks,
Extended Data Fig. 8 | Functional analysis of microbiota samples. To
distinguish samples predicted to increase rates of WBCs, a microbiota potency
score was calculated from posterior coefficients (Fig. 3, Methods) and the
relative abundance of taxa in samples. Bars show linear discriminant analysis
(LDA) scores of MetaCyc pathway profiles from 124 shotgun sequenced
samples that distinguished positive and negative potency samples the most
(LDA-score magnitude in the 95th percentile). Highlighted pathways are
discussed in the main text. For each pathway, we tested whether pathway
presence was enriched (depleted) in positive (negative) potency samples using
one-sided Fisher’s exact test; ***P < 0.001, **P < 0.01, *P < 0.05.
Article
Extended Data Fig. 9 | Abundance profiles of bacterial genera across
analysed samples. a, The relative non-zero abundance of Staphylococcus is
inversely related to microbiome alpha diversity, bold line: regression line from
a linear model of the mean of the log10 Staphylococcus relative abundance,
shaded: 95% confidence intervals (n = 1,381 samples with non-zero
Staphylococcus abundances). b, Abundance profiles of the two genera,
Faecalibacterium and Ruminococcus 2, most strongly associated with WBC
increase; number of times detected (left) and log10 abundance distribution
when above detection (right).
Extended Data Fig. 10 | Survival analysis and confirmation of model results prior for σ in the main Bayesian model. Plotted are the posterior means from
with different priors. a, Kaplan–Meier plot of patient 3-year survival with our main analysis against the equivalent inference with an inverse Gamma prior
sufficient available blood data (Supplementary Information, Extended Data (alpha = 1, beta = 1).
Fig. 1). b, posterior association coefficients do not depend on the choice of
 nature research | reporting summary
Corresponding author(s): Jonas Schluter, Joao B. Xavier
Last updated by author(s): Aug 8, 2020

Reporting Summary
Nature Research wishes to improve the reproducibility of the work that we publish. This form provides structure for consistency and transparency
in reporting. For further information on Nature Research policies, see our Editorial Policies and the Editorial Policy Checklist.

Statistics
For all statistical analyses, confirm that the following items are present in the figure legend, table legend, main text, or Methods section.
n/a Confirmed
The exact sample size (n) for each experimental group/condition, given as a discrete number and unit of measurement
A statement on whether measurements were taken from distinct samples or whether the same sample was measured repeatedly
The statistical test(s) used AND whether they are one- or two-sided
Only common tests should be described solely by name; describe more complex techniques in the Methods section.

A description of all covariates tested

A description of any assumptions or corrections, such as tests of normality and adjustment for multiple comparisons
A full description of the statistical parameters including central tendency (e.g. means) or other basic estimates (e.g. regression coefficient)
AND variation (e.g. standard deviation) or associated estimates of uncertainty (e.g. confidence intervals)

For null hypothesis testing, the test statistic (e.g. F, t, r) with confidence intervals, effect sizes, degrees of freedom and P value noted
Give P values as exact values whenever suitable.

For Bayesian analysis, information on the choice of priors and Markov chain Monte Carlo settings
For hierarchical and complex designs, identification of the appropriate level for tests and full reporting of outcomes
Estimates of effect sizes (e.g. Cohen's d, Pearson's r), indicating how they were calculated
Our web collection on statistics for biologists contains articles on many of the points above.

Software and code

Policy information about availability of computer code
Data collection no software was used for data collection

Data analysis python3.7.0, R3.6.1, DADA2, Humann2, ChocoPhlAn, MetaCyc, PyMC3, sklearn-0.23.2, https://github.com/jsevo/wbcdynamics_microbiome/
For manuscripts utilizing custom algorithms or software that are central to the research but not yet described in published literature, software must be made available to editors and
reviewers. We strongly encourage code deposition in a community repository (e.g. GitHub). See the Nature Research guidelines for submitting code & software for further information.

Data
Policy information about availability of data
All manuscripts must include a data availability statement. This statement should provide the following information, where applicable:
- Accession codes, unique identifiers, or web links for publicly available datasets
- A list of figures that have associated raw data
- A description of any restrictions on data availability
April 2020

The data used in our study is organized in Excel compatible comma separated value files as supplementary tables (data-tables.zip). All sequencing data have been
made available publicly, and the NCBI SRA accession numbers are listed in the tables below:

1. cGENUS.csv: relative taxon abundances in fecal microbiota samples from 12,633 stool samples
2. cHCTMETA.csv: HCT characteristics
3. cINFECTIONS.csv: positive blood culture results
4. cMISAMPLES.csv: NCBI SRA accession number, diversity (inverse Simpson index), total 16S (where available), stool consistency for each fecal microbiota sample
5. cMED.csv: medication data

1
6. cPIDMETA.csv: anonymized patient demographics
7. cWBC.csv: absolute counts of neutrophils, lymphocytes, monocytes, eosinophils, and platelets with indication if included in analyses

nature research | reporting summary

8. cDUKE__GENUS.csv: relative taxon abundances in fecal microbiota samples from 12,633 stool samples
9. cDUKE__WBC.csv: absolute counts of neutrophils, lymphocytes, monocytes, eosinophils, and platelets with indication if included in analyses
10. cDUKE__MED.csv: medication data
11. cFMT_analysis.csv: convenience table for Figure 2

Metadata and processed sequencing data are made available on a public repository via Figshare:

meta data: doi.org/10.6084/m9.figshare.12016986.v4

samples: doi.org/10.6084/m9.figshare.12016983.v4
16S counts: doi.org/10.6084/m9.figshare.12016989.v3
16S taxonomy: doi.org/10.6084/m9.figshare.12016992.v1

Field-specific reporting
Please select the one below that is the best fit for your research. If you are not sure, read the appropriate sections before making your selection.
Life sciences Behavioural & social sciences Ecological, evolutionary & environmental sciences
For a reference copy of the document with all sections, see nature.com/documents/nr-reporting-summary-flat.pdf

Life sciences study design

All studies must disclose on these points even when the disclosure is negative.
Sample size Throughout the manuscript, sample sizes are specified. There are many nested analyses on various subsets, and each analysis specifies the
sample sizes. Sample sizes were not predetermined for retrospective analyses, instead all data from electronic health records from allo-HCT
patients since 2003 were used, with specific exclusion criteria listed in the data exclusion section.

Data exclusions Non-adults, non-engrafted patients were excluded. Data from patients without valid two-day apart sample pairs were excluded.The
supplementary methods provides a detailed flow chart of data inclusion/exclusion.

Replication All analyses can be reproduced with the code and data provided. No experiments were conducted.

Randomization N/A as no trial was conducted for this study. The randomized data was previously published and the randomization procedure is explained in
the relevant reference (Taur et al. 2018)

Blinding N/A. No trial was conducted.

Reporting for specific materials, systems and methods

We require information from authors about some types of materials, experimental systems and methods used in many studies. Here, indicate whether each material,
system or method listed is relevant to your study. If you are not sure if a list item applies to your research, read the appropriate section before selecting a response.

Materials & experimental systems Methods

n/a Involved in the study n/a Involved in the study
Antibodies ChIP-seq
Eukaryotic cell lines Flow cytometry
Palaeontology and archaeology MRI-based neuroimaging
Animals and other organisms
Human research participants
Clinical data
Dual use research of concern

Human research participants

April 2020

Policy information about studies involving human research participants

Population characteristics Tables S1-S5 list the patient population characteristics.

Recruitment No patients were specifically recruited for this work. Allo-HCT patients since 2003 were considered and included or excluded
as detailed in the data inclusion/exclusion section.

Ethics oversight The participants in the auto-FMT trial (NCT02269150) provided written informed consent to participate in the trial (#14-025).

2
 Ethics oversight Participants in the observational cohorts at both Memorial Sloan Kettering Cancer Center and at Duke University School of
Medicine provided written informed consent for the use of their fecal specimens and clinical data. The use and analysis of

nature research | reporting summary

these specimens for the work herein was approved by IRBs at both institutions: MSK (#16-834) and Duke (PRO0006268 and
Pro00050975).
Note that full information on the approval of the study protocol must also be provided in the manuscript.

Clinical data
Policy information about clinical studies
All manuscripts should comply with the ICMJE guidelines for publication of clinical research and a completed CONSORT checklist must be included with all submissions.

Clinical trial registration NIH clinicaltrials.gov : NCT02269150

https://clinicaltrials.gov/ct2/show/NCT02269150

Study protocol https://clinicaltrials.gov/ct2/show/NCT02269150

Data collection This is a randomized, open-label, controlled study designed to assess the efficacy of autologous fecal microbiota transplantation
(auto-FMT) for prevention of Clostridium difficile infection (CDI) in patients who have undergone allogeneic hematopoietic stem cell
transplantation (allo-HSCT). Patients will be enrolled prior to allo-HSCT; feces will be collected and stored from all participating
subjects prior to the initiation of conditioning regimens, analyzed by deep 16S rRNA gene sequencing, and tested by assay for
intestinal pathogens including Clostridium difficile. Later in the course of transplantation, following engraftment (defined as the first
day of three consecutive days, that the absolute blood neutrophil count is at above f 500 mm3), subjects will undergo fecal testing
for presence of Bacteroidetes by 16S PCR. Subjects will be eligible for study if they have a microbiologically diverse pre-transplant
colonic microbiota, and if the post-engraftment specimen contains Bacteroidetes at a prevalence equal to or below (0.1%)

Actual Study Start Date : October 2014

Estimated Primary Completion Date : October 2021

Outcomes Primary Outcome: Clostridium difficile infection (CDI) [ Time Frame: up to 1 year following randomization ]

CDI is defined as diarrheal stool (unformed stool conforming to the shape of a specimen container), and a positive test for toxin-
producing C. difficile (either by toxin B gene PCR or cytotoxicity assay).

April 2020

3
Article

Sex differences in immune responses that

underlie COVID-19 disease outcomes

https://doi.org/10.1038/s41586-020-2700-3 Takehiro Takahashi1,21, Mallory K. Ellingson2,21, Patrick Wong1,21, Benjamin Israelow1,3,21,

Carolina Lucas1,21, Jon Klein1,21, Julio Silva1,21, Tianyang Mao1,21, Ji Eun Oh1, Maria Tokuyama1,
Received: 4 June 2020
Peiwen Lu1, Arvind Venkataraman1, Annsea Park1, Feimei Liu1,4, Amit Meir5, Jonathan Sun6,
Accepted: 19 August 2020 Eric Y. Wang1, Arnau Casanovas-Massana2, Anne L. Wyllie2, Chantal B. F. Vogels2,
Rebecca Earnest2, Sarah Lapidus2, Isabel M. Ott2,7, Adam J. Moore2, Yale IMPACT Research
Published online: 26 August 2020
Team*, Albert Shaw3, John B. Fournier3, Camila D. Odio3, Shelli Farhadian3, Charles Dela Cruz8,
Check for updates Nathan D. Grubaugh2, Wade L. Schulz9,10, Aaron M. Ring1, Albert I. Ko2, Saad B. Omer2,3,11,12 &
Akiko Iwasaki1,13 ✉

There is increasing evidence that coronavirus disease 2019 (COVID-19) produces more
severe symptoms and higher mortality among men than among women1–5. However,
whether immune responses against severe acute respiratory syndrome coronavirus
(SARS-CoV-2) differ between sexes, and whether such differences correlate with the
sex difference in the disease course of COVID-19, is currently unknown. Here we
examined sex differences in viral loads, SARS-CoV-2-specific antibody titres, plasma
cytokines and blood-cell phenotyping in patients with moderate COVID-19 who had
not received immunomodulatory medications. Male patients had higher plasma
levels of innate immune cytokines such as IL-8 and IL-18 along with more robust
induction of non-classical monocytes. By contrast, female patients had more robust
T cell activation than male patients during SARS-CoV-2 infection. Notably, we found
that a poor T cell response negatively correlated with patients’ age and was associated
with worse disease outcome in male patients, but not in female patients. By contrast,
higher levels of innate immune cytokines were associated with worse disease
progression in female patients, but not in male patients. These findings provide a
possible explanation for the observed sex biases in COVID-19, and provide an
important basis for the development of a sex-based approach to the treatment and
care of male and female patients with COVID-19.

SARS-CoV-2 is the novel coronavirus first detected in Wuhan, China, in
November 2019 that causes COVID-196. On 11 March 2020, the World
Health Organization (WHO) declared COVID-19 a pandemic7. A growing
body of evidence reveals that male sex is a risk factor for a more severe dis-
ease, including death. Globally, approximately 60% of deaths from COVID-
19 are reported in men5, and a cohort study of 17 million adults in England
reported a strong association between male sex and the risk of death
from COVID-19 (hazard ratio 1.59, 95% confidence interval 1.53–1.65)8.
Past studies have shown that sex has a considerable effect on the out-
come of infections and has been associated with underlying differences
in immune responses to infection9,10. For example, the prevalence of
hepatitis A and tuberculosis are notably higher in men that in women11.
Viral loads are consistently higher in male patients with hepatitis C virus
and human immunodeficiency virus (HIV)12,13. By contrast, women mount
a more robust immune response to vaccines14. These findings collectively
suggest a more robust ability among women to control infectious agents.
However, the mechanism by which SARS-CoV-2 causes more severe dis-
ease in male patients than in female patients remains unknown.
To determine the immune responses against SARS-CoV-2 infection in
male and female patients, we performed detailed analyses on the sex dif-
ferences in immune phenotypes by the assessment of viral loads, levels
of SARS-CoV-2-specific antibodies, plasma cytokines or chemokines,
and blood-cell phenotypes.
Overview of the study design
Patients who were admitted to the Yale-New Haven Hospital between
18 March and 9 May 2020 and were confirmed positive for SARS-CoV-2

1
Department of Immunobiology, Yale University School of Medicine, New Haven, CT, USA. 2Department of Epidemiology of Microbial Diseases, Yale School of Public Health, New Haven, CT,
USA. 3Department of Medicine, Section of Infectious Diseases, Yale University School of Medicine, New Haven, CT, USA. 4Department of Biomedical Engineering, Yale School of Engineering &
Applied Science, New Haven, CT, USA. 5Boyer Center for Molecular Medicine, Department of Microbial Pathogenesis, Yale University, New Haven, CT, USA. 6Department of Comparative
Medicine, Yale University School of Medicine, New Haven, CT, USA. 7Department of Ecology and Evolutionary Biology, Yale University, New Haven, CT, USA. 8Department of Medicine, Section of
Pulmonary and Critical Care Medicine, Yale University School of Medicine, New Haven, CT, USA. 9Department of Laboratory Medicine, Yale University School of Medicine, New Haven, CT, USA.
10
Center for Outcomes Research and Evaluation, Yale-New Haven Hospital, New Haven, CT, USA. 11Yale Institute for Global Health, Yale University, New Haven, CT, USA. 12Yale School of Nursing,
Yale University, Orange, CT, USA. 13Howard Hughes Medical Institute, Chevy Chase, MD, USA. 21These authors contributed equally: Takehiro Takahashi, Mallory K. Ellingson, Patrick Wong,
Benjamin Israelow, Carolina Lucas, Jon Klein, Julio Silva, Tianyang Mao. *A list of authors and their affiliations appears at the end of the paper. ✉e-mail: akiko.iwasaki@yale.edu

Nature | Vol 588 | 10 December 2020 | 315

Article
by RT–PCR from nasopharyngeal and/or oropharyngeal swabs in a Np swab Saliva b Anti-S1-IgG Anti-S1-IgM
P < 0.0001
CLIA-certified laboratories were enrolled through the IMPACT biore- P = 0.082 2.0 P = 0.0021
10 10 2.5
pository study15. In the IMPACT study, biospecimens including blood,

log10[SARS-CoV-2
8 8 2.0 P = 0.0034
nasopharyngeal swabs, saliva, urine and stool samples were collected 1.5

(copies ml–1)]

A450 nm
6 6 1.5
at study enrolment (baseline denotes the first time point) and longitu- 1.0
4 4 1.0
dinally on average every 3 to 7 days (serial time points). The detailed 0.5
2 2 0.5
demographics and clinical characteristics of these 98 participants are
0 0 0.0 0.0
shown in Extended Data Table 1. Plasma and peripheral blood mono-

F_ W
W

t
Pt

F_ W
W

t
Pt
_P

_P

_P

_P
C
C

C
C
nuclear cells (PBMCs) were isolated from whole blood, and plasma was

F_

F_

F_

F_
M

M
_H

_H

H
M

M
used for titre measurements of SARS-CoV-2 spike S1 protein-specific IgG c
P = 0.082 P = 0.0011 P = 0.024 P = 0.0005 P = 0.027
and IgM antibodies (anti-S1-IgG and -IgM) and cytokine or chemokine 400 300 250

IL-18 (pg ml–1)

measurements. Freshly isolated PBMCs were stained and analysed by 200

IL-6 (pg ml–1)

IL-8 (pg ml–1)

300
200
flow cytometry15. We obtained longitudinal serial time-point samples 150
200
from a subset of these 98 study participants (n = 48; information in 100
100
100 50
Extended Data Table 1). To compare the immune phenotypes between
0 0 0
sexes, two sets of data analyses were performed in parallel—baseline

F_ W
W

t
Pt

F_ W
W

t
Pt

F_ W
W

t
Pt
and longitudinal, as described below. As a control group, healthcare

_P

_P

_P
C
C

C
C

C
C
F_

F_

F_
M

M
_H

_H

_H

H
workers (HCWs) from Yale-New Haven Hospital were enrolled who were

M
P = 0.018 P = 0.0029 P = 0.0015
uninfected with COVID-19. Demographics and background information 8,000 400 40,000
P = 0.0004 P = 0.027 P= 0.0003
for the HCW group and the demographics of HCWs for cytokine assays

CXCL10 (pg ml–1)

CCL5 (pg ml–1)

ml–1)
6,000 300 30,000
and flow cytometry assays for the primary analyses are in Extended

CCL8 (pg
4,000 200 20,000
Data Table 1. Demographic data, time-point information of the samples
defined by the days from the symptom onset (DFSO) in each patient, 2,000 100 10,000

treatment information, and raw data used to generate figures and tables 0 0 0
is in Supplementary Table 1.

F_ W
W

t
Pt

W
W

t
Pt

W
W

t
Pt
_P

_P

_P
C
C

C
C

C
C
F_

F_

F_
M

M
_H

_H

_H

H
F_

F_
M

Baseline analysis
The baseline analysis was performed on samples from the first time
point of patients who met the following criteria: not in intensive care
unit (ICU), had not received tocilizumab, and had not received high
doses of corticosteroids (prednisone equivalent of more than 40 mg)
before the first sample collection date. This patient group, cohort
A, consisted of 39 patients (17 male and 22 female) (Extended Data
Tables 1, 2). Intersex and transgender individuals were not repre-
sented in this study. Figures 1–4 represent analyses of baseline raw
values obtained from patients in cohort A. In cohort A patients, male
and female patients were matched in terms of age, body mass index
(BMI), and DFSO at the first time point sample collection (Extended
Data Fig. 1a). However, there were significant differences in age and
BMI between HCW controls and patients (patients had higher age
and BMI values) (Extended Data Table 1), and therefore an age- and
BMI-adjusted difference-in-differences analysis was also performed
in parallel (Extended Data Table 3).
M
Fig. 1 | Comparison of viral RNA concentrations, titres of anti-SARS-CoV-2
antibodies, and plasma cytokines and chemokine levels at the first
sampling of cohort A patients. a, Comparison of viral RNA measured from
nasopharyngeal (Np) swab and saliva. n = 14 for male and female patients
(M_Pt and F_Pt, respectively) for nasopharyngeal samples, and n = 9 and 12,
respectively, for saliva samples. Dotted lines indicate the detection limit of the
assay (5,610 copies ml−1), and negatively tested data are shown on the x axis. ND,
not detected. b, Titres of specific IgG and IgM antibodies against SARS-CoV-2
S1 protein were measured. n = 13, 74, 15 and 20 for IgG, and n = 3, 18, 15 and
20 for IgM, for male HCW (M_HCW), female HCW (F_HCW), M_Pt and F_Pt,
respectively. The cut-off values for positivity are shown by the dotted lines.
c, Comparison of the plasma levels of representative innate immune cytokines
and chemokines. n = 15, 28, 16 and 19 for M_HCW, F_HCW, M_Pt and F_Pt,
respectively. P values were determined by unpaired two-tailed t-test (a) or
one-way analysis of variance (ANOVA) with Bonferroni multiple comparison
test (b, c). All P values < 0.10 are shown. Data are mean ± s.e.m. The results of all
the cytokines or chemokines measured can be found in Extended Data Fig. 1b.

Longitudinal analysis
As parallel secondary analyses, we performed longitudinal analy-
sis on a total patient cohort (cohort B) to evaluate the difference in
immune response over the course of the disease between male and
female patients. Cohort B included all patient samples from cohort
A (including several time-point samples from the cohort A patients)
as well as an additional 59 patients who did not meet the inclusion
criteria for cohort A. Because cohort B included more severely affected
patients in ICU, the average clinical scores were higher in cohort B than
in cohort A (mean ± s.d.: 1.3 ± 0.5 (female) and 1.4 ± 0.5 (male) for cohort
A, and 2.5 ± 1.5 (female) and 2.7 ± 1.3 (male) for cohort B) (Extended
Data Table 1). This analysis included several time-point samples from
98 participants in total. Data from cohort B were analysed for sex dif-
ferences in immune responses among patients using longitudinal
analysis, controlling for potential confounding by age, BMI, receipt
of immunomodulatory treatment (tocilizumab or corticosteroids),
DFSO and ICU status. Second, we conducted a longitudinal analysis that
compared male and female patients with COVID-19 to male and female
HCWs, controlling for age and BMI. Adjusted least square means differ-
ence over time in immune responses between male and female patients
with COVID-19 (Extended Data Table 4) and adjusted least square means
difference over time in immune responses between male and female
patients with COVID-19 and male and female HCWs (Extended Data
Table 5) were calculated.
Sex differences in cytokines and chemokines
We first compared the concentrations of viral RNA of male and female
patients. For both cohorts A and B, there was no difference by sex in
terms of the viral RNA concentrations in nasopharyngeal swab and
saliva (Fig. 1a, Extended Data Tables 3, 4).
Anti-SARS-CoV-2 S1-specific IgG and IgM (anti-S1-IgG and -IgM)
antibodies were comparable in infected male and female in cohort A
(Fig. 1b) and in cohort B (Extended Data Tables 4, 5). Thus, at baseline
and during the course of the disease, there were no clear differences
in the amount of IgG or IgM generated against the S1 protein between
male and female patients.
Next, we analysed the levels of 71 cytokines and chemokines in the
plasma. Levels of many pro-inflammatory cytokines, chemokines and
growth factors, including IL-1β, IL-6, IL-8, TNF, CCL2, CXCL10 and G-CSF,
are increased in the plasma of patients with COVID-1916. In line with
previous reports, levels of inflammatory cytokine or chemokine were

316 | Nature | Vol 588 | 10 December 2020

 a P = 0.021 b HCW Patients factors were increased in patients compared to HCWs in both sexes,
20 7.82 intMono 6.43 39.6 9.59
and the levels between sexes were comparable (Fig. 1c, Extended Data
B cells (% of live)

(CD14+
CD16+)
15
ncMono Fig. 1b, Extended Data Table 3). However, levels of IL-8 and IL-18 were
10 (CD14dim Male
CD16+) significantly higher in male patients than in female patients in cohort

CD16-Brilliant Violet 786

5 cMono
(CD14+ A (Fig. 1c). In age- and BMI-adjusted analyses of cohort A, we found
61.0 CD16–)
32.5
0 that although IL-8 and IL-18 were no longer significantly higher among
F_ W
W

t
Pt

male patients than in to female patients, IL-8 and CXCL10 were sig-
_P
C
C

F_

105
M
_H

2.43 5.49 19.6 19.1

M

P < 0.0001
104
nificantly increased in male patients compared to male HCWs than in
100 P = 0.0003
female patients compared to female HCWs (difference-in-differences,
T cells (% of live)

80 Female
103
60
Extended Data Table 3). In adjusted analyses of cohort B, although we
40
0
59.7 25.8
did not see significant sex differences in the levels of IL-8 and IL-18, we
-103
20
-103 0 103 104 105 found significantly higher levels of CCL5 in male patients than in female
0
CD14-PE-Cy7
patients over the course of the disease (Extended Data Table 4) and
F_ W
W

t
Pt

significantly increased levels of CCL5 in male patients compared to male

_P
C
C

F_
M
_H

H
M

c HCWs than in female patients compared to female HCWs (Extended

P = 0.0015 P < 0.0001
P = 0.047 P = 0.0081
50 25 20 P = 0.039 20 Data Table 5, difference-in-differences). These data indicated that,
Total Mono (% of live)

P= 0.019
intMono (% of live)

ncMono (% of live)
cMono (% of live)

40 20 15 15 although levels of most of the innate inflammatory cytokines and

30 15
10 10 chemokines were comparable, there were a few factors that are more
20 10
10 5 5 5 robustly increased at the baseline (IL-8 and IL-18) and during the course
0 0 0 0 of the disease (CCL5) in male patients than in female patients.
F_ W
W

t
Pt

F_ W
W

t
Pt

F_ W
W

t
Pt

F_ W
W

t
Pt
_P

_P

_P

_P
C
C

C
C

C
C
F_

F_

F_

F_
M

M
_H

_H

_H

_H

H
M

d e
R = –0.210, P = 0.387 (F)
F Monocyte differences by sex
50 25 M (low-int)
≥90
80 40 20
R = 0.635, P = 0.011 (M) M (high) Next, we examined the immune cell phenotype by flow cytometry. Freshly
BMI (kg m–2)

70
Age (years)

isolated PBMCs were stained with specific antibodies to identify T cells,

DFSO

60 30 15
50 15
40
30
20 10 B cells, natural killer T cells, natural killer cells, monocytes, macrophages
20 10 5
and dendritic cells to investigate the composition of PBMCs (Extended
ncMono (% of live)

10 10
0 0 0
Data Fig. 2b). Consistent with a previous report on a decrease in T cells
t

h
-in

-in

-in
ig

ig

ig

in patients16, in cohort A, the proportion of T cells in the live cells was sig-
H

H
w

w
Lo

Lo

Lo

80 P = 0.030 250 P = 0.050 8,000 P = 0.013 nificantly lower in patients, whereas the proportion of B cells was higher
in both male and female patients than in HCWs (Fig. 2a, Extended Data
ml–1)

CCL5 (pg ml–1)

T cells (% of live)

60 200 6,000 0
150 Table 3). There was no difference in the numbers of B cells across all groups,
IL-18 (pg

40 4,000
100
20 2,000
but the numbers of T cells were lower in patients of both sexes (data not
50
0 0 0 2.8 3.2 3.6
shown). By contrast, in cohort B, we found that male patients had signifi-
log10[CCL5 (pg ml–1)] cantly lower numbers of T cells, both total counts and as a proportion of
h

h
t

t
ig

ig

ig
-in

-in

-in
H

H
w

live cells, over the course of the disease than female patients (Extended
Lo

Lo

Lo

Fig. 2 | Differences in composition of PBMCs between male and female Data Table 4). Next, we found higher populations of monocytes in both
patients in cohort A at the first sampling. a, Comparison on the proportion of sexes in cohort A (Fig. 2b, c, Extended Data Fig. 2b) compared to HCWs.
B cells (top) and T cells (bottom) in live PBMCs. n = 6, 42, 16 and 21 for M_HCW, F_ Although CD14+CD16− classical monocytes were comparable across all
HCW, M_Pt and F_Pt, respectively. b, Representative 2D plots for CD14 and CD16 groups, levels of CD14+CD16+ intermediate monocytes were increased
in monocytes gate (live/singlets/CD19−CD3−/CD56−CD66b−). Numbers in red in patients compared with HCWs, and this increase was more robust in
indicate the percentages of each population in the parent monocyte gate. female patients (Fig. 2b, c). By contrast, male patients had higher levels of
c, Comparison between percentages of total monocytes, classical monocytes CD14loCD16+ non-classical monocytes than controls and female patients
(cMono), intermediate monocytes (intMono) and non-classical monocytes (Fig. 2b, c). These differences were observed in age- and BMI-adjusted
(ncMono) in the live PBMCs. n = 6, 42, 16 and 21 for M_HCW, F_HCW, M_Pt and analyses, too, but were not significant (Extended Data Table 3).
F_Pt, respectively. d, Comparison of age, BMI, DFSO, T cells (percentage of live
We then divided the 17 cohort A male patients into two groups, namely,
PBMCs) and plasma IL-18 and CCL5 levels between male patients who had high
a ‘high’ group who had high percentages of non-classical monocytes
non-classical monocytes and low-intermediate non-classical monocytes. n = 13
(upper quartile 4 patients, all had more than 5% of non-classical mono-
and 4 for ‘low-int’ and ‘high’ group, respectively, for age, BMI and DFSO. n = 12
and 4 for ‘low-int’ and ‘high’ group, respectively, for T cells and IL-18 or CCL5
cytes) and a ‘low-intermediate’ group (others, 13 patients). We compared
levels. e, Correlation between plasma CCL5 levels and non-classical monocytes age, BMI, DFSO, T cells, and plasma levels of IL-18 and CCL5. Although
(percentage of live cells). Pearson correlation coefficients (R) and P values we found no differences in age, BMI or DFSO (Fig. 2d), we noted that the
for each sex are shown. Lines represent linear regression lines and shading group with high levels of non-classical monocytes had significantly lower
represents 95% confidence intervals for each sex. ncMono-high male patients levels of T cells and higher levels of CCL5 in plasma (Fig. 2d). In addition,
(n = 4) are shown with orange open squares, and ncMono-low-int male patients we found a significant correlation between CCL5 levels and abundance
(n = 11) are shown with orange closed squares. n = 19 for female patients (purple in non-classical monocytes only in male patients (Fig. 2e). These findings
circles). Data are mean ± s.e.m. in a, c and d. P values were determined by one-way suggest that the progression from classical to non-classical monocytes
ANOVA with Bonferroni multiple comparison test (a, c) or unpaired two-tailed may be arrested at the intermediate stage in female patients, and that
t-test (d). All P values < 0.10 are shown. increased innate inflammatory cytokines and chemokines are associated
with more robust activation of innate immune cells at the baseline as well
generally higher in patients than in controls (Fig. 1c, Extended Data as more robust longitudinal T cell decrease in male patients.
Figs. 1b, 2a, Extended Data Table 3). The levels of type-I, -II or -III inter-
feron (IFN) were comparable between the sexes in cohort A (Extended
Data Fig. 1b, Extended Data Table 3). However, we found higher levels Higher T cell activation in female
of IFNα2 in female patients than in male patients in cohort B (Extended We further examined the T cell phenotype in patients with COVID-19.
Data Table 4). The levels of many cytokines, chemokines and growth The composition of overall CD4-positive and CD8-positive cells among

Nature | Vol 588 | 10 December 2020 | 317

Article
a CD4 T cells CD8 T cells
b CD4 CD8
100 80 HCW Patient HCW Patient

Percentage of CD3
80 60
60 0.85 0.83 0.42 1.41
40 Male
40
20

HLA-DR-FITC
20
0 0
105

F_ W
W

t
Pt

F_ W
W

t
Pt
_P

_P
C
C

C
C
F_

F_
M

M
_H

_H

H
104

M
0.78 13.7 0.90 27.9
c CD38+HLA-DR+ CD4 CD38+HLA-DR+ CD8 103 Female
P < 0.0001
0
10 P = 0.0002 8 P = 0.036
Percentage of CD3

8 -104 0 104 105

6
6 CD38-Brilliant Violet 711
4
4
2
d CD4 CD8
2
HCW Patient HCW Patient
0 0
0.48 0.55 0.75 2.27
F_ W
W

t
Pt

F_ W
W

t
Pt
_P

_P
C
C

C
C
F_

F_
M

M
_H

_H

H
Male
M

e PD-1+TIM-3+ CD4 PD-1+TIM-3+ CD8

P = 0.016 P < 0.0001

PD-1-PE
6 8
Percentage of CD3

105
6 1.43 10.8 2.67 25.9
4 104
4 Female
103
2
2 0
–103
0 0
–103 0 103 104 105
W
W

t
Pt

F_ W
W

t
Pt
_P

_P
C
C

C
C
F_

F_
M

M
_H

_H

TIM-3-Alexa Fluor 647

F_
M

Fig. 3 | Sex difference in T cell phenotype at the first sampling of cohort
A patients. a, Percentages of CD4 and CD8 in the CD3-positive cells.
b, Representative 2D plots for CD38 and HLA-DR in the CD4 and CD8 T cells.
Numbers in red indicate the percentages of CD38+HLA-DR+ populations in the
parent gate (live/singlets/CD3+/CD4+ or CD8+). c, Percentages of CD38+HLA-DR+
CD4 or CD8 cells in CD3-positive cells. d, Representative 2D plots for PD-1 and
TIM-3 in the CD4 and CD8 T cells. Numbers in red indicate the percentages of
PD-1+TIM-3+ populations in the parent gate (live/singlets/CD3+/CD4+ or CD8+/
CD45RA−). e, Percentages of PD-1+TIM-3+ CD4 or CD8 cells in CD3-positive cells.
n = 6, 45, 16 and 22 for M_HCW, F_HCW, M_Pt and F_Pt, respectively. P values
were determined by one-way ANOVA with Bonferroni multiple comparison
test. Data are mean ± s.e.m. All P values < 0.10 are shown.

T cells were similar between all groups in cohort A (Fig. 3a, Extended Data A, 6 patients of each sex deteriorated during the course of the disease
Fig. 2c, Extended Data Table 3). Detailed phenotyping of T cells for naive (35.3% and 27.3%, respectively), and the intervals between the dates
T cells, central or effector memory T (TCM/TEM) cells, follicular helper at which the patients reached Cmax (DFSO at Cmax) and the first sample
T (TFH) cells, regulatory T (Treg) cells revealed no remarkable differences collection (DFSO at C1) were not significantly different between dete-
in the frequency of these subsets between sexes (Extended Data Fig. 2c). riorated male and female patients (mean ± s.d. = 3.7 ± 4.1 and 4.2 ± 2.7,
However, we observed higher levels of CD38 and HLA-DR-positive acti- respectively; P = 0.81 by unpaired two-tailed t-test).
vated T cells in female patients than in male patients (Fig. 3b, c). In paral- We first examined age, BMI, viral loads and titres of anti-S1-IgG
lel, PD-1- and TIM-3-positive terminally differentiated T cells were more antibodies between the stabilized and deteriorated groups in a
prevalent among female patients than male patients (Fig. 3d, e). These sex-aggregated manner. We found that the deteriorated group had on
findings were seen in both CD4 and CD8 T cells, but the differences average a higher BMI than the stabilized group. Although the age was not
were more robust in CD8 T cells (Fig. 3c, e, Extended Data Table 3). We statistically different, the stabilized group spanned a larger age range
also stained for intracellular cytokines such as IFNγ, granzyme B (GzB), than the deteriorated group, who were generally of a more advanced age.
TNF, IL-6 and IL-2 in CD8 T cells, and IFNγ, TNF, IL-17A, IL-6 and IL-2 in The viral load and antibody titres were comparable (Fig. 4a). Next, we
CD4 T cells. Levels of these cytokines were higher in patients than in examined these factors in a sex-disaggregated manner, and found that
controls, and were generally comparable between sexes in the patients the deteriorated male (M_deteriorated) group was on average signifi-
(Extended Data Fig. 2d). Analyses of T cell phenotypes in cohort B did cantly older than the stabilized male (M_stabilized) group, whereas the
not reveal any significant differences between sexes (Extended Data two female groups (F_deteriorated and F_stabilized) were comparable
Tables 4 and 5). Therefore, female patients with COVID-19 had more in age (Fig. 4b). In addition, BMI was higher for the M_deteriorated than
abundant activated and terminally differentiated T cell populations the M_stabilized group, whereas there was no difference in BMI between
than male patients at baseline in unadjusted analyses. the F_deteriorated and F_stabilized groups (Fig. 4b). By contrast, the
F_deteriorated group had higher viral load in saliva than the F_stabilized
group, whereas there was no difference in the male groups (Fig. 4b).
Sex-dependent immunity and disease course The levels of antibodies were comparable between the deteriorated
We investigated whether certain immune phenotypes were correlated and stabilized groups both in male and female, but stabilized female
with disease trajectory, and whether these phenotypes and factors dif- tended to have higher antibody levels (Fig. 4b).
fered between the sexes. To this end, we evaluated the disease course We further investigated whether the key factors identified in the
of patients in cohort A. The clinical scores at the first sample collection previous analyses correlated with disease progression in male and
(C1) were 1 or 2 for all of the patients in cohort A. The patients were cat- female patients. We observed that regardless of sex, some chemokines
egorized into a ‘deteriorated’ group if the patients marked a score of 3 or and growth factors, such as CXCL10 (also known as IP-10) and M-CSF,
higher after the first sample collection date as their maximum clinical were increased in patients that went on to develop worse disease. How-
scores during admission (Cmax). By contrast, if the patients maintained ever, there were some innate immune factors, such as CCL5, TNFSF10
the score of 1 or 2, they were categorized as ‘stabilized’ (Extended Data (also known as TRAIL) and IL-15, that were specifically increased only
Table 2). Both in male (n = 17) and female (n = 22) patients from cohort in female patients that subsequently progressed to worse disease,

318 | Nature | Vol 588 | 10 December 2020

 a Np swab Saliva Anti-S1-IgG Stabilized
P = 0.051 P = 0.012 Deteriorated
50 10 10 2.5

log10[SARS-CoV-2

log10[SARS-CoV-2
≥90
80 40 8 8 2.0

(copies ml–1)]

(copies ml–1)]
BMI (kg m–2)
70

Age (years)

A450 nm
60 30 6 6 1.5
50
40 20 4 4 1.0
30
20 10 2 2 0.5
10
0 0 0 0 0.0

b M_stabilized M_deteriorated F_stabilized F_deteriorated

Np swab Saliva Anti-S1-IgG
P = 0.0086 P = 0.024 P = 0.024 P = 0.074
50 10 10 2.5

log10[SARS-CoV-2

log10[SARS-CoV-2
≥90

(copies ml–1)]

(copies ml–1)]
BMI (kg m–2)
80 40 8 8 2.0
Age (years)

70

A450 nm
60 30 6 6 1.5
50
40 20 4 4 1.0
30
20 10 2 2 0.5
10
0 0 0 0 0.0

c P = 0.046 P = 0.080 P = 0.014 P = 0.0023 P = 0.011 P = 0.037

40,000 1,500 250 8,000 150

TNFSF10 (pg ml–1)

M-CSF (pg ml–1)
CXCL10 (pg ml–1)

CCL5 (pg ml–1)

200

IL-15 (pg ml–1)

30,000 6,000
1,000 100
150
20,000 4,000
100
500 50
10,000 50 2,000

0 0 0 0 0

d
P = 0.0026 P = 0.025 P = 0.094
10 8 6 8 20
CD38+HLA-DR+ CD4

CD38+HLA-DR+ CD8

PD-1+TIM-3+ CD4

PD-1+TIM-3+ CD8
8 6 6 15
(% of CD3)

(% of CD3)

(% of CD3)

(% of CD3)

(% of CD3)
IFNγ+ CD8
4
6
4 4 10
4
2
2 2 2 5

0 0 0 0 0
CD38 HLA DR CD8

CD38 HLA DR CD8

e f
PD 1TIM3 CD8

PD 1TIM3 CD8

R = 0.103, P = 0.647 (F) R = 0.245, P = 0.272 (F)

Deterioration

Deterioration
Anti-S1-IgG

Anti-S1-IgG

R = –0.523, P = 0.038 (M) R = –0.529, P = 0.035 (M)

Cmax – C1
IFNγ CD8
TNFSF10

Cmax – C1
IFNγ CD8
TNFSF10
CXCL10

CXCL10
Np load

Np load

Sex 15
MCSF

CD38+HLA-DR+CD8 (% of CD3)
MCSF
CCL5

CCL5
IL-15

F
IL-15
BMI
Age

BMI
Age

M
1 1 6

IFNγ+ CD8 (% of CD3)

Age Age
BMI 0.8 BMI 0.8
Np load 0.6 Np load 0.6 10
Anti-S1-IgG 0.4 Anti-S1-IgG
CXCL10 0.4 4
0.2 CXCL10
MCSF MCSF 0.2
TNFSF10 0 TNFSF10 0
CCL5 –0.2 CCL5 –0.2 2 5
Female IL-15 Male IL-15
PD1 TIM3 CD8 –0.4 –0.4
PD1 TIM3 CD8
CD38 HLA DR CD8 –0.6 CD38 HLA DR CD8 –0.6
IFNγ CD8 –0.8 IFNγ CD8 –0.8 0
Deterioration –1 Deterioration 0
Cmax – C1 –1
Cmax – C1
40 60 80 ≥90 40 60 80 ≥90
Age Age

Fig. 4 | Differential immune phenotypes at the first sampling and disease are shown. n = 10, 6, 16 and 6 for M_stabilized, M_deteriorated, F_stabilized and
progression between sexes in cohort A patients. a, b, Sex-aggregated (a) and F_deteriorated group, respectively. e, Pearson correlation heat maps of the
sex-disaggregated (b) comparison of age, BMI, RNA concentration in indicated parameters are shown for each sex. For viral RNA concentrations
nasopharyngeal swab and saliva, and anti-S1-IgG antibodies between the and cytokine or chemokine levels, log-transformed values were used for the
stabilized and deteriorated group. n = 11, 6, 16 and 6 for age and BMI, n = 9, 5, calculation of the correlations. The size and colour of the circles indicate
9 and 5 for nasopharyngeal swab, n = 6, 3, 8 and 4 for saliva, and n = 10, 5, 14 and the correlation coefficient (R), and only statistically significant correlations
6 for anti-S1-IgG antibodies, for M_stabilized, M_deteriorated, F_stabilized and (P < 0.05) are shown. Clinical deterioration from the first time point was scored
F_deteriorated group, respectively. Dotted lines in the viral concentration and by Cmax − C1. n = 17 and 22 for male and female, respectively. f, Correlation
anti-S1-IgG panels indicate the detection limit and cut-off value for positivity, between age and CD38+HLA-DR+ CD8 T cells (left) and IFNγ+CD8 T cells (right).
respectively. c, Cytokine or chemokine comparison between stabilized and Pearson correlation coefficient (R) and P values for each correlation and sex are
deteriorated groups. n = 10, 6, 14 and 5 for the M_stabilized, M_deteriorated, shown. Lines represent linear regression lines and shading represents 95%
F_stabilized and F_deteriorated groups, respectively. d, Comparisons in the confidence intervals for each sex. P values were determined by unpaired
proportions of activated (CD38+HLA-DR+) and terminally differentiated two-tailed t-test in a–d. Data are mean ± s.e.m. All P values < 0.10 are shown.
(PD-1+TIM-3+) CD4 or CD8 T cells, and IFNγ+CD8 T cells in CD3-positive T cells

but this difference was not observed in male patients (Fig. 4c). In the We finally examined the correlations between age, BMI, viral loads,
age- and DFSO-adjusted analysis of cohort A, we also found that CCL5 anti-S1 antibodies, cytokines or chemokines, activated or terminally
was only increased in female patients that progressed to worse disease differentiated or IFNγ-producing CD8 T cells, and clinical disease
compared to the stabilized patients, but no such correlation was found course (‘Cmax − C1’ was used for the deterioration score). The corre-
in male patients (Extended Data Table 6). lation matrix showed that in female patients, higher levels of innate
T cell phenotypes in these groups showed that male patients immunity cytokines, such as TNFSF10 and IL-15, were positively cor-
whose disease worsened had a significantly lower proportion of acti- related with disease progression, whereas there was no association
vated T cells (CD38+HLA-DR+) and terminally differentiated T cells between CD8 T cell status and deterioration (Fig. 4e, results of age-
(PD-1+TIM-3+) and tendencies for fewer IFNγ+ CD8 T cells at the first and DFSO-adjusted analysis in Extended Data Table 6). In particular,
sample collection, compared with their counterpart male who did not CXCL10, M-CSF and IL-15 were positively correlated with IFNγ+CD8
progressed to worse disease (Fig. 4d). However, in female patients, T cells in female patients (Fig. 4d).
the deteriorated group had similar levels of these types of CD8 T cells By contrast, in male patients, progressive disease was associated
compared with the stabilized group (Fig. 4d). with higher age, higher BMI, and poor CD8 T cell activation (Fig. 4e).

Nature | Vol 588 | 10 December 2020 | 319

Article
Poor CD8 T cell activation and poor production of IFNγ by CD8 T cells
were significantly correlated with patients’ age, whereas these cor-
relations were not seen in female patients (Fig. 4e, f ). These differ-
ences seemed to highlight the differences between the sexes in the
immune responses against SARS-CoV-2 as well as the difference of
the potential prognostic or predictive factors for clinical deteriora-
tion of COVID-19.
Discussion
Our results revealed key differences in immune responses during the
disease course of SARS-CoV-2 infection in male and female patients.
First, we found that the levels of several important pro-inflammatory
innate immunity chemokines and cytokines such as IL-8, IL-18 (at base-
line) and CCL5 (longitudinal analysis) were higher in male patients,
which correlated with higher non-classical monocytes (at baseline).
Second, we observed a more robust T cell response among female
patients than male patients at baseline. In particular, activated CD8
T cells were significantly increased only in female patients but not
in male patients compared with healthy volunteers. Analysis of their
clinical trajectory showed that, although poor T cell responses were
associated with future progression of disease in male patients, higher
levels of innate immune cytokines were associated with worsening of
COVID-19 disease in female patients. Notably, the T cell response was
significantly and negatively correlated with patients’ age in male, but
not female, patients. These data indicate key differences in the base-
line immune capabilities in male and female patients during the early
phase of SARS-CoV-2 infection, and suggest a potential immunologi-
cal underpinning of the distinct mechanisms of disease progression
between sexes. These analyses also provide a potential basis for taking
sex-dependent approaches to prognosis, prevention, care, and therapy
for patient with COVID-19.
Although our study provides a strong basis for further investigation
into how COVID-19 disease dynamics may differ between male and
female patients, it is important to note that there are some limitations
to the analyses presented in this Article. First, we acknowledge that
the healthy HCWs used as the control population were not matched to
patients on the basis of age, BMI or underlying risk factors. To account
for this, we performed adjusted analyses for the baseline and longi-
tudinal comparisons between patients (cohort A and the full patient
population, cohort B) and HCWs, controlling for age and BMI. However,
we cannot rule out residual confounding due to underlying risk factors
not available for the HCW controls.
Collectively, these data suggest that vaccines and therapies to
increase T cell immune responses to SARS-CoV-2 might be warranted
for male patients, whereas female patients might benefit from thera-
pies that dampen innate immune activation early during disease. The
immune landscape in patients with COVID-19 is considerably different
between the sexes, and these differences may underlie heightened
disease vulnerability in men.
Online content
Any methods, additional references, Nature Research reporting summa-
ries, source data, extended data, supplementary information, acknowl-
edgements, peer review information; details of author contributions
320 | Nature | Vol 588 | 10 December 2020
and competing interests; and statements of data and code availability
are available at https://doi.org/10.1038/s41586-020-2700-3.
1. Chen, N. et al. Epidemiological and clinical characteristics of 99 cases of 2019 novel
coronavirus pneumonia in Wuhan, China: a descriptive study. Lancet 395, 507–513 (2020).
2. Li, Q. et al. Early transmission dynamics in Wuhan, China, of novel coronavirus-infected
pneumonia. N. Engl. J. Med. 382, 1199–1207 (2020).
3. Yang, X. et al. Clinical course and outcomes of critically ill patients with SARS-CoV-2
pneumonia in Wuhan, China: a single-centered, retrospective, observational study.
Lancet Respir. Med. 8, 475–481 (2020).
4. Meng, Y. et al. Sex-specific clinical characteristics and prognosis of coronavirus
disease-19 infection in Wuhan, China: a retrospective study of 168 severe patients.
PLoS Pathog. 16, e1008520 (2020).
5. Gebhard, C., Regitz-Zagrosek, V., Neuhauser, H. K., Morgan, R. & Klein, S. L. Impact of sex
and gender on COVID-19 outcomes in Europe. Biol. Sex Differ. 11, 29 (2020).
6. Zhu, N. et al. A novel coronavirus from patients with pneumonia in China, 2019. N. Engl. J.
Med. 382, 727–733 (2020).
7. WHO. WHO Director-General’s Opening Remarks at the Media Briefing on COVID-19 - 11
March 2020 https://www.who.int/dg/speeches/detail/who-director-general-s-opening-
remarks-at-the-media-briefing-on-covid-19---11-march-2020 (2020).
8. Williamson, E. J. et al. Factors associated with COVID-19-related death using OpenSAFELY.
Nature 584, 430–436 (2020).
9. Klein, S. L. & Flanagan, K. L. Sex differences in immune responses. Nat. Rev. Immunol. 16,
626–638 (2016).
10. Fischer, J., Jung, N., Robinson, N. & Lehmann, C. Sex differences in immune responses to
infectious diseases. Infection 43, 399–403 (2015).
11. Guerra-Silveira, F. & Abad-Franch, F. Sex bias in infectious disease epidemiology: patterns
and processes. PLoS One 8, e62390 (2013).
12. Moore, A. L. et al. Virologic, immunologic, and clinical response to highly active
antiretroviral therapy: the gender issue revisited. J. Acquir. Immune Defic. Syndr. 32,
452–461 (2003).
13. Collazos, J., Asensi, V. & Cartón, J. A. Sex differences in the clinical, immunological and
virological parameters of HIV-infected patients treated with HAART. AIDS 21, 835–843
(2007).
14. Fink, A. L., Engle, K., Ursin, R. L., Tang, W. Y. & Klein, S. L. Biological sex affects vaccine
efficacy and protection against influenza in mice. Proc. Natl Acad. Sci. USA 115,
12477–12482 (2018).
15. Lucas, C. et al. Longitudinal analyses reveal immunological misfiring in severe COVID-19.
Nature 584, 463–469 (2020).
16. Huang, C. et al. Clinical features of patients infected with 2019 novel coronavirus in
Wuhan, China. Lancet 395, 497–506 (2020).
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in
published maps and institutional affiliations.
© The Author(s), under exclusive licence to Springer Nature Limited 2020
Yale IMPACT Research Team*
Kelly Anastasio14, Michael H. Askenase15, Maria Batsu16, Hannah Beatty16, Santos Bermejo16,
Sean Bickerton17, Kristina Brower2, Molly L. Bucklin1, Staci Cahill14, Melissa Campbell3,
Yiyun Cao1, Edward Courchaine17, Rupak Datta3, Giuseppe DeIuliis8, Bertie Geng16,
Laura Glick16, Ryan Handoko16, Chaney Kalinich2, William Khoury-Hanold1, Daniel Kim1,
Lynda Knaggs16, Maxine Kuang14, Eriko Kudo1, Joseph Lim18, Melissa Linehan1, Alice
Lu-Culligan1, Amyn A. Malik11, Anjelica Martin1, Irene Matos16, David McDonald16, Maksym
Minasyan16, Subhasis Mohanty3, M. Catherine Muenker2, Nida Naushad16, Allison Nelson16,
Jessica Nouws16, Marcella Nunez-Smith19, Abeer Obaid16, Isabel Ott2, Hong-Jai Park16,
Xiaohua Peng16, Mary Petrone2, Sarah Prophet20, Harold Rahming16, Tyler Rice1, Kadi-Ann
Rose16, Lorenzo Sewanan16, Lokesh Sharma8, Denise Shepard16, Erin Silva16, Michael
Simonov16, Mikhail Smolgovsky16, Eric Song1, Nicole Sonnert1, Yvette Strong16, Codruta
Todeasa16, Jordan Valdez16, Sofia Velazquez15, Pavithra Vijayakumar16, Haowei Wang3,
Annie Watkins2, Elizabeth B. White2 & Yexin Yang1
14
Yale Center for Clinical Investigation, Yale University School of Medicine, New Haven,
CT, USA. 15Department of Neurology, Yale University School of Medicine, New Haven, CT,
USA. 16Yale School of Medicine, New Haven, CT, USA. 17Department of Biochemistry and
of Molecular Biology, Yale University School of Medicine, New Haven, CT, USA. 18Yale
Viral Hepatitis Program, Yale University School of Medicine, New Haven, CT, USA. 19Equity
Research and Innovation Center, Yale University, New Haven, CT, USA. 20Department of
Molecular, Cellular and Developmental Biology, Yale University School of Medicine,
New Haven, CT, USA.
Methods
No statistical methods were used to predetermine sample size. The
experiments were not randomized. Investigators were blinded during
experiments in terms of the sex or other clinical background informa-
tion, with the sample labels having de-identified patient IDs that did
not contain any of this information.
Ethics statement
This study was approved by Yale Human Research Protection Program
Institutional Review Boards (FWA00002571, Protocol ID. 2000027690).
Informed consent was obtained from all enrolled patients and health-
care workers.
Patients and HCWs
Adult patients (≥18 years old) admitted to Yale-New Haven Hospital
between 18 March and 9 May 2020, positive for SARS-CoV-2 by RT–PCR
from nasopharyngeal and/or oropharyngeal swabs, and able to provide
informed consent (surrogate consent accepted) were eligible for the
Yale IMPACT Biorepository study, and 198 patients were enrolled in this
period. All patients necessitated hospitalization for their symptoms and
had an WHO score17 of at least 3 at admission (denoting hospitalized,
mild disease). At the initial screening, clinical PCR tests were performed
in CLIA-certified laboratory and only the PCR-positive patients were
enrolled. Only after the confirmation of PCR-positivity, the patients
were enrolled and the first time point samples for this study were col-
lected for each patient. The first time point samples were collected at
11.4 ± 8.1, 10.2 ± 6.3, 11.7 ± 7.2 and 12.1 ± 7.3 (mean ± s.d.) DFSO in cohort A
female, cohort A male, cohort B female and cohort B male, respectively
(Extended Data Fig. 1a, right panel for cohort A, Extended Data Table 1).
Among these patients, we could obtain whole blood for flow cytom-
etry analysis using fresh PBMCs, plasma for cytokine or chemokine
measurements, anti-S1 antibody measurements and nasopharyngeal
swab and saliva from total of 98 individuals for the present study. For
longitudinal analyses, biospecimens (blood, nasopharyngeal swabs,
saliva, urine, and/or stool) were collected at study enrolment (baseline)
and on average every 3 to 7 days while in the hospital in 48 of these 98
patients.
The patients were assessed with a locally developed clinical scor-
ing system for disease severity; (1): admitted and observed without
supplementary oxygen; (2) required ≤ 3 l supplementary oxygen via
nasal canal to maintain SpO2 > 92%; (3) received tocilizumab, which per
hospital treatment protocol required that the patient to require >3 l
supplementary oxygen to maintain SpO2 > 92%, or, required >2 l sup-
plementary oxygen to maintain SpO2 > 92% and had a high-sensitivity
C-reactive protein (CRP) > 70; (4) the patient required ICU-level care;
(5) the patient required intubation and mechanical ventilation. In rela-
tion to the WHO scoring17, our clinical scores 1, 2/3, 4 and 5 largely corre-
spond to WHO scores 3, 4, 5 and 6/7, respectively. Detailed demographic
information for the entire cohort (98 cohort B patients, and several
time-point samples from 54 patients among them) and of cohort A
(39 patients) are shown in Extended Data Tables 1–3. For the patients
who are 90 years old or older, their ages were protected health informa-
tion, and ‘90’ was put as the surrogate value for the analyses. Among
198 total patients enrolled in IMPACT study in this period, we obtained
whole blood, nasopharyngeal swabs or saliva samples from 98 patients
for the present study. Individuals with active chemotherapy against
cancers, pregnant patients, patients with background haematological
abnormalities, patients with autoimmune diseases and patients with a
history of organ transplantation and on immunosuppressive agents,
were excluded from this study.
As a control group, COVID-19-uninfected HCWs from Yale-New Haven
Hospital were enrolled. HCWs were tested every 2 weeks for PCR and
serology. For the control group, the PBMCs and plasma analysis were
done when both tests were negative. That is, if either or both of these tests
were positive, these samples were excluded from the analyses. In some
HCWs, samples were collected for the assays at up to two time points. In
these cases, if the data for a certain type of assay were available for both of
these time points, only the first time point data were used and otherwise
data for either time point were used in the main analyses with cohort A.
Viral RNA measurement
SARS-CoV-2 RNA concentrations were measured from nasopharyngeal
samples and saliva samples by RT–PCR as previously described18,19. In
short, total nucleic acid was extracted from 300 μl of viral transport
media from the nasopharyngeal swab or 300 μl of whole saliva using
the MagMAX Viral/Pathogen Nucleic Acid Isolation kit (ThermoFisher
Scientific) using a modified protocol and eluted into 75 μl of elution
buffer19. For SARS-CoV-2 RNA detection, 5 μl of RNA template was tested
as previously described18, using the US CDC real-time RT–PCR primer/
probe sets for 2019-nCoV_N1, 2019-nCoV_N2, and the human RNase P
(RP) as an extraction control. Virus RNA copies were quantified using
a tenfold dilution standard curve of RNA transcripts that we previ-
ously generated18. If the RNA concentration was lower than the limit
of detection (ND) that was determined previously18, the value was set
to 0 and used for the analyses.
Isolation of plasma
Plasma samples were collected after whole blood centrifugation at
400g for 10 min at room temperature with brake off. The plasma was
then carefully transferred to 15-ml conical tubes and then aliquoted
and stored at −80 °C for subsequent analysis.
SARS-CoV-2 specific antibody measurement
ELISAs were performed as previously described20. In short, Triton X-100
and RNase A were added to serum samples at final concentrations of
0.5% and 0.5 mg ml−1 respectively and incubated at room temperature
for 30 min before use to reduce risk from any potential virus in serum.
Then, 96-well MaxiSorp plates (Thermo Scientific 442404) were coated
with 50 μl per well of recombinant SARS-CoV-2 S1 protein (ACROBiosys-
tems S1N-C52H3-100 μg) at a concentration of 2 μg ml−1 in PBS and were
incubated overnight at 4 °C. The coating buffer was removed, and plates
were incubated for 1 h at room temperature with 200 μl of blocking
solution (PBS with 0.1% Tween-20, 3% milk powder). Serum was diluted
1:50 in dilution solution (PBS with 0.1% Tween-20, 1% milk powder) and
100 μl of diluted serum was added for two hours at room temperature.
Plates were washed three times with PBS-T (PBS with 0.1% Tween-20) and
50 μl of HRP anti-Human IgG Antibody (GenScript A00166, 1:5,000) or
anti-Human IgM-Peroxidase Antibody (Sigma-Aldrich A6907, 1:5,000)
diluted in dilution solution were added to each well. After 1 h of incu-
bation at room temperature, plates were washed six times with PBS-T.
Plates were developed with 100 μl of TMB Substrate Reagent Set (BD
Biosciences 555214) and the reaction was stopped after 12 min by the
addition of 100ul of 2 N sulfuric acid. Plates were then read at a wave-
length of 450 nm and 570 nm.
The cut-off values for sero-positivity were determined as 0.392 and
0.436 for anti-S1-IgG and anti-S1 IgM, respectively. Eighty pre-pandemic
plasma samples were assayed to establish the negative baselines, and
these values were statistically determined with confidence level of 99%.
Cytokine and chemokine measurement
Patients’ sera isolated as above were stored in −80 °C until the meas-
urement of the cytokines. The sera were shipped to Eve Technologies
on dry ice, and levels of 71 cytokines and chemokines were measured
with Human Cytokine Array/Chemokine Array 71-Plex Panel (HD71).
All the samples were measured upon the first thaw.
The shipment of the samples and measurements were done in two
separate batches, but the measurements were performed with the same
assay kits using the same standard curves, therefore minimizing the
batch effects between the measurements.
Article
For the out of range values of the measurements, either the lowest
highest extrapolatable values or the lowest or highest standard curve
were recorded following the instructions of HD71 assay, and included
in the analyses. Among all the samples measured, we found that two
samples had outlier values (beyond 1.5× interquartile range) in more
than half of the 71 cytokines or chemokines measured, suggesting the
technical error and/or poor sample qualities in the measurements.
Therefore, cytokine or chemokine data of these individuals were
excluded from the analyses.
Isolation of PBMCs
The PBMCs were isolated from heparinized whole blood using Histo-
paque density gradient under the biosafety level 2+ facility. To isolate
PBMCs, blood 1:1 diluted in PBS was layered over in Histopaque in a
SepMate tube and centrifuged for 10 min at 1,200g. The PBMC layer was
collected by quickly pouring the content into a new 50-ml tube. The cells
were washed twice with PBS to remove any remaining histopaque and
to remove platelets. The pelleted cells were treated with ACK buffer for
red cell lysis and then counted. The percentage viability was estimated
using Trypan blue staining.
Flow cytometry
Using the freshly isolated PBMCs, the staining was performed in three
separate panels for (1) PBMC cell composition, (2) T cell surface stain-
ing, and (3) T cell intracellular staining. Exact antibody clones and
vendors that were used for flow cytometric analysis are as follows:
BB515 anti-HLA-DR (G46-6), BV785 anti-CD16 (3G8), PE-Cy7 anti-CD14
(HCD14), BV605 anti-CD3 (UCHT1), BV711 anti-CD19 (SJ25C1), BV421
anti-CD11c (3.9), AlexaFluor647 anti-CD1c (L161), Biotin anti-CD141
(M80), PE anti-CD304 (12C2), APCFire750 anti-CD11b (ICRF44), PerCP/
Cy5.5 anti-CD66b (G10F5), BV785 anti-CD4 (SK3), APCFire750 or PE-Cy7
or BV711 anti-CD8 (SK1), BV421 anti-CCR7 (G043H7), AlexaFluor
700 anti-CD45RA (HI100), PE anti-PD1 (EH12.2H7), APC anti-TIM-3
(F38-2E2), BV711 anti-CD38 (HIT2), BB700 anti-CXCR5 (RF8B2), PE-Cy7
anti-CD127 (HIL-7R-M21), PE-CF594 anti-CD25 (BC96), BV711 anti-CD127
(HIL-7R-M21), BV421 anti-IL-17a (N49-653), AlexaFluor 700 anti-TNF
(MAb11), PE or APC/Fire750 anti-IFNy (4S.B3), FITC anti-GranzymeB
(GB11), AlexaFluor 647 anti-IL-4 (8D4-8), BB700 anti-CD183/CXCR3
(1C6/CXCR3), PE-Cy7 anti-IL-6 (MQ2-13A5), PE anti-IL-2 (5344.111), BV785
anti-CD19 (SJ25C1), BV421 anti-CD138 (MI15), AlexaFluor700 anti-CD20
(2H7), AlexaFluor 647 anti-CD27 (M-T271), PE/Dazzle594 anti-IgD (IA6-
2), PE-Cy7 anti-CD86 (IT2.2), APC/Fire750 anti-IgM (MHM-88), BV605
anti-CD24 (M1/69), APC/Fire 750 anti-CD10 (HI10a), BV421 anti-CD15
(SSEA-1), AlexaFluor 700 Streptavidin (ThermoFisher). Freshly isolated
PBMCs were plated at 1 × 106–2 × 106 cells in a 96-well U-bottom plate.
Cells were resuspended in Live/Dead Fixable Aqua (ThermoFisher) for
20 min at 4 °C. Following a wash, cells were then blocked with Human
TruStan FcX (BioLegend) for 10 min at room temperature. Cocktails
of desired staining antibodies were directly added to this mixture for
30 min at room temperature. For secondary stains, cells were washed
and supernatant aspirated; to each cell pellet, a cocktail of secondary
markers was added for 30 min at 4 °C. Before analysis, cells were washed
and resuspended in 100 μl of 4% paraformaldehyde for 30 min at 4 °C.
For intracellular cytokine staining following stimulation, cells were
resuspended in 200 μl cRPMI (RPMI-1640 supplemented with 10% FBS,
2 mM l-glutamine, 100 U ml−1 penicillin, and 100 mg ml−1 streptomycin,
1 mM sodium pyruvate, and 50 μM 2-mercaptoethanol) and stored at
4 °C overnight. Subsequently, these cells were washed and stimulated
with 1× Cell Stimulation Cocktail (eBioscience) in 200 μl cRPMI for 1 h
at 37 °C. Directly to this, 50 μl of 5× Stimulation Cocktail (plus protein
transport inhibitor) (eBioscience) was added for an additional 4 h of
incubation at 37 °C. After stimulation, cells were washed and resus-
pended in 100 μl of 4% paraformaldehyde for 30 min at 4 °C. To quan-
tify intracellular cytokines, these samples were permeabilized with 1×
Permeabilization Buffer from the FOXP3/Transcription Factor Staining
Buffer Set (eBioscience) for 10 min at 4 °C. All further staining cocktails
were made in this buffer. Permeabilized cells were then washed and
resuspended in a cocktail containing Human TruStan FcX (BioLegend)
for 10 min at 4 °C. Finally, intracellular staining cocktails were directly
added to each sample for 1 h at 4 °C. After this incubation, cells were
washed and prepared for analysis on an Attune NXT (ThermoFisher).
Data were analysed using FlowJo software v.10.6 software (Tree Star).
Set of markers used to identify each subset of cells are summarized
in Extended Data Table 7, and gating strategies for the key cell popula-
tions presented in the main figures are shown in Extended Data Fig. 3a–c.
For most samples, all available staining panels were implemented and
analysed. The few exceptions pertained to those samples during which
a mechanical malfunction occurred, which depleted the sample before
acquisition, or to the samples with poor staining qualities. In these cases,
data for these samples or panels were missing and not available. All the
data available were used for the analyses, and the data used to generate
figures and tables can be found in Supplementary Table 1, and the raw fcs
files are available at ImmPort as described in the 'Data Availability' section.
Statistical analysis for the primary analyses
For the primary analyses shown in the main figures, Graph Pad Prism
(v,8.0) was used for all statistical analysis. Unless otherwise noted,
one-way ANOVA with Bonferroni’s multiple comparison test was used
for the comparisons between M_Pt versus F_Pt, M_Pt versus M_HCW,
F_Pt versus F_HCW, and M_HCW versus F_HCW for the comparisons.
For two-group comparisons including the comparison between stabi-
lized group and deteriorated group in each sex (Fig. 4a–d), two-sided
unpaired t-test was used for the comparison. Bioconductor R (v.3.6.3)
package ggplot2 (v.3.3.0) was used to generate heat maps (Extended
Data Fig. 2), X–Y graphs for correlation analyses (Figs. 2e, 4f), and Pear-
son correlation heat maps (Fig. 4e).
Statistical analysis for the secondary analyses
All multivariable analyses were conducted using R v.3.6.1 (for data clean-
ing) and SAS version 9.4 (Cary, NC; for data analysis). The code used
for data cleaning and data analysis is available at https://github.com/
muhellingson/covid_immresp. We conducted longitudinal analyses of
the differences in immune response by sex for patients with COVID-19
and differences in immune response between patients with COVID-19
and HCWs by sex and adjusted linear regression to evaluate differences
in immune response by sex at baseline and the differences in immune
response by sex and patient trajectory.
Longitudinal difference in immune response in all patients positive
for COVID-19 (cohort B) by sex. A marginal linear model was fitted
to evaluate the difference in various immune responses (outcome)
in patients by sex (exposure). We used an auto-regressive correlation
structure to account for correlation between repeated observations
in an individual over time. To account for the small sample size and
unequal follow-up between participants, we used the Morel–Bokossa–
Neerchal (MBN) correction. In addition to sex, the model contained
time-independent terms for age (in years) and BMI and time-dependent
terms for days from symptom onset (self-reported), ICU status (as a
proxy for disease severity) and treatment with either tocilizumab or
corticosteroids. A patient was defined as ‘on tocilizumab’ at a given time
point if they had received the treatment within 14 days before the time
the sample was taken. Patients were defined as ‘on corticosteroids’ if
they had received the treatment on the same day the sample was taken.
The resulting regression coefficients were interpreted as the difference
in the adjusted least square means immune response between female
and male patients.
Difference in immune response between patients with COVID-19
(cohort A) and HCWs by sex at baseline. We used linear regression
to evaluate the difference in immune response between female and
male patients at the first time point for those patients who had not
received corticosteroids or tocilizumab before enrolment (cohort
A). The model contained terms for sex, patient trajectory (worsened
versus stable), age, BMI, and an interaction term for sex and group
(patient versus HCWs). We calculated the least square means for each
group (female patients who worsened, female patients who stabi-
lized, male patients who worsened and male patients who stabilized)
and evaluated the differences in the least square means of the dif-
ferent immune response outcomes by group and sex. P values and
95% confidence intervals were calculated with a Tukey correction
for multiple pairwise comparisons. The regression coefficient of
the interaction term between sex and group was interpreted as the
difference-in-differences of the two comparisons by sex or by group
(for example, the difference-in-differences between female and male
patients and female and male HCWs).
Longitudinal difference in immune response between all patients
with COVID-19 (cohort B) and HCWs by sex. We used a marginal linear
model with a compound symmetric correlation structure and the MBN
correction to evaluate the difference in immune responses between
patients and HCWs by sex, controlling for age and BMI. We calculated
the least square means for each group (female patients, female HCWs,
male patient, male HCWs) and evaluated the differences in adjusted
least square means to compare study groups by sex (female patients
versus male patients, female HCWs versus male HCWs, female patients
versus female HCWs and male patients versus male HCWs). P values
and 95% confidence intervals were corrected using the Tukey correc-
tion for multiple pairwise comparisons. The regression coefficient of
the interaction term between sex and study group was interpreted as
the difference-in-differences between the two comparisons by sex or
by group.
Multivariable patient trajectory analysis. We used linear regression
to evaluate the difference in baseline immune response between pa-
tients who worsened after the baseline sample was taken and those
who stabilized by sex. The model contained terms for sex, patient
trajectory (worsened versus stable), age, days from symptom onset
and an interaction term for sex and patient trajectory. We calculated
the adjusted least square means for each group (female patients who
worsened, female patients who stabilized, male patients who wors-
ened and male patients who stabilized) and evaluated the differences
in least square means of the different immune responses by patient
trajectory and sex using the Tukey correction for multiple compari-
sons. The regression coefficient of the interaction term between sex
and patient trajectory was interpreted as the difference-in-differences
between the two patient trajectories by sex or sex by the two patient
trajectories.
Reporting summary
Further information on research design is available in the Nature
Research Reporting Summary linked to this paper.
Data availability
All of the background information of HCWs, clinical information of
patients, and raw data used in this study are included in the Supplemen-
tary Table 1. In addition, all of the raw fcs files for the flow cytometry
analysis are uploaded in ImmPort (https://www.immport.org/shared/
home, study ID: SDY1648).
17. WHO. R&D Blueprint Novel Coronavirus COVID-19 Therapeutic Trial Synopsis https://www.
who.int/blueprint/priority-diseases/key-action/COVID-19_Treatment_Trial_Design_Master_
Protocol_synopsis_Final_18022020.pdf (2020).
18. Vogels, C. B. F. et al. Analytical sensitivity and efficiency comparisons of SARS-CoV-2
RT–qPCR primer–probe sets. Nat. Microbiol. 5, 1299–1305 (2020)
19. Wyllie, A. L. et al. Saliva or nasopharyngeal swab specimens for detection of SARS-CoV-2.
N. Engl. J. Med. https://doi.org/10.1056/NEJMc2016359 (2020).
20. Amanat, F. et al. A serological assay to detect SARS-CoV-2 seroconversion in humans.
Nat. Med. 26, 1033–1036 (2020).
Acknowledgements We thank M. Linehan for technical and logistical assistance. This work
was supported by the Women’s Health Research at Yale Pilot Project Program (A.I., A.M.R.),
Fast Grant from Emergent Ventures at the Mercatus Center, Mathers Foundation, the
Beatrice Kleinberg Neuwirth Fund, Yale Institute for Global Health, and the Ludwig Family
Foundation. IMPACT received support from the Yale COVID-19 Research Resource Fund. A.I.
is an Investigator of the Howard Hughes Medical Institute. CBFV is supported by NWO
Rubicon 019.181EN.004. A.M. is supported by NIH grant R37AI041699.
Author contributions A.I., S.B.O. and A.I.K. conceived the study. C.L., P.W., J.K., J. Silva, T.M. and
J.E.O. defined parameters for flow cytometry experiments, collected and processed patient
PBMC samples. P.W. acquired and analysed the flow cytometry data. B.I., J.K., C.L. and C.D.O.
collected epidemiological and clinical data. F.L., A.M., J. Sun, E.Y.W. and A.M.R. acquired and
analysed ELISA data. A.L.W., C.B.F.V., I.M.O., R.E., S.L., P.L., A.V., A.P. and M.T. performed the
virus RNA concentration assays. N.D.G. supervised viral RNA concentration assays. A.C.-M. and
A.J.M. processed and stored patient specimens, J.B.F., C.D.C. and S.F. assisted in patient
recruitment, W.L.S. supervised clinical data management, A.S. coordinated and secured
funding for PBMC collection. T.T. designed the analysis scheme, analysed and interpreted the
data for the baseline analyses. M.K.E. and S.B.O. designed the analysis scheme, and
interpreted the data for the longitudinal analyses. M.K.E. analysed the longitudinal data. T.T.,
M.K.E. and A.I. drafted the manuscript. A.I., A.M.R. and S.B.O. revised the manuscript. A.I.
secured funds and supervised the project. Authors from the Yale IMPACT Research Team
contributed to collection and storage of patient samples, as well as the collection of the
patients’ epidemiological and clinical data.
Competing interests The authors declare no competing interests.
Additional information
Supplementary information is available for this paper at https://doi.org/10.1038/s41586-020-
2700-3.
Correspondence and requests for materials should be addressed to A.I.
Peer review information Nature thanks Petter Brodin, Malik Peiris and the other, anonymous,
reviewer(s) for their contribution to the peer review of this work. Peer reviewer reports are
available.
Reprints and permissions information is available at http://www.nature.com/reprints.
Article
Extended Data Fig. 1 | Comparison of basic clinical parameters of cohort A
patient samples and plasma levels of 71 cytokines and chemokines at the
first sampling of cohort A. a, Comparisons of age, BMI and DFSO at the first
sampling between male and female patients in cohort A. n = 17 and 22 for M_Pt
and F_Pt, respectively. b, Comparison of the plasma levels of 71 cytokines and
chemokines. n = 15, 28, 16 and 19 for M_HCW, F_HCW, M_Pt and F_Pt,
respectively. Data are mean ± s.e.m. P values were determined by unpaired
two-tailed t-test (a) or one-way ANOVA with Bonferroni multiple comparison
test (b). All P values < 0.10 are shown.
Extended Data Fig. 2 | Heat maps of cytokines and chemokines, PBMC
composition, T cell subsets, and T cell cytokine expression at the first
sampling of cohort A patients. a, A heat map of the plasma levels (pg ml−1) of
71 cytokines and chemokines. n = 15, 28, 16 and 19 for M_HCW, F_HCW, M_Pt and
F_Pt, respectively. b, A heat map for the composition of PBMCs (percentage in
live PBMCs). n = 6, 42, 16 and 21 for M_HCW, F_HCW, M_Pt and F_Pt, respectively.
c, A heat map for the T cell subsets (percentage in CD3+ cells). n = 6, 45, 16
and 22 for M_HCW, F_HCW, M_Pt and F_Pt, respectively. d, A heat map for the
intracellular cytokine staining of T cells (percentage in CD3+ cells). n = 6, 43, 16
and 22 for M_HCW, F_HCW, M_Pt and F_Pt, respectively. In all of these heat
maps, log-transformed values were used for heat map generation.
Article

Extended Data Fig. 3 | Flow cytometry gating strategy. a–c, Gating strategy used for monocytes (a), CD38+HLA-DR+ and PD-1+TIM-3+ CD4 or CD8 T cells (b), and
T cell intracellular staining for IFNγ+ CD8 T cells (c).
Extended Data Table 1 | Demographic and clinical characteristics of cohort A, cohort B and HCW comparison groups

*Cells may not sum to total owing to missing data.

†Categories not mutually exclusive.
‡Grey areas indicate that data are not available or not applicable.
Article
Extended Data Table 2 | Background and sample information of 39 cohort A patients

*Ethnicity: (1) American Indian/Alaskan native; (2) Asian; (3) Black/African American; (4) Native Hawaiian/Pacific Islander; (5) White; (6) Hispanic; (9) Multiple; (98) Unknown/unavailable.
†COVID-related risk factors: (0) No; (1) cancer treatment within 1 year; (2) chronic heart disease; (3) hypertension; (4) chronic lung disease (asthma, chronic obstructive pulmonary disease
(COPD) and interstitial lung disease (ILD)); (5) immunosuppression.
‡C1, clinical score at the first sample collection date.
§Days from symptom onset at the first sample collection.
||Cmax, maximum clinical score during the admission after the first time point sample collection.
¶Days from symptom onset at the first day Cmax was recorded in deteriorated patients.
#Collected sample or data types at the first sample collection date.
E, plasma cytokine/chemokine ELISA; F1, flow cytometry PBMC cell composition staining; F2, flow cytometry T cell surface staining; F3, flow cytometry T cell intracellular staining; G, plasma
anti-S1-IgG; M, plasma anti-S1-IgM; N, nasopharyngeal viral load; S, saliva viral load.
Extended Data Table 3 | Adjusted least square means difference in immune response at baseline between male and female
patients with COVID-19 in cohort A and male and female HCW controls

*Adjusted for age and BMI.

†A450 nm; nPT_F = 20, nPt_M = 15, nHCW_F = 74, nHCW_M = 13.
‡A450 nm; nPT_F = 20, nPt_M = 15, nHCW_F = 18, nHCW_M = 3.
§log10(pg ml−1); nPT_F = 19, nPt_M = 16, nHCW_F = 28, nHCW_M = 15.
||As percentage of live cells, unless otherwise indicated; nPT_F = 21, nPt_M = 16, nHCW_F = 51, nHCW_M = 6.
¶nPT_F = 33, nPt_M = 40, nHCW_F = 51, nHCW_M = 6.
#As a percentage of CD3-positive cells; nPT_F = 21, nPt_M = 16, nHCW_F = 51, nHCW_M = 6.
P values were determined using two-sided t-test with Tukey correction for multiple pairwise comparisons.
Article
Extended Data Table 4 | Adjusted least square means difference over time in immune response between male and female
patients with COVID-19 in cohort B

*Adjusted for age, BMI, days from symptom onset, tocilizumab treatment, corticosteroid treatment and ICU status.
†log10(SARS-CoV-2 copies per ml); nasopharyngeal nPT_F = 33, nPt_M = 30; saliva nPT_F = 20, nPt_M = 18.
‡
OD450; nPT_F = 44, nPt_M = 39. §log10(pg ml−1); nPT_F = 48, nPt_M = 43.
||As a percentage of live cells, unless indicated otherwise; nPT_F = 46, nPt_M = 42.
¶
nPT_F = 33, nPt_M = 40.
#As a percentage of CD3-positive cells; nPT_F = 49, nPt_M = 42.
P values were determined using two-sided t-test and Morel–Bokossa–Neerchal correction.
Extended Data Table 5 | Adjusted least square means difference over time in immune response between male and female
patients with COVID-19 in cohort B and male and female healthy HCW controls

*Adjusted for age and BMI.

†A450 nm; nPT_F = 44, nPt_M = 39, nHCW_F = 74, nHCW_M = 13.
‡OD450; nPT_F = 44, nPt_M = 39, nHCW_F = 18, NHCW_M = 3.
§log10(pg ml−1); nPT_F = 48, nPt_M = 43, nHCW_F = 28, nHCW_M = 15.
||As a percentage of live cells, unless otherwise indicated; nPT_F = 46, nPt_M = 42, nHCW_F = 51, nHCW_M = 6.
¶nPT_F = 33, nPt_M = 40, nHCW_F = 51, nHCW_M = 6.
#As a percentage of CD3-positive cells; nPT_F = 49, nPt_M = 42, nHCW_F = 51, nHCW_M = 6.
P values were determined using two-sided t-test with Tukey correction for multiple pairwise comparisons.
Article
Extended Data Table 6 | Adjusted least square means difference between male and female patients with COVID-19 in cohort
A by patient trajectory

*Adjusted for age and days from symptom onset.

†As percentage of CD3-positive cells unless otherwise indicated. nF_deteriorated = 6, nM_deteriorated = 6, nF_stabilized = 16, nM_stabilized = 11.
‡As a percentage of live cells.
§log10(pg ml−1).
P values were determined using two-sided t-test with Tukey correction for multiple comparisons.
Extended Data Table 7 | Definitions of each cell subset in flow cytometry with specific markers
 12345656762589 56 5317
!"#$%&'( A55j1""5
)"$#"$*
*+"
+ #$%&'(A#!#$kk9lmlm

,/"$#,"$0%1%!-
$
# .."  +
$.2$%#0*3$+
4$%15$%"$1
1#*3%67%4.2$#0$#40$0+"$""0+
$!684#$%4."$
 /"
 $#,"0%309#:$"3;30"$%:$"3;30+%053$6

7425
-$"$$0
8"33$"$$0"3""3+904.$%"$$%4331!$."$ $%4!#3!9$"*33!9."$<$9=$%0$6
>" 4.
n 7%<"0$".3 ?&@'4"0%<.$"3!#>0$9!2" ""
 0$#.*"#$4."#.$
n A$"$.$1%  $%."#.$1$"54.$0$".3 1%$%$%".".31"."#"$3+
n7 %$"$$0"3$$&'#A/B1%$%$%+"C$1C
D@EFGHIJJI@GKLMKMGMNIOEPGQLGPLMHRSQLPGMIELEFGQFG@TJLUGPLMHRSQLGJIRLGHIJVELWGKLHN@SXOLMGS@GKNLGYLKNIPMGMLHKSI@Z
n A0$ 4"3302""$$$
n A0$ 4"+"#.$ 00$9#0%" "$
$ 4."3$+""[#$.$4.#3$30."
n A4 #330$ 4$%$"$$0"3"".$03#!0$"3$0+&6!6."'$%*"0$."$&6!6!0440$'
A/B2""$&6!6$""2"$'"0"$$."$ 4#0$"$+&6!6040$2"3'
n 8#3 3%+$%$$!9$%$$$"$$0&6!6\9K9R'1$%040$2"39440$?9!
^S_LG]G_TEOLMGTMGLWTHKG_TEOLMG̀NL@L_LRGMOSKTQELZ
44."]2"3#$
n 8a"+"""3+94."$ $
 %0%0 4"="520%"=$"3$$!
n 8%"0%0"3"0.3<!9$40"$ 4$%""$3234$$"4#33$!
4#$0.
n :$."$
4440$?&6!6%bP9;"bR'90"$!%1$%+10"30#3"$
DORG̀LQGHIEELHKSI@GI@GMKTKSMKSHMGcIRGQSIEIdSMKMGHI@KTS@MGTRKSHELMGI@GJT@FGIcGKNLGVIS@KMGTQI_LZ
-4$1""0
;30+4."$"*#$"2"3"*3$+
40.#$0
B"$"0330$ :;j:g,4$1"&$0$2:=,21"030"3"$""!!!"$'",:B"o6p6q&030"3"$""!!!"$'6
B"$"""3+ ,&2p6q6p'9!!3$l&2p6p6m'9f"%";.&2r'9-A-&2o6s'9A$$#/<7&2p6k6l'9831t&2km6q97
-$"'6
8."#0$#$3?!0#$."3!$%.4$1"$%"$"0$"3$$
 %"0%*#$$+$0*
#*3%3$"$#94$1".#$*."
 "2"3"*3$
$
 $"
216e
e$!3+0#"!0$"0..#$+$+&6!6f$g#*'6-$%/"$#,"0%!#34#*.$$!0h4$1"44#$%4."$6

B"$"
;30+4."$"*#$"2"3"*3$+
4"$"
A33."#0$.#$03#""$""2"3"*3$+$"$.$7%$"$.$%#32$%4331!4."$91%"30"*3(
CA0009#i#$491*354#*303+"2"3"*3"$"$
CA3$44!#$%"$%"2"0"$"1"$"
CA0$4"+$0$

 "$""2"3"*3$+


A33$%4310+$.$+"$"133*"
 2"3"*32"j..;$1*$&$#+jB(-Bukqsr'6A33$%"$"&2"33"9"$*+$$9:)j-A"$"'#$
$!
 "$4!#"
$"*3$%$#+"03# $%-#3.$"+j4."$7"*3k
k6

0
8;3"33C0$$%0*431$0     $  !

12345656762589 56 5317425
%"$$%*$4$4+#"0%6j4+#"$#9"$%""$0$*4."5!+#30$6
n )400 a%"2#"3h0"300 :03!0"3923#$"+h2.$"300
8"400+4$%0#.$1$%"330$9"$#60.>0#.$>C$!C#.."+C43"$64

)A334$ 0  0   $#  +   ! 
#.#$03$%$21%$%03#!"$26
-".3? /$"$$0"3.$%1#$0"30#3"$$%".3?6-".3?1"$.*"$%#.*4"$$".$$$u"3C
/1g"2g$"3&u/gg'*$1="0%kr$%"="+o$%$%"$133"0$1$%$%0#$$#+#j,a"gj
"2$03wlmmmmlxqom6;"$$1$4$%#!%0!4:=,04$$"333.$6j2#"31$%"0$2
0%.$%"+"!"$0"09!"$"$$9"$$1$%*"05!#%."$3!0"3"*."3$9"$$1$%"#$..#"
&6!6%#."$"$%$'9""$$1$%"%$+4!"$"3"$"$"..##2"!$1$03#9"
$$"3or"$$103#$%$#+6j4.0$1"*$"*+$"$"44"".30330$0..0.."$3+
#$#+33.$630"30.10330$"<."$3+2+s"+1%"2#"3y030"3$"$#.$$9"1"
0$##$3"$$0%"!<"$6
B"$"<03# 71<$.#$3".340+$5:)j-A$%$$#+6="#.$4.$%2#"31#$3&*+k6z<$%
$i#"$3"!'.$%"%"344$%0+$5."#67%$!3+#!!$$%"$"$0%0"300##!$%$1
<.$".3i#"3$+6)51944310+$.$+9"$".+3"3"7033#4"0$"!"34$12#"3
1#$3.$%"%"344$%"".$."#91%0%#!!$$%$"!i#"3$+#!$%<.$67%#9"$"4
$%".31<03#4.$%""3+6
,30"$ 7%."#.$1$30"$C3!$#"3""3+4".34.%#."2#"36
,".?"$ 7%$32"$"$%"*2"$"3$#+6
a3! A$$%$.4".3"0i#$"0!90$$10.3$3+#"1"4$%"$$y0$6a3"0i#$
4."0*+"""$$".6j4."$4"$$y0$"$"2"3"*3#$3"4$0!"""3+!"1
"$"*+4310+$.$+":)j-A6A030"3$".9""$4.$%<.$"3$".94.0%"$21$$."$$y
32"$$"$$06+$5"4"0""3+1*36;"$$030"34."$"030"300!13+2"3"4$
"$"0330$6

,ei#4$."$!4 040."$"39+$.".$%
4."#$%"*#$.$+4."$"39<.$"3+$.".$%#."+$#6g90"$1%$%"0%."$"39
+$..$%3$32"$$+#$#+6j4+#"$#4"3$$."3$+#"0%9"$%""$0$*430$!"6
="$"3h<.$"3+$. =$%
>" j232$%$#+ >" j232$%$#+
n A$* n %j;Ci
n :#5"+$00333 n 8310+$.$+
n ;"3"$3!+""0%"3!+ n =,jC*"#."!!
n A."3"$%!".
n g#.""0%"$0"$
n 30"3"$"
n B#"3#"0%400

A$*
A$*# A33"$*#$%$#+""!"$%#."$6aazkz"$C%g)ACB,&fsqCq'&k(smm'&aBa00'9a{xrz"$%Bkq
&pfr'&k(kmm'&a)!'9;:C+x"$C%Bks&gBks'&k(pmm'&a)!'9a{qmz"$C%Bp&|g7k'&k(pmm'&a)!'9


a{xkk"$C%Bko&-tlzk'&k(pmm'&aBa00'9A3<"83#qsx"$C%Bk0&)kqk'&k(kzm'&a)!'9a$"$C%Bksk&=rm'
&k(kzm'&a)!'9;:CB"??3zos"$C%Bzq&gBzq'&k(pmm'&a)!'9;:"$C%Bpms&kll'&k(pmm'&a)!'9A;8xzm
"$C%Bkk*&j,8ss'&k(kmm'&a)!'9;;>+z6z"$C%Bqq*&fkm8z'&k(lmm'&aBa00'9a{xrz"$C%Bs&-}p'&k(lmm'
&a)!'9A;8xzm;:C+xa{xkk"$C%Br&-}k'&k(lmm'&a)!'9a{slk"$C%,x&fmspgx'&k(zm'&a)!'9
A3<"83#xmm"$C%Bsz,A&gjkmm'&k(lmm'&aBa00'9;:"$C%;Bk&:gkl6lgx'&k(lmm'&a)!'9A;"$C%7j=p
&8prCl:l'&k(zm'&a)!'9a{xkk"$C%Bpr&gj7l'&k(lmm'&a)!'9aaxmm"$C%~,z&,8ral'&k(zm'&aBa00'9
;:+x"$C%Bklx&gj)Cx,C=lk'&k(zm'&a)!'9;:C8zos"$C%Blz&aoq'&k(lmm'&aBa00'9a{xkk"$C%Bklx
v
 &gj)Cx,C=lk'&k(zm'&aBa00'9a{slk"$C%j)kx"&/soCqzp'&k(kmm'&aBa00'9A3<"83#xmm"$C%7/8"&=A*kk'
&k(kmm'&a)!'9;:A;>8xzm"$C%j8/+&s-6ap'&k(qm'&a)!'98j7"$C%f"?+.a&fakk'&k(lmm'&a)!'9
A3<"83#qsx"$C%j)Cs&rBsCr'&k(kmm'&a)!'9aaxmm"$C%Bkrp>~,p&kq>~,p'&k(kmm'&aBa00'9;:C+x

12345656762589 56 5317425
"$%j)Cq&=€lCkpAz'&k(zm'&a)!'9;:"$C%j)Cl&zpss6kkk'&k(zm'&aBa00'9a{xrz"$C%Bko&-tlzk'&k(pmm'
&a)!'9a{slk"$C%Bkpr&=jkz'&k(pmm'&a)!'9A3<"83#xmm"$C%Blm&lgx'&k(lmm'&a)!'9A3<"83#qsx
"$C%Blx&=C7lxk'&k(pzm'&a)!'9;:>B"??3zos"$C%j!B&jAqCl'&k(smm'&a)!'9;:C+x"$C%Brq&j7l6l'&k(kmm'
&a)!'9A;>8xzm"$C%j!=&=g=Crr'&k(lzm'&a)!'9a{qmz"$C%Bls&=)z'&k(lmm'&a)!'9a{slk"$C%Bkm
&gjkm"'&k(lmm'&a)!'9a{slk"$CB%kz&--:ACk'&k(lmm'&a)!'9A3<"83#xmm-$$"2&k(pmm'&7%.8%'9
a{qmz-$$"2&k(pmm'&a)!'6
{"3"$ A33"$*#$%$#+"0..0"33+"2"3"*39""33%"2*2"3"$*+$%."#4"0$#"#*+$%
#*30"$6)5191$$"$$%"$*"00!$#1#$"!0$67%4331!12"3"$$%
4331!0(aazkz"$C%g)ACB,&fsqCq'&aBa00'&g#."9,%#9+.3!#9a"*'9a{xrz"$C%Bkq&pfr'
&a)!'&g#."9A40"f9a"*9"#0%=5+9%."?9+.3!#9=".$9;!$"3="0"i#9,%#9
-$+="!"*+9-i#3=5+'9;:C+x"$C%Bks&gBks'&a)!'&g#."'9a{qmz"$C%Bp&|g7k'&a)!'
&g#."9%."?'9a{xkk"$C%Bko&-tlzk'&aBa00'&g#."'9A3<"83#qsx"$C%Bk0&)kqk'&a)!'&g#."9
A40"f9a"*9+.3!#9,%#'9a$"$C%Bksk&=rm'&a)!'&g#."9A40"f9a"*'9;:CB"??3zos
"$C%Bzq&gBzq'&a)!'&g#."9A40"f9a"*9+.3!#9,%#'9;:"$C%Bpms&kll'&a)!'
&g#."'9A;8xzm"$C%Bkk*&j,8ss'&a)!'&g#."9A40"f9a"*9%."?9..=".$9
+.3!#9,%#9-1'9;;>+z6z"$C%Bqq*&fkm8z'&aBa00'&g#."'9a{xrz"$C%Bs&-}p'&a)!'
&g#."'9A;8xzm;:C+xa{xkk"$C%Br&-}k'&a)!'&g#."9C,"0$2$+(A40"f9%."?9
+.3!#9;!$"3="0"i#9,%#9-$+="!"*+'9a{slk"$C%,x&fmspgx'&a)!'&g#."9A40"f9
a"*9+.3!#9,%#'9A3<"83#xmm"$C%Bsz,A&gjkmm'&aBa00'&g#."'9;:"$C%;Bk&:gkl6lgx'
&a)!'&g#."9A40"f9a"*9%."?9..=".$9+.3!#9,%#9-i#3=5+'9A;
"$%7j=p&8prCl:l'&a)!'&g#."'9a{xkk"$C%Bpr&gj7l'&a)!'&g#."9%."?9g'9aaxmm"$C%~,z
&,8ral'&aBa00'&g#."'9;:C+x"$C%Bklx&gj)Cx,C=lk'&a)!'&g#."'9;:C8zos"$C%Blz&aoq'&aB
a00'&g#."9,%#9+.3!#9a"*'9a{xkk"$C%Bklx&gj)Cx,C=lk'&aBa00'&g#."'9a{slk"$C%j)Ckx"
&/soCqzp'&aBa00'&g#."'9A3<"83#xmm"$C%7/8"&=A*kk'&a)!'&g#."9"$9C,"0$2$+(%."?9
a"*9+.3!#9,%#9;!$"3="0"i#9-$+="!"*+9-1'9;:A;>8xzm"$C%j8/+&s-6ap'&a)!'
&g#."9C,"0$2$+(%."?9a"*9+.3!#9,%#'98j7"$C%f"?+.a&fakk'&a)!'&g#."9=#9
C,"0$2$+(,"$'9A3<"83#qsx"$C%j)Cs&rBsCr'&a)!'&g#."9C,"0$2$+(%."?9a"*9+.3!#9
,%#'9aaxmm"$C%Bkrp>~,p&kq>~,p'&aBa00'&g#."9,%#9+.3!#9a"*'9;:C+x"$Cj)Cq
&=€lCkpAz'&a)!'&g#."'9;:"$C%j)Cl&zpss6kkk'&aBa00'&g#."'9a{xrz"$C%Bko&-tlzk'&a)!'
&g#."'9a{slk"$C%Bkpr&=jkz'&a)!'&g#."'9A3<"83#xmm"$C%Blm&lgx'&a)!'&g#."9a"*9"#0%
=5+9%."?9+.3!#9;!$"3="0"i#9,%#9-i#3=5+'9A3<"83#qsx"$C%Blx&=C7lxk'&a)!'
&g#."9C,"0$2$+(a"*9+.3!#9,%#'9;:>B"??3zos"$C%j!B&jAqCl'&a)!'&g#."'9;:C+x"$C%Brq
&j7l6l'&a)!'&g#."9A40"f9a"*9"#0%=5+9..=".$9$$C$7"."9%."?9
+.3!#9,%#'9A;>8xzm"$C%j!=&=g=Crr'&a)!'&g#."9A40"f9a"*9+.3!#9,%#'9a{qmz
"$C%Bls&=)z'&a)!'&g#."9C,"0$2$+(%."?'9a{slk"$C%Bkm&gjkm"'&a)!'&g#."9A40"
f9a"*9"#0%.5+9%."?9+.3!#9,%#'9a{slk"$C%Bkz&--:ACk'&a)!'&g#."'9A3<"83#
xmm-$$"2&k(pmm'&7%.8%'9a{qmz-$$"2&k(pmm'&a)!'6
g#.""0%"$0"$
;30+4."$"*#$$#232!%#.""0%"$0"$
;#3"$0%""0$$0 84$+C4."3&"!qs6mkq6o'"sx."3&"!qk6okq6x'"$$103#67%$"3.!"%04."$
0"*4#:<$B"$"7"*3k6
,0#$.$ ;"$$".$$$$%u"3/1g"2g$"3&u/gg'*$1$%kr$%4="0%$%#!%$%o$%4="+lmlm91
0#$$$%u"3j=;A7$#+&j.3.$!=0"3";#*30g"3$%A0$A!"$"2#7'"4$$$!
$24-A,-C{l*+i,7C;,6&3!+1"4#$%04.4"33"$$33'6;"$$1$4
$%#!%0!4:=,04$$"333.$1$%3430$6j4.0$1"*$"*+$"
$"44"".30330$0..0.."$3+#$#+33.$630"30.10330$
"<."$3+2+s"+1%"2#"3y030"3$"$#.$$9"1"0$##$3"$$0%"!
<"$6
:$%02!%$ u"3g#.","0%;$0$;!".j$$#$"3,21a"6j4.0$1*$"4."3333
"$$"%"3$%0"156‚ƒ#"0%$031"21""2*+$%u"3-0%34=0j,a"
gj&wlmmmmlxqom'6j4.0$1"*$"*+$"$"44"0."$"#"0%"$"*"4$%
#"$4#$#+67%1.03#$%$#+6


/$$%"$4#334."$$%"2"34$%$#+$03.#$"3*2$%."#0$6


831+$.$+

12345656762589 56 5317
;3$
4.$%"$(
n 7%"<3 "*3$"$$%."5"43#0%.#&6!6BsC8j7'6
n 7%"<0"3 "03"3+2*36j03##.*"3!"<3+4*$$.34$3$4!#&
&"b
"!#b
"
""
 "3+
4$0"3."5'6
n A333$"0$#3$1$%#$3 #033$6
n A#.0"32"3 #4#.*4033 0$"!&1$%$"$$0'
26

7425
=$%3!+
-".3""$ 8%3+3"$;a=1$"432""."59*3051$%g#."7#-$"80~9$"4#4"0
."5"$%4<1$%;8As 68$"033#3"0+$5$"!4331!$.#3"$90331#4"0$"9
1"%"4< ss ;8A6A4$."*3?"$1$%k k~;
~ ."*3?"$a#440331$"4$"033#3"
0+$5""3+6
j$#.$ 331"0i# "
"A
 $$#/~7&7%.8%'6
-4$1" B"$"1""3+#!831t4$1"2km6q4$1"&7-$"'6
33#3"$"*#"0 33#3"$"*#"0(33#3"$1$ 2"#4."$03#!" ""
 #.*00$"$4$%
"$$y*3".3&<kmq033>.)'9""
 $ 4329!3;a=& & 4)2'9""
 $
4""$!"$&&
4Bs70339 4=0+$9$06'67%4#33!"$!"$%403"40"$03# $%<$4!#6
f"$!$"$!+ --CA"8-CA"".$1#$ $
 30$3#50+$4. 3"$;a=6)2""03314*"
"i#"$"!6-!3$1""$*" -
 ->8-"".$6)#50+$1!"$*" $
$$4+
3+.%0+$&Bp>Bs>Br>Bko>Bzq."5'9!"#30+$&Bkq9Bks9g)ACB,."5'"B9"0B&Bpms9
Bk09Bksk'67,C"0$2"$703397."33+C44$"$70339""$"3#*$614#!g)ACB,9
Bpr9,x9Bklx9;Bk97j=Cp9~,z9Bsz,A9Blz6j$"033#3"7033!"$!$"$!+$ $$4+Bs">Br7033
0$!7/8"9j8/C+9j)Cq9j)Cl9f"?+.a9j)Cs9">j)Ckx14#!$%04."5(Bp9Bs9Br97/89
j8/9j)Cq9j)Cl9j)Cs9j)Ckx"!"?+.aa6
n 705$%*<$
$0
 4.$%"$"4!#<.34+!$%!"$!$"$!+
2
$%-#3.$"+j4."$6



„
Article

Tunable dynamics of B cell selection in gut

germinal centres

https://doi.org/10.1038/s41586-020-2865-9 Carla R. Nowosad1,2,4, Luka Mesin1,4, Tiago B. R. Castro1,2, Christopher Wichmann2,3,

Gregory P. Donaldson2, Tatsuya Araki1, Ariën Schiepers1, Ainsley A. K. Lockhart2,
Received: 7 March 2020
Angelina M. Bilate2, Daniel Mucida2 ✉ & Gabriel D. Victora1 ✉
Accepted: 29 July 2020

Published online: 28 October 2020

Germinal centres, the structures in which B cells evolve to produce antibodies with
Check for updates high affinity for various antigens, usually form transiently in lymphoid organs in
response to infection or immunization. In lymphoid organs associated with the gut,
however, germinal centres are chronically present. These gut-associated germinal
centres can support targeted antibody responses to gut infections and immunization1.
But whether B cell selection and antibody affinity maturation take place in the face of
the chronic and diverse antigenic stimulation characteristic of these structures under
steady state is less clear2–8. Here, by combining multicolour ‘Brainbow’ cell-fate
mapping and sequencing of immunoglobulin genes from single cells, we find that
5–10% of gut-associated germinal centres from specific-pathogen-free (SPF) mice
contain highly dominant ‘winner’ B cell clones at steady state, despite rapid turnover
of germinal-centre B cells. Monoclonal antibodies derived from these clones show
increased binding, compared with their unmutated precursors, to commensal
bacteria, consistent with antigen-driven selection. The frequency of highly selected
gut-associated germinal centres is markedly higher in germ-free than in SPF mice, and
winner B cells in germ-free germinal centres are enriched in ‘public’ clonotypes found
in multiple individuals, indicating strong selection of B cell antigen receptors even in
the absence of microbiota. Colonization of germ-free mice with a defined microbial
consortium (Oligo-MM12) does not eliminate germ-free-associated clonotypes, yet
does induce a concomitant commensal-specific B cell response with the hallmarks of
antigen-driven selection. Thus, positive selection of B cells can take place in
steady-state gut-associated germinal centres, at a rate that is tunable over a wide
range by the presence and composition of the microbiota.

Our intestines are constantly exposed to large amounts of antigens
derived from diet and commensal microbes. The interaction of these
antigens with the immune system takes place primarily in gut-associated
secondary lymphoid structures, including gut-draining mesenteric
lymph nodes (mLNs) and Peyer’s patches, where gut-associated ger-
minal centres (gaGCs) provide a site for the hypermutation of immu-
noglobulin genes even under steady state5,9. B cell antigen receptor
(BCR)-driven selection and affinity maturation of antibodies occur
efficiently in gaGCs upon oral immunization6,10. However, given previ-
ous reports that steady-state gaGCs can form in a BCR-independent
fashion2,6, show little evidence of BCR-driven selection of specific
antibodies at the sequence level3, and are associated instead with the
selection of polyreactive immunoglobulins4, it has been postulated that
gaGCs may act predominantly as diversifiers of the immunoglobulin
repertoire, rather than fostering affinity maturation towards com-
mensal microbes (reviewed in refs. 5,8). We thus sought to determine
the extent to which germinal-centre selection and antibody affinity
maturation occur in the midst of chronic antigenic stimulation, and
to define the impact of the microbiota on these processes.
To estimate the rate of B cell selection in steady-state gaGCs, we first
used in situ photoactivation of mice engineered to express photoac-
tivatable green fluorescent protein (PA-GFP)11,12 (Fig. 1a and Extended
Data Fig. 1a) to sequence B cell immunoglobulin heavy chain genes
(Igh) from 20 individual gaGCs from various mLNs of 5 mice housed
under SPF conditions. Clonal diversity in SPF gaGCs spanned a wide
range, with a median of 33 clones per germinal centre (using the Chao1
estimator function), a D50 value (the fraction of clones accounting for
50% of sequenced cells) of 0.20, and 30% of B cells belonging to the
largest clone in the germinal centre (Fig. 1b, c). One of the 20 samples
sequenced contained a highly dominant clone that accounted for 64%
of cells in that germinal centre (Fig. 1b, c). Analysis of somatic muta-
tions within this clone (Fig. 1d) showed the nested expansion of nodes
with increasing numbers of mutations (indicated by arrows) typical of
sequential positive selection.

Laboratory of Lymphocyte Dynamics, The Rockefeller University, New York, NY, USA. 2Laboratory of Mucosal Immunology, The Rockefeller University, New York, NY, USA. 3Mucosal Immunology
1

Group, Department of Pediatrics, University Medical Center Rostock, Rostock, Germany. 4These authors contributed equally: Carla R. Nowosad, Luka Mesin. ✉e-mail: mucida@rockefeller.edu;
victora@rockefeller.edu

Nature | Vol 588 | 10 December 2020 | 321

Article
a b c Clonal Clonal Size of
SPF1 SPF2 SPF3 SPF4 SPF5 richness diversity largest clone
PA–GFP Photoactivate Dissect

germinal centre (Chao1)

SPF hi J J I I CC CC I I CC I I CC I I CC 103 0.4 100

germinal centre (%)

90
mLN

Number of cells sequenced

0.3 80
80

Cells in the
Clones per
102
70 60

D50
0.2
60 40
101
50 0.1
FDCs (anti-CD35-Cy3) 20
40
Sort PA+ cells 30 100 0 0
Igh sequencing
20
d UA
10
Inferred precursor 1 mutation
0 n n cells with same sequence
Dominant clone Expanded clones Singletons

e Steady-state gaGCs 2
3
Δt 2 2

AicdaCreERT2/+.Rosa26Confetti/Confetti 3
Tamoxifen (2×)
f Caecal–colonic mLN, day 7 post-tamoxifen Caecal–colonic mLN, day 15 post-tamoxifen Caecal–colonic mLN, day 21 post-tamoxifen
L  L  L 

L
L LL
L

LL LL  LL  LL 

LL

g Dominance × density

germinal-centre B cells (%)

Coloured cell density Colour dominance h
1.0 100 1.0 100
Half-life 1.9 weeks
Cells per 100 μm2

0.8 80 0.8 80 (95% CI 1.1–3.8 weeks)

Fate-mapped
Fluorescent
B cells (%)

0.6 60 0.6 60
NDS

0.4 40 0.4 40
0.2 20 0.2 20
0.0 0 0.0 0
0 5 10 15 20
14 7
21 15
3
35
42
49
56

14 7
21 15
3
35
42
49
56

14 7
21 5
3
35
42
49
56
–2

–2

–1
–2
–

Weeks post-tamoxifen
Days post-tamoxifen Days post-tamoxifen Days post-tamoxifen

Fig. 1 | Kinetics of clonal selection in steady-state gaGCs. a, Experimental
setup for in situ photoactivation of gaGCs. mLNs from photoactivatable (PA),
GFP-transgenic, SPF mice are photoactivated and dissected; photoactivated
(PA+) cells are sorted and Igh genes sequenced. FDC, follicular dendritic cell,
visualized with an antibody against Cy3-labelled CD35. b, Clonal composition
of individual germinal centres from five mice (SPF1–5), obtained as in a.
J, jejunal mLN; I, ileal mLN; C, caecal-colonic mLN. c, Quantification of data in
b. Each symbol represents one germinal centre; medians indicated by centre
lines. d, Relationships between Igh sequences of B cells derived from the
largest B cell clone in b. Arrows indicate putative positive selection events.
UA, unmutated ancestor. e, Experimental setup for ‘Brainbow’ fate-mapping.
Steady-state gaGCs from AicdaCreERT2/+.Rosa26Confetti/Confetti mice are recombined
to produce different colours by treatment with tamoxifen and followed over
time (Δt). f, Representative multiphoton images of mLNs from SPF mice at
different times after tamoxifen treatment. Blue is collagen (second harmonics);
white is autofluorescence; other colours are from the Confetti allele. Numbers
in parentheses represent normalized dominance scores (NDS; that is, the
frequency of B cells in a germinal centre that carry the dominant colour after
normalization for coloured cell density). Scale bars represent 200 μm in main
images and 50 μm in close-ups. g, Quantification of multiple images as in f. Left,
density of coloured cells (that is, fluorescent cells in the germinal-centre dark
zone). Centre, colour dominance (that is, frequency of the most-common
colour). Right, NDS (excludes germinal centres with density < 0.4). Each symbol
represents one germinal centre. Only germinal centres with a density of more
than 0.4 fluorescent cells per 100 μm2 are included in the NDS calculation.
Filled symbols indicate data from two mice in which tamoxifen was
administered intraperitoneally. Medians are indicated. Data are from 3–5 mice
per time point at days 14–35, and 1 to 2 mice for day 7 and later time points.
h, SPF S1pr2CreERT2 × Rosa26Stop-tdTomato mice were given two doses of tamoxifen,
two days apart, to label germinal-centre B cells. The graph shows the proportion
of tdTomato+ mLN B cells assayed by flow cytometry at the time points
indicated, counting from the first dose of tamoxifen. Each symbol represents
one mouse. Half-lives were quantified using a one-phase exponential decay
function (black line). CI, confidence interval. Gating strategies are detailed in
Extended Data Fig. 1a–c.

This was confirmed in a much larger number of germinal centres progressively after labelling, probably as a result of germinal-centre B
using Brainbow13 multicolour fate-mapping11 (see Supplementary Infor- cells being replaced by incoming unlabelled clones as the response
mation). We fate-mapped steady-state gaGC B cells in AicdaCreERT2/+. evolved (Fig. 1f, g). Because density estimations using Brainbow are sen-
Rosa26Confetti/Confetti (AID-Confetti) mice held under SPF conditions by admin- sitive to dropout of low-fluorescence germinal centres, we measured
istering two doses of tamoxifen, two days apart (Fig. 1e). This labelled germinal-centre turnover by flow cytometry using the S1pr2CreERT2 BAC
roughly 50% of B cells in both mLNs and Peyer’s patches, as estimated by transgene14 crossed to the Rosa26Stop-tdTomato reporter. Pulse-labelling using
the density of coloured cells in the germinal centres and by flow-cytometry this model allowed us to place the half-life of gaGC B cells at roughly two
experiments (Fig. 1f, g and Extended Data Fig. 1b). This fraction decreased weeks under SPF conditions (Fig. 1h and Extended Data Fig. 1c).

322 | Nature | Vol 588 | 10 December 2020

 a UA S120.U
2
(0.76)
5
mLN
88/88 2
12
UA
S120
S078.U 6
4

(0.83) 7 S078 11
2

mLN 2 3
32/44 YFP n n cells with
RFP same sequence
CFP + RFP Inferred precursor
1 mutation

b Secondary only MG053 (negative control) S078 (mutated) S078.U (reverted) S078 (P = 0.016)

antibody binding (%)

105 105 105 105 0.15
0 0.002 0.049 0.004

Monoclonal
104 104 104 104
0.10
103 103 103 103
0.05
102 102 102 102
0 0 0 0
0.00
0102 103 104 105 0102 103 104 105 0102 103 104 105 0102 103 104 105 M U
Secondary only G200 (negative control) S120 (mutated) S120.U (reverted) S120 (P = 0.016)
8

antibody binding (%)

105 105 105 105
0.46 0.96 6.22 1.21
6

Monoclonal
104 104 104 104

103 103 103 103 4

FSC

102 102 102 102 2

0 0 0 0
0
0 103 104 105 0 103 104 105 0 103 104 105 0 103 104 105 M U
Anti-human IgG1 AF647

c ED38 (polyreactive d 1.0

4 positive control) S078 S120
1.5
S078.U 0.8 S120.U
3 S210
MG053 MG053
1.0 0.6
A450

A450

A450

2
S078 0.4
S212 0.5
1 0.2
S120
] S116/S118/S080/
0 0.0 0.0
MG053 (negative control)
100 101 102 103 100 101 102 103 100 101 102 103
Monoclonal antibody (nM) Monoclonal antibody (nM) Monoclonal antibody (nM)

Fig. 2 | Selection of commensal-binding clones in steady-state gaGCs.
a, Relationships among Igh sequences from B cells of high-NDS germinal
centres, sorted as in Extended Data Fig. 2a. Left, images and pie charts show
clones (inner ring, with grey representing the major clone) and Brainbow
colour distributions (outer ring) for each germinal centre. Numbers within
images are NDS values; numbers in pie charts are numbers of cells in the major
clone by the total number of cells sequenced. Right, phylogenies for major
clones (grey in the pie charts). Arrows indicate ‘clonal burst’ points; names are
indicated whenever a recombinant monoclonal antibody was generated from a
sequence (see Supplementary Table 1). CFP, cyan fluorescent protein; RFP, red
fluorescent protein; YFP, yellow fluorescent protein; UA, unmutated ancestor.
Scale bars, 50 μm. Additional trees are in Extended Data Fig. 2b, c. b, Binding of
monoclonal antibodies to faecal bacteria from specific-pathogen-free (SPF)
mice. Gated on SYTO BC+, DAPI– live bacteria (Extended Data Fig. 2d;
see Methods). Clones MG053 and G200 (see below) are non-bacteria-reactive
negative controls. The graphs at the right summarize seven independent
experiments (M, mutated; U, unmutated); background (percentage positive in
secondary only) is subtracted from all data points. P values obtained from
two-tailed Wilcoxon paired samples test. c, Binding of monoclonal antibodies
to faecal bacteria, assessed by ELISA. Lines show means of three assays. A450,
absorbance at 450 nm. d, As in c, but showing monoclonal antibodies S078 and
S120 as well as their unmutated ancestors.

Despite this rapid turnover, the normalized dominance score (NDS,
an estimate of the frequency of B cells in a germinal centre that carry
the dominant colour11,15; see Supplementary Information) in gaGCs
increased progressively to day 23 after tamoxifen, when 11% of germinal
centres (6 of 57) scored 0.75 or higher (Fig. 1f, g). The strongest clonal
expansions that occur in mLN germinal centres are therefore large and
rapid enough to generate dominant lineages, despite the replacement
of labelled clones with incoming unlabelled B cells. In germinal centres
within Peyer’s patches, clonal selection progressed at a slower rate
(Extended Data Fig. 1d, e), as expected from their much larger size and
similar rate of turnover (Extended Data Fig. 1f, g). However, germinal
centres with NDS values of more than 0.5 could occasionally be detected
as early as day 14 after tamoxifen, peaking at day 23 after tamoxifen,
when 15% of Peyer’s patches (3 of 20) had reached NDS values of more
than 50% (Extended Data Fig. 1e). We conclude that clonal selection is
detectable in gaGCs, despite chronic exposure to a high burden and
diversity of foreign antigens and the rapid turnover of B cell clones.
To understand the relationship between clonal selection and affinity
maturation in gaGCs, we used vibratome slicing of agarose-embedded
AID-Confetti lymph nodes11 (Extended Data Fig. 2a) to isolate gaGCs
containing ‘winner’ clones, where antigen-driven selection is most
likely to have occurred11. Sequencing of Igh from B cells sorted from

Nature | Vol 588 | 10 December 2020 | 323

Article
a mLN, day 14 post-tamoxifen Day 20 post-tamoxifen b Coloured cell density Colour dominance Dominance × density
c
P < 0.0001
0.90 0.87 0.89 100 100
1.0 1.0
80

NDS > 0.75 (%)

0.8 0.8 75
0.6 60 0.6

mLN
50
0.4 40 0.4

Fluorescent cells per 100 μm2

25

Fluorescent B cells (%)

0.2 20 0.2
Peyer’s patch, day 20 post-tamoxifen 0.0 0 0.0 0
SPF GF

14 7
8

14 7
8

14 7
8

3
NDS
i (0.91)

5–
–1

–2

5–
–1

–2

5–
–1

–2
20

20

20
P = 0.0084
1.0 100 1.0 100
i
0.8 80 0.8 75

NDS > 0.5 (%)

Peyer’s patches
0.6 60 0.6
ii 50
0.4 40 0.4
ii (0.98)
20 25
0.2 0.2
0.0 0 0.0 0
SPF GF

3
–1

–2

–1

–2

–1

–2
14

20

14

20

14

20
Days 20–23
Days post-tamoxifen

d UA e UA
M218.U
(0.98)

G226 16 M218
5
(0.76) 4
6 6

PP mLN
78/81 RFP Inferred RFP Inferred
precursor 31/39 precursor
n n cells with 1 mutation n n cells with 1 mutation
identical identical
sequence sequence

f g h j
M218 M216 M220 M222 M224 5 2.0
Secondary B.o. C.i. 100 ED38 M218
80 M218 4 1.5 M218.U
MG053 (negative control) B.c. L.r.
M218.U 3 MG053

A450
E.f. 60 M220

A450
A.m. MG053 1.0
ED38 (positive control) M.i. 40 2 M218
F.p. 20 1 M216 0.5
M216 C.c. MG053
Ak.m. 0 0 0.0
0 103 104 105 100 101 102 103 100 101 102 103
M218 Bl.c.
Anti-hIgG1 AF647
M220 i M216
100 M216 100 M218 100 M220 1.2 3 M220
M222 M216.U
80 M216.U 80 M218.U 80 M220.U 1.0 M220.U
M224 MG053 MG053 0.8 MG053 2 MG053
60 60 60 MG053
A450

A450
0.6
M228 40 40 40 0.4 1
20 20 20 0.2
M232
0 0 0 0.0 0
0 103 104 105 0 103 104 105 0 103 104 105 0 103 104 105 100 101 102 103 100 101 102 103
Anti-human IgG1 AF647 Anti-human IgG1 AF647 Monoclonal antibody (nM)

Fig. 3 | Accelerated selection in gaGCs of germ-free and and Oligo-MM12 (e) germinal centres with high NDS values. Details as in Fig. 2a.
Oligo-MM12-colonized mice. a, Representative multiphoton images of Additional trees are in Extended Data Fig. 6a, b. Scale bars, 50 μm. f, Flow
germ-free mLNs and Peyer’s patches at different times after treatment with cytometry showing the binding of monoclonal antibodies to faecal bacteria
tamoxifen. Blue represents collagen (second harmonics); white shows from Oligo-MM12-colonized mice, detected using anti-human IgG1 Alexa Fluor
autofluorescence; other colours are from the Confetti allele. Scale bars 647. ED38, polyreactive positive control monoclonal antibody; MG053,
represent 200 μm (Peyer’s patch overview) and 50 μm in close-ups. Numbers in negative control. The dotted line placed at 102 is for reference purposes. g, Dot
yellow are NDS values. b, Quantification of multiple images as in a for mLNs and blot showing the binding of Oligo-MM12 monoclonal antibodies to a subset of
Peyer’s patches. Left, density of coloured cells (fluorescent cells in the cultured Oligo-MM12 strains (black font; see Supplementary Table 4).
germinal-centre dark zone). Centre, colour dominance (frequency of the Oligo-MM12 strains tested are Acutalibacter muris (A.m.), Clostriudium
most-common colour). Right, NDS (frequency of B cells in a germinal centre innocuum (C.i.), Enterococcus faecalis (E.f.), Muribaculum intestinale (M.i.),
that carry the dominant colour). Each symbol represents one germinal centre; Flavonifractor plautii (F.p.), Clostridium clostridioforme (C.c.), Akkermansia
medians indicated. Only germinal centres with a density of more than 0.4 muciniphila (Ak. M.) and Blautia coccoides (B.c.). Bacteroides ovatus (B.o.) and
fluorescent cells per 100 μm2 are included in the NDS calculations. c, Proportion Bacteroides caccae (B.c.) are negative controls (blue font). Arrows indicate E.f.,
of germinal centres with NDS values of more than 0.75 in mLNs (top) and more which is bound only by M218. h, i, Binding of monoclonal antibody M218 to
than 0.5 in Peyer’s patches (bottom) under SPF and germ-free conditions at cultured E. faecalis (h) and of three Oligo-MM12 antibodies to faecal bacteria
20–23 days after tamoxifen. For SPF and germ-free mLN gaGCs, n = 57 and 27, from Oligo-MM12-colonized mice (i), as measured by flow cytometry. Gated on
respectively. For Peyer’s patch gaGCs, n = 20 and 9, respectively. SPF data are SYTO BC+, DAPI– live bacteria (Extended Data Fig. 2d). Data in f–i are
from Fig. 1g and Extended Data Fig. 1e. P values are from two-tailed Fisher’s representative of experiments carried out on at least two separate occasions.
exact tests. Error bars represent exact binomial 95% confidence intervals. Data j, Binding of monoclonal antibodies to faecal bacteria, measured by ELISA.
for b, c are from 3–5 mice per time point, except for days 5–7 which were from 1 Lines show means of two assays.
mouse. d, e, Relationship among Igh sequences from B cells of germ-free (d)

such germinal centres showed evidence of ‘clonal bursts’—jackpot-type to bacterial flow cytometry, two of seven antibodies produced from
positive selection events in which multiple B cells descending from a burst-associated immunoglobulin sequences reproducibly bound
single somatic hypermutation (SHM) variant account for a large frac- faecal bacteria (Fig. 2b and Extended Data Fig. 2d, e). Binding followed
tion of cells in a germinal centre11 (Fig. 2a and Extended Data Fig. 2a–c). different patterns: whereas monoclonal antibody S078 bound strongly
Because clonal bursts are regularly associated with the acquisition of to a small population of bacteria, S120 bound with moderate intensity
affinity-enhancing mutations11, we produced recombinant monoclonal to a much larger cohort (Fig. 2b). These two antibodies—as well as two
antibodies16 using burst-point sequences to probe for binding to com- other clones (S210 and S212)—reacted with bacteria-rich centrifuga-
mensal bacteria (Supplementary Table 1). Despite the variation inherent tion fractions, as measured by enzyme-linked immunosorbent assay

324 | Nature | Vol 588 | 10 December 2020

 a VH1–47/JH4 VH–12/JH3 they did under SPF conditions (Fig. 3a–c). By day 23 after tamoxifen,
4 4
3 3 roughly 56% of mLN germinal centres had reached an NDS of 0.75
Bits

Bits
2 2
1
0
1
0
or higher, compared with around 11% in SPF mice (Fig. 3c). This was
1 2 3 4 5 6 7 8 9 10 11 12 1 2 3 4 5 6 7
Position Position confirmed at the Igh sequence level: 11 of 14 mLN germinal centres
b GF1 GF2 GF3 GF4 GF5 GF6 GF7 MM12 1 MM12 2 MM12 3 picked at random from vibratome slices had dominant clones that
C I CC P P P J J J P P I DD J J I I C D J I C P I C P CCC C P
90 accounted for more than 50% of all B cells in the germinal centre, com-
80
pared with 1 of 20 germinal centres in SPF conditions (Fig. 1a–c and
Cells sequenced

70
60
50 Extended Data Fig. 5a–d). Faster selection of Brainbow colours was also
40
30
observed in germ-free Peyer’s patches, where 6 of 9 germinal centres
20 exceeded an NDS of 0.5 by day 20–23 after tamoxifen, compared with
10
0 3 of 20 under SPF conditions (Fig. 3a–c). Germinal-centre selection in
Individual germinal centres
(AID-Confetti)
mLN slices
(wild type)
Whole organs
(wild type)
Individual germinal centres
(AID-Confetti)
Oligo-MM12-colonized mice fell between the SPF and germ-free rates
VH1–47/ VH1-47/ Other VH1-47 VH1-12/ Other VH1–12 All other V (Extended Data Fig. 5e–g). Therefore, selection in gaGCs is not depend-
ARGSNY.../JH4 ARGSNY.../JH2 AREGFAY/JH3 segments
ent on a fully diverse microbiota, and in fact becomes accelerated in
c VH1–47/ARGSNY VH1–12/AREGF(A/D)Y d All public clonotypes Larger expansions the absence of commensal bacteria.
(exact match) (30 total wells or more)
SPF 6 1 Germ-free 6 1 SPF 6 1 Germ-free 6 1
Clonal phylogenies of single-coloured germinal centres from
5 2 5 2 5 2 5 2
germ-free and Oligo-MM12-colonized mice revealed strong clonal
4

4
bursting, as shown by the presence of large expansions of B cells with
3

3
3

3
4

4
identical variable heavy chain (VH) immunoglobulin sequences and
2

2
5

5
1

1
6

6
multiple inferred descendants (Fig. 3d, e and Extended Data Fig. 6a, b,
7

7
1

1
6 6 6 6

arrows). We produced monoclonal antibodies from 16 immunoglobulin

2 2 2 2
5 3 5 3 5 3 5 3
4 4 4 4

Oligo-MM12 1,000 clone*wells
Fig. 4 | Prominent public clonotypes in gaGCs of germ-free and
Oligo-MM12-colonized mice. a, CDR H3 amino-acid sequence logos for B cells
bearing public VH1–47/JH4 or JH2 rearrangements of 11 or 12 amino acids (left) or
VH1–12/JH3 rearrangements of 7 amino acids (right), sequenced from germ-free
gaGCs. b, Frequency across mice or samples of B cells that fit public-clonotype
criteria or carry other VH1–47 or VH1–12 rearrangements. Each bar represents
one sample. Data are for 7 germ-free mice (GF1–7) and 3 Oligo-MM12-colonized
mice (MM12 1–3). D, duodenal mLN; I, ileal mLN; C, caecal-colonic mLN; P, Peyer’s
patch. c, Circos plots showing the distribution of VH1–47 and VH1–12 public
clonotypes in a second cohort of sex- and age-matched SPF, germ-free and
Oligo-MM12-colonized mice. Each segment represents pooled germinal-centre
B cells from the mLNs and Peyer’s patches of one mouse, with clones ordered
clockwise from largest to smallest. Only clones containing identical CDR H3s are
linked (see Methods for a full description). Samples were gated as in Extended
Data Fig. 9b; all samples were sequenced in a single experiment. d, Circos plots
as in c, showing all public IgH clonotypes shared across mice housed under the
indicated conditions. Lines connect clonotypes with the same VH/CDR H3/JH
amino-acid sequence. Left, all clones; right, only those clones spanning 30 or
more wells in total. In c, d, only clones that were found in at least two wells from
the same mouse are linked.
(ELISA; Fig. 2c). Only S210 and S212 showed a mild degree of polyreac-
tivity using standard measures4,17 (Extended Data Fig. 2f). Reversion of
somatic mutations in S078 and S120 resulted in decreased binding to
bacteria by both flow cytometry and ELISA (Fig. 2b, d). Thus, commensal
binding with the characteristic features of antigen-driven maturation is
detectable in steady-state gaGCs when analysis is focused on strongly
selected gaGC winner clones.
To investigate the influence of commensal diversity on gaGC selection
dynamics, we rederived AID-Confetti mice into germ-free conditions,
in which germinal centres still form18, as well as into stable vertical colo-
nization with a consortium of 12 bacterial strains representing major
phyla present in the mouse gut19 (Oligo-MM12; Extended Data Fig. 3).
When compared with SPF mice, germ-free mice had higher frequen-
cies of germinal-centre B cells in jejunal and ileal mLNs20, and lower
frequencies in duodenal and caecal-colonic mLNs and Peyer’s patches
(Extended Data Fig. 4a, b). Germ-free germinal centres were strongly
skewed away from IgG2b and IgA towards IgG1 (Extended Data Fig. 4c).
Colonization with Oligo-MM12 microbiota partly restored the phenotype
of SPF mice, increasing germinal-centre B cell frequency in distal Peyer’s
patches and IgG2b in proximal Peyer’s patches (Extended Data Fig. 4d).
Multicolour fate-mapping showed that strongly selected germinal
centres accumulated at a markedly faster rate in germ-free mice than
Oligo-MM12 1,000 clone*wells sequences that were strongly selected under each condition (includ-
ing those indicated by named arrows in Fig. 3d, e and Extended Data
Fig. 6a, b; Supplementary Table 1). Of seven monoclonal antibodies
cloned from Oligo-MM12, three (M216, M218, and M220) bound faecal
bacteria fractions from Oligo-MM12-colonized mice (Fig. 3f, i, j), and
one (M218) bound to cultured Enterococcus faecalis, as assessed by
both flow cytometry and dot blotting (Fig. 3g, h). Reversion of somatic
mutations resulted in larger decreases in binding for M216 and M220
and a modest decrease for M218, as shown by flow cytometry and ELISA
(Fig. 3h–j). Thus, as with SPF microbiota, vertical colonization with
Oligo-MM12 triggers efficient antigen-driven maturation towards com-
mensals in steady-state gaGCs.
We subjected the nine monoclonal antibodies obtained from
germ-free winner clones to an array of assays that covered major poten-
tial sources of antigen, including food, autoantigens (anti-nuclear
antibody and intestinal tissue antigens), faecal bacteria, and a standard
polyreactivity panel. None of the germ-free antibodies reacted above
background levels in any of these assays (Extended Data Fig. 6c–f). To
determine whether germ-free germinal centres were indeed populated
in a BCR-dependent manner, or simply stochastically owing to a lack
of antigenic stimulation, we searched for commonalities in the Igh
sequences of winner clones, along with additional sequences obtained
from single germinal-centre B cells from mLN vibratome slices and
whole mLNs and Peyer’s patches of wild-type germ-free mice. This
revealed substantial overlap of clonotype ‘themes’ across individu-
als, which we regarded as unlikely to be random given the small pool
of germinal centres sampled. Two themes were particularly preva-
lent (Fig. 4a, b). One used the relatively rare VH1–47 segment, coupled
to joining segments JH4 or JH2 via an 11–12-amino-acid heavy-chain
complementarity-determining-region 3 (CDRH3) sequence that begins
with the consensus sequence ARGSNY (Fig. 4a). No commonalities in
light-chain usage were detected at this sampling depth. After allow-
ing a one-amino-acid substitution in the ARGSNY motif, this theme
was present in 5 of 7 germ-free mice, accounting for 16.6% of all cells
sequenced (Fig. 4b and Extended Data Fig. 7a). Clones with these char-
acteristics represented only 0.00006% of all reads (1 in roughly 17,000)
in a previously published database of naive B cell Igh sequences from
C57BL6 mice, containing more than 30 million reads representing
2.5 million unique rearrangements from 5 mice21 (P < 2.2 × 10−16 com-
pared with germ-free gaGCs).
A second public clonotype was encoded by the rare VH1–12 segment,
with the stricter seven-amino-acid consensus CDRH3 sequence, AREG-
FAY, followed by JH3 (Fig. 4a). Again, no patterns of light-chain usage
were identified. Allowing a one-amino-acid substitution in CDRH3, this

Nature | Vol 588 | 10 December 2020 | 325

Article
clonotype was present in 2 of 7 germ-free mice, representing roughly
3% of all cells sequenced. This clonotype was also heavily dominant in
all 3 single-coloured germinal centres sorted from different organs
of one of three Oligo-MM12-colonized mice, accounting for 171 of 191
cells sequenced from this mouse (Fig. 4b and Extended Data Fig. 8a).
Despite its short length, the VH1–12/AREGFAY/JH3 combination was
seen only 7 times (in 1 of 5 mice) in the more than 30 million reads of
the naive B cell database21, and only twice more in one other mouse if
a single amino-acid substitution in the AREGFAY motif was allowed
(P < 2.2 × 10−16 compared with germ-free gaGCs). Both clonotypes accu-
mulated somatic mutations that converged across mice and between
SPF and Oligo-MM12 conditions (Extended Data Fig. 7b). In agreement
with their failure to bind food protein extracts (VH1–47 and VH1–12
clonotypes are represented by monoclonal antibodies G082/G226 and
M228/M232, respectively) (Fig. 3d–j and Extended Data Fig. 6), both
clonotypes were also detected in mice fed a custom-made protein-free
chow formulated from purified ingredients (Extended Data Fig. 8b, c
and Supplementary Table 2).
To assess whether reliance on public clonotypes is broadly charac-
teristic of germ-free gaGCs when sampled in an unbiased manner, we
developed a multiwell incidence-based approach to measure clonal
overlaps between mice with high confidence. In total, we sequenced
roughly 80 thousand cells from gaGCs in the mLNs and Peyer’s patches
of 6 germ-free, 6 SPF and 7 Oligo-MM12-colonized mice (Extended
Data Fig. 9a–d). Confirming our initial findings, the VH1–47/ARGSNY
theme (regardless of JH usage) was present in 6 of 6 germ-free and 5 of
7 Oligo-MM12-colonized mice, corresponding to 4.4% (173 of 3,929)
and 2.4% (98 of 4,043) of clone*wells (a number obtained by multiply-
ing each clone by the number of wells it was found in) (Extended Data
Fig. 9a–c) in each condition (Fig. 4c). One such clone was also detected
in a single SPF mouse (5 of 3,629 clone*wells) (Fig. 4c), and 3 others (one
using the JH1 segment) were observed in our photoactivation data (see
Fig. 1 and Supplementary Spreadsheet 1), indicating that full bacterial
colonization is not sufficient to completely exclude such clonotypes
from gaGCs. Clonotype VH1–12/AREGFAY was present, again at a lower
frequency, in 5 of 6 germ-free mice (44 of 3,929 clone*wells) (Fig. 4c).
Analysis of Levenshtein distances showed that, in germ-free and
Oligo-MM12-colonized but not SPF mice, CDRH3 sequences of VH1–47
and VH1–12 clones were closer in sequence to the ARGSNY and AREGFAY
motifs, respectively, than were CDRH3 sequences of clones using other
V segments (Extended Data Fig. 9e). Of note, both VH1–47/ARGSNY/
JH4 and VH1–12/AREGFAY/JH3 rearrangements were found recurrently
in Peyer’s patches of germ-free mice in independent work22 published
while this paper was under review, further underscoring the public
nature of these two clonotypes. Finally, public clonotypes in general
(defined as any recurrent VH/JH combination with exactly matching
CDRH3 amino-acid sequences) were markedly more frequent across
germ-free and Oligo-MM12 gaGCs than across SPF mice (Fig. 4d and
Extended Data Fig. 9f). Although almost all germ-free-associated clo-
notypes were either eliminated or reduced to below detection levels
by SPF colonization, the Oligo-MM12 gaGC repertoire overlapped more
substantially with that of germ-free mice, indicating that Oligo-MM12
colonization is insufficient to completely replace the germ-free
response (Fig. 4d and Extended Data Fig. 9g). Thus, rather than being
stochastically populated, gaGCs display stringent selection driven by
BCR specificity under conditions of low antigenic diversity, resulting
in rapid focusing of germinal centres on dominant lineages and pro-
nounced reliance on clonotypes found repeatedly across different mice.
We have shown that gut-associated germinal centres undergo clonal
selection and antigen-driven maturation in the absence of infection
or immunization, and that the rate of selection varies markedly
326 | Nature | Vol 588 | 10 December 2020
depending on the presence and complexity of the gut microbiota.
Under microbial-replete conditions, selection of highly dominant
clones is relatively rare and is associated with improved binding of
commensal-derived antigens. At the other extreme, gaGC selection
accelerates precipitously in the absence of microbes, leading to strong
convergent selection of IgH clonotypes across mice. Thus, clonal selec-
tion in steady-state gaGCs is a tunable process (see Supplementary
Information for further discussion). The ability to generate specific,
affinity-matured responses to commensals would allow targeted con-
trol of individual bacterial species and may thus play a part in maintain-
ing the composition of the gut microbiota.
Online content
Any methods, additional references, Nature Research reporting sum-
maries, source data, extended data, supplementary information,
acknowledgements, peer review information; details of author con-
tributions and competing interests; and statements of data and code
availability are available at https://doi.org/10.1038/s41586-020-2865-9.
1. Bergqvist, P. et al. Re-utilization of germinal centers in multiple Peyer’s patches results in
highly synchronized, oligoclonal, and affinity-matured gut IgA responses. Mucosal
Immunol. 6, 122–135 (2013).
2. Casola, S. et al. B cell receptor signal strength determines B cell fate. Nat. Immunol. 5,
317–327 (2004).
3. Yeap, L. S. et al. Sequence-intrinsic mechanisms that target AID mutational outcomes on
antibody genes. Cell 163, 1124–1137 (2015).
4. Bunker, J. J. et al. Natural polyreactive IgA antibodies coat the intestinal microbiota.
Science 358, eaan6619 (2017).
5. Reboldi, A. & Cyster, J. G. Peyer’s patches: organizing B cell responses at the intestinal
frontier. Immunol. Rev. 271, 230–245 (2016).
6. Bemark, M. et al. Somatic hypermutation in the absence of DNA-dependent protein
kinase catalytic subunit (DNA-PK(cs)) or recombination-activating gene (RAG)1 activity.
J. Exp. Med. 192, 1509–1514 (2000).
7. Biram, A. et al. B cell diversification is uncoupled from SAP-mediated selection forces in
chronic germinal cnters within Peyer’s patches. Cell Rep. 30, 1910–1922 (2020).
8. Bunker, J. J. & Bendelac, A. IgA responses to microbiota. Immunity 49, 211–224 (2018).
9. Casola, S. & Rajewsky, K. B cell recruitment and selection in mouse GALT germinal
centers. Curr. Top. Microbiol. Immunol. 308, 155–171 (2006).
10. Biram, A. et al. BCR affinity differentially regulates colonization of the subepithelial dome
and infiltration into germinal centers within Peyer’s patches. Nat. Immunol. 20, 482–492
(2019).
11. Tas, J. M. et al. Visualizing antibody affinity maturation in germinal centers. Science 351,
1048–1054 (2016).
12. Victora, G. D. et al. Germinal center dynamics revealed by multiphoton microscopy with a
photoactivatable fluorescent reporter. Cell 143, 592–605 (2010).
13. Livet, J. et al. Transgenic strategies for combinatorial expression of fluorescent proteins in
the nervous system. Nature 450, 56–62 (2007).
14. Shinnakasu, R. et al. Regulated selection of germinal-center cells into the memory B cell
compartment. Nat. Immunol. 17, 861–869 (2016).
15. Meyer-Hermann, M., Binder, S. C., Mesin, L. & Victora, G. D. Computer simulation of
multi-color Brainbow staining and clonal evolution of B cells in germinal centers. Front.
Immunol. 9, 2020 (2018).
16. Tiller, T., Busse, C. E. & Wardemann, H. Cloning and expression of murine Ig genes from
single B cells. J. Immunol. Methods 350, 183–193 (2009).
17. Meffre, E. et al. Surrogate light chain expressing human peripheral B cells produce
self-reactive antibodies. J. Exp. Med. 199, 145–150 (2004).
18. Pollard, M. in Germinal Centers in Immune Responses (eds Odartchenko, N. et al.)
343–348 (Springer, 1967).
19. Brugiroux, S. et al. Genome-guided design of a defined mouse microbiota that confers
colonization resistance against Salmonella enterica serovar Typhimurium. Nat. Microbiol.
2, 16215 (2016).
20. Esterházy, D. et al. Compartmentalized gut lymph node drainage dictates adaptive
immune responses. Nature 569, 126–130 (2019).
21. Greiff, V. et al. Systems analysis reveals high genetic and antigen-driven
predetermination of antibody repertoires throughout B cell development. Cell Rep.
19, 1467–1478 (2017).
22. Chen, H. et al. BCR selection and affinity maturation in Peyer’s patch germinal centres.
Nature 582, 421–425 (2020).
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in
published maps and institutional affiliations.
© The Author(s), under exclusive licence to Springer Nature Limited 2020
Methods
Mice and treatments
Mice were housed at a temperature of 72 °F and humidity of 30–70% in a
12-h light/dark cycle with ad libitum access to food and water. Male and
female mice aged 8–12 weeks at the start of the experiment were used.
PA-GFP transgenic mice were generated in our laboratory12. Rosa26Confetti
(B6.129P2-Gt(ROSA)26Sortm1(CAG-Brainbow2.1)Cle/J) and Rosa26Stop-tdTomato (B6.
Cg-Gt(ROSA)26Sortm14(CAG-tdTomato)Hze/J)23 mice were obtained from Jackson
Laboratories. Rosa26Confetti mice were backcrossed to C57BL/6 mice for
several generations in our laboratory and restricted to the Ighb/b haplo-
type. AicdaCreERT2 (Aicdatm1.1(cre/ERT2)Crey) mice24 were a gift from C.-A. Rey-
naud and J.-C. Weill (Institut Necker, Paris). S1pr2CreERT2 (Tg(S1pr2-cre/
ERT2)#Kuro) BAC-transgenic mice14 were a gift from T. Okada (Riken,
Yokohama) and T. Kurosaki (Univ. Osaka). Mice were bred and main-
tained either in SPF facilities or in germ-free isolators at the Rockefeller
University animal facility. The Oligo-MM12 consortium was a gift from K.
McCoy (Univ. Calgary). We colonized germ-free AID-Confetti breeders
with a single gavage of Oligo-MM12 and monitored colonization (includ-
ing the presence of the entire consortium in successive generations)
by specific amplification of individual bacterial members by quantita-
tive polymerase chain reaction (qPCR; see below). Mice were bred and
maintained in isolators. Vertically colonized AID-Confetti Oligo-MM12
mice were used for all experiments. To deplete protein antigen from the
diet, we used a custom solid diet containing free amino acids (Modified
TestDiet 9GCV with 5% cellulose; composition details are in Supple-
mentary Table 1). Diets were irradiated at more than 45 kGy to ensure
sterility for germ-free conditions and were provided to mice from one
week of age until the time of analysis.
Recombination of floxed alleles in both AID-Confetti and
S1PR2-Tomato mice was induced by two gavages of 10 mg tamoxifen
(Sigma, catalogue number T5648) dissolved in corn oil at 50 mg ml−1,
2 days apart. To ensure that selection was not a function of the route
of administration of either tamoxifen or corn oil, we also injected two
AID-Confetti mice intra-peritoneally once with 10 mg tamoxifen for
analysis at the SPF 21–23 day time point. In germ-free AID-Confetti
mice, tamoxifen was prepared and administered under sterile condi-
tions to mice housed in individually ventilated isocages (TecniPlast).
Sample sizes were not calculated a priori. Given the nature of the
comparisons (mice born under differing microbial colonization sta-
tus), mice were not randomized into each experimental group and
investigators were not blinded to group allocation.
All animal procedures were approved by the Institutional Animal
Care and Use Committee of the Rockefeller University.
Oligo-MM12 qPCR
Colonization of mice by the Oligo-MM12 consortium was confirmed and
monitored over generations by qPCR, using primer pairs specific to
each species as previously validated (individual strain primer sequences
are in Supplementary Table 5, adapted from ref. 19; universal bacterial
qPCR primers are as follows: UNIF340-ACTCCTACGGGAGGCAGCAGT
and UNIR44R-ATTACCGCGGCTGCTGGC). DNA was extracted from fae-
cal samples using the ZR Fecal DNA kit (Zymo Research) according to the
manufacturer’s instructions. Quantitative PCR was performed with the
Power SYBR Green master mix (Applied Biosystems). The average cycle
threshold (Ct) value of two technical replicates was used to quantify
the relative abundance of each species’ 16S ribosomal RNA using the
∆∆Ct method, with the universal 16S rRNA primers serving as controls
between samples. Relative abundance was corrected according to the
genome copy number of 16S rRNA for each species.
Multiphoton imaging and photoactivation
mLNs and Peyer’s patches were collected and imaged as previously
described11. In brief, adipose tissue and excess epithelium were removed
under a dissecting microscope, and mLNs and Peyer’s patches were
mounted in phosphate-buffered saline (PBS) between two coverslips
held together with vacuum grease, as previously described25. Mounted
mLNs and Peyer’s patches were imaged on an Olympus FV1000 upright
microscope with a ×25 Plan water-immersion objective (numerical
aperture (NA) 1.05) and a Mai-Tai DeepSee titanium-sapphire laser
(Spectraphysics). Confetti alleles were imaged at an excitation wave-
length of λ = 930 nm. Fluorescence emission was collected in three
channels, using a pair of CFP (480/40 nm) and YFP (525/50 nm) filters
separated by a 505-nm dichroic mirror to detect CFP, GFP and YFP,
and a dedicated RFP filter (605/70 nm). In situ photoactivation was
performed as previously described12,26. PA-GFP transgenic mice were
injected intravenously with 5 μg of a non-blocking antibody to CD35
(clone 8C12, produced in house) conjugated to Cy3 to label networks of
follicular dendritic cells (FDCs). Clusters of CD35+ cells were identified
by imaging at λ = 950 nm, where photoactivation does not take place,
and three-dimensional (3D) regions of interest were photoactivated by
higher-power scanning at λ = 830 nm. Fragments of lymph nodes were
then processed for flow cytometry as described below.
Image analysis
Colour dominance in AID-Confetti germinal centres was determined in
3D data sets reconstructed using ImageJ software and the Bio-formats
plugin. Cells of each colour or colour combination were counted manu-
ally in two or more 2D slices, at least 20 μm apart, using the Cell Counter
plugin. For overviews of Peyer’s patches, colours were counted in the
same way, but from a single imaging plane. The normalized dominance
score (NDS) was calculated as before11. In brief, we first estimated the
fraction of recombined cells in each germinal centre by calculating
the density of fluorescent cells per area unit (100 μm2) in anatomically
defined dz areas in which cell distribution was homogeneous, then
multiplied the density by the fraction of coloured cells accounted for by
the dominant colour in the germinal centre. Fully recombined germinal
centres have a cell density of very close to 1 (ref. 11), making an NDS of 1 a
good approximation of 100% occupancy by the dominant colour. The
sizes of germinal centres in mLNs and Peyer’s patches were calculated
as the cross-sectional area of the largest available z-section. All image
analysis was carried out in ImageJ.
Lymphocyte flow cytometry and sorting
To evaluate the dynamics of germinal centres and the distribution
of isotypes across space and time, we isolated individual mLNs that
drain the duodenum, jejunum, ileum or caecum/colon as previously
described20. Pairs of Peyer’s patches were isolated from the most proxi-
mal part of the duodenum or the most distal part of the ileum. Cells
were isolated by maceration using disposable micropestles (Axygen)
in 100 μl of PBS supplemented with 0.5% bovine serum albumin (BSA)
and 1 mM EDTA (constituting PBE buffer), and single-cell suspensions
were obtained by two passes through a 70-μm mesh. Cells were stained
with antibodies against B220 protein, T cell antigen receptor (TCR) α/β
chains (or a ‘dump’ mixture containing antibodies against CD4, CD3,
CD8 and NK1.1), CD38, Fas, IgM, IgG1, IgG2b and IgA, supplemented with
Fc block (see Supplementary Table 3) for 30 min on ice. Samples were
run on a FACS LSRII or FACS Symphony (BD Biosciences).
To sort B cells from single mLN germinal centres, we first deter-
mined the localization of single-coloured germinal centres (see ‘Mul-
tiphoton imaging and photoactivation’ section above). As previously
described11, we embedded selected mLNs in 4% low-melt NuSieve GTG
agarose in PBS that had been heated to boiling then cooled to 37 °C
before embedding. We then cut lymph nodes into 300-μm slices using
a Leica VT1000A vibratome. Slices were further dissected under a Leica
M165FC fluorescence stereomicroscope using a double-edged razor
blade to isolate single germinal centres from slices in which several
germinal centres were present. Slice fragments were placed in micro-
centrifuge tubes containing 100 μl PBE, macerated using disposable
micropestles and dissociated into single-cell suspensions by gentle
Article
vortexing. We then added 100 μl of 2× antibody stain (comprising
antibodies against CD38, Fas, B220 and TCR-αβ supplemented with
Fc block; see Supplementary Table 3) to the cell suspension, which was
incubated on ice for 30 min. Single cells were sorted as described below
using FACS Aria II or III cell sorters. Cells positive for any fluorescent
Confetti colour, detected as previously described11, were index-sorted
into 96-well plates containing 5 μl TCL buffer (Qiagen) supplemented
with 1% β-mercaptoethanol11,27. The precise assignment of colours from
index-sorted cells was carried out post-acquisition using Diva software,
version 8.0.2, using all four channels.
For the isolation of single follicles from non-fluorescent germ-free
mice, mice were injected with anti-CD35 (8C12) Alexa Fluor 594, individ-
ually dissected, imaged, and sliced as above. Slices of 250 μm thickness
were examined under a stereomicroscope and manually microdissected
into slice fragments comprising roughly one follicle each, as defined by
staining of FDCs (see the section ‘Multiphoton imaging and photoacti-
vation’ above). Two slice fragments were prepared per mLN, isolated
from slices that were at least 1,000 μm (four 250-μm slices) apart in
the intact node. Slice fragments were prepared for single-cell sorting
as above. For the isolation of single germinal centres from fluorescent
germ-free mice, we combined Confetti fate-mapping with anti-CD35
follicle staining. Germ-free AID-Confetti mice were treated with tamox-
ifen as described above, and, 24 h before imaging, were injected with
5 μg anti-CD35-Cy3 antibody. Mesenteric lymph nodes were sliced
and imaged and fluorescent clusters were manually excised, avoid-
ing any other follicles (marked with anti-CD35 antibody). In this case,
both fluorescent and non-fluorescent germinal-centre B cells were
single-sorted for sequencing.
For single-cell sequencing from whole mLNs and Peyer’s patches,
organs from non-fluorescent germ-free mice were isolated and pro-
cessed into single-cell suspensions by maceration using disposable
micropestles. After straining through a 70-μm mesh, cells were stained
and sorted as above. For incidence-based sequencing experiments,
samples from mLNs and Peyer’s patches were stained as above, and
sorted at 100 cells per well into 16–32 individual wells in a 96-well plate
containing 10 μl TCL buffer per well (Qiagen) supplemented with 1%
β-mercaptoethanol. Analysis of data from flow-cytometry experiments
was carried out using FlowJo software, version 10.5.3 (Tree Star Inc.).
Immunoglobulin sequencing
RNA from single cells or 100-cell pools was reverse-transcribed using
oligo-dT primers as in refs. 11,16. PCR primers were used as previously
described11,28 with the addition of IgA-specific amplification primers
Cα outer (ATCAGGCAGCCGATTATCAC) and Cα inner (GAGCTCGTGGG
AGTGTCAGTG)16. Pooled PCR products were then purified using SPRI
beads (0.7× volume ratio), gel purified and sequenced with a 500-cycle
Reagent Nano kit v2 for single-cell libraries and with a 600-cycle Rea-
gent kit v3 for 100-cell pool libraries on the Illumina Miseq platform.
Sequence analysis
For single-cell Igh analyses, raw paired-end sequences were merged at
the overlapping regions using PANDAseq (version 2.11) for full amplicon
reconstruction29, then processed with the FASTX toolkit. Only those
sequences with high counts for each single cell/well were analysed. For
annotation of V(D)J gene rearrangement, the Ig sequences obtained
were submitted to both HighV-QUEST (version 1.6.9)30 and Vbase2
(ref. 31) databases, choosing the assignment that yielded the lowest
number of somatic mutations in case of discrepancy. Sequences with
common VH/JH genes and the same CDRH3 length were grouped into
clonal lineages when CDRH3 nucleotide identity was 75% or more. Igk
sequences were determined using the same method when needed for
the cloning of monoclonal antibodies. Clonal lineage trees were plot-
ted using GCtree32, with the unmutated V gene sequence of the V(D)J
rearrangement used as an outgroup. Logo plots for public clonotypes
were created by first using the T-Coffee algorithm33 to align all CDRH3
sequences bearing VH1–47 and JH2 or JH4 connected by a 11-amino-acid
or 12-amino-acid CDRH3, or VH1–12 and JH3 connected by a 7-amino-acid
CDRH3. The results of this alignment were then processed using the
WebLogo3 web server34. Public clones were searched in the Greiff
et al.21 database using R. Matching required exact VH, JH and CDRH3
length matches and up to one-amino-acid mismatch in the ARGSNY or
AREGFAY motifs. Dendrograms were generated using ClustalX (version
2.1) and FigTree (version 1.4.4) and branches were coloured in Adobe
Illustrator according to the sequence annotations.
For incidence-based sequencing, raw paired-end sequences were
merged as above and submitted to HighV-QUEST (version 1.6.9) for
annotation30. The output database was then processed in the R envi-
ronment to remove non-functional and out-of-frame sequences. Any
sequences with less than six reads were discarded. Expanded clonal
populations were defined using Change-O35. Briefly, sequences within
the same mouse that shared VH and JH genes and with a maximum CDRH3
hamming distance of four nucleotides were grouped into a single
clone*well of size equivalent to the number of wells this sequence was
found in and retaining the full list of CDRH3 sequences in the cluster
(generating a CDR3list for that clone*well). To remove the possibility
of errors due to sequence misassignment and interwell contamina-
tion, sequences found in a single well within their mouse of origin were
eliminated. CDRH3 sequences were translated to amino-acid sequences
and non-CDRH3 sequences were discarded before assignment of public
clones. Our definition of a public clone required that clones have the
same VH and JH segments, and an exact match between any of the CDRH3
sequences in the clone*wells’ CDR3lists. Public clonotypes VH1–47/
ARGSNY and VH1–12/AREGFAY were identified in this database using
the same criteria as above but without restriction of JH segment usage.
Circular ideogram plots were created using Circos (version 0.69-9), with
each individual mouse represented by a single ideogram bar36. The full
analysis pipeline is available at https://github.com/victoraLab/MIBS.
Levenshtein distances were calculated in R using the stringdist package.
Wells assigned to the same clone within the same mouse were counted
only once to avoid overrepresentation due to clonal expansion. All
sequences with 15 or more reads were used in this analysis, regardless
of the number of wells in which a clone was present. For the VH1–47/
ARGSNY clonotype, distances were calculated starting from the string
ARGSNYXXXXDY and plotted as the resulting difference minus 4 to
correct for the 4 ‘X’ characters.
Production of monoclonal antibodies
Heavy- and light-chain sequences obtained from the same expanded
nodes in somatic hypermutation (SHM) phylogenies of germinal-centre
B cells, as well as their deduced unmutated ancestors, were produced
and assembled into custom mammalian expression vectors (modi-
fied from ref. 16) encoding the human IgG1 and IgK constant regions
(Twist Biosciences). Plasmids were transfected into Freestyle 293-F
suspension cells (obtained from Life Technologies and tested for myco-
plasma contamination in our laboratory), and monoclonal antibodies
were purified using protein-G affinity chromatography, as previously
described11,37. Integrity of all monoclonal antibodies was assayed/quan-
tified by SDS-PAGE and biolayer interferometry on an Octet Red96
instrument using protein-G-coated sensors (FortéBio).
Assays for monoclonal antibody binding
Bacterial flow cytometry was carried out using protocols adapted from
refs. 4,38. Freshly collected faeces from SPF or Oligo-MM12 mice from our
own facility were macerated with micropestles in 100 μl of ice-cold PBS
per 10 mg of faeces and vortexed for 5 min. Large debris was removed
by spinning at 400g for 5 min at 4 °C. Supernatant containing bacteria
was removed and pelleted by centrifugation at 8,000g before staining
with SYTO BC bacterial DNA stain (Thermo Fisher; 1:5,000). For staining
of cultured bacteria, overnight cultures were pelleted at 8,000g and
resuspended in SYTO BC at approximately 109 colony-forming units
(CFU) per millilitre (a density similar to that of faecal bacteria). Stained
bacteria were incubated with monoclonal antibodies at 10 μg ml−1 for
faeces, or 1 μg ml−1 for cultures in PBS supplemented with 0.25% BSA,
for 1 h on ice before washing and incubating with Alexa Fluor 647 con-
jugated goat anti-human IgG1 secondary antibody (Thermo Fisher,
catalogue number A-21445; 1:2,000) with 5% v/v normal goat serum
in PBS plus 0.25% BSA for 30 min. Bacteria were washed in PBS plus
0.25% BSA for 15 min, and 0.25 μg ml−1 4′,6-diamidino-2-phenylindole
(DAPI) was added immediately before sample acquisition. Samples
were run on a FACS Symphony (BD Biosciences), with forward scatter
and side scatter set to logarithmic mode. Within each experiment,
the same number of events was acquired for all samples. Binding was
evaluated in SYTO BC+ DAPI– live bacteria. Data were analysed using
FlowJo software version 8.7 or 10.5.3 (Tree Star Inc.).
Faecal bacterial ELISAs were performed with bacteria isolated from
freshly collected faeces of either Oligo-MM12 or SPF mice. To regulate
the bacterial diversity sampled in SPF mice, faeces from multiple cages
of C57BL/6 mice were pooled and prepared as a single sample. Fae-
ces were macerated with disposable micropestles and centrifuged as
described in the paragraph above for bacterial flow cytometry. Bacteria
were fixed in 0.5% paraformaldehyde for 20 min with continual rotation
to avoid clumping, then washed three times in PBS. Poly-l-lysine-coated
high-binding ELISA plates were incubated overnight with bacterial prep-
arations at an optical density at 600 nm (OD600) of 0.2 for Oligo-MM12
or an OD600 of 0.35 for SPF. Plates were blocked with PBS plus 1% BSA
for 2 h, washed with PBS plus 0.05% Tween, and incubated for 1 h with
serial dilutions of monoclonal antibodies, starting at a concentration of
900 nM. Plates were washed and incubated with a horseradish peroxi-
dase (HRP)-conjugated goat anti-human IgG (Fcɣ-specific) secondary
antibody ( Jackson Immuno Research, catalogue number 109-035-098)
at 1 μg ml−1 for 1 h. Assays were developed using a chromogenic substrate
(Sigma) according to the manufacturer’s instructions. Absorbance
was read at 450 nm using an AccuScan plate reader (Fisher Scientific).
Polyreactivity ELISAs were carried out as previously described39,
with the following modifications. Briefly, high-binding ELISA plates
(Costar) were coated with single-stranded DNA, double-stranded DNA,
lipopolysaccharide (LPS) and keyhole limpet haemocyanin (KLH) at
10 μg ml−1 and insulin at 5 μg ml−1 in PBS overnight at 4 °C. Plates were
washed with PBS plus 0.05% Tween (PBST) and incubated in PBST
for 1 h at room temperature. Plates were incubated with serial dilu-
tions of monoclonal antibodies in PBST for 2 h at room temperature,
starting at a concentration of 10 μg ml−1. Plates were incubated with
HRP-conjugated secondary antibodies, developed and absorbances
read as above. Self-reactivity was tested using a QUANTA Lite ANA ELISA
kit (Inova Diagnostics) according to the manufacturer’s instructions.
For testing of reactivity to food proteins, crude protein extracts were
prepared by grinding 10 g of germ-free autoclaved chow to a powder in
a pestle and mortar, then shaking at 120 r.p.m. overnight in 40 ml PBS
at 37 °C. Extracts were clarified by spinning at 4,000g for 10 min and
then 21,130g for 1 h at 4 °C before filtering through a 0.22-μm filter.
Final protein concentrations were determined using a bicinchoninic
acid (BCA) assay (Pierce). ELISA plates were coated overnight with
100 μg ml−1 of food extract, blocked with PBS plus 2% BSA, and incubated
with serial dilutions of monoclonal antibody in PBS plus 2% BSA start-
ing at 10 μg ml−1 for 2 h. Assays were incubated with HRP-conjugated
secondary antibodies, developed, and absorbances read as above.
For dot blots of bacterial lysates, overnight cultures of bacteria were
centrifuged at 8,000g for 8 min to pellet bacteria and washed once with
brain–heart infusion (BHI) medium, then resuspended at a 10× concen-
tration in BHI. These concentrates were normalized to an OD600 of 4.0,
then mixed 1:1 with Laemmli sample buffer and boiled for 10 min to lyse
cells. Next, 2 μl of prepared bacterial lysates were spotted on nitrocel-
lulose and dried at room temperature overnight. Blots were blocked
in PBST (1× PBS and 0.1% Tween) with 1% BSA for 2 h, washed briefly
with PBST, then stained with 4 μg ml−1 primary antibody (monoclonal
human IgG1s) in PBST with 1% BSA for 2 h. Blots were washed for 5 min
in PBST three times, stained with 1:1,000 peroxidase-conjugated goat
anti-human IgG ( Jackson ImmunoResearch, catalogue number 109-035-
098) in PBST with 1% BSA for 1 h, then washed for 5 min in PBST three
times. Blots were dabbed dry and then incubated for 5 min with the
Clarity enhanced chemiluminescence (ECL) substrate (Bio-Rad); chemi-
luminescence images were acquired with an ImageQuant LAS4000
(GE) using the same exposure time for all blots.
Self-reactivity was assessed in western blots that probed whole-
tissue lysates. The small intestine was dissected from wild-type SPF
mice and a 1-cm segment of the distal ileum was removed, cut open
longitudinally, and washed thoroughly with PBS to remove lumenal
content. The tissue was snap-frozen and then bead-beat directly in Lae-
mmli buffer to pulverize the tissue, lyse the cells, and reduce/denature
proteins in one step. Samples were boiled for 10 min and then centri-
fuged at 16,00g for 5 min. Supernatants were run on 4–20% Tris-glycine
gels with SDS, and transferred onto polyvinyldifluoride (PVDF)
membranes. Blots were blocked and stained using the procedure
described above for dot blots.
Bacterial culture
A list of bacterial strains is provided in Supplementary Table 4. Bacteria
were grown in an anaerobic atmosphere of 10% carbon dioxide, 5%
hydrogen, and 85% nitrogen. Members of the Oligo-MM12 consortium
were cultured in BHI (Becton Dickinson) supplemented with 5 μg ml−1
hemin (Sigma), 5 μg ml−1 vitamin K1 (Sigma), 250 mg l−1 cysteine,
250 mg l−1 sodium sulfide, and 4 g l−1 porcine mucin type 3 (Sigma)
(the latter only for Akkermansia muciniphila YL44), with the excep-
tion of Lactobacillus reuteri I49, which was grown in MRS agar (Becton
Dickinson).
Statistical analysis
No statistical methods were used to predetermine sample size. Unless
otherwise noted, statistical calculations were performed using the tests
described in the figure legends in GraphPad Prism version 8.3.0. The
Chao1 formula40 was used to estimate total clonal richness in photo-
activated germinal centres, calculated using the EstimateS package41.
Proportions of clones bearing public-clonotype motifs in our sample
versus the Greiff et al.21 database were compared using Fisher’s exact
test calculated in R. For the ARGSNYXXXXDY CDRH3, Levenshtein dis-
tances were compared between conditions; for the AREFGAY CDRH3,
distances were compared between VH1–12 and all clones using a Mann–
Whitney test. Exact binomial confidence intervals were calculated
using the JavaStat online tool at https://statpages.info/confint.html.
Reporting summary
Further information on research design is available in the Nature
Research Reporting Summary linked to this paper.
Data availability
Incidence-based sequencing raw and processed data are available
through BioProject (https://www.ncbi.nlm.nih.gov/bioproject/; identi-
fication code PRJNA647715); the analysis pipeline is available at https://
github.com/victoraLab/MIBS.
23. Madisen, L. et al. A robust and high-throughput Cre reporting and characterization
system for the whole mouse brain. Nat. Neurosci. 13, 133–140 (2010).
24. Dogan, I. et al. Multiple layers of B cell memory with different effector functions. Nat.
Immunol. 10, 1292–1299 (2009).
25. Liu, K. et al. In vivo analysis of dendritic cell development and homeostasis. Science 324,
392–397 (2009).
26. Shulman, Z. et al. T follicular helper cell dynamics in germinal centers. Science 341,
673–677 (2013).
27. Trombetta, J. J. et al. Preparation of single-cell RNA-seq libraries for next generation
sequencing. Curr. Protoc. Mol. Biol. 107, 4.22.1-17 (2014).
28. Mesin, L. et al. Restricted clonality and limited germinal center reentry characterize
memory B cell reactivation by boosting. Cell 180, 18–20 (2020).
Article
29. Lefranc, M. P. et al. IMGT, the international ImMunoGeneTics information system. Nucleic
Acids Res. 37, D1006–D1012 (2009).
30. Masella, A. P., Bartram, A. K., Truszkowski, J. M., Brown, D. G. & Neufeld, J. D. PANDAseq:
paired-end assembler for illumina sequences. BMC Bioinformatics 13, 31 (2012).
31. Retter, I., Althaus, H. H., Münch, R. & Müller, W. VBASE2, an integrative V gene database.
Nucleic Acids Res. 33, D671–D674 (2005).
32. DeWitt, W. S., III, Mesin, L., Victora, G. D., Minin, V. N. & Matsen, F. A., IV. Using
genotype abundance to improve phylogenetic inference. Mol. Biol. Evol. 35,
1253–1265 (2018).
33. Notredame, C., Higgins, D. G. & Heringa, J. T-Coffee: a novel method for fast and accurate
multiple sequence alignment. J. Mol. Biol. 302, 205–217 (2000).
34. Crooks, G. E., Hon, G., Chandonia, J. M. & Brenner, S. E. WebLogo: a sequence logo
generator. Genome Res. 14, 1188–1190 (2004).
35. Gupta, N. T. et al. Change-O: a toolkit for analyzing large-scale B cell immunoglobulin
repertoire sequencing data. Bioinformatics 31, 3356–3358 (2015).
36. Krzywinski, M. et al. Circos: an information aesthetic for comparative genomics. Genome
Res. 19, 1639–1645 (2009).
37. Pasqual, G., Angelini, A. & Victora, G. D. Triggering positive selection of germinal center B
cells by antigen targeting to DEC-205. Methods Mol. Biol. 1291, 125–134 (2015).
38. Palm, N. W. et al. Immunoglobulin A coating identifies colitogenic bacteria in
inflammatory bowel disease. Cell 158, 1000–1010 (2014).
39. Mouquet, H. et al. Polyreactivity increases the apparent affinity of anti-HIV antibodies by
heteroligation. Nature 467, 591–595 (2010).
40. Chao, A. Nonparametric-estimation of the number of classes in a population. Scand. J.
Stat. 11, 265–270 (1984).
41. Colwell, R. K. EstimateS: Statistical estimation of species richness and shared species
from samples. Version 9. http://purl.oclc.org/estimates (2013).
Acknowledgements We thank all members of the Victora and Mucida laboratories, past and
present, for assistance with experiments, fruitful discussions and critical reading of the
manuscript. In particular we thank: A. Rogoz and G. Fayzikhodjaeva for maintaining
gnotobiotic mice; S. Gonzalez for maintaining SPF mice; K. Gordon and K. Chhosphel for FACS;
the Rockefeller University Genomics Center for RNA sequencing; and employees of the
Rockefeller University for continuous assistance. We thank A. Vale (Universidade Federal do
Rio de Janeiro, Brazil) for help with analysis of public clonotypes; K. McCoy (University of
Calgary, Canada), S. Y. Wong and K. Cadwell (New York University, USA) for providing
Oligo-MM12 strains; J. Faith (Mount Sinai School of Medicine, USA) for other bacterial strains;
and J Däbritz and E. Wirthgen (University of Rostock, Germany) for contributing to the training
and supervision of C.W. This work was supported by National Institutes of Health (NIH)/
National Institute of Allergy and Infectious Diseases (NIAID) grants R01AI119006 and
R01AI139117 (to G.D.V.), and NIH/NIAID/National Institute of Diabetes and Digestive and Kidney
Diseases (NIDDK) grants R01DK093674, R01DK113375 and R21AI144827 (to D.M.), with
additional support from NIH grant DP1AI144248 (Pioneer Award) and from the Robertson
Foundation to G.D.V. and NIH grant R01DK116646 (Transformative Award) to D.M. C.R.N. is a
Human Frontier of Science Program postdoctoral fellow. G.P.D. is a Robert Black Fellow of the
Damon Runyon Cancer Research Foundation. A.S. is a Boehringer-Ingelheim Fonds PhD fellow.
G.D.V. and D.M. are Burroughs-Wellcome Investigators in the Pathogenesis of Infectious
Disease. G.D.V. is a Searle Scholar, a Pew-Stewart Scholar, and a MacArthur Fellow.
Author contributions C.R.N. and L.M. performed all mouse and antibody-sequencing
experiments, with help from T.A. C.W. and C.R.N. produced and assayed the reactivity of
monoclonal antibodies by ELISA, with help from A.S. G.P.D. stained monoclonal antibodies in
cultured bacteria and carried out and dot and western blots. A.M.B. optimized and performed
flow cytometry of faecal bacteria. A.A.K.L. established the protein-free-diet protocol. L.M. and
T.B.R.C. designed and performed all bioinformatics analyses. C.R.N., D.M. and G.D.V.
conceptualized the study, designed all experiments, and wrote the manuscript with input from
all authors.
Competing interests The authors declare no competing financial interests.
Additional information
Supplementary information is available for this paper at https://doi.org/10.1038/s41586-020-
2865-9.
Correspondence and requests for materials should be addressed to D.M. or G.D.V.
Peer review information Nature thanks Ramy Arnaout, Rachael Bashford-Rogers and Hai Qi for
their contribution to the peer review of this work.
Reprints and permissions information is available at http://www.nature.com/reprints.
Extended Data Fig. 1 | Clonal replacement in steady-state gaGCs. a, Gating
strategy for PA-GFP mice used in Fig. 1a–d. GC, germinal centre. b, Gating
strategy and efficiency of labelling in germinal centres of AID-Confetti mice
seven days after the administration of tamoxifen, as in Fig. 1e. The labelling
efficiency is calculated as 100% minus the product of the percentage of
unlabelled cells in the GFP/YFP, RFP and CFP channels. c, Gating strategy for the
S1pr2CreERT2/tdTomato fate-mapping experiments shown in g and Fig. 1h. All
flow plots are representative of multiple experiments. d, Multiphoton images
of Peyer’s patches from SPF mice at different times after tamoxifen treatment
(see Fig. 1f). Values in parentheses in images are NDS values. e, Quantification
of multiple images as exemplified in d (see also Fig. 1g). Data are from three to
five mice per time point at days 14–35, and one to two mice per time point for
day 7 and later times. f, Size of germinal centres in Peyer’s patches (PPs) versus
mLNs, calculated from samples obtained seven days after tamoxifen treatment
as in d and Fig. 1f, plotted as the cross-sectional area of the largest available
z-section. Each symbol represents one germinal centre. Lines show medians;
P values are from two-tailed Mann–Whitney U tests. Data are pooled from
multiple mLNs and Peyer’s patches of two mice from two independent
experiments. g, Turnover of B cell clones in germinal centres of Peyer’s patches
from S1pr2CreERT2 × Rosa26Stop-tdTomato mice (see Fig. 1h).
Article
Extended Data Fig. 2 | Binding characteristics of ‘winner’ gaGC clones from
SPF mice. a, Gating strategy for isolating AID-Confetti single germinal centres
shown in b, c, Figs. 2a, 3d, e and Extended Data Figs. 6a, b, 8a. CR, CFP and/or
RFP; non CR, non-CFP, non-RFP, GY, GFP and/or YFP. b, c, Additional Igh
sequence relationships among B cells from high-NDS germinal centres (b) and
one low-NDS germinal centre (c) (see Fig. 2a). Scale bars, 50 μm. In c, each tree is
for a separate clone (defined as a unique V(D)J rearrangement). Only clones
with more than five cells are shown (grey slices in pie charts). d, Gating strategy
for bacterial flow cytometry, performed in e, Figs. 2b, 3f, h, i and Extended Data
Fig. 6d. e, Flow-cytometry analysis of the binding of recombinant monoclonal
antibodies to faecal bacteria isolated from SPF mice. Plots gated as in d. All
plots are representative of data obtained from at least two separate
experiments. f, Summary of the reactivity of SPF monoclonal antibodies,
assayed by ELISA against food protein extracts, autoantigens (anti-nuclear
antibody, ANA), and a five-antigen polyreactivity panel comprising
single-stranded DNA, double-stranded DNA, keyhole limpet haemocyanin
(KLH), insulin and LPS. Shown are background-subtracted OD450 values. Data
representative of assays repeated in at least three separate experiments.
 a Quantitative PCR, total 16S b Quantitative PCR, species-specific 16S c Quantitative PCR, species-specific 16S
100 50
SPF Species
10 Oligo-MM12 40 LOD Clostridium clostridioforme YL32
Clostridium innocuum I46

Ct value
30
1 Acutalibacter muris KB18
20 Akkermansia muciniphila YL44
0.1
10 Bacteroides caecimuris I48
Blautia coccoides YL58
0.01 0
Feces Cecum Enterococcus faecalis KB1

6S
6

27
8
9
KB 1
18

YL 2

YL 1
YL 2
YL 4
YL 5
58
Flavonifractor plautii YL31

I4
I4
I4
KB

YL

3
3
4
4
YL
l1
ta
Muribaculum intestinale YL27

To
Lactobaccilus reuteri I49
Turicimonas muris YL45
P F1

Extended Data Fig. 3 | Stable vertical transmission of the Oligo-MM12
consortium. a–c, qPCR of total (a) and strain-specific (b, c) 16S DNA from
faecal samples of mice stably colonized with the Oligo-MM12 consortium. In
a, ΔCt values were calculated in respect to a reference SPF sample, marked
by the black filled symbols, with which all other values were compared.
In c, Ct values were used to quantify the relative abundance of each species
(see Methods). LOD, limit of detection. F1 refers to the first generation after the
parental strain (P, colonized by gavage). Note that Bifidobacterium animalis
(YL2) is usually undetectable in faeces19.
Article
Extended Data Fig. 4 | Frequency and isotype distribution of gaGCs in
germ-free and Oligo-MM12-colonized mice. a, Gating strategy for analysing
the frequency of germinal centres and distribution of isotypes (results shown
in b–d). b, Frequency of cells with the phenotype of germinal centres (CD38–
FAShi) among total B220+ B cells in the indicated organs of mice raised under
the indicated conditions. Each symbol represents one mouse. SPF, n = 25;
germ-free (GF), n = 16; Oligo-MM12, n = 11. c, Frequency of germinal-centre B
cells positive for the indicated surface BCR isotype in different organs of mice
raised under the indicated conditions. Data are from at least three mice per
group, as in d. Data are presented as means ± s.e.m. d, Statistical analysis of
selected isotypes and anatomical locations, using data from c. Each symbol
represents one mouse. Lines indicate medians; P values are obtained from
two-tailed Kruskall–Wallis tests carried out on each trio, with Dunn’s multiple
comparisons post-test. All P values below 0.05 are reported.
 a Individual GC gating strategy Oligo-MM12 mLN, day 21 post-tamoxifen Peyer’s patch, day 21 post-tamoxifen

All cells Lymphocytes Single cells B cells

250K 250K 5 4
10 10
Lymphocytes

TCRβ APC-Cy7
200K 200K Single Cells 4
33.8 10 B cells

FAS PE-Cy7
98.2 3
150K 150K 86.4 10
3 GC
100K 100K 10 4.9
SSC-A

FSC-H
0 0
50K 50K
-103
0 0
0 100K 200K 0 100K 200K 4 4
0 10 0 10
FSC-A FSC-A B220 BV421 CD38 APC

b AicdaCreERT2/+.Rosa26Confetti/Confetti (day 7 post-tamoxifen), individually dissected GCs

GF1 GF2 GF 3 GF4 GF5

J J J C C J J C C C J C J J J f Oligo-MM12 g Oligo-MM12 day 20-23
80
Colored cell Color Density x dominance PSPFxMM12 = 0.055
70 density dominance (density > 0.4 only) PGFxMM12 = 0.208
Number of cells sequenced

60 1.0 100 1.0 100

Normalized dominance score

Fluorescent cells/100 µm2

% of fluorescent B cells
50 0.8 80 0.8

% GCs > NDS 0.75

75

Mesenteric LN
40
0.6 60 0.6
30 50
0.4 40 0.4
20
25
10 0.2 20 0.2

0 0.0 0 0.0 0
Dominant clone Expanded clones Singletons

2
F

F
M1
-1

-2

-1

-2

-1

-2

SP

G
14

20

14

20

14

20

M
Colored cell Color Density x dominance PSPFxMM12 = 0.015
c Clonal Clonal Size of d density dominance (density > 0.4 only) PGFxMM12 > 0.999
richness diversity largest clone P < 0.0001 100 100
D50 (fraction of clones accounting

103 0.4 100 100 1.0 1.0

Normalized dominance score

% GCs with largest clone > 50%

Fluorescent cells/100 µm2

for 50% of sequenced cells)

% of fluorescent B cells
80
Clones per GC (Chao1)

80 0.8 0.8 75
% of cells in the GC

%GCs > NDS 0.5

Peyer’s patches
0.3 75
102 0.6 60 0.6
60
50
0.2 50 40
0.4 0.4
40
101 25
0.1 25 0.2 20 0.2
20
0.0 0 0.0 0
100 0 0 0

2
8

F
M1
mLN GCs mLN GCs mLN GCs
-1

-2

-1

-2

-1

-2

SP

G
14

20

14

20

14

20
(P PF

lic GF

M
)
A)

es
S

Days post-tamoxifen
(s

Extended Data Fig. 5 | Clonal selection in germ-free and Centre bars represent the proportion in the sample; error bars show the exact
Oligo-MM12-colonized mice. a, Gating strategy for germ-free AID-Confetti binomial 95% confidence interval. e, Multiphoton images of Oligo-MM12 mLNs
single germinal centres used in b–d. b–d, Sequencing of Igh genes from B cells and Peyer’s patches at different times after treatment with tamoxifen. Blue
obtained from individual mLN germinal centres. Germinal-centre B cells were represents collagen (second harmonics); white shows autofluorescence; other
single-cell-sorted from fragments of vibratome slices containing single colours are from the Confetti allele. Scale bars, 200 μm (overviews), 50 μm
germinal centres. To avoid biased selection of germinal centres based on NDS (close-ups). N/D, NDS not determined owing to a low density of coloured cells.
or loss of germinal centres with a low density of coloured cells, mLNs were f, Quantification of images as in e for mLNs (top) and Peyer’s patches (bottom).
harvested at five to seven days after treatment with tamoxifen, before Each symbol represents one germinal centre. Medians are indicated. Only
extensive selection or clonal turnover; both fluorescent and non-fluorescent germinal centres with a density of more than 0.4 fluorescent cells per 100 μm2
cells were included in the sample. This unbiased selection ensures that data are are included in the NDS calculations. g, Proportion of germinal centres with
comparable to those obtained using in situ photoactivation (Fig. 1a–d), which NDS values of more than 0.75 in mLNs (top) and more than 0.5 in Peyer’s patches
we could not perform because the photoactivatable GFP-transgenic strain is (bottom) under SPF, germ-free and Oligo-MM12 conditions at 20–23 days
not available under germ-free status. b, Clonal composition of individual after tamoxifen; SPF and germ-free data are as in Fig. 3c. For SPF, Oligo-MM12
germinal centres from five mice (GF1–GF5). C, caecal-colonic mLN; J, jejunal and germ-free mLN gaGCs, n = 57, 16 and 27, respectively; for gaGCs from
mLN. c, Quantification of data from b. Each symbol represents one germinal Peyer’s patches, n = 20, 10 and 9, respectively. P values obtained by two-tailed
centre. d, Proportion of germinal centres in which the largest clone accounts Fisher’s exact tests. Error bars represent exact binomial 95% confidence
for more than 50% of all B cells in mLNs of SPF mice (data from Fig. 1b) and intervals. All data are from three to five mice per time point.
germ-free mice (data from b). P values are from two-tailed Fisher’s exact tests.
Article
a Germ-free UA Oligo-MM12
UA
M220.U
M220
13
7 RFP
2 3 2 2 3 n n cells with
2 identical
2 sequence
2 12
3 Inferred
mLN mLN precursor
86/88 24/44 1 mutation

c GF mAbs: ELISA panel d GF mAbs, flow cytometry on SPF fecal bacteria e GF mAbs, ELISA on SPF fecal bacteria f GF mAbs, WB on SPF ileum protein extract

-)
G082 2.0

9 +)

82 (
G 053
ED ry
3H 38(
G196 G200 2.0

2
G 6
G 8
G 0
G 2
G 4
G 6
G 8
G 6
M 8
2ndary

a
Abs. 450 nm

GF mAbs (n = 9)

23
19
16
20
20
20
20
20
22
22
2 nd

G
G198 1.5

G
G202 ED38
G200 1.5

Abs. 450 nm
1.0 MG053 (- ctrl.) MG038
G202
G204 G204
0.5 1.0
G206 G082
0 G206
G208 0.5
G226 G196 G208
ED38(+) 0.0
G198 G226 10-1 100 101 102 103
MG053(-)
5 5
0 10 3 10 4 10 0 10 3 10 4 10
mAb concentration (nM)
od

ss NA
ds NA
A
In S
lin
H

Anti-human IgG1 Alexa Fluor 647

N
LP

KL
su
Fo
A
D
D

Extended Data Fig. 6 | Characteristics of ‘winner’ gaGC clones from
germ-free and Oligo-MM12-colonized mice. a, b, Additional Igh sequence
relationships among B cells from high-NDS germinal centres of germ-free (a)
and Oligo-MM12-colonized (b) mice. Details are as in Fig. 2a. Scale bars, 50 μm.
c, Reactivity summary of germ-free monoclonal antibodies assayed by ELISA
against food protein extracts, autoantigens (anti-nuclear antibody, ANA), and a
five-antigen polyreactivity panel. Shown are background subtracted OD450
values. d, Flow-cytometry analysis of the binding of monoclonal antibodies
from germ-free mice to faecal bacteria from SPF mice. Details are as in Fig. 2b.
e, ELISA analysis of the binding of monoclonal antibodies from germ-free mice
to faecal bacterial fractions from SPF mice. MG053 was assayed at three
dilutions only. Other monoclonal antibodies were assayed at dilutions
indicated on the x-axis. Lines show the means of two assays. f, Western blot
(WB) analysis of the binding of monoclonal antibodies from germ-free mice to a
protein extract from mouse ileum tissue, run on a single-well 4–15% gel and
blotted using a multiwell mask. Monoclonal antibody 3H9 is a DNA-specific
negative control. Data in c–f are representative of two or more independent
experiments.
 a b VH1-47 VH1-12
Cell # Cell #

9
Replacements Replacements

45
17
19
31
87
51

59
17
10
9-A . PTPP. 10-E DD.
ARRSN(Y/F)/12 10-E AD. DDD 30-T I . A
11-L . Q. V V I 31-S NNN
Clonotype VH1-47 34-M I VI
12-V . . L . L M
ARGSNY/12 35-H
13-K N. RER. . YY
JH 4 16-A . T T DD. 37-V I . I
19-K . MR R . . 41-P T. S
ARGGFY/11
23-K . R. ER. 43-Q E. K
JH2 28-T . . I I NS 50-A GGV
33-P . R. SF . 52-Y F. H
ARGSNY/11 ARGSSY/11 34-I . V ML . M 55-N NDD

Targeted AA
35-E D. . HD. 58-T A. N
ARGTNY/12 37-M . I L I V. 60-Y N. F
38-K RR. R. . 62-Q P. H
ARGSNF/(11/12) 39-Q . RR. HR 63-K R. Q
40-N . SCSS. 66-G DDD
41-H . L . L P. 67-K M. R
42-G EEEER. 72-V . I A
ARGSNY/12 43-K . EEER. 77-S C. N
50-N SS. SSD 78-T . KK
54-Y . FFFSS 79-A . VV
ARGSNY/12 55-N SDDDDD
0.01 80-Y F. F
57-D EANEE. 83-L

Targeted AA
F VF
58-T I AI AAI 84-S N. T
59-K . NNNNQ
88-S F. F
60-Y . CC. F C
90-D E NE
61-N . S. SDD
94-Y S. S
62-E . . . ADD
63-K MN E N N N

G 5
M F7
12
M1
F
65-K ERRRR.

G
Clonotype VH1-12 66-G . ADDAD
Mouse
AREGFVY 67-K RR. RRR
69-T . S. AA.
VREGFAY 70-L V . . MMM
72-V . . . I AA
73-E . DAD. D >50

Frequency %
74-K . . I . . R 40
TREGFAY 77-S CNNNNN 30
78-T . . . SSK 20
AREGFTY Mouse:
AREGFAH 80-Y . F SNF F 10
AREGFAF GF 1 82-E . DA DGD 0
GF 2 83-L . V. V. V

2
F
M1
G
0.01 GF 3 84-S . C. GG.

M
GF 5 86-L . F. VI S
GF 7 87-T . I . I AI
92-A . . . GDV
MM12 1 93-V . I I I I I
MM12 2 94-Y . C. F F .
95-Y . FFFFF
G 1
G 2
F3
G 5
M F7
32
F
F

F
M1
G

Mouse

Extended Data Fig. 7 | Mutational patterns in germ-free/Oligo-MM12 public
clonotypes. a, Dendrograms showing the sequence relationships between
VH1–47 and VH1–12 clones in different mice. All clones with up to two-amino-acid
differences from the public-clonotype CDR H3 motifs are included. b, Heat
maps showing the frequency of amino-acid replacements along the VH1–47 and
VHH1–12 families in germ-free (blue) and Oligo-MM12 (green) mice, using the
same data as in Fig. 4b. Only mice with more than two cells within the specified
clone were included in the analysis. The number of cells analysed per mouse is
indicated at the top of each column. Only those amino acids mutated in at least
three (VH1–47) or two (VH1–12) mice are listed on the left, using Immunogenetics
(IMGT; http://www.imgt.org) numbering; to the right, the most frequent
amino-acid replacement in each mouse is given. Arrows indicate recurrent
amino-acid mutations found in five of six mice (VH1–47) or three of three mice
(VH1–12).
Article
Oligo-MM12 -colonized mouse

VH1-12/AREGFAY/JH3

UA

Cecal mLN
16 PP (M232) Ileal mLN
2 3 4 6 3
16 31

2
12 (M228) 27
3
CFP YFP
3 3
RFP CPF/YFP
n n cells with identical
sequence
Inferred precursor
1 mutation
PFC 1 PFC 2 PFC 3
c
D I C PP I C PP C
b GC frequency 80 VH1-47/ARGSNY.../JH4
VH1-47/ARGSNY.../JH2
10 30 70
SPF VH1-47/ARGSNY.../JH3
Cells sequenced

GF 60 Other VH1-47 H3
8
PFC
50 VH1-12/AREGFAY/JH3
20
% of B cells

6 Other VH1-12
40
Other V segments
4 30
10
20
2
10

0 0 0
D mLN J mLN I mLN C mLN pPP dPP
Whole organs (WT)

Extended Data Fig. 8 | Stereotypical germ-free IgH clonotypes are present
in Oligo-MM12 and germ-free/dietary-protein-free conditions. a, Massive
expansion of a public VH1–12 clonotype across different secondary lymphoid
organs of mouse MM12 1 (from Fig. 4b), at 21 days after tamoxifen treatment.
Multiphoton images show all three germinal centres sequenced from this
mouse (yellow dotted boxes), magnified in the side panels. Scale bars, 200 μm
(overviews) and 50 μm (close-ups). mLN close-ups are from different image
acquisitions of the same germinal centre. A clonal tree of all cells from this
clone is shown at the bottom right. Arrowheads indicate clonal bursts and the
organ of origin of cells with that particular sequence. b, Frequency of cells with
a germinal-centre phenotype (CD38dim FAShi) among total B220+ B cells in the
indicated organs of mice raised on protein-free chow (PFC). Data for SPF and
germ-free mice are reproduced from Extended Data Fig. 4b. Each symbol
represents one mouse. For PFC, n = 8 mice. c, Clonal distribution of germinal-
centre B cells sequenced from the indicated tissues of three separate mice
(PFC1–3), with public clonotypes colour-coded. See also Fig. 4b. C, caecal
colonic mLN; D, duodenal mLN; I, ileal mLN; PP, Peyer’s patch.
Extended Data Fig. 9 | See next page for caption.
Article
Extended Data Fig. 9 | Multiwell incidence-based Igh sequencing reveals
clonal overlap among individual mice and between microbial colonization
conditions. a, Overview of the incidence-based Igh sequencing method used
for c–g and Fig. 4c, d. To identify expanded public clonotypes among gaGC
samples from multiple mice with high confidence, we developed an incidence-
based sequencing strategy based on repeated sampling of the same germinal-
centre B cell population. We sorted multiple samples of 100 germinal-centre B
cells (usually 32 for mLN and 16 for Peyer’s patches) from 6 germ-free, 6 SPF, and
7 Oligo-MM12-colonized mice, and sequenced all BCRs in each sample, for a
total of roughly 80 thousand input B cells, plus 32 wells each of non-germinal-
centre B cells from the mLN of 3 germ-free and 3 SPF mice as controls. To avoid
counting as ‘public’ sequences that were spuriously present in different mice
owing to barcode misassignment or DNA contamination, we included in our
analysis only those clones that were represented by more than five reads in
any single well and found in at least two wells from the same sample. Key
bioinformatics steps are described in the figure; see Methods for a full
description of the bioinformatic pipeline. b, Gating strategy used for data in
c–g and Fig. 4c, d, described in a. c, Number of distinct clones per well, after
collapsing sequences with matching VH, JH, and CDR H3 nucleotide sequences.
Each symbol represents one well. Boxes represent medians and interquartile
ranges. As expected, non-germinal-centre B cell samples had many more total
clones per well than did germinal-centre B cells. d, Proportion of expanded
clones (present in more than one well per sample) in germinal-centre and non-
germinal-centre samples from mLNs and Peyer’s patches of mice held under
the specified conditions. e, Histograms showing Levenshtein distances
between the indicated consensus CDR H3 sequence and the CDR H3 sequence of
all clones in the indicated category. For ARGSNYXXXXDY, distances are plotted
for clones carrying the ‘correct’ VH1–47 gene or two ‘control’ VH regions with
similar usage frequency in our sample. P values were obtained by Kruskall–
Wallis test comparing all three conditions. Owing to the very low number of
total VH1–12 clones outside of the germ-free condition, distances to the
AREGFAY CDR H3 are compared between VH1–12 clones and all clones. P values
obtained by two-tailed Mann–Whitney U test. f, Fraction of clone*wells
containing public clonotypes in each condition, pooled from all mice. P values
were obtained by Fisher’s exact test. g, Venn diagram showing the number of
clones per condition (pooled from all mice) and overlap between conditions.
The clone in the centre of the graph (SPF/Oligo-MM12/germ-free overlap)
corresponds to the VH1–47 public clonotype. In f, g, data are as in Fig. 4d.
 nature research | reporting summary
Gabriel Victora
Corresponding author(s): Daniel Mucida
Last updated by author(s): Jul 20, 2020

Reporting Summary
Nature Research wishes to improve the reproducibility of the work that we publish. This form provides structure for consistency and transparency
in reporting. For further information on Nature Research policies, see Authors & Referees and the Editorial Policy Checklist.

Statistics
For all statistical analyses, confirm that the following items are present in the figure legend, table legend, main text, or Methods section.
n/a Confirmed
The exact sample size (n) for each experimental group/condition, given as a discrete number and unit of measurement
A statement on whether measurements were taken from distinct samples or whether the same sample was measured repeatedly
The statistical test(s) used AND whether they are one- or two-sided
Only common tests should be described solely by name; describe more complex techniques in the Methods section.

A description of all covariates tested

A description of any assumptions or corrections, such as tests of normality and adjustment for multiple comparisons
A full description of the statistical parameters including central tendency (e.g. means) or other basic estimates (e.g. regression coefficient)
AND variation (e.g. standard deviation) or associated estimates of uncertainty (e.g. confidence intervals)

For null hypothesis testing, the test statistic (e.g. F, t, r) with confidence intervals, effect sizes, degrees of freedom and P value noted
Give P values as exact values whenever suitable.

For Bayesian analysis, information on the choice of priors and Markov chain Monte Carlo settings
For hierarchical and complex designs, identification of the appropriate level for tests and full reporting of outcomes
Estimates of effect sizes (e.g. Cohen's d, Pearson's r), indicating how they were calculated
Our web collection on statistics for biologists contains articles on many of the points above.

Software and code

Policy information about availability of computer code
Data collection BD FACSDiva software v8.0.2 was used for flow cytometry data acquisition. Sequencing data was acquired using the Illumina MiSeq
platform. Microscopy was acquired on the Olympus FluoView v4.2 acquisition software.

Data analysis Data was analysed using FlowJo (TreeStar) v8.7 and v10.5.3, Prism (GraphPad) v8.3.0, ImageJ v1.51 and R v3.6.3. Sequencing analysis was
carried out using PANDASeq v.2.11, HighVQUEST v. 1.6.9, VBASE2, FASTX Toolkit v0.0.13, Change-O v0.4.6, the T-coffee algorithm
(Notredame et. al. 2000) and GCtree v1 (deWitt et. al. 2018). Circular ideogram plots were created using Circos v. 0.69-9. Dendograms
were generated using clustalx v2.1 and FigTree v1.4.4. Data was presented using Adobe Illustrator v23.0.4 and 15.1.0.
For manuscripts utilizing custom algorithms or software that are central to the research but not yet described in published literature, software must be made available to editors/reviewers.
We strongly encourage code deposition in a community repository (e.g. GitHub). See the Nature Research guidelines for submitting code & software for further information.

Data
Policy information about availability of data
All manuscripts must include a data availability statement. This statement should provide the following information, where applicable:
- Accession codes, unique identifiers, or web links for publicly available datasets
- A list of figures that have associated raw data
October 2018

- A description of any restrictions on data availability

Single cell sequences are available as a supplementary spreadsheet. These refer to Figures 2, S2, 3, 4, 5 and S6 as labeled in the spreadsheet. Incidence-based
sequencing data is available at https://github.com/victoraLab/MIBS.

1
 nature research | reporting summary
Field-specific reporting
Please select the one below that is the best fit for your research. If you are not sure, read the appropriate sections before making your selection.
Life sciences Behavioural & social sciences Ecological, evolutionary & environmental sciences
For a reference copy of the document with all sections, see nature.com/documents/nr-reporting-summary-flat.pdf

Life sciences study design

All studies must disclose on these points even when the disclosure is negative.
Sample size No statistical method were used to determine sample size. Each experiment was repeated across multiple animals.

Data exclusions Data only excluded for technical reasons. In bacterial flow cytometry data was excluded if the background secondary antibody binding was too
high and no mAb binding could be observed (experiments appear in figures 2, S2 and 4.)

Replication All experiments were reproducible and were repeatable as detailed in figure legends. Sequencing in Figure 5c-g was carried out once, but on
a large sample size (6 or 7 animals per group in 3 groups).

Randomization Mice of either sex were used for most studies. For sequencing experiments in Figure 5c-g age and sex matched mice were used. Mice were
allocated into groups based on genotype and colonization status (not randomized)

Blinding Investigators were not blinded to group allocation; all of the measurements reported objectively quantifiable.

Reporting for specific materials, systems and methods

We require information from authors about some types of materials, experimental systems and methods used in many studies. Here, indicate whether each material,
system or method listed is relevant to your study. If you are not sure if a list item applies to your research, read the appropriate section before selecting a response.

Materials & experimental systems Methods

n/a Involved in the study n/a Involved in the study
Antibodies ChIP-seq
Eukaryotic cell lines Flow cytometry
Palaeontology MRI-based neuroimaging
Animals and other organisms
Human research participants
Clinical data

Antibodies
Antibodies used B220- BV421, BioLegend, #103240, RA3-6B2, Lot: B288312, 1:200 dilution
B220- BV605, BioLegend, #563708, RA3-6B2, Lot: 8201934, 1:200 dilution
B220- BV711, BioLegend, #103255, RA3-6B2, Lot: B267109, 1:200 dilution
CD38- APC, BioLegend, #17-0381-82, 90, Lot: 2068810, 1:200 dilution
CD38- PerCP-Cy5.5, BD, #562770, 90, Lot: 9112983, 1:200 dilution
FAS- PE-Cy7, BD, #557653, Jo2, Lot: 9039631, 1:400 dilution
FAS- BV421, BD, #562633, Jo2, Lot: 9029848, 1:400 dilution
CD16/32 (Fc block)- Bio-X-Cell, BE0307, 2.4G2, 1:200 dilution
TCR β- APC-e780, invitrogen, #47-5961-82, H57-597, Lot: 2114197, 1:200 dilution
CD3- BV785, BioLegend, #100232, 17A2, Lot: B277518, 1:200 dilution
CD4- BV785, BioLegend, #100552, RM4-5, Lot: B264992, 1:200 dilution
CD8- BV785, BioLegend, #100749, 53-6.7, Lot: B258589, 1:200 dilution
Nk1-1- BV785, BioLegend, #108749, PK136, Lot: B279624, 1:200 dilution
IgM- PE-Cy7, eBioscence, #25-5790-81, Il/41, Lot: 2039912, 1:200 dilution
October 2018

IgG2b- AF488, Southern Biotech, #1090-30, Lot: A2513-X665G, 1:500 dilution

IgA- Biotin, Southern Biotech, #1040-08, Lot: I5613-P366E, 1:500 dilution
IgA- PE, eBioscience, #12-4204-83, mA-6E1, Lot: E01650-1634, 1:200 dilution
IgG1- APC, BioLegend, #406610, RMG-1, Lot: B247242, 1:200 dilution
Streptavidin- APC-e780, Invitrogen, #47-4317-82, Lot: 2005846, 1:200 dilution
goat anti-human IgG1- Alexa Fluor 647, Thermo Fisher, #A-21445, Lot: 1962791, 1:2,000 dilution
goat anti-human IgG (Fcɣ specific)- HRP, Jackson Immuno Research, #109-035-098, Lot: 13741, 1:1,000 dilution

MAbs produced in-house:

2
 CD35, clone 8C12
S078.U

nature research | reporting summary

S078
S080
S116
S118
S120.U
S120
S210
S212
G082
G196
G198
G200
G202
G204
G206
G208
G226
M216.U
M216
M218.U
M218
M220.U
M220
M222
M224
M228
M232

Validation All fluorescent antibodies validated as described on the manufacturers website. HRP-conjugated antibodies validated in-house
by ELISA measuring full length IgG1 antibody concentration of commercially purchased standards. mAbs produced by us in this
study were validated by SDS-PAGE, ELISA, spectrophotometry (nanodrop) and bio-layer interferometry (Octet Red 96) to ensure
proper expression, folding and concentration.

Eukaryotic cell lines

Policy information about cell lines
Cell line source(s) ATCC, HEK 293F cells

Authentication Cell lines were not authenticated; validation of functionality was established measuring the quantity and quality of the
produced antibody.

Mycoplasma contamination All cell lines tested negative for mycoplasma contamination.

Commonly misidentified lines None used.

(See ICLAC register)

Animals and other organisms

Policy information about studies involving animals; ARRIVE guidelines recommended for reporting animal research
Laboratory animals Mice both sexes were used in all studies. 8-12 week old mice were used in all studies and strains.

Rosa26.Confetti (013731) and Rosa26.Stop.TdTomato (007914) mice were from The Jackson Laboratory. AicdaCreERT2 mice
were provided by Claude-Agnès Reynaud and Jean-Claude Weill (Institut Necker). S1pr2CreERT2 BAC transgenic mice provided
by T. Okada (RIKEN Yokohama) and T. Kurosaki (U. Osaka). PA-GFP mice were generated by G. Victora and M. Nussenzweig
(Rockefeller University).

Wild animals The study did not involve wild animals.

Field-collected samples This study did not involve samples collected from the field.
October 2018

Ethics oversight All animal procedures were approved by the Institutional Animal Care and Use Committee of the Rockefeller University.

Note that full information on the approval of the study protocol must also be provided in the manuscript.

3
Flow Cytometry

nature research | reporting summary

Plots
Confirm that:
The axis labels state the marker and fluorochrome used (e.g. CD4-FITC).
The axis scales are clearly visible. Include numbers along axes only for bottom left plot of group (a 'group' is an analysis of identical markers).
All plots are contour plots with outliers or pseudocolor plots.
A numerical value for number of cells or percentage (with statistics) is provided.

Methodology
Sample preparation Cells were isolated by maceration with disposable micropestles (Axygen) in 100 μl of PBS supplemented with 0.5% BSA and 1
mM EDTA (PBE), and single cell suspensions obtained by two passes through a 70 μm mesh. Cells were stained with fluorescently
labeled antibodies on ice for 30 minutes.

Instrument Samples were run on a FACS LSRII or FACS Symphony (BD). For cell sorted samples were run on a FACS ARIA (BD).

Software BD FACSDiva software v8.0.2 was used for flow cytometry data acquisition. Analyzed using FlowJo software package (Tri-Star,
USA) v10.5.3 and v8.7.

Cell population abundance Most cells were single cell index sorted into 96-well PCR plates, with single-cell precision. For bulk sorting, 100 cells per well were
sorted with single-cell precision.

Gating strategy
All positive and negative populations were determined by compensation with single color controls. For sorting and analysis, all
lymphocytes were first gated based on SSC-A vs FSC-A, followed by 2 singlet gates (FSC-H vs FSC-A and SSC-H vs SSC-A). For GC
gating, cells were gated on either TCRbeta or Dump-, B220+, CD38-, Fas+ and interrogated for IgM, IgG1, IgG2b or IgA. For AID-
confetti sorting experiments cells were gated on SSC-A vs FSC-A in the same way as GC cells. TCRbeta-B220+CD38-Fas+ cells (GC
B cells) were then plotted as follows: CFP vs RFP, GFP vs YFP and all colored GC cells were single-cell index sorted. For PA-GFP
sorting experiments cells were gated as above for GC cells, then GFP+ (photoactivated) GC cells were single-cell index sorted. For
GF single GC experiments (AID-Confetti), cells were gated as described, but fluorescent and non-fluorescent TCRbeta-B220
+CD38-Fas+ cells were index sorted. For non-fluorescent sequencing analysis, cells were gated as above for GC cells and single-
cell index sorted. For bacterial flow cytometry, cells were gated on SSC-A vs FSC-A with only far outliers removed. Then, SYTO
+DAPI- live bacteria were assayed for mAb binding.

Tick this box to confirm that a figure exemplifying the gating strategy is provided in the Supplementary Information.

October 2018

4
Article

Receptor binding and priming of the spike

protein of SARS-CoV-2 for membrane fusion

https://doi.org/10.1038/s41586-020-2772-0 Donald J. Benton1,6 ✉, Antoni G. Wrobel1,6 ✉, Pengqi Xu2,3, Chloë Roustan4, Stephen R. Martin1,
Peter B. Rosenthal5, John J. Skehel1 & Steven J. Gamblin1 ✉
Received: 1 July 2020

Accepted: 11 September 2020

Infection with severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) is
Published online: 17 September 2020
initiated by virus binding to the ACE2 cell-surface receptors1–4, followed by fusion of
Check for updates
the virus and cell membranes to release the virus genome into the cell. Both
receptor binding and membrane fusion activities are mediated by the virus spike
glycoprotein5–7. As with other class-I membrane-fusion proteins, the spike protein is
post-translationally cleaved, in this case by furin, into the S1 and S2 components that
remain associated after cleavage8–10. Fusion activation after receptor binding is
proposed to involve the exposure of a second proteolytic site (S2′), cleavage of which
is required for the release of the fusion peptide11,12. Here we analyse the binding of
ACE2 to the furin-cleaved form of the SARS-CoV-2 spike protein using cryo-electron
microscopy. We classify ten different molecular species, including the unbound,
closed spike trimer, the fully open ACE2-bound trimer and dissociated monomeric S1
bound to ACE2. The ten structures describe ACE2-binding events that destabilize the
spike trimer, progressively opening up, and out, the individual S1 components. The
opening process reduces S1 contacts and unshields the trimeric S2 core, priming the
protein for fusion activation and dissociation of ACE2-bound S1 monomers. The
structures also reveal refolding of an S1 subdomain after ACE2 binding that disrupts
interactions with S2, which involves Asp61413–15 and leads to the destabilization of the
structure of S2 proximal to the secondary (S2′) cleavage site.

Recognition of the ACE2 receptor by the membrane spike glycoprotein
of SARS-CoV-2 is a major determinant of virus infectivity, pathogenesis
and host range. Previous structural studies on the spike glycoproteins of
coronaviruses6,16–22 have shown that the spike trimer consists of a central
helical stalk—comprising three interacting S2 components—that is cov-
ered at the top by S1. Each S1 component consists of two large domains,
the N-terminal domain (NTD) and receptor-binding domain (RBD), each
associated with a smaller intermediate subdomain. In virus membranes,
spike glycoproteins exist in a closed form, in which the RBDs cap the
top of the S2 core and are inaccessible to ACE2, and in an open form,
in which one S1 component has opened to expose the RBD for ACE2
binding6,16,18,23. Recent structural studies7,24,25 on the isolated RBD of
the SARS-CoV-2 spike protein in complex with ACE2 have provided
a molecular description of the receptor-binding interface. Although
some comparisons can be inferred from the previous cryo-electron
microscopy studies on the spike protein of SARS-CoV12,18,19,23, structures
of intact trimeric SARS-CoV-2 spike with bound ACE2 are needed to
determine the effects of binding on the overall spike conformation.
To examine this interaction between the SARS-CoV-2 spike protein
and its receptor, we mixed the ectodomains of furin-cleaved spike
with the ectodomains of ACE2 and incubated them for around 60 s
before plunge-freezing the mixture in liquid ethane for examination
by cryo-electron microscopy. In the images that we obtained, we could
resolve ten distinct species of spike and spike–ACE2 complexes (Fig. 1
and Extended Data Fig. 1), ranging from tightly closed, unbound trim-
ers to open trimers that formed complexes with three ACE2 molecules
and dissociated monomeric S1–ACE2 complexes. Of the spike trimers
analysed, two thirds were bound to ACE2 (Extended Data Fig. 1). Of
the unbound species, we observe good-quality particles in the closed
unbound conformation, equally compact to those reported in our
previous study26 and slightly more so than those described in previous
reports6,16. There are also considerable numbers (16% of all trimers)
of unbound particles with one erect RBD, as well as some (4%) in an
intermediate conformation, a less-compact closed form, with a single
disordered RBD, which have also been reported in a previous study of
the furin-cleaved spike protein26.
Of the spike trimers bound to the receptor, half accommodate one
ACE2 receptor. As previously reported for the SARS-CoV spike pro-
tein12,23, the ACE2-bound RBD occupies a range of tilts with respect to
the long axis of the trimer (Extended Data Fig. 2a). Of the two RBDs per
trimer that are not engaged with the receptor, either both are closed or
one of the RBDs remains closed and one (either clockwise or anticlock-
wise to the bound S1 (Extended Data Fig. 1)) is in the open conformation.
We were also able to identify, reconstruct and refine trimers to which
two or three ACE2 receptors were bound, in successively more open
structures (Fig. 1 and Extended Data Fig. 1).

Strutural Biology of Disease Processes Laboratory, Francis Crick Institute, London, UK. 2Precision Medicine Center, The Seventh Affiliated Hospital, Sun Yat-sen University, Shenzhen, China.
1

3
Francis Crick Institute, London, UK. 4Structural Biology Science Technology Platform, Francis Crick Institute, London, UK. 5Structural Biology of Cells and Viruses Laboratory, Francis Crick
Institute, London, UK. 6These authors contributed equally: Donald J. Benton, Antoni G. Wrobel. ✉e-mail: donald.benton@crick.ac.uk; antoni.wrobel@crick.ac.uk; steve.gamblin@crick.ac.uk

Nature | Vol 588 | 10 December 2020 | 327

Article
Fig. 1 | Sequential steps in ACE2 binding of the
Monomeric SARS-CoV-2 spike protein. Surface representation
S1–ACE2 Closed of the spike, with monomers coloured in blue, rosy
brown and gold, and ACE2 coloured in green. Each
Open step shows two views of the spike complexes: a trimer
One erect RBD axis vertical view (left) and an orthogonal top-down
view along the axis (right). Clockwise from the top, we
show structures for closed, open but unbound RBD,
followed by sequential ACE2-binding events until
Three bound reaching the fully open, three-ACE2-bound spike
ACE2 protein state. From this final trimeric species, we
show dissociation into monomeric S1–ACE2, which
ACE2 may also occur for the one- or two-ACE2-bound
species.

ACE2
Two bound
One bound

One bound
One erect RBD

Comparison of the trimers with one erect RBD that is either bound approximately 5.5 Å away from the trimer axis, the NTD-associated and
or unbound by an ACE2 receptor revealed two things. First, ACE2 RBD-associated subdomains of the same monomer shift around 1.9 Å
binding alters the position of the open RBD by a rigid-body rotation and about 2.3 Å, respectively (Extended Data Fig. 2c), and at the same
of the domain that moves its centre of mass on average a further time the NTDs of all three S1 components move by around 1.5–3.0 Å

b a
R634
Y837 S2
Y636 F318

P295
S1
W633 K854
D614

RBD

Putative FP

R815 NTD
8Å
Closed S1

c
R634
W633 F318 S2
Unfolded
827–855 3Å
Y636 P295
D614
S2

Putative FP

R815

ACE2-bound S1

Fig. 2 | Structural rearrangements between the closed and the ACE2-bound
states of the spike protein. a, Surface representation of a monomer of S2 in
the one-ACE2-bound, two-RBD-closed state coloured in light pink with the S1
subunit of the adjacent monomer in ribbon representation; the S1 of the
one-ACE2-bound, two-RBD-closed state is shown in green and the three-RBD-
closed state (PDB 6ZGE 26) is shown in blue. The atoms on the surface of S2 that
contact the S1 intermediate domains are coloured in red. The arrows indicate
the direction of movements of the intermediate domains, and of the RBD,
between the closed and ACE2-bound conformations of the spike. b, Ribbon
representations of the NTD-associated intermediate domain in blue and the
moiety of the S2 chain that it interacts with (in red) in the closed conformation
of the spike. Essential residues that participate in the interaction are labelled;
of particular note is the salt bridge between Asp614 (S1, chain A) and Lys854
(S2, chain B). c, Ribbon representation of the same intermediate domain as
in b, but in the conformation observed in the ACE2-bound structure of the spike
(in green), in which the movement and refolding of the domain leads to a loss of
interaction with S2, which becomes disordered. The putative fusion peptide
(FP) and the S2′ site of the second protease cleavage at R815 adjacent to the
region that undergoes unfolding are shown in dark red.

328 | Nature | Vol 588 | 10 December 2020


Closed Three ACE2 bound
Fig. 3 | Structural basis of S2 unsheathing by ACE2 binding. The spike
protein is shown as a space-filling representation for S1, with each monomer
coloured blue, rosy brown and gold, and as a ribbon representation for S2
coloured in red for all three monomers. Left, top-down and side-on views of the
trimer in the closed conformation. Right, the same views for the fully open
three-ACE2-bound species.
(Extended Data Fig. 2d). Similar changes in the domain orientation are
observed in the recent structure of the SARS-CoV-2 spike complex with
C105 Fab27 (Extended Data Fig. 2e), which binds at the ACE2-binding site.
However, the molecular basis of both of these sets of changes remains
unclear. Binding of more than one ACE2 molecule does not induce
any substantial further changes in the average positioning of the RBD
(Extended Data Fig. 2e). Second, our data suggest that ACE2 binding
favours the open conformation of the RBD. The relatively high-affinity
interaction of RBD with ACE2 generates an RBD–ACE2 structure that
cannot be accommodated in a closed trimer—the bound state does not
have access to the closed conformation. In addition, the fact that ACE2
binding induces a more-open conformation of the spike RBD suggests
that some of the binding energy is used to drive the new conformation
of S1, which is then further excluded from a closed state.
The successive steps, from closed unbound trimer to the fully open,
three-ACE2-bound trimer, are associated with a substantial reduction
in the contact area that each S1 makes with both its neighbouring S1
monomers and with the S2 trimeric core (Extended Data Table 1). For
the fully, three-ACE2-bound species, each S1 makes 1,400 Å2 less con-
tact with both its S1 trimer neighbours and 1,300 Å2 less contact with
the S2 core than in the fully closed trimer conformation; all of these
rearrangements are driven by the energetics of the three ACE2-binding
events. The movements of the RBD and NTD domains of S1 that are
associated with the opening of the structure and stabilization of the
new arrangement by ACE2 binding, as described above, leave a trim-
eric ring of S1 molecules that are attached to the S2 core only through
contacts with its two small intermediate subdomains (Fig. 2a). Com-
paring the ACE2-bound, open form (the open-unbound structure is
S1–ACE2
S1–ACE2
dissociated
trimer
monomer
Fig. 4 | ACE2-bound S1 subunit as a part of the spike trimer and as an isolated
monomer. Space-filling representations of the spike protein with one monomer
coloured polychromatically. NTD, yellow; NTD-associated subdomain, blue;
RBD-associated subdomain, pink; RBD, rosy brown; S2, red; ACE2, green. The
remainder of the trimer on the left is coloured grey. The structure on the right is
aligned on the RBD:ACE2 moiety of the trimer complex on the left. The arrow
indicates the direction of movement of the NTD and NTD-associated subdomain
on the transition from the trimer (left) to the monomer species (right).
similar but of poorer local resolution) with the fully closed trimer, the
RBD-associated intermediate subdomain moves about 8 Å, whereas
the NTD-associated intermediate subdomain moves by 3 Å (Fig. 2a).
The latter also undergoes a partial restructuring with possibly impor-
tant implications for the mechanism of fusion activation of spike. In
the closed form, one edge of the NTD-associated intermediate subdo-
main interacts with a short helix and a loop from S2 of the neighbour-
ing monomer (Fig. 2b). Notably, two components of this interaction
comprise a series of side-chain π-stacking interactions in the closed
structure26: Tyr636, Phe318 and Arg634 of S1 with Tyr837 of S2; and
a salt bridge formed by Asp614 of S1 with Lys854 of S2. By contrast,
in the ACE2-bound form, Tyr636, Phe318 and Trp633 refold to the
side of the domain further away from the symmetry axis (as viewed in
Fig. 2c), leaving a channel to accommodate a new segment of α-helix
that forms downstream of Asp614 from polypeptide chain that was
previously disordered. As a consequence, the interactions between S1
and S2 described above for the closed form are lost in the ACE2-bound
form and the segment comprising residues 827–855 of S2 becomes
disordered (Fig. 2c). This part of S2 is immediately C-terminal to the
putative fusion peptide of S211, the N terminus of which is defined by
Arg815 at the S2′ cleavage site9,11. The opening of the ACE2-stabilized
S1 therefore leads to the destabilization of the S2 structure just after
the putative fusion peptide, potentially activating it for exposure in
the next stages of membrane fusion. Notably, Asp614, which forms salt
bridges to Lys854 of S2 in the closed form, is frequently substituted13–15
by a glycine residue and it has been suggested that this substitution
reduces shedding of S1 (and increases the number of spike proteins
on the virus surface)13. We also propose that this substitution would
remove a key salt bridge, and that the unique stereochemistry available
to glycine may facilitate the formation of the new segment of α-helix,
which is also incompatible with the S2 interaction. Furthermore, it could
lead to reduced stability of the closed form of the spike protein, which
in turn would increase the likelihood of the RBDs adopting the open
conformation and hence the ability of the spike protein to bind to ACE2.
The opening up, and out, from the trimer axis of the S1 domains
after ACE2 binding gives rise to an unshielding of the top surface of
Nature | Vol 588 | 10 December 2020 | 329
Article
the helix–loop–helix (approximately residues 980–990 within the
HR1 region20,22,28,29) at the top of the S2 domain (Fig. 3). In the closed
form, these helices and their connecting turns are tightly shielded
by the RBDs; each S2 monomer is predominantly covered by its
anticlockwise-related S1 trimer neighbour. In the fully open state, the
S1 domains move in such a way as to generate a cavity with a diameter
of 50 Å around the trimer axis that is about 65 Å deep. At the bottom
of this cavity is the now solvent-exposed, central portion of HR1. For
membrane fusion to occur—in comparison with other class-I fusion
proteins and as described in coronavirus post-fusion structures22,28,29—
the S2 component is likely to undergo a major helical rearrangement,
in which the long trimer interface helix (spanning residues 990–1035)
grows and extends, by incorporating the refolded turn and helix from
the N-terminal portion of HR1, and projects the fusion peptide towards
the host cell membrane. In this process, opening up of all three S1 mono-
mers and their subsequent dissociation would enable the concerted
helical refolding, as the cooperative displacement of the capping
portions of the protein will probably be required for the extension of
the helical coil, as has recently been observed for the haemagglutinin
protein of influenza30. The stoichiometry of S1 subunit–ACE2 interac-
tions required for effective cell-surface contact or for priming is not
addressed by our experiments. However, as the affinity of individual
monomers for ACE2 appears to be sufficient for cellular association, it
may be that more than one subunit is required to be in the open form
for efficient priming of these rearrangements in S2 that occur in the
process of membrane fusion. It seems reasonable to propose that the
likelihood of triggering the fusion conformation increases with the
number of ACE2 receptors bound.
In addition to the range of species of trimeric spike described above,
the largest single population of particles that we were able to identify
and reconstruct represent ACE2 bound to a S1 monomer (Fig. 4). The
interaction between ACE2 and the RBD, and the interaction of the latter
with its associated intermediate subdomain, are very similar between
the monomeric and trimer versions and with previously determined
crystal and electron microscopy structures of ACE2 and RBD7,24,25.
However, there are increasingly large rearrangements between the
two intermediate subdomains and then with the NTD. By applying
non-uniform refinement, the highest resolution was achieved for the
reconstruction of the ACE2–RBD interaction (Extended Data Fig. 4),
in part because of the tight interaction but also probably because of
the dominant influence of this part of the structure on the alignment
process. Nevertheless, it is clear that there are both increasingly large
changes in the interfaces between domains on moving towards the NTD
and a range of subpopulations of related but variable conformations.
The high proportion of ACE2–S1 monomers, and the limited contact
areas between the trimeric S1 ring interactions with S2, suggest that
the fully open ACE2-bound spike complex is probably metastable.
Taken together, our structural data enable mechanistic suggestions
for the early stages of SARS-CoV-2 infection of cells. The SARS-CoV-2
spike protein is produced in a compact closed form in which the
helices in the S2 membrane fusion component are capped by the
RBD of neighbouring monomers. After cleavage by furin between
the S1 and S2 domains, the proportion of the spike trimers that is
able to accommodate RBD in an open, ACE2-binding conformation
increases26. Binding of the ACE2 receptor to an open RBD leads to
a more-open trimer conformation. The geometry of ACE2 binding
is incompatible with the RBD adopting a closed conformation and
leads to our observation of several two-open-RBD conformations as
well as the three-RBD-bound conformation. Successive RBD open-
ing and ACE2 binding lead to a fully open and ACE2-bound form in
which the trimeric S1 ring remains bound to the core S2 trimer by
limited contacts through the intermediate subdomains of S1. This
arrangement leaves the top of the S2 helices fully exposed. In the
process, the interaction of the closed form of S1 with a segment of
the S2 chain that precedes the putative fusion peptide region, in the
330 | Nature | Vol 588 | 10 December 2020
open form, is lost. We suggest that in this form the S trimer is primed
for the helical rearrangements of S2 that are required for fusion of
the viral and host cell membranes28.
Online content
Any methods, additional references, Nature Research reporting sum-
maries, source data, extended data, supplementary information,
acknowledgements, peer review information; details of author con-
tributions and competing interests; and statements of data and code
availability are available at https://doi.org/10.1038/s41586-020-2772-0.
1. Zhou, P. et al. A pneumonia outbreak associated with a new coronavirus of probable bat
origin. Nature 579, 270–273 (2020).
2. Wan, Y., Shang, J., Graham, R., Baric, R. S. & Li, F. Receptor recognition by the novel
coronavirus from Wuhan: an analysis based on decade-long structural studies of SARS
coronavirus. J. Virol. 94, e00127-20 (2020).
3. Li, F., Li, W., Farzan, M. & Harrison, S. C. Structure of SARS coronavirus spike
receptor-binding domain complexed with receptor. Science 309, 1864–1868 (2005).
4. Li, W. et al. Angiotensin-converting enzyme 2 is a functional receptor for the SARS
coronavirus. Nature 426, 450–454 (2003).
5. Li, F. Structure, function, and evolution of coronavirus spike proteins. Annu. Rev. Virol. 3,
237–261 (2016).
6. Walls, A. C. et al. Structure, function, and antigenicity of the SARS-CoV-2 spike
glycoprotein. Cell 181, 281–292 (2020).
7. Shang, J. et al. Structural basis of receptor recognition by SARS-CoV-2. Nature 581,
221–224 (2020).
8. Belouzard, S., Chu, V. C. & Whittaker, G. R. Activation of the SARS coronavirus spike
protein via sequential proteolytic cleavage at two distinct sites. Proc. Natl Acad. Sci. USA
106, 5871–5876 (2009).
9. Millet, J. K. & Whittaker, G. R. Host cell entry of Middle East respiratory syndrome
coronavirus after two-step, furin-mediated activation of the spike protein. Proc. Natl
Acad. Sci. USA 111, 15214–15219 (2014).
10. Hoffmann, M. et al. SARS-CoV-2 cell entry depends on ACE2 and TMPRSS2 and is blocked
by a clinically proven protease inhibitor. Cell 181, 271–280 (2020).
11. Lai, A. L., Millet, J. K., Daniel, S., Freed, J. H. & Whittaker, G. R. The SARS-CoV fusion
peptide forms an extended bipartite fusion platform that perturbs membrane order in a
calcium-dependent manner. J. Mol. Biol. 429, 3875–3892 (2017).
12. Song, W., Gui, M., Wang, X. & Xiang, Y. Cryo-EM structure of the SARS coronavirus spike
glycoprotein in complex with its host cell receptor ACE2. PLoS Pathog. 14, e1007236 (2018).
13. Zhang, L. et al. The D614G mutation in the SARS-CoV-2 spike protein reduces S1
shedding and increases infectivity. Preprint at https://doi.org/10.1101/2020.06.12.148726
(2020).
14. Hu, J. et al. D614G mutation of SARS-CoV-2 spike protein enhances viral infectivity.
Preprint at https://doi.org/10.1101/2020.06.20.161323 (2020).
15. Korber, B. et al. Tracking changes in SARS-CoV-2 spike: evidence that D614G increases
infectivity of the COVID-19 virus. Cell 182, 812–827 (2020).
16. Wrapp, D. et al. Cryo-EM structure of the 2019-nCoV spike in the prefusion conformation.
Science 367, 1260–1263 (2020).
17. Tortorici, M. A. et al. Structural basis for human coronavirus attachment to sialic acid
receptors. Nat. Struct. Mol. Biol. 26, 481–489 (2019).
18. Yuan, Y. et al. Cryo-EM structures of MERS-CoV and SARS-CoV spike glycoproteins reveal
the dynamic receptor binding domains. Nat. Commun. 8, 15092 (2017).
19. Kirchdoerfer, R. N. et al. Stabilized coronavirus spikes are resistant to conformational
changes induced by receptor recognition or proteolysis. Sci. Rep. 8, 15701 (2018).
20. Pallesen, J. et al. Immunogenicity and structures of a rationally designed prefusion
MERS-CoV spike antigen. Proc. Natl Acad. Sci. USA 114, E7348–E7357 (2017).
21. Walls, A. C. et al. Cryo-electron microscopy structure of a coronavirus spike glycoprotein
trimer. Nature 531, 114–117 (2016).
22. Cai, Y. et al. Distinct conformational states of SARS-CoV-2 spike protein. Science 369,
1586–1592 (2020).
23. Gui, M. et al. Cryo-electron microscopy structures of the SARS-CoV spike glycoprotein
reveal a prerequisite conformational state for receptor binding. Cell Res. 27, 119–129
(2017).
24. Lan, J. et al. Structure of the SARS-CoV-2 spike receptor-binding domain bound to the
ACE2 receptor. Nature 581, 215–220 (2020).
25. Yan, R. et al. Structural basis for the recognition of SARS-CoV-2 by full-length human
ACE2. Science 367, 1444–1448 (2020).
26. Wrobel, A. G. et al. SARS-CoV-2 and bat RaTG13 spike glycoprotein structures inform on
virus evolution and furin-cleavage effects. Nat. Struct. Mol. Biol. 27, 763–767 (2020).
27. Barnes, C. O. et al. Structures of human antibodies bound to SARS-CoV-2 spike reveal
common epitopes and recurrent features of antibodies. Cell 182, 828–842 (2020).
28. Walls, A. C. et al. Tectonic conformational changes of a coronavirus spike glycoprotein
promote membrane fusion. Proc. Natl Acad. Sci. USA 114, 11157–11162 (2017).
29. Fan, X., Cao, D., Kong, L. & Zhang, X. Cryo-EM analysis of the post-fusion structure of the
SARS-CoV spike glycoprotein. Nat. Commun. 11, 3618 (2020).
30. Benton, D. J., Gamblin, S. J., Rosenthal, P. B. & Skehel, J. J. Structural transitions
in influenza haemagglutinin at membrane fusion pH. Nature 583, 150–153 (2020).
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in
published maps and institutional affiliations.
© The Author(s), under exclusive licence to Springer Nature Limited 2020
Methods
Constructs design, protein expression and purification
The ectodomains of ACE2 (19–615) and stabilized, ‘2P’ mutant
(K986P and V987P) of SARS-CoV-2 spike (residues 1–1208) with intact
furin-cleavage site were prepared as described in a recent study26. In
brief, the proteins were expressed in Expi293F cells (Gibco), collected
twice after 3–4 and 6–7 days, and purified with affinity chromatography
(spike using CoNTA resin from TAKARA, ACE2 with Streptactin XT resin
from IBA Lifesciences), followed by gel filtration into a buffer containing
20 mM Tris pH 8.0 and 150 mM NaCl. As previously described26, the puri-
fied spike was then incubated for 5 h with exogenous furin (New England
Biolabs), after which the reaction was stopped by addition of EDTA.
Electron microscopy sample preparation and data collection
R2/2 200-mesh Quantifoil grids were glow-discharged for 30 s at
25 mA to prepare them for freezing. The furin-treated SARS-CoV-2
spike was mixed with octyl glucoside as previously described26 and,
45–60 s before ultimately plunge-freezing the grid, with concentrated
ACE2 at a 1:2 final molar ratio of trimeric spike:ACE2, aiming to obtain
a final concentration of spike of 0.5 mg ml−1 and octyl glucoside of
0.1%. Then, 4 μl of the obtained reaction mixture was applied on a grid
pre-equilibrated at 4 °C in 100% humidity, blotted with filter paper for
4–4.5 s using Vitrobot Mark III, and plunge-frozen in liquid ethane.
Data were collected using EPU software on a Titan Krios microscope
operating at 300 kV. Micrographs were collected using a Gatan K2 detec-
tor mounted on a Gatan GIF Quantum energy filter operating in zero-loss
mode with a slit width of 20 eV. Exposures were 8 s, fractionated into
32 frames with an accumulated dose of 54.4 e−Å−2, with a calibrated pixel size
of 1.08 Å. Images were collected at a range of defoci between 1.5 and 3.0 μm.
Electron microscopy data processing
Movies were aligned using MotionCor231 implemented in RELION32, fol-
lowed by contrast transfer function (CTF) estimation using Ctffind433.
Particles were picked using crYOLO34 using a manually trained model.
Particles were subjected to multiple rounds of two-dimensional clas-
sification using cryoSPARC35. Classes that displayed a clear secondary
structure were retained and split into subsets, which either resem-
bled spike trimers or S1 monomers bound to ACE2. Initial models were
made using the ab initio reconstruction in cryoSPARC. Different spe-
cies containing trimeric spike proteins were separated by extensive
three-dimensional classification in RELION as shown in Extended Data
Fig. 3. Before the final refinement, particles corresponding to each
of these species were subjected to Bayesian polishing in RELION36
followed by homogeneous refinement in cryoSPARC coupled to CTF
refinement. The monomeric S1–ACE2 complex was classified as in
Extended Data Fig. 4a and refined using non-uniform refinement in
cryoSPARC coupled to CTF refinement. The final particles from the
S1–ACE2 complex were subjected to an unmasked refinement in RELION
to better resolve less-ordered domains, with an overall lower global
resolution (Extended Data Fig. 4b, c). Local resolution was estimated
using blocres37 implemented in cryoSPARC. Maps were locally filtered
and globally sharpened38 in cryoSPARC (Extended Data Figs. 5, 6).
Model building
The model for the monomeric S1–ACE2 complex was based on the previ-
ously determined crystal structure (PDB: 6M0J)24, with additional parts of
the RBD and intermediate domain taken from a previous structure of the
closed trimer (PDB: 6ZGE)26. Models of the trimer structures were built
using structures from our previous study26 for the closed trimer (PDB:
6ZGE) and the one-erect-RBD structure (PDB: 6ZGG). The RBD–ACE2
parts of the model were built using the structure from the high resolu-
tion S1–ACE2 complex from this study. Models were manually adjusted
using COOT39. The models of S1–ACE2 and the one-ACE2-bound closed
structure were refined and validated using PHENIX real space refine40.
The other, lower resolution models were refined using NAMDINATOR41
and geometry minimization and validation in PHENIX (Extended Data
Table 2). Measurements were made using Chimera42, CCP4MG43 and
PISA44, with structures aligned on the large helix of S2 (residues 986–1032).
Reporting summary
Further information on research design is available in the Nature
Research Reporting Summary linked to this paper.
Data availability
Maps and models have been deposited in the Electron Microscopy Data
Bank (EMD) and the Protein Data Bank (PDB) with the following acces-
sion codes: EMD-11681 and PDB 7A91 (dissociated S1 domain bound to
ACE2 (non-uniform refinement)); EMD-11682 and PDB 7A92 (dissoci-
ated S1 domain bound to ACE2 (unmasked refinement)); EMD-11683
and PDB 7A93 (SARS-CoV-2 spike with two RBDs erect); EMD-11684 and
PDB 7A94 (SARS-CoV-2 spike with one ACE2 bound); EMD-11685 and
PDB 7A95 (SARS-CoV-2 spike with one ACE2 bound and one RBD erect
in clockwise direction); EMD-11686 and PDB 7A96 (SARS-CoV-2 spike
with one ACE2 bound and one RBD erect in anticlockwise direction);
EMD-11687 and PDB 7A97 (SARS-CoV-2 spike with two ACE2 bound);
EMD-11688 and PDB 7A98 (SARS-CoV-2 spike with three ACE2 bound).
31. Zheng, S. Q. et al. MotionCor2: anisotropic correction of beam-induced motion for
improved cryo-electron microscopy. Nat. Methods 14, 331–332 (2017).
32. Scheres, S. H. W. RELION: implementation of a Bayesian approach to cryo-EM structure
determination. J. Struct. Biol. 180, 519–530 (2012).
33. Rohou, A. & Grigorieff, N. CTFFIND4: fast and accurate defocus estimation from electron
micrographs. J. Struct. Biol. 192, 216–221 (2015).
34. Wagner, T. et al. SPHIRE-crYOLO is a fast and accurate fully automated particle picker for
cryo-EM. Commun. Biol. 2, 218 (2019).
35. Punjani, A., Rubinstein, J. L., Fleet, D. J. & Brubaker, M. A. cryoSPARC: algorithms for rapid
unsupervised cryo-EM structure determination. Nat. Methods 14, 290–296 (2017).
36. Zivanov, J., Nakane, T. & Scheres, S. H. W. A Bayesian approach to beam-induced motion
correction in cryo-EM single-particle analysis. IUCrJ 6, 5–17 (2019).
37. Cardone, G., Heymann, J. B. & Steven, A. C. One number does not fit all: mapping local
variations in resolution in cryo-EM reconstructions. J. Struct. Biol. 184, 226–236 (2013).
38. Rosenthal, P. B. & Henderson, R. Optimal determination of particle orientation, absolute
hand, and contrast loss in single-particle electron cryomicroscopy. J. Mol. Biol. 333,
721–745 (2003).
39. Emsley, P., Lohkamp, B., Scott, W. G. & Cowtan, K. Features and development of Coot.
Acta Crystallogr. D 66, 486–501 (2010).
40. Adams, P. D. et al. PHENIX: a comprehensive Python-based system for macromolecular
structure solution. Acta Crystallogr. D 66, 213–221 (2010).
41. Kidmose, R. T. et al. Namdinator - automatic molecular dynamics flexible fitting of structural
models into cryo-EM and crystallography experimental maps. IUCrJ 6, 526–531 (2019).
42. Pettersen, E. F. et al. UCSF Chimera—a visualization system for exploratory research and
analysis. J. Comput. Chem. 25, 1605–1612 (2004).
43. McNicholas, S., Potterton, E., Wilson, K. S. & Noble, M. E. M. Presenting your structures:
the CCP4mg molecular-graphics software. Acta Crystallogr. D 67, 386–394 (2011).
44. Krissinel, E. & Henrick, K. Inference of macromolecular assemblies from crystalline state.
J. Mol. Biol. 372, 774–797 (2007).
Acknowledgements We thank A. Nans of the Structural Biology Science Technology Platform
for assistance with data collection, P. Walker and A. Purkiss of the Structural Biology Science
Technology Platform and the Scientific Computing Science Technology Platform for
computational support, and L. Calder, P. Cherepanov, G. Kassiotis and S. Kjaer for discussions.
This work was funded by the Francis Crick Institute, which receives its core funding from
Cancer Research UK (FC001078 and FC001143), the UK Medical Research Council (FC001078
and FC001143) and the Wellcome Trust (FC001078 and FC001143). P.X. is also supported by
the 100 Top Talents Program of Sun Yat-sen University, the Sanming Project of Medicine in
Shenzhen (SZSM201911003) and the Shenzhen Science and Technology Innovation
Committee (grant no. JCYJ20190809151611269).
Author contributions D.J.B., A.G.W., P.X., C.R. and S.R.M. performed research, collected and
analysed data; D.J.B., A.G.W., P.B.R., J.J.S. and S.J.G. conceived and designed research and
wrote the paper.
Competing interests The authors declare no competing interests.
Additional information
Supplementary information is available for this paper at https://doi.org/10.1038/s41586-020-
2772-0.
Correspondence and requests for materials should be addressed to D.J.B., A.G.W. or S.J.G.
Peer review information Nature thanks Stephen Harrison and the other, anonymous,
reviewer(s) for their contribution to the peer review of this work.
Reprints and permissions information is available at http://www.nature.com/reprints.
Article

Extended Data Fig. 1 | Surface representation of obtained structures. The three monomers of S in each trimer are coloured in blue, rosy brown and gold with
ACE2 shown in green. Relative percentages of all trimeric S particles used to calculate electron microscopy maps are shown.
Extended Data Fig. 2 | Features of the obtained spike structures. a, Two
three-dimensional classes, obtained by further classification of the
one-ACE2-bound closed state from Fig. 1, representative of the range of
motion of the RBD with bound ACE2, tilting away from the trimer axis of the
spike trimer. The tilt of the RBD and ACE2 is indicated with a dashed line.
b, Representative density of different obtained electron microscopy maps for
residues 996–1030 of S2. Built model shown in pink, with EM density shown as a
mesh. c, d, Comparison of spike structures for the open one-erect-RBD
structure (purple) with the one-ACE2-bound structure (orange). c, S1 domains
shown to highlight domain shifts of the RBD and RBD-associated intermediate
domain. d, Outwards movements of spike domains (excluding RBDs).
e, Comparison of RBD displacements of one-bound, two-bound and
three-bound RBDs after binding of ACE2 to the unbound open structure of the
spike protein (beige). These are compared to the RBD displacement after
binding of the C105 Fab fragment 27, which binds at the ACE2 interface of the
RBD (PDB: 6XCM).
Article

Extended Data Fig. 3 | Cryo-electron microscopy data processing scheme. final maps shown at the bottom. The global resolution, final particle number
Classes of particles used to obtain the final spike trimer structures, unbound and percentage for each trimer species are shown at the bottom.
and in complex with ACE2, are surrounded by a box of the same colour as the
Extended Data Fig. 4 | Monomeric S1 bound to ACE2. a, Classification particles. Domains are coloured as follows: green, ACE2; yellow, NTD; rosy
scheme for the S1–ACE2 complex. b, c, Maps are shown of orthogonal views of brown, RBD; pink, RBD ganymede; blue, NTD ganymede; cream, disseminated
the non-uniform refinement (b) and unmasked refinement (c) of the final S1 density in b.
Article

Extended Data Fig. 5 | Fourier shell correlation graphs for each of the determined structures. FSC, Fourier shell correlation.
Extended Data Fig. 6 | Maps and models of determined structures. Top, orthogonal views of electron microscopy density (grey) and ribbon diagram
representation of the models. Bottom, electron microscopy maps coloured by local resolution shown below.
Article
Extended Data Table 1 | Buried interface surface area between monomers in different conformations

Different confirmations of unbound and ACE2-bound trimers were analysed. The interface area was calculated using PISA. In the open and ACE2-bound conformations, chain A is the one to
open first and to bind the receptor first, then B follows, if the second RBD changes the conformation. Chain B is the chain anticlockwise to A when looking down the symmetry axis with the
membrane-proximal part at the bottom. The unbound and three-ACE2-bound molecules are of C3 symmetry.
Extended Data Table 2 | Cryo-electron microscopy data collection, refinement and validation statistics
 nature research | reporting summary
Donald Benton, Antoni Wrobel, Steven
Corresponding author(s): Gamblin
Last updated by author(s): Aug 30, 2020

Reporting Summary
Nature Research wishes to improve the reproducibility of the work that we publish. This form provides structure for consistency and transparency
in reporting. For further information on Nature Research policies, see Authors & Referees and the Editorial Policy Checklist.

Statistics
For all statistical analyses, confirm that the following items are present in the figure legend, table legend, main text, or Methods section.
n/a Confirmed
The exact sample size (n) for each experimental group/condition, given as a discrete number and unit of measurement
A statement on whether measurements were taken from distinct samples or whether the same sample was measured repeatedly
The statistical test(s) used AND whether they are one- or two-sided
Only common tests should be described solely by name; describe more complex techniques in the Methods section.

A description of all covariates tested

A description of any assumptions or corrections, such as tests of normality and adjustment for multiple comparisons
A full description of the statistical parameters including central tendency (e.g. means) or other basic estimates (e.g. regression coefficient)
AND variation (e.g. standard deviation) or associated estimates of uncertainty (e.g. confidence intervals)

For null hypothesis testing, the test statistic (e.g. F, t, r) with confidence intervals, effect sizes, degrees of freedom and P value noted
Give P values as exact values whenever suitable.

For Bayesian analysis, information on the choice of priors and Markov chain Monte Carlo settings
For hierarchical and complex designs, identification of the appropriate level for tests and full reporting of outcomes
Estimates of effect sizes (e.g. Cohen's d, Pearson's r), indicating how they were calculated
Our web collection on statistics for biologists contains articles on many of the points above.

Software and code

Policy information about availability of computer code
Data collection CryoEM data collected using Thermo Scientific EPU v2.7

Data analysis CryoEM data processed using following packages: RELION-3.1, cryoSPARC v2.14, CTFfind4 v.4.1.10, MotionCor2 v.1.2.6, crYOLO v1.4,
Coot v.0.9, PHENIX v.1.17, UCSF Chimera v.1.12, UCSF ChimeraX v.0.5, CCP4MG v2.10, PISA v1.52
For manuscripts utilizing custom algorithms or software that are central to the research but not yet described in published literature, software must be made available to editors/reviewers.
We strongly encourage code deposition in a community repository (e.g. GitHub). See the Nature Research guidelines for submitting code & software for further information.

Data
Policy information about availability of data
All manuscripts must include a data availability statement. This statement should provide the following information, where applicable:
- Accession codes, unique identifiers, or web links for publicly available datasets
- A list of figures that have associated raw data
- A description of any restrictions on data availability

Maps and models have been deposited in the Electron Microscopy Data Bank, http://www.ebi.ac.uk/pdbe/emdb/ and the Protein Data Bank, https://
October 2018

www.ebi.ac.uk/pdbe/ with the following accession codes: EMD-11681 and PDB 7A91 (Dissociated S1 domain bound to ACE2 [Non-Uniform Refinement]);
EMD-11682 and PDB 7A92 (Dissociated S1 domain bound to ACE2 [Unmasked Refinement]); EMD-11683 and PDB 7A93 (SARS-CoV-2 S with 2 RBDs Erect);
EMD-11684 and PDB 7A94 (SARS-CoV-2 S with 1 ACE2 Bound); EMD-11685 and PDB 7A95 (SARS-CoV-2 S with 1 ACE2 Bound and 1 RBD Erect in Clockwise
Direction); EMD-11686 and PDB 7A96 (SARS-CoV-2 S with 1 ACE2 Bound and 1 RBD Erect in Anticlockwise Direction); EMD-11687 and PDB 7A97 (SARS-CoV-2 S with
2 ACE2 Bound); EMD-11688 and PDB 7A98 (SARS-CoV-2 S with 3 ACE2 Bound).

1
 nature research | reporting summary
Field-specific reporting
Please select the one below that is the best fit for your research. If you are not sure, read the appropriate sections before making your selection.
Life sciences Behavioural & social sciences Ecological, evolutionary & environmental sciences
For a reference copy of the document with all sections, see nature.com/documents/nr-reporting-summary-flat.pdf

Life sciences study design

All studies must disclose on these points even when the disclosure is negative.
Sample size All cryoEM datasets consist of several thousand images. The number of images were sufficient to achieve the reported resolution, according
to the most commonly reported resolution measure in cryoEM described in Rosenthal and Henderson 2003, as cited in the manuscript

Data exclusions CryoEM single particles were included and excluded within the image processing workflow using standard image processing techniques such
as 2D and 3D classifications, as detailed in Extended Data Figures 3 and 4.

Replication Structures were determined using independent half datasets, according to standard procedures in cryoEM. Images were collected from three
independent replicate prepared grids, which all produced similar images both by low resolution visual inspection and high resolution class
averages. There were no unsuccessful replications.

Randomization Not applicable to this study, as samples were not assigned to experimental groups and data were collected and processed according to
standard techniques for cryoEM.

Blinding Not applicable to this study, as there was no experimental group allocation in data collection and analysis.

Reporting for specific materials, systems and methods

We require information from authors about some types of materials, experimental systems and methods used in many studies. Here, indicate whether each material,
system or method listed is relevant to your study. If you are not sure if a list item applies to your research, read the appropriate section before selecting a response.

Materials & experimental systems Methods

n/a Involved in the study n/a Involved in the study
Antibodies ChIP-seq
Eukaryotic cell lines Flow cytometry
Palaeontology MRI-based neuroimaging
Animals and other organisms
Human research participants
Clinical data

Eukaryotic cell lines

Policy information about cell lines
Cell line source(s) expi293F cells were purchased from Thermo Scientific and used for protein expression

Authentication Cell line used was not authenticated

Mycoplasma contamination Cell line was not tested for mycoplasma contamination

Commonly misidentified lines Name any commonly misidentified cell lines used in the study and provide a rationale for their use.
(See ICLAC register)
October 2018

2
Article

A metastasis map of human cancer cell lines

https://doi.org/10.1038/s41586-020-2969-2 Xin Jin1 ✉, Zelalem Demere1, Karthik Nair1, Ahmed Ali1,2, Gino B. Ferraro3, Ted Natoli1,
Amy Deik1, Lia Petronio1, Andrew A. Tang1, Cong Zhu1, Li Wang1, Danny Rosenberg1,
Received: 17 December 2018
Vamsi Mangena4, Jennifer Roth1, Kwanghun Chung1,4, Rakesh K. Jain3,5, Clary B. Clish1,
Accepted: 26 August 2020 Matthew G. Vander Heiden1,2,6 & Todd R. Golub1,5,6 ✉

Published online: 9 December 2020

Check for updates Most deaths from cancer are explained by metastasis, and yet large-scale metastasis
research has been impractical owing to the complexity of in vivo models. Here we
introduce an in vivo barcoding strategy that is capable of determining the metastatic
potential of human cancer cell lines in mouse xenografts at scale. We validated the
robustness, scalability and reproducibility of the method and applied it to 500 cell
lines1,2 spanning 21 types of solid tumour. We created a first-generation metastasis
map (MetMap) that reveals organ-specific patterns of metastasis, enabling these
patterns to be associated with clinical and genomic features. We demonstrate the
utility of MetMap by investigating the molecular basis of breast cancers capable of
metastasizing to the brain—a principal cause of death in patients with this type of
cancer. Breast cancers capable of metastasizing to the brain showed evidence of
altered lipid metabolism. Perturbation of lipid metabolism in these cells curbed brain
metastasis development, suggesting a therapeutic strategy to combat the disease and
demonstrating the utility of MetMap as a resource to support metastasis research.

Human cancer cell lines have been a driving force in cancer research,
leading to the discovery of oncogenic mechanisms and therapeutic
targets1–4. However, large-scale characterization of cell lines has been
limited to rudimentary readouts such as viability in cell culture, because
more complex phenotypes—such as behaviours in vivo—have not been
tractable at scale. By contrast, most studies of metastasis rely on only
a small number of experimental models5–9, thereby making it difficult
to extrapolate findings to genetically diverse human tumours10.
Ideally, it would be possible to construct a map of organ-specific
metastatic potential of hundreds of human cancer cell lines using
xenograft models, so that the molecular features of the cell lines could
be related to their ability to survive and proliferate in organ-specific
microenvironments. However, the prospect of in vivo testing of each
cell line individually is unattractive, because it is labour-intensive and
expensive, as well as because of the difficulty in sufficiently controlling
for variability between animal experiments. We proposed that if cell
lines were labelled with molecular barcodes and injected into recipi-
ent mice as a pool, internally controlled, metastatic potential could be
assessed in a highly scalable manner.
Pilot study with breast cancer
To test the feasibility and reliability of in vivo barcoding to monitor
growth in different tissues in mice, we performed a pilot study using
four breast cancer cell lines (Fig. 1a, Extended Data Fig. 1, Supple-
mentary Note 1). Each cell line was engineered to express a unique
26-nucleotide barcode, together with luciferase for in vivo imaging
and either GFP or mCherry to facilitate subsequent cell sorting and
measurement of reproducibility within a single mouse (Extended Data
Fig. 1a, Supplementary Table 1). The 8 barcoded lines were injected
1
Broad Institute of MIT and Harvard, Cambridge, MA, USA. 2Koch Institute for Integrative Cancer Research, Department of Biology, Massachusetts Institute of Technology, Cambridge, MA, USA.
3
Edwin L. Steele Laboratories, Department of Radiation Oncology, Massachusetts General Hospital, Boston, MA, USA. 4Institute for Medical Engineering and Science, Picower Institute for
Learning and Memory, Department of Chemical Engineering, Massachusetts Institute of Technology, Cambridge, MA, USA. 5Harvard Medical School, Boston, MA, USA. 6Dana-Farber Cancer
Institute, Boston, MA, USA. ✉e-mail: xjin@broadinstitute.org; golub@broadinstitute.org
as a pool into the left ventricle of 5–6-week-old NOD-SCID-gamma
(NSG) mice so as to focus our analysis on the ability of tumour cells
to exit circulation and undergo expansion in distant organs. Biolu-
minescence imaging (BLI) revealed metastatic lesions throughout
the body (Extended Data Fig. 1b). Five weeks after injection, brain,
lung, liver, kidney and bone were collected, human tumour cells were
isolated by fluorescence-activated cell sorting (FACS) using GFP
or mCherry, and barcodes were quantified using RNA sequencing
(RNA-seq) (Extended Data Fig. 1c–g). Whereas barcode abundances
were similar pre-injection, some barcodes were enriched in specific
organs (Extended Data Fig. 1h). Different cell lines exhibited distinct
patterns of metastatic spread, but each cell line showed highly similar
pattern of spread across multiple mice independent of whether GFP
or mCherry versions were used, demonstrating the reproducibility of
this pooled approach (Extended Data Fig. 1d). For example, HCC1954
was most strongly detected in brain, whereas extracranial metastases
were dominated by MDAMB231. Barcodes quantified by bulk RNA-seq
were independently validated by quantitative PCR with reverse tran-
scription (RT–qPCR) and single-cell RNA-seq (Extended Data Fig. 1i–m,
Supplementary Note 1).
Having validated the method, we next characterized the metastatic
behaviours of all 21 basal-like breast cancer cell lines in the Cancer Cell
Line Encyclopedia (CCLE) (Extended Data Fig. 1a–d). Basal-like breast
cancers are known to have diverse metastatic abilities in patients11.
Reflecting this diversity, the cell lines showed disparate metastatic pat-
terns: pan-metastatic, metastatic preferentially to particular organs or
not metastatic (Fig. 1b, Supplementary Table 2). Notably, one cell line
(BT20) was detected in multiple organs, but at very low abundance in all
of them, reflecting its ability to colonize but not expand. To validate the
patterns of metastasis observed in the pooled in vivo system, we selected

Nature | Vol 588 | 10 December 2020 | 331

Article
a Barcoded cell Metastasis Isolate cells such competition. A similarly reduced diversity was observed in the
line pool formation from organs RNA-seq
Brain Tissue
Barcode
transcripts
Cancer cell
composition
Inferred cell
number
orthotopic setting, where injection of a pool of nine breast cancer cell
dissociation
Pilot Lung lines into the mammary fat pad resulted in a single cell line dominating
Group 1 Liver Mouse cell Cancer In vivo
Kidney
depletion transcripts expression
Metastatic
the resulting tumour (Extended Data Fig. 3d).
Bone
Group 2 FACS potential
To determine whether the MetMap reflects the metastatic behaviour
b MDAMB231 HCC1187 JIMT1 HCC1806 HMC18 HCC70 of human cancers, we analysed available clinical annotations of the cell
lines (Fig. 3a–e, Extended Data Fig. 4). We found statistically significant
associations with tumour lineage, the site from which the cell line was
derived (primary tumour versus metastatic lesion) and patient age.
HCC1954 MDAMB468 BT20 MDAMB436 HCC1599 DU4475
There was no association between metastatic potential and gender
or ethnicity. As expected, metastatic potential was higher in certain
tumour types, such as melanoma and pancreatic cancer, which also
HCC1395 MDAMB157 HCC1569 HCC38 HDQP1 HCC1937 tend to develop metastasis in the human disease setting13. By contrast,
cell lines derived from brain tumours were generally non-metastatic,
reflective of their tendency to not undergo haematogenous spread14,15.
BT549 CAL851 HCC1143 Petal plot Similarly, the DU145 prostate cancer cell line, derived from a brain
Brain
105 metastasis lesion16, exhibited brain metastasis in our experimental
ne

Lun
Bo

100 system. Cell lines derived from metastases showed higher metastatic
g

Ki
dn
ey Liv
e r
Organ Metastatic potential
Fig. 1 | Scalable in vivo metastatic potential mapping with barcoded cell
line pools. a, A schematic of the experiment determining the feasibility of
in vivo metastatic potential profiling using barcoded cell line pools. Barcode
abundance reflecting cancer cell compositions was determined by RNA-seq,
and the cell number of each cell line was inferred by cancer cell composition
and total cancer cell counts isolated from the target organ. b, Petal plots
displaying the metastatic patterns of 21 basal-like breast cancer cell lines. Petal
length represents metastatic potential, quantifying the mean of inferred
cancer cell numbers detected from the target organs. Data are mean ± 95%
confidence interval. Petal width shows penetrance, quantifying percentage of
mice detected with the cell line.
eight cell lines for individual characterization, and observed similar
results from the pooled and individual screens (Extended Data Fig. 1n, o).
A metastasis map of 500 human cancer cell lines
Having demonstrated its feasibility in breast cancer, we attempted
to expand the mapping of metastatic potential to human cancer cell
lines from diverse lineages. To facilitate higher-throughput profil-
ing, we used cell lines barcoded for use with the PRISM method, which
was developed for in vitro drug-sensitivity screening12. A simplified
workflow enabled the quantitative detection of barcodes from crude
tissue lysates without the need for FACS-based tumour cell purification
(Extended Data Fig. 2, Supplementary Note 2). We applied this method
to 503 cell lines spanning 21 lineages to develop a first-generation
Metastasis Map (MetMap) (Fig. 2a). The data and interactive visualiza-
tion are publicly accessible at https://pubs.broadinstitute.org/metmap.
To test the robustness of the MetMap dataset, we tested cell lines in
two formats: in one, we injected all 498 cell lines as a single pool; in the
other, we injected 5 pools of 25 lines, with each pool being injected into
different mice (referred to as MetMap500 and MetMap125, respectively)
(Fig. 2b). We similarly varied cell numbers, mouse age and cohort size to
determine whether results varied substantially with these parameters.
We observed strong correlation of the metastatic potential despite
differences in experimental conditions (Fig. 2c), suggesting that the
approach is extremely robust. We also note that intracardiac injec-
tion enabled the evaluation of many more cell lines in vivo compared
with subcutaneous injection. Specifically, we recovered an average of
197 cell lines per mouse following intracardiac injection, whereas an
average of 42 cell lines were recovered following subcutaneous injection
(Extended Data Fig. 3a–c). We suspect that this difference is explained
by the local competition for nutrients and other microenvironmental
factors in the subcutaneous setting, whereas the spatial separation
of tumour cells delivered through the intracardiac route minimizes
1
0 potential than lines derived from primary tumours, although lines
Penetrance derived from primary tumours known to later give rise to metastases
in patients were metastatic in the MetMap (Fig. 3b), consistent with
previously reported suggestions that metastatic potential is already
encoded in primary tumours17–19. The association between decreased
metastatic potential and increased patient age was unexpected (Fig. 3c),
and its basis remains to be determined.
Perhaps most importantly, extensive variation in metastatic potential
was observed within individual lineages, making it possible to search
for associations between metastasis propensity and genomic features
of the tumours. Of note, metastatic potential was not simply explained
by proliferation rate or mutational burden (Fig. 3f–h, Extended Data
Fig. 4f, g), suggesting that more subtle molecular determinants of
metastasis were involved.
Molecular correlates of brain metastasis
To develop mechanistic insights, we focused on breast cancer and its
potential for brain metastasis (Fig. 1b), because brain metastasis is a
feature of some—but not all—breast cancers, and little is known about
the underlying factors that could inform therapeutic approaches20,21.
We therefore undertook a systematic and unbiased comparison of
the molecular features that distinguished brain metastatic versus
non-metastatic lines, using genomic data available for each of the
cell lines.
At the level of somatic mutations, PIK3CA was the top associated cor-
relate: 4 out of 7 brain metastatic lines contained a PIK3CA mutation,
compared with 0 out of 14 non-metastatic or weakly metastatic lines
(false discovery rate (FDR) = 0.0034) (Fig. 4a, Extended Data Fig. 5a).
A fifth line, HCC70, has a loss-of-function mutation in PTEN. PI3K is a prin-
cipal downstream mediator of ERBB2 (also known as HER2), which itself
has been reported to be associated with brain metastasis in humans11,20.
Indeed, two of the brain metastatic cell lines ( JIMT1 and HCC1954) also
contain typical ERBB2 gene amplifications (Extended Data Fig. 5a).
At the level of DNA copy number, we observed an association between
metastatic potential and deletions of chromosome 8p12–8p21.2
(referred to as 8p) (FDR = 0.0017) (Fig. 4b). Five out of seven brain
metastatic breast cancer cell lines contained deletions in this region,
compared with 0 out of 14 non-metastatic lines (Extended Data Fig. 5a).
A sixth metastatic line, JIMT1, has a small deletion within this commonly
deleted region.
To ascertain the clinical relevance of these associations, we analysed
clinical breast cancer datasets for which metastasis information was
available18. We observed a strong correlation between 8p copy number
and gene expression in the METABRIC and TCGA datasets22,23 (Extended
Data Fig. 6a), thereby validating 8p expression as a surrogate for copy
number in datasets for which copy number data were not available.

332 | Nature | Vol 588 | 10 December 2020

 a PRISM cell Metastasis Mouse stroma depleted Barcode b
line pool formation tissue lysates quantification MetMap500 MetMap125
Brain
Lung Relative
metastatic
Liver
potential
Kidney
Bone
498 lines in 1 pool 25 lines × 5 pools

Neuroblastoma
Head and neck

Mesothelioma
Endometrium
~500 cells per line ~10k cells per line

Oesophageal
Pancreatic
15 mice ~5 mice per pool

Melanoma

Colorectal

Sarcoma

Bile duct

Prostate
Bladder
Ovarian
8–10 weeks old, female 5–6 weeks old, female

Thyroid
Gastric

Kidney
Breast

Bone
Brain
Lung
5 organs 4 organs

Liver
120 cell lines
Primary 4 organs
shared
tumour
Metastasis

c Overall metastatic potential - 4 organs combined

Pearson’s r = 0.80 Pearson’s r = 0.63

ES2 P = 9.8 × 10–15
P < 2.2 × 10–16

Brain
A2780
HEYA8
102
SKHEP1
Penetrance HCC1806 DMS273
0% LU99
20% A673
SNU407 MELHO
40%
SW620
KP4 Pearson’s r = 0.75
60% HSC3
MELJUSO
80% P < 2.2 × 10–16
SKMEL5 A549
100% SNU869
8505C
ASPC1

Lung
G361 MHHES1
CAL62 UACC62
100 LN229
CORL23
HS766T SW480
YAPC OVCAR8
HMC18 S117 KYSE510 NCIH841
MetMap500

JIMT1 YD15
SKMEL24 HEC1A
DU145 MDAMB435S
YKG1 HARA
RKO
SNU840 HS294T PC14 Pearson’s r = 0.84
RMUGS MESSA P < 2.2 × 10–16
NCIH322 RERFLCAD2
HT144 SNU1041
KYSE30 KYSE410 SNUC2A
T24 NCIH1437

Liver
AN3CA HCC827
SH4 SCABER NCIH1703
NCIH2172 AGS L33
10−2 647V NCIH1975
NCIH2030
ISTMES1ISHIKAWAHERAKLIO02ER
J82 NCIH1355 YD10B 22RV1
G401 HLF
GB1 LS411N MIAPACA2
KURAMOCHI Pearson’s r = 0.60
U251MG CAL12T CAPAN2
SF295
KYSE70
OE19 P = 5.1 × 10–13
CAL78 SNU719
YH13 NCIH1339 NCIN87
SNU761
G292CLONEA141B1 UBLC1NCIH1623 KNS62 SW837
CAL54 SNU601 PANC0327 U2OS
Bone

OAW42 NCIH2052 T98G COV362

SNU46 GI1
JHH4 5637 PLCPRF5 ONCODG1 HGC27 WM793
NCIH1793 PANC1005 KALS1 CAL29
10−4 UO31 SW948 GAMG
769P

10−4 10−2 100 102

MetMap125

Fig. 2 | Drafting MetMap for 500 human cancer cell lines. a, A schematic of primary tumour or metastasis. b, Comparison of experimental conditions
the workflow using pan-cancer PRISM cell line pools for high-throughput between MetMap500 and MetMap125. c, Scatter plots showing overall and
metastatic potential profiling. Relative metastatic potential was quantified by organ-specific metastatic potential as determined in MetMap500 and
deep sequencing of PRISM barcode abundance from tissue. The cancer lineage MetMap125. Strong correlation is observed between the two experiments.
distribution of the profiled 500 cancer cell lines is presented, with each dot Each dot represents a cell line. Cancer lineage is colour-coded as in a.
representing a cell line, and showing whether the cell line was derived from

Coordinated expression of 8p genes stratified tumours into two clus-

ters, with the low-expressing cluster showing enrichment in brain Lipid metabolism and brain metastasis
metastasis and lower brain metastasis-free survival (Extended Data Confirming these genetic findings, expression analysis revealed enrich-
Fig. 6b). Whereas 8p loss was more frequent in basal-subtype breast ment of PI3K and ERBB2 signatures in the brain metastatic cell lines
cancer (known to have poor prognosis), 8p loss remained significantly (Fig. 4c). Furthermore, we observed a strong association between
associated with brain metastases within basal tumours. A similar trend brain metastatic potential and a lipid-synthesis signature (Fig. 4c),
was seen in other subtypes, but the sample size was too small to reach which has been reported in association with both PI3K activation and
statistical significance. Concordant with these findings, the 8p-low 8p-deletion27,28. To investigate a potential role of lipid metabolism in
signature was strongly enriched in brain metastasis lesions compared breast-cancer brain metastatic potential, we analysed the abundance
with extracranial metastases or primary tumours24 (Extended Data of lipid metabolites across the cell lines29. We observed increased
Fig. 6c, d). Similarly, we observed that response signatures25,26 indicat- levels of cholesterol species in highly brain metastatic cells (Fig. 4d).
ing PI3K activation are associated with brain metastases (Extended In addition to cholesterols, membrane lipids including phosphatidyl-
Data Fig. 6e–g). The PI3K-high signature tended to co-occur with the choline and sphingomyelin were similarly more abundant, as were
8p-low signature, and the overlapping events captured the majority metabolites associated with the pentose phosphate pathway30, which
of patients with brain metastases (Extended Data Fig. 6h, i). These can support cholesterol and lipid synthesis. By contrast, we observed
results established the validity of the MetMap experimental system global decreases in levels of triacylglycerols (TAGs) in brain metastatic
for discovery. cells (Fig. 4d). Non-brain metastatic cells had higher levels of TAGs and

Nature | Vol 588 | 10 December 2020 | 333

Article
a Cancer type: P = 2.2 × 10–7 To investigate the role of SREBF1 in affecting the lipid phenotype
102 observed in brain metastatic cells, we performed lipidomics after
Metastatic

knocking out SREBF1 in JIMT1 and HCC1806 cells using CRISPR–Cas9.

potential

100

10–2 SREBF1 knockout resulted in a marked shift in intracellular lipid con-

10–4 tent, including a decrease in levels of cholesterol, membrane lipids and
b Derived from: P = 0.00028 P = 5.5 × 10–5 diacylglycerols (Fig. 4h). SREBF1 knockout also resulted in an increase
102 102
P = 0.055 P = 0.28
in intracellular TAG levels, presumably by scavenging TAGs from the
lipid-rich serum added to the culture medium. To test this hypoth-
Metastatic
potential

100 100

10–2 10–2 esis, we repeated the experiment in culture medium prepared with
10–4 10–4
delipidated serum, which prevented the increase in TAGs observed in
Primary tumour Metastasis Primary tumour Primary with
metastasis
Metastasis
SREBF1-knockout cells (Extended Data Fig. 7).
c Age: P = 0.0077
102
To further explore the role of SREBF1, we performed RNA-seq fol-
lowing SREBF1 knockout and found SCD35 to be the most consistently
Metastatic
potential

100

10–2
downregulated gene (Fig. 4i). Consistent with this, SCD was the top
10–4
co-dependency of SREBF1 across 688 cell lines in the genome-wide
0–10 10–20 20–30 30–40 40–50 50–60 60–70 70–80 80–90 CRISPR–Cas9 viability screens (Fig. 4j). The next highest scoring SREBF1
d Gender: P = 0.55 e Ethnicity: P = 0.91
co-dependency was SCAP, which encodes the upstream activator of
102 102
SREBF135. Comparison of gene expression in breast cancer cells grown
Metastatic
potential

100 100
in vitro or in the brain similarly showed that in the brain, cells adopted
10–2 10–2
gene-expression signatures of adipogenesis, fatty acid metabolism
10–4 10–4
Male Female Asian African american Caucasian and xenobiotic metabolism (Extended Data Fig. 8, Supplementary
f Doubling time (h): P = 0.058 g Mutation burden: P = 0.52 h Aneuploidy: P = 0.23
Note 3). The enrichment of lipid-metabolism signatures (including
102 102 102
upregulation of SREBF1 and SCD) was unique to brain compared with
other sites of metastasis. Similar upregulation was also observed in
Metastatic
potential

100 100 100

10-2 10-2 10-2

10-4 10-4
0 100 200 300 0 1000
Fig. 3 | Clinical correlates of metastatic potential. a–e, Single-variate
correlation of different clinical parameters with overall metastatic potential
from MetMap500 data. Primary with metastasis indicates that the cell line was
derived from the primary tumour and the donor developed metastasis at
diagnosis or later. In box plots, boxes display quartiles of the data; outlier
points extend beyond 1.5× interquartile ranges from either hinge. Cancer
lineage is colour-coded as in Fig. 2a. f–h, Single-variate correlation of cell
doubling, mutation burden and aneuploidy status with overall metastatic
potential from MetMap500 data. f, Doubling time in hours. g, Mutation burden
quantified by somatic mutations from exon-sequencing data. h, Aneuploidy
quantified by chromosome-arm-level events from exon-sequencing data. Each
dot represents a cell line.
contained a fatty acid oxidation signature (Fig. 4c). Metabolite profiling
of normal mouse tissues31 showed that the brain has markedly lower
levels of TAGs compared with other tissues (Fig. 4e). This reflects brain
physiology, whereby instead of storing fatty acids as TAGs, the brain
accumulates specialized lipids to support neural activity and brain
function32. One possibility is that for breast cancer cells to survive in the
brain microenvironment, where TAGs and other storage lipids present
in other tissues are not abundant, they must access lipids via de novo
synthesis or another route, in line with the seed-and-soil hypothesis33.
To further investigate the characteristics of breast cancer cell lines
capable of brain metastasis, we analysed genome-wide CRISPR–Cas9
viability-screening data34 to identify gene vulnerabilities associated with
the brain metastatic state. We identified SREBF1 as the top-correlated
dependency with brain metastasis (FDR = 0.001) (Fig. 4f). SREBF1 is a
pivotal transcription factor that mediates lipid synthesis downstream
of the PI3K pathway27,35. SREBF1 was selectively required for growth of
brain metastatic lines in culture compared with breast cancer lines with
low or no brain metastatic potential. The association was specific to
brain, as no association was observed between SREBF1 essentiality and
metastasis to other organs (Fig. 4f). This SREBF1–breast-cancer brain
metastasis association was also recovered in the MetMap500 data-
set, indicating strong reproducibility of the finding (Extended Data
Fig. 5b, c). Of note, the SREBF1 paralogue SREBF2 showed no association
between its essentiality in culture and metastatic potential (Fig. 4g).
brain metastases from patients compared with extracranial metastases
or their matched primary tumours36 (Extended Data Fig. 9). Further-
10-4
2,000 0 20 40 more, the requirement for SREBF1, SCD, SCAP and other members of
the lipid-metabolism pathway for brain metastasis formation was con-
firmed in both mini-pool and individual gene-knockout experiments
(Fig. 5a–c, Supplementary Note 4). Together, these genetic, metabolic,
transcriptomic and functional genomic evidence all point to an associa-
tion between SREBF1-mediated lipid metabolism and brain metastasis.
Given the observation that SREBF1 knockout resulted in a viability
defect in vitro (Extended Data Fig. 10a), we compared the relative effect
of knockout on metastasis to different organs, to determine whether the
viability defect was preferentially observed in brain (Fig. 5d). Five weeks
following intracardiac injection of SREBF1-knockout cells, we observed a
marked defect in brain metastasis (196-fold reduction), compared with a
modest defect in other organs (9–21 fold) (Fig. 5d). Histologic examina-
tion of brains from xenografted mice revealed large metastatic lesions in
mice receiving wild-type cells, whereas those receiving SREBF1-knockout
cells contained micrometastases (Extended Data Fig. 10b), suggesting
that SREBF1 is not required for seeding the brain, but rather for prolifera-
tion in the brain microenvironment. Consistent with this hypothesis,
injection of tumour cells into the carotid artery increased the probability
of seeding the brain, but nevertheless a marked growth defect was still
observed in SREBF1-knockout cells (Fig. 5e).
To determine the generality of the SREBF1 requirement for breast
cancer growth in the brain, we knocked out SREBF1 in additional
brain metastatic lines including HCC1954, MDAMB231 and HCC1806
using CRISPR–Cas9. As with JIMT1, a significant inhibition in brain
metastatic growth was also observed in these lines, although the
magnitude and duration of growth inhibition varied (Extended Data
Fig. 10c, d). The least responsive cell line was HCC1806, in which
SREBF1-knockout cells displayed a brain growth defect for the first
week, but then assumed a growth trajectory that paralleled wild-type
cells. This restoration of growth was not explained by reversion of the
genome editing, as brain metastases at the end of the experiment
showed evidence of editing at the SREBF1 locus and minimal SREBF1
protein expression (Extended Data Fig. 10e, f). Instead, we found that
the SREBF1-independent growth was associated with upregulation
of the fatty acid transporter CD36 and the fatty acid-binding protein
FABP6 (Extended Data Fig. 10g). Of note, culture of HCC1806 in mouse
brain-slice-conditioned medium similarly resulted in upregulation of
SCD and CD36 expression (Extended Data Fig. 10h, i). JIMT1 cells did

334 | Nature | Vol 588 | 10 December 2020

 a Mutation b Copy number alteration c Expression signature
6 PIK3CA 8 Enriched 10-5
ADAM28-WRN (chr8p) Burton: adipogenesis peak at 8 h
6

−log10(P value)
Hallmark: PI3K–AKT–MTOR signaling

−log10(P value)
4
GO: ERBB signalling pathway
4
GO: ERBB2 signalling pathway
2 GO: carnitine metabolic process
KRAS 2
TP53 Reactome: mitochondrial fatty acid β-oxidation
BRCA1 BRAF
GO: short chain fatty acid metabolic process
0 BRCA2 0
10–5 Depleted
0 1,000 2,000 0 5,000 10,000
Gene ranks Gene ranks

st
t
rg tes
Sk inte

hi at
t
fa
La l in

us
Sp oc

W nf
Ki al
en

ey
us

Th s

te
e

ym
G t
tr

rta

e
n

ow
al
ng

Ad r
r

re
d e

dn
le

le

in
st

ve
ea

as
ai
Metabolite

Sm
Ao
So

Lu
Br

Br
Te

Li
H
Enriched 10–3 log2(FC) CE
PPP PPP DAG
LPC
CE CE LPE
PC PC PC
Lipid species

PE
SM SM SM
LPC LPC
LPE LPE
TAG
DAG DAG
TAG TAG
10–3 Depleted Abundance z-score –2 2
Brain met Non/weakly
f CRISPR dependency Brain Lung Liver g metastatic
SREBF1 HCC1806
104 JIMT1 HCC1954 SREBF1

CRISPR-dependency
Metastatic potential

6
MDAMB231
103
−log10(P value)

4
102 Kidney Bone

2 SREBF2
101

0 100
0 5,000 10,000 15,000 –3 –2 –1 0 SREBF1 gene effect –3 –2 –1 0 1
Gene ranks Gene effects
h Lipidomics: SREBF1-KO vs WT i RNA-seq: SREBF1-KO vs WT j SREBF1 co-dependency
Enriched 10–3 6
SREBF1 SCD
TAG TAG SCD
Ceramide Ceramide 75
−log10(P value)

MAG MAG SCAP

LPE LPE 4 −log10(P value)
PI PI 50
SM SM
PE PE
CE 2
CE 25
DAG DAG
LPC LPC
PC PC 0
0
10–3 Depleted log2(FC) –3 –2 –1 0 1 0 5,000 10,000 15,000
log2(FC) Gene ranks

Fig. 4 | An altered lipid-metabolism state associates with brain metastatic diacylglycerol; PPP, pentose phosphate pathway metabolites. e, Heat map
potential in basal-like breast cancer. a, Somatic mutations that associate presenting distribution of lipid species measured by mass spectrometry from
with brain metastatic potential in the basal-like breast cancer cohort. The top different mouse tissues. Gastroc, gastrocnemius. f, CRISPR gene dependencies
correlate, PIK3CA, reaches statistical significance (FDR = 0.0034, highlighted that associate with brain metastatic potential. The top gene, SREBF1
in bold). All PIK3CA mutations are activating. Positive correlations are in red, (FDR = 0.001), is a selective dependency in highly brain metastatic lines.
negative correlations are in blue. Selected known oncogenes or tumour Positive correlations are in red, negative correlations are in blue.
suppressors in basal-like breast cancer are presented for comparison. g, Distribution of SREBF1 (top) and SREBF2 (bottom) dependencies across 688
b, Alterations in copy number that associate with brain metastatic potential. human cancer cell lines. The positions of highly brain metastatic (met) breast
The top correlates cluster in chr 8p12–8p21.2 (FDR = 0.0017, highlighted in lines are highlighted in red, whereas weakly metastatic or non-brain metastatic
bold). c, Gene-expression signatures that associate with brain metastatic breast lines are highlighted in blue. h, Consensus alterations in lipid species
potential. Bars indicate P values. Expression signature scores were projected abundance upon SREBF1 knockout (KO) in JIMT1 and HCC1806, two brain
for each cell line with their in vitro RNA-seq data and used for regression metastatic cell lines. Bars indicate adjusted P values. Lipid metabolites
analysis. GO (Gene Ontology), Hallmark, Reactome and Burton are gene sets in measured by mass spectrometry were grouped by species, and enrichment
the MSigDB gene set enrichment analysis (GSEA) collection. d, Lipid-metabolite analysis of the species was performed using GSEA. WT, wild type. i, Consensus
species that associate with brain metastatic potential. Bars indicate P values. gene-expression changes upon SREBF1 knockout in JIMT1, HCC1806, HCC1954
Lipid metabolites measured by mass spectrometry were grouped by species, and MDAMB231, four brain metastatic cell lines. The two top genes are SREBF1
and enrichment analysis of the species was performed using GSEA. CE, and SCD (FDR <0.05, highlighted in bold). j, Co-dependencies of SREBF1 across
cholesterol ester; PC, phosphatidylcholine; SM, sphingomyelin; LPC, 688 human cancer cell lines in genome-wide CRISPR viability screen. The two
lysophosphatidylcholine; LPE, lysophosphatidylethanolamine; DAG, top genes are SCD and SCAP (FDR < 1 × 10 −60, highlighted in bold).

not upregulate CD36 or FABP6 expression following SREBF1 knockout

(Extended Data Fig. 10g), perhaps explaining their inability to survive Discussion
in the brain. Together, these results further demonstrate the relation- This work describes MetMap as an approach for large-scale in vivo char-
ship between lipid metabolism and brain metastasis, as cells under the acterization of human cancer cell lines. The MetMap resource (available
selective pressure of SREBF1 loss must acquire lipids by other means at https://pubs.broadinstitute.org/metmap) currently includes metas-
to survive in the brain microenvironment. tasis profiles of 500 cell lines spanning 21 tumour types, providing a

Nature | Vol 588 | 10 December 2020 | 335

Article
a JIMT1-Cas9 Arrayed guide infection Pooling b PMVK 2. Ghandi, M. et al. Next-generation characterization of the Cancer Cell Line Encyclopedia.
... 8 Nature 569, 503–508 (2019).
3. Garnett, M. J. et al. Systematic identification of genomic markers of drug sensitivity in

–log10 (adj. P)
...

...
... 6
cancer cells. Nature 483, 570–575 (2012).
UBIAD1 IRX3 4 4. Behan, F. M. et al. Prioritization of cancer therapeutic targets using CRISPR–Cas9 screens.
SCD SCAP
ACLY 2 Nature 568, 511–516 (2019).
Guide dropout Guide amplification Intracranial
SREBF1 5. Kang, Y. et al. A multigenic program mediating breast cancer metastasis to bone. Cancer
quantification from brain lesion injection
–5 –2.5 0
Cell 3, 537–549 (2003).
c JIMT1 intracranial Effect size 6. Chen, S. et al. Genome-wide CRISPR screen in a mouse model of tumor growth and
metastasis. Cell 160, 1246–1260 (2015).
107 WT SREBF1-KO SCAP-KO SCD-KO ACLY-KO PMVK-KO IRX3-KO
7. Malladi, S. et al. Metastatic latency and immune evasion through autocrine inhibition of
WNT. Cell 165, 45–60 (2016).
BLI

104
8. van der Weyden, L. et al. Genome-wide in vivo screen identifies novel host regulators of
g1 g2 g1 g2 g1 g2 g1 g2 g1 g2 g1 g2 g1 g2
107 metastatic colonization. Nature 541, 233–236 (2017).
106 9. Tasdogan, A. et al. Metabolic heterogeneity confers differences in melanoma metastatic
105 potential. Nature 577, 115–120 (2020).
104 10. Robinson, D. R. et al. Integrative clinical genomics of metastatic cancer. Nature 548,
103 297–303 (2017).
0 10 20 30 40 11. Kennecke, H. et al. Metastatic behavior of breast cancer subtypes. J. Clin. Oncol. 28,
Days post injection
3271–3277 (2010).
d JIMT1 intracardiac e JIMT1 intracarotid
Brain Lung Liver Bone Whole body
12. Yu, C. et al. High-throughput identification of genotype-specific cancer vulnerabilities in
mixtures of barcoded tumor cell lines. Nat. Biotechnol. 34, 419–423 (2016).
13. Budczies, J. et al. The landscape of metastatic progression patterns across major human

BLI
BLI

cancers. Oncotarget 6, 570–583 (2015).

14. Müller, C. et al. Hematogenous dissemination of glioblastoma multiforme. Sci. Transl.
105 107 105 107 106 108 106 108 106 108 105 107
Med. 6, 247ra101 (2014).
101 101
BLI fold change

196-fold 10-fold 21-fold 9-fold 9-fold 111-fold 15. Fonkem, E., Lun, M. & Wong, E. T. Rare phenomenon of extracranial metastasis of
100 100
glioblastoma. J. Clin. Oncol. 29, 4594–4595 (2011).
10–1 10–1
10–2
16. Stone, K. R., Mickey, D. D., Wunderli, H., Mickey, G. H. & Paulson, D. F. Isolation of a human
10–2
10–3 10–3
prostate carcinoma cell line (DU 145). Int. J. Cancer 21, 274–281 (1978).
WT KO
WT KO 17. Ramaswamy, S., Ross, K. N., Lander, E. S. & Golub, T. R. A molecular signature of
metastasis in primary solid tumors. Nat. Genet. 33, 49–54 (2003).
Fig. 5 | Investigation of lipid-metabolism genes in breast cancer brain 18. Zhang, X. H.-F. et al. Selection of bone metastasis seeds by mesenchymal signals in the
metastasis. a, A schematic of an in vivo CRISPR screen investigating relative primary tumor stroma. Cell 154, 1060–1073 (2013).
19. Campbell, P. J. et al. The patterns and dynamics of genomic instability in metastatic
gene fitness in brain metastasis outgrowth. b, Volcano plot showing the result
pancreatic cancer. Nature 467, 1109–1113 (2010).
of a mini-pool in vivo CRISPR screen targeting 29 lipid-metabolism-related 20. Witzel, I., Oliveira-Ferrer, L., Pantel, K., Müller, V. & Wikman, H. Breast cancer brain
genes. Thirteen genes scored at FDR < 0.05, with selective hits highlighted. metastases: biology and new clinical perspectives. Breast Cancer Res. 18, 8 (2016).
c, Individual gene validation of six hits by intracranial injection of JIMT1 edited 21. Kodack, D. P., Askoxylakis, V., Ferraro, G. B., Fukumura, D. & Jain, R. K. Emerging strategies
for treating brain metastases from breast cancer. Cancer Cell 27, 163–175 (2015).
cells. Cell outgrowth in brain metastasis was monitored by real-time BLI. Two
22. Curtis, C. et al. The genomic and transcriptomic architecture of 2,000 breast tumours
independent guides per gene were tested, with one guide per-mouse. d, BLI and reveals novel subgroups. Nature 486, 346–352 (2012).
quantification of relative fold change in metastasis load in the organs of mice 23. Cancer Genome Atlas Network. Comprehensive molecular portraits of human breast
receiving intracardiac injection of wild-type (WT) or SREBF1-knockout (KO) tumours. Nature 490, 61–70 (2012).
JIMT1 cells. Data are mean ± s.e.m. Each group contains five mice. e, BLI and 24. Razavi, P. et al. The Genomic Landscape of Endocrine-Resistant Advanced Breast
Cancers. Cancer Cell 34, 427–438 (2018).
quantification of relative fold change in brain metastasis load in mice receiving 25. Gatza, M. L. et al. A pathway-based classification of human breast cancer. Proc. Natl Acad.
intracarotid injection of wild-type or SREBF1-KO JIMT1 cells. Data are Sci. USA 107, 6994–6999 (2010).
mean ± s.e.m. n = 7 (wild-type) and n = 8 (knockout) mice. 26. Creighton, C. J. et al. Proteomic and transcriptomic profiling reveals a link between the
PI3K pathway and lower estrogen-receptor (ER) levels and activity in ER+ breast cancer.
Breast Cancer Res. 12, R40 (2010).
large repertoire of models for exploration of metastasis mechanisms. 27. Ricoult, S. J. H., Yecies, J. L., Ben-Sahra, I. & Manning, B. D. Oncogenic PI3K and K-Ras
stimulate de novo lipid synthesis through mTORC1 and SREBP. Oncogene 35,
A limitation of the use of human cell lines for such experiments is that 1250–1260 (2016).
they require the use of immunodeficient mice. The extent to which the 28. Cai, Y. et al. Loss of chromosome 8p governs tumor progression and drug response by
immune system has a role in mediating patterns of metastasis remains altering lipid metabolism. Cancer Cell 29, 751–766 (2016).
29. Li, H. et al. The landscape of cancer cell line metabolism. Nat. Med. 25, 850–860 (2019).
to be determined37. 30. Patra, K. C. & Hay, N. The pentose phosphate pathway and cancer. Trends Biochem. Sci.
We followed up only a small proportion of the MetMap findings— 39, 347–354 (2014).
specifically, breast cancer metastasis to brain. Multiple lines of experi- 31. Jain, M. et al. A systematic survey of lipids across mouse tissues. Am. J. Physiol.
Endocrinol. Metab. 306, E854–E868 (2014).
mental and clinical evidence pointed to a role of lipid metabolism in 32. Piomelli, D., Astarita, G. & Rapaka, R. A neuroscientist’s guide to lipidomics. Nat. Rev.
governing the ability of cells to survive in the brain microenvironment. Neurosci. 8, 743–754 (2007).
The importance of lipid metabolism in cancer has been highlighted by 33. Paget, S. The distribution of secondary growths in cancer of the breast. 1889. Cancer
Metastasis Rev. 8, 98–101 (1989).
a number of studies, but its role in brain metastasis has, to our knowl- 34. Dempster, J. M. et al. Agreement between two large pan-cancer CRISPR–Cas9 gene
edge, not been fully appreciated38–41. The possibility that interfering dependency data sets. Nat. Commun. 10, 5817 (2019).
with lipid or cholesterol metabolism might abrogate metastatic growth 35. Horton, J. D., Goldstein, J. L. & Brown, M. S. SREBPs: activators of the complete program
of cholesterol and fatty acid synthesis in the liver. J. Clin. Invest. 109, 1125–1131 (2002).
in the brain is particularly intriguing. More generally, this work illus- 36. Varešlija, D. et al. Transcriptome characterization of matched primary breast and brain
trates the complex interplay between cancer cell growth and the tissue metastatic tumors to detect novel actionable targets. J. Natl. Cancer Inst. 111, 388–398
microenvironment. (2019).
37. Angelova, M. et al. Evolution of metastases in space and time under immune selection.
Cell 175, 751–765 (2018).
38. Zhang, M. et al. Adipocyte-derived lipids mediate melanoma progression via FATP
Online content proteins. Cancer Discov. 8, 1006–1025 (2018).
39. Zou, Y. et al. Polyunsaturated fatty acids from astrocytes activate PPARγ signaling in
Any methods, additional references, Nature Research reporting sum- cancer cells to promote brain metastasis. Cancer Discov. 9, 1720–1735 (2019).
maries, source data, extended data, supplementary information, 40. Pascual, G. et al. Targeting metastasis-initiating cells through the fatty acid receptor
acknowledgements, peer review information; details of author con- CD36. Nature 541, 41–45 (2017).
41. Sullivan, M. R. et al. Quantification of microenvironmental metabolites in murine cancers
tributions and competing interests; and statements of data and code reveals determinants of tumor nutrient availability. eLife 8, e44235 (2019).
availability are available at https://doi.org/10.1038/s41586-020-2969-2.
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in
published maps and institutional affiliations.
1. Barretina, J. et al. The Cancer Cell Line Encyclopedia enables predictive modelling of
anticancer drug sensitivity. Nature 483, 603–607 (2012). © The Author(s), under exclusive licence to Springer Nature Limited 2020

336 | Nature | Vol 588 | 10 December 2020

Methods
No statistical methods were used to predetermine sample size. The
experiments were not randomized. The investigators were not blinded
to allocation during experiments and outcome assessment.
Breast cancer cell lines and barcoding
Breast cell lines were cultured under the recommended conditions from
CCLE (https://portals.broadinstitute.org/ccle). Cell line identities were
confirmed by SNP fingerprinting as well as RNA-seq, and compared to
the CCLE results. All cell lines were tested negative for mycoplasma. The
fluorescence-luciferase-barcode (FLB) construct was engineered using
the FUW lentiviral vector backbone (a gift from D. Baltimore; Addgene
plasmid no. 14882). Barcodes 26 nucleotides in length were designed
using barcode_generator.py (v.2.8; http://comailab.genomecenter.
ucdavis.edu/index.php/), and cloned into the landing pad C-terminal to
the TGA stop codon of fluorescence luciferase using Gibson assembly
(New England Biolabs). Lentivirus preparation and cell infection were
performed according to published protocols available at http://www.
broadinstitute.org/rnai. Infected cells were analysed by FACS with a
fixed gate for GFP or mCherry, using a Sony SH4800 sorter.
Animal studies
Animal work was performed in accordance with a protocol approved
by the Broad Institute Institutional Animal Care and Use Committee
(IACUC). NSG female mice (The Jackson Laboratory) at 5–6 weeks of age
were used. Cancer cells were suspended in PBS, 0.4% BSA and 100 μl of
cell suspensions were injected into the left ventricle of anaesthetized
mice (ketamine 100 mg kg−1; xylazine 10 mg kg−1). In vivo metastasis
progression was monitored via real-time BLI using the IVIS SpectrumCT
Imaging System (PerkinElmer) on a weekly basis. Mice were anaesthe-
tized with inhaling isoflurane, injected intraperitoneally with D-Luciferin
(150 mg kg−1), and imaged with the auto exposure setting in prone and
supine positions. At the end point, ex vivo BLI was performed by sub-
merging the excised organs in DMEM/F12 medium (Thermo Fisher
Scientific) containing D-Luciferin for 10 min and imaged with the auto
exposure setting. BLI analysis was performed using Living Image soft-
ware (v.4.5, PerkinElmer). In the case of breast cancer cohort study (pilot,
group 1 and group 2 in Fig. 1, Extended Data Fig. 1), cell lines were mixed
at an equal ratio immediately before animal injection, and cell line pools
containing 2 × 104 cells per barcoded line were injected. In the case of
single breast cell line validation (Extended Data Fig. 1n), cell lines were
injected individually at a density of 2 × 104 cells, to be comparable with
the pooled experiments. In the case of MetMap125 (Fig. 2, Extended
Data Fig. 2), PRISM pools of 25 cell lines were used, and 2.5 × 105 total
cells were injected per mouse, corresponding to 1 × 104 cells per bar-
coded line. Five PRISM pools were injected separately into cohorts of
5–6-week-old NSG mice. In the case of MetMap500, 20 PRISM pools of
25 cell lines were combined to form a large pool of 498 cell lines. The large
pool was injected into a cohort of 8–10-week-old NSG mice, with 2.5 × 105
cells per mouse, equivalent to a density of 500 cells per line. Mammary
fat pad and subcutaneous injections were performed with Matrigel
(Corning) support, at a matching density to their intracardiac assays,
respectively (Extended Data Fig. 3). For all pooled cell line experiments,
mice were euthanized 5 weeks after injection, in a time-matched manner,
unless they displayed severe paralysis or poor body condition, in which
case they were euthanized earlier. Intracartoid injection of JIMT1 was
performed following a published protocol42, at a density of 1 × 105 cells
per mouse, similar to the intracardiac injection (Fig. 5e). Intracranial
injection was performed as previously described43, at a density of 1 × 103
cells per perturbation per animal (Fig. 5a–c, Extended Data Fig. 10c, d).
Tissue processing and cancer cell isolation from organs
Organs including brain, lung, liver, kidney were dissociated using
gentleMACS Octo Dissociator with Heaters (Miltenyi Biotec).
The optimized dissociation solutions and programs (Miltenyi Bio-
tec) are listed in Supplementary Table 9. Bones (from both hind limbs)
were chopped into fine pieces and incubated in the dissociation buffer
with vigorous shaking. The dissociated cell suspensions were filtered
using 100-μm filters, and washed with DMEM/F12 twice. Cell suspen-
sions were then washed with staining buffer (PBS, 2 mM EDTA, 0.5%
BSA), and incubated with mouse cell depletion beads according to
the instructions (Miltenyi Biotec). Cell suspensions were subjected
to negative selection using autoMACS Pro Separator (Miltenyi Bio-
tec) to deplete mouse stroma. Brains were subjected to an additional
myelin-debris-depletion step using myelin removal beads II (Miltenyi
Biotec). The resultant cell suspensions were then analysed by FACS
using a Sony SH4800 sorter, with the fixed gate for GFP or mCherry.
DAPI staining was used to exclude dead cells. For bulk RNA-seq, cells
were sorted to a single tube in PBS, 0.4% BSA and RNasin Plus RNase
Inhibitor (Promega), centrifuged at 1,500 rpm for 10 min, and cell pel-
lets were frozen at −80 °C for downstream use. For single-cell RNA-seq,
single cells were sorted into 96-well plates containing cold TCL buffer
(Qiagen) containing 1% β-mercaptoethanol, snap frozen on dry ice,
and then stored at −80 °C. Ninety single cells were sorted per plate,
the rest wells on the plate were used for negative and positive controls.
RNA extraction, library preparation and sequencing
Individual cell lines, cell line pools before injection, and cells isolated
from metastases were analysed by RNA-seq. RNA extraction was per-
formed using Quick-RNA MicroPrep according to the manufacturer’s
instructions (Zymo Research). RNA was quantified using an RNA 6000
Pico Kit on a 2100 Bioanalyzer (Agilent). RNA samples from cell num-
bers lower than 500 were not measured but all were used as input for
library preparation. cDNA was synthesized using Clontech SmartSeq
v.4 reagents from up to 2 ng RNA input according to the manufacturer’s
instructions (Clontech). Full-length cDNA was fragmented to a mean
size of 150 bp with a Covaris M220 ultrasonicator and Illumina libraries
were prepared from 2 ng of sheared cDNA using Rubicon Genomics
Thruplex DNaseq reagents according to the manufacturer’s protocol.
The finished double-stranded DNA (dsDNA) libraries were quantified
by Qubit fluorometer, Agilent TapeStation 2200, and RT–qPCR using
the Kapa Biosystems library-quantification kit. Uniquely indexed librar-
ies were pooled in equimolar ratios and sequenced on Illumina Next-
Seq500 runs with paired-end 75-bp reads at the Dana-Farber Cancer
Institute Molecular Biology Core Facilities. RT–qPCR quantification of
barcodes was performed using Maxima First Strand cDNA Synthesis
Kit, Taqman Fast Advanced Master Mix, custom synthesized Taqman
probes, and QuantStudio 6 PCR System (ThermoFisher Scientific).
Single-cell RNA-seq was performed as previously described44.
Bioinformatic analysis
Barcode quantification from RNA-Seq of metastases. Because the
RNA-seq library preparation sheared the cDNA randomly into small
pieces, demultiplexed RNA-seq reads were mapped to the barcode
references using Bowtie 2 local mode45 for barcode detection and
quantification. Mapped reads were filtered with the criteria that reads
(either 5′ or 3′) must cover over 50% of the barcodes from either end,
and counted using samtools. Barcode percentage corresponding to
cell composition was calculated for single cell lines, pre-injected cell
mixtures and in vivo metastasis samples.
Metastatic potential quantification and feature associations. For
breast cohort study, metastatic potential of cell line j targeting organ
1 n
i, Mi,j was calculated as: Mi , j = n ∑ k =1 cipj , in which ci is the total cancer
cell number isolated from organ i, pj is the fractional proportion of cell
line j estimated by barcode quantification, and n is the number of rep-
licates of mice. To identify features that associate with brain meta-
static potential, a two-class comparison method was used46. The
analysis was performed on mutation, copy number, expression,
Article
metabolite, and CRISPR-gene dependency (available at https://depmap.
org/portal/). Copy number data were binarized using a cutoff of ≤−1
(loss) and ≥1 (gain).
Cancer transcriptomic analysis from RNA-seq of metastases. Po-
tential mouse contaminating reads were removed by competitive map-
ping to the human/mouse hybrid genome using BBSplit (https://source-
forge.net/projects/bbmap/). Reads that uniquely mapped to the human
genome were then used as input for mapping and gene-level counting
with the RSEM package47. Gene count estimates were normalized using
the TMM method in edgeR48. For differential analysis, to properly ac-
count for the cancer cell composition differences in each in vivo sam-
ple, an in silico modelled in vitro mixture was generated first. For each
in silico metastasis model, the estimated expression ĝ of gene i is com-
puted as a weighted average of the cell lines present in the correspond-
M
ing in vivo sample: gˆi = ∑ j =1 gi , j pj , in which gi,j is the baseline in vitro
expression of gene i in cell line j and pj is the fractional proportion of
cell line j in the in vivo sample, as estimated by barcode quantification,
and M is the number of cell lines present in the in vivo sample. The in vivo
and in silico counterpart were then compared using a paired design for
each organ in voom-limma46. GSEA was performed using camera or
GSEA-preranked method implemented in fgsea46,49. Single sample GSEA
signature projection was performed using gsva package 50.
Gene-signature datasets were from MSigDB (https://www.gsea-msigdb.
org/).
PRISM in vivo assay
PRISM pool preparation. PRISM cell lines (source of each available
at https://depmap.org/portal/) were adapted to the same culture
conditions in phenol-red-free RPMI1640 medium (ThermoFisher
Scientific) and barcoded as previously described12. SNP fingerprint-
ing authentication was performed before and after barcoding. My-
coplasma contamination was examined (MycoScope, Genlantis) and
only negative lines were used for experiments. These included eight
oestrogen-receptor-positive breast cancer cell lines. Despite the lack
of phenol red (a weak oestrogen) these breast cell lines maintained
ESR1 positivity and expression of a downstream marker of its activity,
FOXA1. This is probably explained by the remaining oestrogens in the
fetal bovine serum (FBS). PRISM cell lines were pooled on the basis of
their in vitro doubling bins, at equal number, in the format of 25 lines
per pool, and cryopreserved until use. Cells were thawed and recovered
for 48 h before in vivo injection. To form the large pool of 498 cell lines,
20 PRISM pools were mixed at equal total number immediately before
injection.
Tissue processing, library preparation and sequencing. After in vivo
experiments, organs were subjected to tissue dissociation, mouse
stroma depletion, and the dissociated cell pellets were frozen at −80 °C
as described above. The pellets (≤50 mg dry weight) were lysed in 200 μl
freshly prepared lysis buffer with proteinase K, heat digested at 60 °C,
and denatured at 95 °C for 10 min. Twenty microlitres of the lysates was
used for barcode amplification per 100 μl PCR volume (multiple tech-
nical replicates per sample). PCR was performed using the following
conditions: 95 °C for 3 min; 98 °C for 20 s, 57 °C for 15 s, 72 °C for 10 s (30
cycles); 72 °C for 5 min; 4 °C stop. PCR libraries were pooled, purified
using Select-a-Size DNA Clean & Concentrator Kit (Zymo Research), and
quantified using Qubit dsDNA HS Assay Kit (ThermoFisher Scientific)
and a 2100 Bioanalyzer (Agilent). The purified 2 nM of libraries with
20% spike-in PhiX DNA were sequenced on Illumina MiSeq or HiSeq at
800 K mm−2 cluster density.
Metastatic potential quantification. Demultiplexed sequencing reads
were mapped to the barcode reference to generate a table of cell line
barcode counts for each sample/condition. Sequencing-depth normal-
ized read counts were used for calculation of relative metastatic
potential. Relative metastatic potential of cell line j targeting organ i,
1 n 1 m
rMi,j, was defined as: rMi , j = n ∑ k =1 ci , j / m ∑ k =1 pj , in which ci,j is the read
counts of cell line j from organ i, pj is the read counts of cell line j from
pre-injected population, n is the number of replicate samples of mice,
m is the number of replicates of pre-injected population. Confidence
intervals were calculated using bootstrap resampling.
In vivo CRISPR screen and gene validation
CRISPR–Cas9 versions of cell lines were generated by infecting lucifer-
ized cells with Cas9-Blast lentivirus and selecting in 5 μg ml−1 blasticidin
for 10 days with continuous passaging until non-infected controls were
killed. For pooled in vivo screen, JIMT1–Cas9 cells were infected with a
CRISPR guide library (Supplementary Table 10) in an arrayed-fashion in
6-well plates, and selected in 2 μg ml−1 puromycin for 4 days. At this time,
non-infected controls were killed, and no growth defect was observed in
the perturbed cell lines. Post antibiotic selection, cells were pooled and
subjected to intracranial injection at 6 × 104 cells per mouse in 1 μl PBS.
This was equivalent to 1 × 103 cells per guide on average per mouse. Intrac-
ranial growth was allowed to progress for 4 weeks, and brain tissues
were processed adopting the workflow of PRISM in vivo assay, except
that guides were amplified using primers targeting the guide vector.
Demultiplexed sequencing reads were mapped to the guide reference
to generate a table of barcode counts for each guide for each sample.
Sequencing-depth was normalized using the upper-quartile method
and relative depletion was quantified using a linear model in limma46.
For individual gene validation (Fig. 5c, Extended Data Fig. 10c, d),
Cas9-expressing cells of different cell lines were infected with cor-
responding guides, selected in 2 μg ml−1 puromycin for 4 days, and
subjected to intracranial injection at 1 × 103 cells per mouse in 1 μl PBS.
Two independent guides per gene were tested, with one mouse per
guide. Intracranial growth was monitored by BLI following injection.
Liquid chromatography–mass spectrometry lipidomics
Positive ion mode analyses of polar and nonpolar lipids (C8-pos) were
conducted using a liquid chromatography–mass spectrometry (LC–MS)
system composed of a Shimadzu Nexera X2 U-HPLC (Shimadzu) cou-
pled to an Exactive Plus orbitrap mass spectrometer (ThermoFisher
Scientific). Cellular extracts were collected from 6-well plate culture,
in LC-MS-grade isopropanol (Sigma-Aldrich) containing an internal
standard 1,2-didodecanoyl-sn-glycero-3-phosphocholine (Avanti Polar
Lipids). Extracts were centrifuged for 10 min at 10,000g to remove resid-
ual cellular debris. After centrifugation, supernatants were injected
directly onto a 100 × 2.1 mm, 1.7-μm ACQUITY BEH C8 column (Waters).
The column was eluted isocratically with 80% mobile phase A (95:5:0.1
v/v/v 10 mM ammonium acetate/methanol/formic acid) for 1 min fol-
lowed by a linear gradient to 80% mobile phase B (99.9:0.1 v/v methanol/
formic acid) over 2 min, a linear gradient to 100% mobile phase B over
7 min, then 3 min at 100% mobile phase B. Mass spectrometry analyses
were performed using electrospray ionization in the positive ion mode
using full scan analysis over 200 to 1,000 m/z at 70,000 resolution and 3
Hz data acquisition rate. Other mass spectrometry settings were as fol-
lows: sheath gas 50, in source collision-induced dissociation 5 eV, sweep
gas 5, spray voltage 3 kV, capillary temperature 300 °C, S-lens RF 60,
heater temperature 300 °C, microscans 1, automatic gain control target
106, and maximum ion time 100 ms. Lipid identities were determined on
the basis of comparison to reference standards and reference plasma
extracts and were denoted by the total number of carbons in the lipid
acyl chain(s) and total number of double bonds in the lipid acyl chain(s).
Western blot
Protein lysates were prepared in RIPA lysis buffer (ThermoFisher Sci-
entific) with cOmplete Mini EDTA-free Protease Inhibitor Cocktail
(Roche). Western blot was performed using NuPAGE gel (ThermoFisher
Scientific) with wet tank blotting (Bio-Rad) and Odyssey detection
system (LI-COR). SREBF1 primary antibody (14088-1-AP, Proteintech),
SCD (CD.E10) antibody (ab19862, Abcam), GAPDH (D16H11) XP rabbit
monoclonal antibody (5174S, Cell Signaling), β-actin (8H10D10) mouse
monoclonal antibody (3700S, Cell Signaling), and IRDye 800CW goat
anti-mouse IgG (926-32210, LI-COR), IRDye 680RD goat anti-rabbit IgG
(926-68071, LI-COR) secondary antibodies were used. Western blot was
performed for cells cultured in different medium conditions. These
include RPMI1640 with 10% FBS, with 10% delipidated FBS, with 10%
human cerebrospinal fluid (991-19-P-5, Lee BioSolutions), or with 1%
SM1 supplement (05711, STEMCELL Tech), or brain-slice-conditioned
medium. Brain-slice-conditioned medium was prepared by submerg-
ing brain slices (150 μm) in RPMI1640 (no serum) for 48 h. Delipidated
FBS was prepared as described51.
Clinical data analysis
METABRIC, TCGA and MSK targeted-sequencing breast cancer datasets
were downloaded from cBioPortal52. The EMC-MSK dataset including
615 primary tumours (GSE2035, GSE2603, GSE5327 and GSE12276)
and the 65-metastasis-sample dataset (GSE14020) were collected and
processed as previously described18. Paired primary breast tumour
and brain metastasis RNA-seq was obtained from ref. 36. To exclude the
confounding effect of brain stroma contamination in this dataset, a con-
tamination indicator generated from GSE52604 was applied, and the
contaminating effect was regressed out, generating a corrected gene
matrix. PI3K-response signatures were from refs. 25,26. Signature analysis
was conducted as described7. Hierarchical clustering and heatmaps
were generated using gplots package. Other plots were generated using
ggplot2. log-rank tests of survival curve difference were calculated
using survival package. A multivariate Cox proportional hazards model
was built using the coxph function (Extended Data Fig. 6h). Significance
of overlap was calculated using chisq.test or fisher.test function.
Reporting summary
Further information on research design is available in the Nature
Research Reporting Summary linked to this paper.
Data availability
MetMap data and interactive visualization can be accessed at https://
pubs.broadinstitute.org/metmap. RNA-seq data generated from
this study have been deposited in the Gene Expression Omnibus
(GEO) under accession numbers GSE148283 and GSE148372. Addi-
tional datasets used in this study include METABRIC, TCGA and
MSK-targeted-sequencing breast cancer datasets from cBioPortal,
the EMC-MSK dataset (GSE2035, GSE2603, GSE5327 and GSE12276),
the 65-metastasis-sample dataset (GSE14020), paired primary tumour
and brain metastasis RNA-seq from ref. 36, and GSE52604. Source data
are provided with this paper.
Code availability
Custom codes used for this study are accessible at the MetMap portal
(https://pubs.broadinstitute.org/metmap).
42. Zhang, C., Lowery, F. J. & Yu, D. Intracarotid cancer cell injection to produce mouse
models of brain metastasis. J. Vis. Exp. 120, e55085 (2017).
43. Ozawa, T. & James, C. D. Establishing intracranial brain tumor xenografts with subsequent
analysis of tumor growth and response to therapy using bioluminescence imaging. J. Vis.
Exp. 41, e1986 (2010).
44. Tirosh, I. et al. Dissecting the multicellular ecosystem of metastatic melanoma by
single-cell RNA-seq. Science 352, 189–196 (2016).
45. Langmead, B. & Salzberg, S. L. Fast gapped-read alignment with Bowtie 2. Nat. Methods
9, 357–359 (2012).
46. Ritchie, M. E. et al. limma powers differential expression analyses for RNA-sequencing
and microarray studies. Nucleic Acids Res. 43, e47 (2015).
47. Li, B. & Dewey, C. N. RSEM: accurate transcript quantification from RNA-seq data with or
without a reference genome. BMC Bioinformatics 12, 323 (2011).
48. Robinson, M. D., McCarthy, D. J. & Smyth, G. K. edgeR: a Bioconductor package for
differential expression analysis of digital gene expression data. Bioinformatics 26,
139–140 (2010).
49. Korotkevich, G., Sukhov, V. & Sergushichev, A. Fast gene set enrichment analysis. Preprint
at https://doi.org/10.1101/060012 (2019).
50. Hänzelmann, S., Castelo, R. & Guinney, J. GSVA: gene set variation analysis for microarray
and RNA-seq data. Bmc Bioinformatic 14, 7 (2013).
51. Hosios, A., Li, Z., Lien, E. & Heiden, M. Preparation of lipid-stripped serum for the study of
lipid metabolism in cell culture. Bio Protoc. 8, e2876 (2018).
52. Cerami, E. et al. The cBio cancer genomics portal: an open platform for
exploring multidimensional cancer genomics data. Cancer Discov. 2, 401–404
(2012).
Acknowledgements We thank J. L. Goldstein for suggestions; A. Regev, N. Marjanovic and
A. Bankapur for assistance with single-cell RNA-seq and analysis; Z. Herbert for assistance with
RNA-seq; T. Mason for assistance with next-generation sequencing; S. Kim and S. Roberge for
assistance with animal work; B. Wong for suggestions on figure and portal designs; and G. Wei,
U. Ben-David, S. Corsello, P. Tsvetkov, I. Tirosh, R. Hosking and C. Mader for discussions. X.J.
and G.B.F. were Susan G. Komen Fellows. A.A. is a HHMI Medical Research Fellow. This work
was supported by BroadNext10, Broad SharkTank grants (X.J.), HHMI (T.R.G.), and in part by
Koch Institute/DFHCC Bridge project grant (M.G.V.H. and R.K.J.). R.K.J. acknowledges support
from the NIH (R35CA197742, R01CA208205 and U01CA224173), National Foundation for
Cancer Research; the Ludwig Center at Harvard; the Jane’s Trust Foundation; the Advanced
Medical Research Foundation and by the U.S Department of Defense Breast Cancer Research
Program Innovator Award W81XWH-10-1-0016.
Author contributions X.J. conceptualized the project, conducted experiments, collected
data and analysed results. Z.D. assisted with experiments. K.N. and T.N. assisted with
bioinformatic and RNA-seq analysis. A.A., A.D., C.B.C. and M.G.V.H. performed lipidomics
and data interpretation. G.B.F. and R.K.J. performed intracranial injection experiments and
data analysis. L.P. and A.A.T. assisted with petal plot and portal development. C.Z., L.W., D.R.
and J.R. assisted with PRISM assay and data generation. V.M. and K.C. performed tissue
imaging, data acquisition and analysis. T.R.G. supervised the research. X.J. and T.R.G. wrote
the manuscript.
Competing interests T.R.G. receives research funding unrelated to this project from Bayer
HealthCare, Novo Ventures and Calico Life Sciences; holds equity in FORMA Therapeutics;
is a consultant to GlaxoSmithKline; and is a founder of Sherlock Biosciences. M.G.V.H. is a
scientific advisory board member for Agios Pharmaceuticals, Aeglea Biotherapeutics,
Auron Therapeutics and iTeos Therapeutics. R.K.J. received a honorarium from Amgen;
consultant fees from Chugai, Merck, Ophthotech, Pfizer, SPARC and SynDevRx; owns equity
in Accurius, Enlight, Ophthotech and SynDevRx; and serves on the Boards of Trustees of
Tekla Healthcare Investors, Tekla Life Sciences Investors, Tekla Healthcare Opportunities
Fund and Tekla World Healthcare Fund. No reagents or funding from these organizations
were used in this study. X.J. and T.R.G. are named as inventors on pending PCT Patent
Application No. PCT/US20/29584 filed by The Broad Institute, which describes
compositions and methods for characterizing the metastatic potential of cancer cell lines.
The other authors declare no competing interests.
Additional information
Supplementary information is available for this paper at https://doi.org/10.1038/s41586-020-
2969-2.
Correspondence and requests for materials should be addressed to X.J. or T.R.G.
Peer review information Nature thanks Roger Gomis, Jason Locasale, Ultan McDermott and
the other, anonymous, reviewer(s) for their contribution to the peer review of this work.
Reprints and permissions information is available at http://www.nature.com/reprints.
Article

Extended Data Fig. 1 | See next page for caption.

Extended Data Fig. 1 | An in vivo barcoding approach to establish
multiplexed cancer metastasis xenografts and validation using
orthogonal assays. a, Principal component analysis (PCA) of transcriptomic
expression of the breast cancer collection from CCLE, and the pooling schemes
focusing on basal-like breast cancer. G, GFP; R, mCherry. The linked numbers
indicate the labelling barcodes. b, Real-time BLI monitoring of the overall
metastasis progression from pilot, group1, group2 cell line pools. Data are
mean ± s.e.m. n = 5 (pilot), n = 8 (group1), n = 7 (group2) mice. c, Total cancer cell
numbers isolated by FACS from each target organ from pilot, group1, group2
pools. Each dot represents an animal. Box plots display quartiles of the data.
d, Cancer cell composition of metastases from different organs as determined
by barcode abundance from pilot, group1, group2 pools. pilot: G portion
samples are highlighted in green, R portion samples are highlighted in red.
preinj, pre-injected population. Data c, d were used to quantify the metastatic
potential presented in Fig. 1b. e, An example of the gating strategy to isolate
GFP+ barcoded cancer cells for the pilot pool. Infected cell lines expressed GFP
at different levels as shown in the histogram, and a fixed gate was used to enrich
cells with closer expression level. Numbers correspond to cell percentages
within the gate. f, An example of barcode mapping result visualized by
Integrative Genomics Viewer (IGV). g, Distribution of the barcode read counts
versus all gene transcript counts. Barcodes are among the top 10% highly
expressed genes, allowing robust quantification. h, An example of barcode
read quantification in the pre-injected and metastasis samples from pilot pool.
Barcodes are listed as in a. cpk, counts per thousand. i, Taqman assay on in vitro
cultured barcoded cell lines from the pilot pool. The signal is very specific to
each barcode and there is no cross detection. j, Quantification of barcode
abundance and cancer cell composition using the Taqman RT–qPCR assay in
the pre-injected and metastasis samples from the pilot pool. The results agree
with barcode quantification from bulk RNA-Seq (Extended Data Fig. 1d).
k, Single cell RNA-Seq of metastases from different organs from the pilot pool.
Single cancer cells isolated from each organ were sorted into 96-well plates,
with 90 cells per plate (rest 6 wells for positive and negative controls), and
subjected to Smart-Seq2. PCA revealed that PC1 maximally separated the
cancer cells into 2 clusters (CLs), with CL1 enriched in cells isolated from brain,
and CL2 enriched in cells isolated from lung, liver and bone. Heat map on the
right shows gene expression that associates with PC1 and clustering of cells.
Based on marker expression, CL1 corresponds to HCC1954 (ERBB2+, CDH1+)
and CL2 corresponds to MDAMB231 (CDKN2A loss, VIM+). l, Projection of
marker gene expression on the PCA plot. m, Cancer cell composition based on
single cell RNA-Seq data. The results agree with barcode quantification from
bulk RNA-Seq (Extended Data Fig. 1d). n, Real-time BLI monitoring of
metastasis progression of the 8 cell lines that were individually tested. Each
plot highlights one of the 8 lines. Data are mean ± s.e.m. Each group contains
4 mice. o, Scatter plot showing the correlation of overall metastatic potential
(5 organs combined) from pooled cell line experiments with whole body BLI of
metastases measured individually. Pearson’s correlation coefficient and its
test P value are presented.
Article

Extended Data Fig. 2 | See next page for caption.

Extended Data Fig. 2 | Using PRISM cell line pools for metastatic potential
profiling. a, Optimizing the workflow of metastatic potential mapping using
PRISM. A PRISM pool of 25 cell lines was used for testing the need of GFP
labelling and cancer cell purification. The barcode abundance altered
compared to the unlabelled population after GFP labelling as shown by the pie
chart. b, A detailed line-by-line view of barcode abundance before and after GFP
labelling. The unlabelled cell pool had more even distribution. Post labelling,
several lines showed noticeable dropout, but all lines were detectable.
c, Scatter plot comparing barcode enrichment after normalizing to the pre-
injected input from the two experiments. Pearson’s correlation coefficient and
its test P value are presented. Strong positive correlation is observed, with the
exception of one cell line U2OS. d, Quality control of MetMap500 and
MetMap125 datasets showing initial barcode abundance in the pre-injected
populations. MetMap500, 1 large pool containing 498 cell lines was profiled,
with 10 cell lines showing low initial abundance. These 10 cell lines were not
detected in any in vivo sample, and were excluded from subsequent analysis.
MetMap125, 5 pools of 25 cell lines were profiled separately and data were
combined for analysis. e, Quality control of MetMap500 and MetMap125
datasets showing scatter plots of raw barcode abundance from in vivo organs
versus the data normalized to the pre-injected input (in d). A strong linear
relationship was observed.
Article
Extended Data Fig. 3 | Subcutaneous injection of PRISM cell line pool.
a, The same PRISM pool of 498 cell lines used for MetMap500 profiling was
tested using subcutaneous injection on a cohort of 6 mice. Survival curves
compare animal survival difference between subcutaneous and intracardiac
(IC) injections, P value calculated using two-sided, log-rank test. b, Total
numbers of cell lines detected in animals from the subcutaneous and IC
injections. Detected lines are coloured in pink and non-detected lines are
coloured in light-blue. P value calculated using two-sided t-test. c, Scatter plot
showing barcode-quantified tumorigenic potential and metastatic potential
from subcutaneous and IC experiments respectively. d, Group1 of basal breast
cancer pool (Extended Data Fig. 1a) was subjected to mammary fat pad
injection, barcode quantitation through RNA-Seq, and cell number inference.
Extended Data Fig. 4 | Association of overall metastatic potential with
clinical parameters. a. Bar plots showing significance of single variate and
multi variate association analysis with metastatic potential in MetMap500.
P values are calculated using linear regression and Anova (type II) of the linear
models. The dotted lines indicate 0.05 cutoff. b. Box plots showing metastatic
potential of cell lines stratified by metastasis status in the corresponding
patients and cancer lineage. Box plots display quartiles of the data. Outlier
points extend beyond 1.5 × interquartile ranges from either hinge. c, Scatter
plots showing correlation of metastatic potential with patient age, stratified by
cancer lineage. An inverse correlation was observed in several cancer types.
d–g, Correlation of overall metastatic potential with derived site (d), time
length in culture to derive the cell lines (e), mutation burden (f) and cell
doubling (g) in the 21 basal breast cancer cohort. d, P value calculated using
two-sided t-test. e–g, Pearson’s correlation coefficients and test P values are
presented.
Article
Extended Data Fig. 5 | Genetic correlates of brain metastatic potential in
basal-like breast cancer. a. A line-by-line view of brain metastatic potential
and its associated features at genetic, expression, metabolite, and gene
dependency levels. Mutation: mutant (MUT), wild-type (WT). Copy number:
data are binarized, with deletion (DEL) cutoff < = -1 and amplification (AMP)
cutoff > = 1. Expression signatures: 1. Hallmark: PI3K/AKT/MTOR signalling,
2. GO: ERBB signalling pathway, 3. GO: ERBB2 signalling pathway, 4. Burton:
adipogenesis peak at 8hr, 5. GO: carnitine metabolic process, 6. Reactome:
mitochondrial fatty acid beta oxidation, 7. GO: short chain fatty acid metabolic
process. Data not available for the cell lines are marked with X. b, c, Scatter
plots showing the correlation of SREBF1 in vitro dependency and brain
metastatic potential in MetMap500 (a) and MetMap125 (b). Strong inverse
correlation was observed for breast cancer in both datasets. Each dot
represents a cell line.
Extended Data Fig. 6 | See next page for caption.
Article
Extended Data Fig. 6 | Association of chr 8p gene copy number status and
PI3K-response signatures with brain metastasis in clinical breast cancer
specimens. a, Heat maps showing coordinated expression of chr 8p genes
mirrored their copy number status in the two large breast cancer datasets,
METABRIC and TCGA. The 8plow cluster is defined by CNA data. The right panel
shows distribution of 8plow cluster in different breast cancer subtypes and its
association with disease specific survival. P values calculated using two-sided,
log-rank tests. CNA, Copy Number Alteration. Exp, RNA-Seq Expression.
b, Hierarchical clustering of primary breast tumours by 8p gene expression in
the EMC-MSK dataset. The 8plow cluster is enriched in tumours that developed
brain metastasis, but not lung or bone metastasis. The right panel shows organ-
specific metastasis free survival curves stratified by 8plow status. The 8plow
cluster displays poorer brain metastasis compared to the 8pWT cluster. Brain
metastasis free survival curves stratified by 8plow status in different subtypes
is also presented. P values calculated using two-sided, log-rank tests.
c, Hierarchical clustering of breast cancer metastases by 8p gene expression,
with the 8plow cluster being enriched in brain metastases. d, Chr 8p CNA status
determined by Targeted Seq in the MSK metastatic breast cancer dataset. Brain
metastases are enriched in chr 8p deletion compared to primary tumour, local
recurrence and metastases at other sites. The 8plow cluster predicts poor brain
metastasis free survival. P values calculated using two-sided, log-rank tests. LN,
lymph node. e, Heat maps showing co-regulated patterns of two independent
PI3K-response signatures in METABRIC and TCGA breast cancer datasets.
PI3Ksig.1 was generated by overexpression of PIK3CAmut in breast epithelial
cells. PI3Ksig.2 was generated by PI3K inhibitor treatment in the CMap
database. The right panel shows distribution of PI3Ksighigh cluster in different
breast cancer subtypes and its association with disease specific survival.
P values calculated using two-sided, log-rank tests. f, Hierarchical clustering of
primary breast tumours by PI3K signatures in the EMC-MSK dataset. The
PI3Ksighigh cluster is enriched in tumours that developed brain metastasis. The
right panel shows organ-specific metastasis free survival curves stratified by
PI3K signatures. The PI3Ksighigh cluster displayed poorer brain metastasis.
Brain metastasis free survival curves stratified by PI3K signatures in different
subtypes is also presented. P values calculated using two-sided, log-rank tests.
g, Hierarchical clustering of breast cancer metastases by PI3K signatures, with
the PI3Ksighigh cluster being enriched in brain metastases. h, Heat maps
showing significant yet non-complete overlap between 8plow and PI3Ksighigh
clusters in the EMC-MSK dataset. 8plow and PI3Ksighigh clusters co-capture a
subset of patients with the worst brain metastasis prognosis. P values
calculated using two-sided, log-rank tests. The lower panel presents a Cox
proportional-hazards model of brain metastasis free survival using multi
variates – 8p, PI3Ksig, and breast cancer subtype. The 8plow/PI3Ksighigh cluster
is the most associated with brain metastasis. i. 8plow and PI3Ksighigh clusters co-
capture the majority of brain metastasis samples.
Extended Data Fig. 7 | Lipid metabolite profile changes upon SREBF1 SREBF1-KO of JIMT1 (PIK3CAmut) and HCC1806 (8plow) were used. Lipid species
knockout. Heat maps showing relative lipid abundance in cells cultured in groupings and lipid desaturation levels are also presented. WT, wild-type; KO,
medium supplemented with serum or delipidated serum. SREBF1-WT and knockout.
Article
Extended Data Fig. 8 | Analysis of multiplexed breast cancer metastasis
in vivo transcriptomes. a, A schematic of the differential analysis approach
for in vivo transcriptomes with mixtured cancer cell lines. An in silico
transcriptome was modelled based on single cell line in vitro transcriptomes
and cell line composition (comp.) of the metastasis sample. The in silico profile
was then compared with the actual in vivo data in a paired-wise manner.
b, Comparison of in silico modelled profiles to the actual pre-injected or in vivo
metastasis sample profiles. The pre-injected populations are direct mixtures
of in vitro cell lines and show tight correlation with in silico data. In vivo
samples show large fold changes. c, Box plots showing log 2 fold changes of
MUCL1 and SCGB2A2 in in vivo metastasis samples and pre-injected cells. Each
point represents a sample. Box plots display quartiles of the data. Outlier
points extend beyond 1.5 × interquartile ranges from either hinges. d, Heat map
showing log 2 fold change of lung metastasis genes (Minn et al.) in lung, liver,
kidney and bone metastasis samples from the pilot study, where MDAMB231
dominated the population. e, Correlation of gene expression changes in
different metastasis sites. Pre-injected population had no expression change
and thus showed no correlation with in vivo samples. Brain metastases showed
weaker correlations with extracranial metastases. f, Bubble plot showing
enrichment of Hallmark gene pathways (MSigDB) comparing in vivo expression
of metastases at different organ sites to their in vitro counterparts. g, Bubble
plot showing in vivo upregulation of SREBF1, SCD and SREBF1-response
signature in brain metastases. h, i, GSEA analysis of lipid metabolism gene sets
using in vivo RNA-Seq profiles combined by metastasis organ sites irrespective
of sample or cell line composition (h). Gene sets related to lipid metabolism are
selectively enriched on top in the brain but not in other organs or in vitro.
Restricting analysis to JIMT1-dominant samples revealed a similar result. No
enrichment was seen in normal brain when analysis was performed on GTEX
normal tissue (i). Each tick represents a lipid metabolism gene set from
MSigDB. ***, P = 0.001; ** = 0.01.
Extended Data Fig. 9 | See next page for caption.
Article
Extended Data Fig. 9 | Expression of TGFβ signalling, EMT status,
inflammatory response and lipid metabolism genes in clinical breast
cancer metastasis specimens. a, Comparison of brain metastasis versus
extracranial metastasis clinical samples. Lower expression of TGFβ signature
genes and EMT signature genes in brain metastases than other metastasis sites.
Enriched expression of selective SREBF1 target genes (including FASN, SCD,
SREBF1 itself) and Pentose Phosphate Pathway (PPP) genes in brain metastases.
b, c, A strategy to remove brain stroma contamination effect from brain
metastasis expression profiles when performing comparison of paired
primary breast tumour and brain metastasis clinical specimens. A gene
signature indicating brain stroma contamination was derived from
comparison of brain with breast and breast cancer brain metastasis (b).
Arrowheads indicate a few brain metastasis samples with noticeable brain
stroma contamination. A brain contamination score was calculated and its
effect was regressed out in the RNASeq data of matched primary tumours and
brain metastases (c). The heat map shows expression of brain stroma indicator
before and after removal of the contamination effect. d, e, Paired comparison
of primary breast tumour and brain metastasis clinical specimens after
removal of brain stroma contamination. d, Lipid metabolism genes and PPP
genes. e, Signature scores were projected for each sample using the corrected
RNA-Seq profiles. P, Primary breast tumour; M, brain Metastasis; upregulation
in red, downregulation in blue. P values calcutated using paired, two-sided
t-tests.
Extended Data Fig. 10 | See next page for caption.
Article
Extended Data Fig. 10 | In vivo and in vitro effects of SREBF1 knockout.
a, Growth kinetics of SREBF1-WT and -KO cells in in vitro culture medium with
10% serum or 10% delipidated serum. Cell growth was monitored by Incucyte
real-time imaging. WT, wild-type, in black; KO, knockout, in red. Two
independent guides were used per group. b, Fluorescence imaging of
metastases in serial brain sections from mice receiving intracardiac injection
of JIMT1 SREBF1-WT or -KO cells (Fig. 5d). Confocal tile scans of representative
sections are presented at the lower panel. GFP+ signals indicate cancer lesions.
Circles highlight macro-metastatic lesions and arrows indicate micro lesions.
c, d, One-by-one assessment of lipid metabolism gene fitness in additional
brain metastatic cell lines through intracranial injection. SREBF1 was tested for
HCC1954, MDAMB231 (c) and HCC1806. Additional genes were tested for
HCC1806 (d). Cell outgrowth in brain metastasis was monitored by real-time
BLI. Two independent guides per gene were tested, in a one guide one mouse
fashion. e–g, Outgrowing (HCC1806) or residual (JIMT1) SREBF1-KO cells from
brain metastases were derived for CRISPR-seq (e), western blot (f), and RT–
qPCR (g) assays. e, CRISPR-seq quantifying SREBF1 gene editing efficiencies of
brain-derived and pre-injected cells. f, Western blot quantifying SREBF1
protein levels. g, RT–qPCR quantifying relative expression of SREBF1, SCD,
CD36, FABP6 in brain-derived versus pre-injected cells. Pre-injected WT
HCC1806 was used as reference. h, i, Brain-derived and pre-injected HCC1806
cells were cultured in brain-slice-conditioned medium (CM) or medium
supplemented with cerebrospinal fluid, or serum, or delipidated serum, or SM1
supplement, and western blot (h) or RT–qPCR was performed (i). SREBF1, SCD
and CD36 were upregulated when cells were cultured in brain slice CM,
cerebrospinal fluid, and delipidated serum. Brain-derived SREBF1-KO cells
were better at inducing SCD and CD36, in comparison to pre-injected SREBF1-
KO cells. Experiments were performed twice independently with similar
results.
 nature research | reporting summary
Corresponding author(s): Todd R Golub; Xin Jin
Last updated by author(s): Jul 21, 2020

Reporting Summary
Nature Research wishes to improve the reproducibility of the work that we publish. This form provides structure for consistency and transparency
in reporting. For further information on Nature Research policies, see Authors & Referees and the Editorial Policy Checklist.

Statistics
For all statistical analyses, confirm that the following items are present in the figure legend, table legend, main text, or Methods section.
n/a Confirmed
The exact sample size (n) for each experimental group/condition, given as a discrete number and unit of measurement
A statement on whether measurements were taken from distinct samples or whether the same sample was measured repeatedly
The statistical test(s) used AND whether they are one- or two-sided
Only common tests should be described solely by name; describe more complex techniques in the Methods section.

A description of all covariates tested

A description of any assumptions or corrections, such as tests of normality and adjustment for multiple comparisons
A full description of the statistical parameters including central tendency (e.g. means) or other basic estimates (e.g. regression coefficient)
AND variation (e.g. standard deviation) or associated estimates of uncertainty (e.g. confidence intervals)

For null hypothesis testing, the test statistic (e.g. F, t, r) with confidence intervals, effect sizes, degrees of freedom and P value noted
Give P values as exact values whenever suitable.

For Bayesian analysis, information on the choice of priors and Markov chain Monte Carlo settings
For hierarchical and complex designs, identification of the appropriate level for tests and full reporting of outcomes
Estimates of effect sizes (e.g. Cohen's d, Pearson's r), indicating how they were calculated
Our web collection on statistics for biologists contains articles on many of the points above.

Software and code

Policy information about availability of computer code
Data collection Bioluminescence imaging data was acquired with Living Image software (v4.5, PerkinElmer). Lipidomics mass spectrometry data was
acquired using a LC-MS system composed of a Shimadzu Nexera X2 U-HPLC (Shimadzu Corp) coupled to a Exactive Plus orbitrap mass
spectrometer (ThermoFisher Scientific).

Data analysis Following softwares were used for data analysis: Living Image software (v4.5), Bowtie 2 (v2.2.8), samtools (v 1.3.1), BBSplit (https://
sourceforge.net/projects/bbmap/), RSEM (v1.3.1), R statistical software (v3.6.2), ggplot2 (3.3.0), limma (3.42.2), edgeR (3.28.0), gsva
(1.34.0), gplots (3.0.1.2), survival (3.1-8), fgsea (1.12.0), GSEA (v3.0), GenePattern (v2.0).
For manuscripts utilizing custom algorithms or software that are central to the research but not yet described in published literature, software must be made available to editors/reviewers.
We strongly encourage code deposition in a community repository (e.g. GitHub). See the Nature Research guidelines for submitting code & software for further information.

Data
Policy information about availability of data
All manuscripts must include a data availability statement. This statement should provide the following information, where applicable:
- Accession codes, unique identifiers, or web links for publicly available datasets
- A list of figures that have associated raw data
October 2018

- A description of any restrictions on data availability

MetMap data and interactive visualization can be accessed at pubs.broadinstitute.org/metmap. RNA-seq data generated from this study have been deposited to
Gene Expression Omnibus (GEO), at accession numbers GSE148283 and GSE148372. Additional datasets used in this study include METABRIC, TCGA, and MSK-
targeted-sequencing breast cancer datasets downloadable from cBioPortal, EMC-MSK dataset (GSE2035, GSE2603, GSE5327, GSE12276), 65 metastasis sample
dataset (GSE14020), paired primary tumor and brain metastasis RNA-Seq from Vareslija et al, and GSE52604.

1
 nature research | reporting summary
Field-specific reporting
Please select the one below that is the best fit for your research. If you are not sure, read the appropriate sections before making your selection.
Life sciences Behavioural & social sciences Ecological, evolutionary & environmental sciences
For a reference copy of the document with all sections, see nature.com/documents/nr-reporting-summary-flat.pdf

Life sciences study design

All studies must disclose on these points even when the disclosure is negative.
Sample size Sample sizes were determined to be adequate for minimal n required for statistical tests, or consistency of measurable differences between
groups following guidance and experience from Ref. 5, 7, 18.

Data exclusions Failed RNA-Seq samples were excluded from analysis presented in the manuscript. In MetMap500 experiment (Fig. 2), one animal died early
and organs could not be collected in time, and is excluded from analysis.

Replication Cell culture based experiments (including growth assay, RT-qPCR, western blot) were performed twice independently. Animal experiments
were validated using completely independent methods instead of direct repeat (Pooled experiment vs individual injection in Fig. 1a, Extended
Data Fig. 2g; MetMap500 vs MetMap125 in Fig. 2c; mini-pool CRISPR screen vs one-by-one testing in Fig. 5a-c).

Randomization Randomization was not applicable to experiments in this study. In MetMap profiling, we varied pooling format, cell density, cohort size,
animal age to account for these potential covariates.

Blinding Blinding to group allocations was not applicable to experiments in this study.

Reporting for specific materials, systems and methods

We require information from authors about some types of materials, experimental systems and methods used in many studies. Here, indicate whether each material,
system or method listed is relevant to your study. If you are not sure if a list item applies to your research, read the appropriate section before selecting a response.

Materials & experimental systems Methods

n/a Involved in the study n/a Involved in the study
Antibodies ChIP-seq
Eukaryotic cell lines Flow cytometry
Palaeontology MRI-based neuroimaging
Animals and other organisms
Human research participants
Clinical data

Antibodies
Antibodies used SREBF1 primary antibody (14088-1-AP, Proteintech)
SCD (CD.E10) antibody (ab19862, Abcam)
GAPDH (D16H11) XP® Rabbit mAb (5174S, Cell Signaling)
β-Actin (8H10D10) Mouse mAb (3700S, Cell Signaling)
IRDye® 800CW Goat anti-Mouse IgG (926-32210, LI-COR)
IRDye® 680RD Goat anti-Rabbit IgG secondary antibodies (926-68071, LI-COR).

Validation SREBF1 primary antibody (14088-1-AP, Proteintech): validated by manufacturer, and by this study (Extended Data Fig. 11f,h), and
cited in publications, suitable for western blot
SCD (CD.E10) antibody (ab19862, Abcam): validated by manufacturer, and by this study (Extended Data Fig. 11f,h), suitable for
western blot
GAPDH (D16H11) XP® Rabbit mAb (5174S, Cell Signaling): validated by manufacturer and cited in publications, suitable for
October 2018

western blot
β-Actin (8H10D10) Mouse mAb (3700S, Cell Signaling): validated by manufacturer and cited in publications, suitable for western
blot

2
Eukaryotic cell lines

nature research | reporting summary

Policy information about cell lines
Cell line source(s) All cell lines listed in Supplementary Table 1 and 3 were obtained from CCLE.

Authentication Cell lines were authenticated by DNA fingerprinting analysis. The breast cell line identities were also confirmed by RNA-Seq
and compared to CCLE RNA-Seq profiles.

Mycoplasma contamination All cell lines were confirmed to be mycoplasma free using the MycoAlertTM Mycoplasma Detection Kit (Lonza).

Commonly misidentified lines PC-14 was identical to PC-9 as reported before (https://web.expasy.org/cellosaurus/CVCL_1640; https://
(See ICLAC register) www.sigmaaldrich.com/catalog/product/sigma/cb_90071810?lang=en&region=US). To keep consistent with CCLE
nomenclature, PC14_LUNG was used. KPL-1 was found to be a MCF-7 derivative (https://web.expasy.org/cellosaurus/
CVCL_2094). To keep separate from MCF-7 and consistent with CCLE nomenclature, KPL1_BREAST was used.

Animals and other organisms

Policy information about studies involving animals; ARRIVE guidelines recommended for reporting animal research
Laboratory animals NOD scid gamma (NSG) female mice (The Jackson Laboratory) of 5~6 or 8~10 weeks were used for metastasis xenograft studies.
Broad Vivarium’s housing conditions for NSG mice include sterilized, individually ventilated cages with cellulose bedding. Water
bottles are supplied with acidified, reverse osmosis water. The holding room is maintained under positive pressure, temperature
70°F (+/-2°F), humidity 40% (+/- 10%), lighting 12 on/12 off light cycle.

Wild animals No wild animals were used in the study.

Field-collected samples No field collected samples were used in the study.

Ethics oversight Animal work was performed in accordance with a protocol approved by the Broad Institute Institutional Animal Care and Use
Committee (IACUC).
Note that full information on the approval of the study protocol must also be provided in the manuscript.

Flow Cytometry
Plots
Confirm that:
The axis labels state the marker and fluorochrome used (e.g. CD4-FITC).
The axis scales are clearly visible. Include numbers along axes only for bottom left plot of group (a 'group' is an analysis of identical markers).
All plots are contour plots with outliers or pseudocolor plots.
A numerical value for number of cells or percentage (with statistics) is provided.

Methodology
Sample preparation Organs were dissociated using dissociation protocols listed in Supplementary Table 9 with gentleMACS Octo Dissociator (Miltenyi
Biotec). Dissociated cell suspensions were filtered using 100 μm filters, and washed with DMEM/F12 twice. Cell suspensions
were then washed with staining buffer (PBS + 2mM EDTA + 0.5% BSA), and incubated with mouse cell depletion beads according
to the instructions (Miltenyi Biotec). Cell suspensions were subjected to negative selection using autoMACS Pro Separator
(Miltenyi Biotec) to deplete mouse stroma. Brains were subjected to an additional myelin debri depletion step using myelin
removal beads II (Miltenyi Biotec). In vitro cultured cells were trypsinized and resuspended as single cell suspensions. DAPI
staining was used to exclude dead cells.

Instrument SONY SH4800

Software SH4800S and FlowJo (v10.2)

Cell population abundance Data is presented in Extended Data Fig 1c and Source Data.
October 2018

Gating strategy Gating strategy is illustrated in Extended Data Fig. 1e to select for single cells with the fixed gate for GFP or mCherry.

Tick this box to confirm that a figure exemplifying the gating strategy is provided in the Supplementary Information.

3
Article

A map of cis-regulatory elements and 3D

genome structures in zebrafish

https://doi.org/10.1038/s41586-020-2962-9 Hongbo Yang1,11, Yu Luan1,11, Tingting Liu1,11, Hyung Joo Lee2, Li Fang3, Yanli Wang4,
Xiaotao Wang1, Bo Zhang4, Qiushi Jin1, Khai Chung Ang5, Xiaoyun Xing2, Juan Wang1, Jie Xu1,
Received: 20 July 2019
Fan Song4, Iyyanki Sriranga1, Chachrit Khunsriraksakul4, Tarik Salameh4, Daofeng Li2,
Accepted: 17 September 2020 Mayank N. K. Choudhary2, Jacek Topczewski6,7, Kai Wang3, Glenn S. Gerhard8,
Ross C. Hardison9, Ting Wang2, Keith C. Cheng5 & Feng Yue1,10 ✉
Published online: 25 November 2020

Check for updates

The zebrafish (Danio rerio) has been widely used in the study of human disease and
development, and about 70% of the protein-coding genes are conserved between the
two species1. However, studies in zebrafish remain constrained by the sparse
annotation of functional control elements in the zebrafish genome. Here we
performed RNA sequencing, assay for transposase-accessible chromatin using
sequencing (ATAC-seq), chromatin immunoprecipitation with sequencing,
whole-genome bisulfite sequencing, and chromosome conformation capture (Hi-C)
experiments in up to eleven adult and two embryonic tissues to generate a
comprehensive map of transcriptomes, cis-regulatory elements, heterochromatin,
methylomes and 3D genome organization in the zebrafish Tübingen reference strain.
A comparison of zebrafish, human and mouse regulatory elements enabled the
identification of both evolutionarily conserved and species-specific regulatory
sequences and networks. We observed enrichment of evolutionary breakpoints at
topologically associating domain boundaries, which were correlated with strong
histone H3 lysine 4 trimethylation (H3K4me3) and CCCTC-binding factor (CTCF)
signals. We performed single-cell ATAC-seq in zebrafish brain, which delineated 25
different clusters of cell types. By combining long-read DNA sequencing and Hi-C, we
assembled the sex-determining chromosome 4 de novo. Overall, our work provides an
additional epigenomic anchor for the functional annotation of vertebrate genomes
and the study of evolutionarily conserved elements of 3D genome organization.

The zebrafish has been an important vertebrate model system for
several decades because of its high fecundity, external embryogen-
esis, rapid embryonic development and nearly transparent embryos.
These features have made it an ideal system for the study of verte-
brate development and ageing2, comparative genomics3 and human
disease modelling. However, there is no comprehensive annotation
of the cis-regulatory elements in the zebrafish genome. Although
previous genomic studies in zebrafish have provided critical biologi-
cal insights4–8, most used whole embryos and our understanding of
tissue-specific regulators remains limited.
To profile the transcribed regions, chromatin accessibility
and DNA methylation patterns in the zebrafish genome, we per-
formed strand-specific RNA sequencing (RNA-seq), ATAC-seq9 and
whole-genome bisulfite sequencing (WGBS) in up to eleven zebrafish
adult tissues and two embryonic tissues (Fig. 1a, b, Supplementary
Table 1). Because histone modifications have been used to predict
different classes of potential regulatory elements such as enhancers
and repressors10,11, we also performed chromatin immunoprecipitation
followed by DNA sequencing (ChIP-seq) for a panel of histone modifica-
tions, including H3K4me3, H3 lysine 27 acetylation (H3K27ac), H3 lysine
9 dimethylation (H3K9me2) and H3 lysine 9 trimethylation (H3K9me3).
To study higher-order chromatin structure and link distal enhancers to
their target genes, we performed Hi-C experiments in adult brain and
muscle (Fig. 1a, b). Although chromosome 4 is regarded as the ‘rudimen-
tary’ sex chromosome in zebrafish12, the quality of its current assembly
is poor owing to the heavy presence of heterochromatin. Therefore,
we performed three long-read DNA sequencing experiments (nano-
pore, 10X Genomics and Bionano optical mapping) in one Tübingen
female zebrafish to generate a de novo assembly of chromosome 4. To
investigate the cell types and their regulatory elements in the zebrafish

1
Department of Biochemistry and Molecular Genetics, Feinberg School of Medicine Northwestern University, Chicago, IL, USA. 2Department of Genetics, The Edison Family Center for Genome
Sciences and Systems Biology, Washington University School of Medicine, St Louis, MO, USA. 3Raymond G. Perelman Center for Cellular and Molecular Therapeutics, Children’s Hospital of
Philadelphia, Philadelphia, PA, USA. 4Bioinformatics and Genomics Program, The Pennsylvania State University, State College, PA, USA. 5Department of Pathology and Penn State Zebrafish
Functional Genomics Core, College of Medicine, The Pennsylvania State University, Hershey, PA, USA. 6Department of Pediatrics, Northwestern University Feinberg School of Medicine,
Chicago, IL, USA. 7Stanley Manne Children’s Research Institute, Ann and Robert H. Lurie Children’s Hospital of Chicago, Chicago, IL, USA. 8Department of Medical Genetics and Molecular
Biochemistry, Lewis Katz School of Medicine at Temple University, Philadelphia, PA, USA. 9Department of Biochemistry and Molecular Biology, Pennsylvania State University, University
Park, PA, USA. 10Robert H. Lurie Comprehensive Cancer Center of Northwestern University, Chicago, IL, USA. 11These authors contributed equally: Hongbo Yang, Yu Luan, Tingting Liu.
✉e-mail: Yue@northwestern.edu

Nature | Vol 588 | 10 December 2020 | 337

Article
a Epiderm Mesoderm Endoderm Embryonic b
chr4:22,000,000-28,520,000
Brain
RNA-seq
ATAC-seq

Hi-C
H3K27ac
H3K4me3
H3K9me3
H3K9me2 chr4:22.80 Mb-23.32 Mb
cd36 magi2
WGBS
Hi-C

H3K9me2 H3K9me3 ATAC-seq H3K4me3 H3K27ac

Brain

n
e

nk
ey

e
in

on
is

r
n

en
rt
Muscle

ai
cl

ve

in
oo
ai

ea
st

Sk

dn

tru
br
us

le

ol
st
Br
20

Li
Te

Bl
H

E-
Sp

C
Ki

E-
M

te
c d Liver

In
3 0
Zebrafish 2 Brain-specific genes
log(TPM+1) myh6
1 20
0
0 0

5 3 20
Human Human orthologues
log(TPM+1)

MYH6 2 0
1
0 0 5
0
Te ain

e
Sk d
Ki lon
C e

C ne
Te ain

e
Sp ney

Sk d
H in
us t
Bl en
In Liv s

Ki lon
st r

Sp ney

H in
us t
Bl en
In Liv s
te er
M ear
te e

cl

M ar
in

cl
oo
i

oo
i
st

st

i
le
Br

le
st
o

Br

e
o
d

d
e f 5
A predicted novel transcript RNA-seq H3K4me3
Novel transcripts 0

Intensity

WGBS
E -brain H3K4me3
n = 8,311

H3K27ac 1
RNA seq 0
10

RNA-seq
E-trunk
0
20 10
0 0
chr1:660,000-671,000 –2 kb TSS +3 kb –2 kb TSS +3 kb

Fig. 1 | Identification of cis-regulatory elements in the zebrafish genome.
a, Tissues and analyses performed in this study. H3K27ac, H3K4me3, H3K9me3
and H3K9me2 represent ChIP-seq analyses with the indicated antibody.
b, Snapshot of an example region, showing Hi-C, ChIP-seq, ATAC-seq, WGBS
and RNA-seq data in adult zebrafish brain, muscle and liver, using WashU
Epigenome Browser. Plots show relative amount, normalized to the range of
values. The values on the y-axis for ChIP-seq analyses were normalized to the
input. Data range for the Hi-C heat map is 0–40 raw read counts. c, Expression
of myh6 in zebrafish and the paralogue MYH6 in human are heart-specific
(n = 2). The values for human expression were from GTEx. Data are
mean ± s.e.m. TPM, transcripts per million base pairs. d, Box plot of the
expression of brain-specific genes in zebrafish (top) (n = 2,481) and the
expression of their orthologues in human (bottom) (n = 2,481). The y-axis
shows the gene expression value: log10(TPM+1). e, An example of a predicted
novel transcript. Vertical scale: 0–20 (H3K27ac and H3K4me3), 0–10 (RNA-seq).
f, RNA-seq and H3K4me3 ChIP-seq signals for the predicted 8,311 novel
transcripts across all the tissues. In all box plots, horizontal line shows the
median, the box encompasses the interquartile range, and whiskers extend to
5th and 95th percentiles.

brain, we performed single-cell ATAC-seq (scATAC-seq). In total, we
generated 161 genomic datasets that comprised over 10 billion reads.
To our knowledge, this is the most comprehensive analysis of candidate
cis-regulatory elements in zebrafish to date and represents a major
resource for comparative genomics and the study of gene regulation
in this vertebrate model organism.
Transcriptome analysis
We detected 39,188 transcripts across all tissues using RNA-seq, 14,764
of which exhibited tissue-specific patterns (Extended Data Fig. 1a–c).
We identified 13,285 previously unknown transcripts, 8,311 of which
were also supported by H3K4me3 peaks at the promoter regions
(Fig. 1e, f, Extended Data Fig. 1d, Supplementary Table 2). These 8,311
transcripts include 976 long noncoding RNAs (lncRNAs), 3,596 previ-
ously unknown isoforms and 3,739 potential previously uncharacter-
ized protein-coding genes.
Next, we examined whether the expression patterns for the ortho-
logues of tissue-specific genes were conserved between zebrafish and
human. Among the 14,764 tissue-specific zebrafish transcripts, 3,737
have a one-to-one human orthologue, 1,747 of which (47%) also show
tissue-specific patterns in human (Fig. 1c, d, Supplementary Table 3),
suggesting that these genes might have a critical and conserved role
in the tissues in which they are uniquely expressed.
in each tissue and merged them into a list of 436,036 non-redundant
peaks across all tissues (Fig. 2a, Supplementary Table 5). Of these
peaks, 116,353 were previously unidentified in previous work in whole
embryos15–20 (Extended Data Fig. 2a and Supplementary Table 6). As
expected, distal ATAC-seq peaks show higher tissue specificity than
proximal peaks (Fig. 2b, c).
We then defined the cis-regulatory elements with the following com-
binations of histone modifications and ATAC-seq peaks: active pro-
moter (H3K27ac, H3K4me3 and ATAC-seq), weak promoter (H3K4me3
and ATAC-seq), active enhancer (distal H3K27ac and ATAC-seq) and
heterochromatin (H3K9me2 or H3K9me3 sites). Across all the tissues,
we predicted 25,593 active promoters, 40,220 weak promoters, 58,065
active enhancers and 112,445 heterochromatin sites (Extended Data
Fig. 2b, c and Supplementary Tables 7–10). A total of 40.9% of the pre-
dicted promoters and 62.5% of the predicted enhancers reported in
this study were not identified in previous reports6,21–25 (Extended Data
Fig. 2a). Of the enhancers, 71.3% were tissue-specific and also showed
tissue-specific ATAC-seq signals (Fig. 2d, e). Gene ontology (GO) analy-
sis showed that they were located near genes important for relevant
tissue-specific functions (Extended Data Fig. 2d).
To validate the predicted enhancers and their tissue specificities,
we used a GFP-based zebrafish embryo reporter assay. Of the 32
tissue-specific enhancers tested, 87.5% (28 out of 32) showed restricted
GFP expression (Fig. 2f, Extended Data Fig. 3, Supplementary Table 11).

Accessibility, promoter and enhancer dynamics Single-cell ATAC-seq in zebrafish brain

Chromatin accessibility is associated with a wide range of regulatory We performed single-cell ATAC-seq (scATAC-seq) in adult zebrafish
elements13,14, thus we performed ATAC-seq in all 11 adult tissues. We brain, generating 654 million usable reads from 19,955 cells in the
identified 66,771–180,788 ATAC-seq peaks (Supplementary Table 4) adult brain. We identified 268,268 non-redundant peaks, covering

338 | Nature | Vol 588 | 10 December 2020

 a b c No. of tissues
3′UTR Proximal ATAC-seq peaks Distal ATAC-seq peaks ≥5
5 4 2

ATAC-seq peaks (×105)

ATAC-seq peaks (×105)

ATAC-seq peaks (×104)

5′UTR 4

Number of proximal
Distal

Number of distal
Exon 3
2

Number of
Intron
Promoter 1

0 0 0

Br les

on

on

on
H le

H e

H le

L d
Li d
C ne

C ne

C ne
Sp ney

Sp ey

Sp ey
S s

d t

is

is
Ki art

Ki art
Bl n

Bl n
Te ain

Bl en

Te in

Te in
M kin

M in
M kin
st r

te r

te r
In Liv d
Ki ear

In ve
te e

In ive
cl
i

oo
oo
e

e
c

c
oo
st

st
st
a

Sk
ol

ol

ol
dn

dn
i

i
p

le

le
le

e
us

us

us
st

st
Br

Br
S
am
ls
d Al Clustering of enhancers e ATAC-seq signal f
Brain enhancer 5 5 dpf Heart enhancer 5 3 dpf
Intensity

Intensity
3 2

1 elavl3 gata6
0 chr3:48,934,231-48,935,968 chr2:4,311,945-4,313,944
–2 0

Muscle enhancer 7 5 dpf Kidney enhancer 1 5 dpf

musk zgc153722
chr10:13,126,525-13,128,100 chr23:16,855,310-16,856,895

n n s n d r n e t e n y
ai ai sti ee oo ive lo tin ar cl ki ne
Li d
Te ain

S le
te on
H ne
S p stis

us t
In ol r
B l en

ey
K i kin
M ear
Br -br Te pl Bl L Co tes He us SKid
C e
oo

c
v

dn
le
Br

st

E S In M
g h i j MA0077.1_SOX9
20 0.5
PCC = 0.878 Bulk vs scATAC-seq
40 18 16 8

Motif enrichment
0
Bulk ATAC-seq

24 9
20 1 15
25 21
6
t-SNE2

2,625 195,004 73,264 0 14 23

3 5 MA0678.1_OLIG2
22 17 19
−20 11 20 0.5
12 4 7
−40 2 13
0
Bulk Single-cell 10
8
8 scATAC-seq 20 –40 –20 0 20 40 60 0 25
t-SNE1 Clusters

Fig. 2 | Characterization of tissue-specific cis-regulatory elements. the 63 surviving embryos showed similar patterns. For kidney enhancer 1, 47
a, Number of ATAC-seq peaks predicted in each tissue and their genomic out of the 82 surviving embryos showed similar patterns. Scale bar, 200 μm.
distribution. b, c, Tissue specificity of proximal and distal ATAC-seq peaks in dpf, days post-fertilization. g, Pearson correlation coefficient between
11 adult tissues. d, Clustering analysis identified tissue-specific enhancers. aggregated signals of scATAC-seq and bulk ATAC-seq data. Values are the sums
Values in the heat map were input-normalized H3K27ac intensity (n = 58,226 of the reads in continuous 10-kb bins, normalized by sequencing depth. h, The
enhancers). e, Normalized ATAC-seq intensity in the corresponding enhancer overlap between peaks predicted in bulk and scATAC-seq data. i, t-distributed
elements shown in d. f, Examples of validated tissue-specific enhancers by GFP stochastic neighbour embedding (t-SNE) analysis identified 25 clusters in the
reporter assay in zebrafish embryos. For brain enhancer 5, 112 out of 143 scATAC-seq data in zebrafish adult brain (n = 19,955). j, Examples of enriched
surviving embryos showed similar patterns. For heart enhancer 7, 61 out of the motifs in different clusters from scATAC-seq peaks (n = 19,955).
67 surviving embryos showed similar patterns. For muscle enhancer 5, 53 out of

98.7% of the bulk ATAC-seq peaks in the brain (Fig. 2g, h, Extended Short interspersed nuclear elements (SINEs) were enriched in both
Data Fig. 4a–c). Among them, 73,264 peaks were detected only by H3K9me2 and H3K9m3 sites, whereas long terminal repeats (LTRs) were
scATAC-seq, suggesting that there are potentially more regulatory enriched only in H3K9me2 sites (Fig. 3b). Although both H3K9me2 and
elements in the zebrafish genome than we predicted on the basis of H3K9me3 sites were depleted of active marks within the same tissue,
the bulk tissue results. 20% of these sites overlapped with ATAC-seq peaks or other active
Using the scATAC-seq data, we identified 25 clusters of cells in the marks in other tissues (Fig. 3a, Extended Data Fig. 5e, g), suggesting
zebrafish brain (Fig. 2i). By identifying the key cell-type-specific tran- that heterochromatin regions in one tissue may be active regulatory
scription factor motifs, we inferred the potential cell type of each cluster, elements in other tissues.
such as oligodendrocyte progenitor cells and prefrontal cortex cells To study DNA methylation patterns in zebrafish, we performed
(Extended Data Fig. 4d, e). We quantitatively determined the enrich- WGBS in 11 adult tissues with approximately 30× coverage in each
ment of transcription factor motifs in each of the 25 clusters (Fig. 2j, dataset. Genome-wide CpG methylation levels were approximately
Extended Data Fig. 4f). Many neuronal transcription factors (such 80% across different tissues, with the exception of the testis, which
as SOX9 and OLIG2) were enriched in different clusters, suggesting exhibited higher CpG methylation levels (Extended Data Fig. 6a, b). We
potential roles in cell-type-specific regulation in the zebrafish brain. also detected increased levels of methylation at the CAC trinucleotide
in brain compared with other tissues (Extended Data Fig. 6c), similar
to reports in human and mouse26. Unmethylated CpGs were found
Heterochromatin and DNA methylation mostly in CpG islands, gene promoters and 5′ untranslated regions
We performed ChIP-seq for H3K9me2 and H3K9me3 in 11 adult (Extended Data Fig. 6d, e), whereas CpGs in gene bodies and differ-
zebrafish tissues (Extended Data Fig. 5a). Across all tissues, we identi- ent classes of repetitive elements were heavily methylated (Extended
fied 73,777 non-redundant H3K9me2 sites and 68,798 non-redundant Data Fig. 6e). We also identified unmethylated regions (UMRs) and
H3K9me3 sites. While both H3K9me2 and H3K9me3 are heterochro- low-level-methylated regions (LMRs) (Supplementary Table 12). Most
matic marks, they were located in different parts of the genome, with UMRs overlapped with candidate promoters and proximal ATAC-seq
overlap of about 10% in the same tissue (Extended Data Figs. 5b–d). peaks, whereas LMRs overlapped more with candidate enhancers and

Nature | Vol 588 | 10 December 2020 | 339

Article
a b We therefore performed nanopore, 10X Genomics and BioNano opti-
Enhancer H3K9me3 H3K9me2
Brain Brain cal mapping in an individual female zebrafish of the Tübingen strain.

Log2(FC)
Blood Blood
Colon Colon By combining these long-read DNA-sequencing results with the Hi-C
Heart Heart
Intestine Intestine 2 data from brain, we de novo assembled a new version of chromosome
Kidney Kidney 1
Liver Liver
0
4 (Methods, Extended Data Fig. 7c, Supplementary Dataset 1). With the
Muscle Muscle
Skin
Spleen
Skin
Spleen
–1 newly assembled genome, we reprocessed the Hi-C data and observed
Testis Testis that most of the aberrant signals were no longer visible on the Hi-C map

Unknown

Unknown
0 Percentage100 0 Percentage 100

Satellite
Unknown

RC

RC
DNA

DNA
SINE

Satellite

SINE
LINE

LINE
LTR

LTR
Satellite
RC
DNA
LINE
LTR

SINE
(Fig. 3e, f). We reprocessed the BioNano optical-mapping data and also
Heterochromatin Active marks in Active marks in
same tissue other tissues observed fewer structural-variation events (Extended Data Fig. 7d, e).
c d
DNA
Examples of tissue-specific hypoDMRs
This newly assembled chromosome 4 will serve as a resource to study
methylation H3K27ac ATAC-seq
sex determination and other processes in zebrafish that involve genes
Brain dazl elavl4
H3K4me3 Testis on this chromosome.
H3K27ac
ATAC-seq
WGBS Conservation of cis-regulatory elements
RNA
Functional elements are often conserved during evolution27. We
–5 kb +5 kb
H e

chr19:20,815,000-20,820,000 chr8:15,964,500-15,984,000
first examined the sequence conservation of different classes of
rt

r
cl

ve
ea

Intensity
us

Li

0 1 0 5 0 5
M

e f cis-regulatory elements. Promoters had the highest degree of sequence

10 Mb
conservation, while enhancers had a much lower but still significant
GRCz11 chr4

level of sequence conservation (Fig. 4a, Extended Data Fig. 8b). Of

= 17 = 16 = 17
GRCz10 GRCz11 De novo assembly
chr4: 0-76.6 Mb chr4: 0-78.1 Mb chr4: 0-74.5 Mb
Fig. 3 | Analysis of heterochromatin and repetitive elements and de novo
assembly of zebrafish chromosome 4. a, Comparison of H3K9me2 and
H3K9me3 sites with active marks (ATAC-seq, H3K4me3 or H3K27ac peaks) from
the same tissue (left) and active marks in other tissues (right). The number of
heterochromatin regions in each tissue: testis, 36,672; spleen, 20,813; skin,
24,687; muscle, 25,692; liver, 29,117; kidney, 21,821; intestine, 24,072; heart,
14,706; colon, 22,426; blood, 19,082; and brain, 21,596. b, Repetitive-elements
enrichment analysis for predicted enhancers, H3K9me3 and H3K9me2 sites.
Colour and size indicate fold enrichment. c, DNA methylation levels, H3K27ac
ChIP-seq and ATAC-seq signals for tissue-specific hypoDMRs. Tissue-specific
hypoDMRs cluster size: muscle, 1,912; heart, 1,708; and liver, 3,386. d, Examples
of tissue-specific hypoDMRs. Vertical scale: 0–300 for H3K4me3, 0–100 for
H3K27ac and ATAC-seq, 0–1 for WGBS, 0–40 for RNA-seq. ChIP-seq data are
normalized by input. ATAC-seq and RNA-seq are normalized by total
sequencing depths. For WGBS, scale is the methylation level. e, Brain Hi-C data
mapped to GRCz10, GRCz11 and the de novo assembled chromosome 4.
Aberrant Hi-C signals were observed when Hi-C reads were mapped to the
GRCz10 or GRCz11 reference genome but were not visible when mapped to the
de novo assembled chromosome 4. f, Alignment of the de novo assembled
chromosome 4 to the GRCz11 reference genome (alignment of 2-kb bins using
LASTZ).
De novo assembled chr4
the predicted enhancers with sequences that were not conserved in
humans or mouse, 88.6% were conserved in other fish species (Fig. 4c,
Extended Data Fig. 8c). Furthermore, 60–90% of zebrafish enhancers
with sequences that were conserved in human were also predicted as
candidate enhancer elements by NIH ENCODE and Roadmap Epigenet-
ics Project (Fig. 4b, Extended Data Fig. 8a). Notably, even for zebrafish
enhancers with no detectable sequence conservation in humans, we
found that they were conserved in other fish species with increased fish
PhyloP scores (Fig. 4c), suggesting that they might be used as enhanc-
ers in other fish.
Previous efforts have identified several thousand ultra-conserved
noncoding elements (UCNEs) in vertebrates28. There are 2,405, 4,337
and 4,351 UCNEs in zebrafish, mouse and human, respectively. We
found that 69% of the zebrafish UCNEs overlapped with the predicted
cis-regulatory elements: 15% as promoters, 53% as enhancers and 1% as
noncoding RNAs. One such example is shown in Fig. 4d: the zebrafish
UCNE SALL3_Anna is conserved in both human and mouse, and this
element is predicted as an enhancer sequence in all three species on the
basis of the enrichment of H3K27ac signals (another example is shown
in Extended Data Fig. 8d). Furthermore, this enhancer has also been
validated by transgenic reporter assays in mouse embryos29. Overall,
enhancers localized in the ultra-conserved regions were more likely
to be predicted as enhancers in mouse and human (30% versus 5.68%).

Linking distal elements to target genes

distal ATAC-seq peaks (Extended Data Fig. 6f). We identified differen-
tially methylated regions (DMRs) and tissue-specific hypomethylated
DMRs (hypoDMRs) (Extended Data Fig. 6g, Supplementary Table 13).
Tissue-specific hypoDMRs were enriched with tissue-specific H3K27ac
and ATAC-seq signals (Fig. 3c, d), suggesting that they can potentially
identify tissue-specific cis-regulatory elements.
De novo assembly of chromosome 4
Chromosome 4 has been regarded as the rudimentary sex chromosome
in zebrafish12. However, extensive heterochromatin and transposable
elements have made the analysis of this chromosome challenging.
Indeed, we observed strong enrichment of H3K9me3, H3K9me2 and
DNA methylation on the long arm of chromosome 4 (Extended Data
Fig. 7a, b). When we examined the Hi-C data (Fig. 3e) mapped to the
GRCz10 and GRCz11 genome, there were many aberrant off-diagonal
signals, indicating the presence of structural variations between the
reference genomes and the genome of the fish used in this study, owing
to either assembly error or inter-individual variations.
To link distal ATAC-seq peaks to their target genes, we adopted a
correlation-based strategy30. We generated 340,527 ATAC-seq peak-
to-gene links with false discovery rate (FDR) below 0.01, which
contain 144,886 distal ATAC-seq peaks and 18,792 genes (Extended
Data Fig. 9a). Using a similar strategy, we predicted 96,540 enhancer-
to-gene links, 37,241 of which were also supported by ATAC-seq
peak-to-gene links (Fig. 4f, Extended Data Fig. 9b, c, Supplemen-
tary Table 14). The enhancer-to-gene links contain 33,728 putative
enhancers and 16,935 genes. We observed higher Hi-C reads for
the predicted enhancer-to-gene pairs than expected values at each
genomic distance (Fig. 4g, Extended Data Fig. 9d), supporting the
linkages between genes and distal elements on the basis of activity
correlation.
Tissue-specific transcriptional regulatory network
To identify putative key transcription factors in each tissue, we
performed motif analysis in each group of tissue-specific enhanc-
ers and identified a set of motifs that were enriched in the same

340 | Nature | Vol 588 | 10 December 2020

a b Enhancer in other tissues c Enhancers not conserved
Sequence conservation Enhancer in same tissue in human Higher-order genome structure in zebrafish
in human Not an enhancer in human Random sequence
100
100
0.10
To study higher-order chromatin structure in zebrafish, we gener-
Percentage (%)

Percentage (%)

Fish phyloP
ated high-resolution Hi-C data (10 kb) in the adult brain and muscle
(Extended Data Fig. 10a), with about 2.1 and 1.4 billion paired-end
reads, respectively. Replicates of the Hi-C experiments were highly
0 0
r r
on ote ce om
n t e y
ai od lon ar in ne
e
cl kin en
0.05
–10 kb Enhancer +10 kb reproducible32. We predicted the A/B compartments and found that

r
Br Blo Co He testKid us S ple

ve
Ex om han and

Li
Pr En R In M S their genomic coverages in these two tissues were similar. H3K27ac,
d e Human
H3K4me3 and ATAC-seq signals were enriched in the A compartment,
UCNE: SALL3_Anna Zebrafish
20
RFX
OLIG2
whereas H3K9me2 and H3K9me3 signals were enriched in the B com-
Zebrafish

H3K27ac
NEUROD partment (Extended Data Fig. 10b). We identified 5,348 regions with
Brain

20 H3K4me3 ATOH1
chr19:22,457,896-22,461,885 TLX
ETV
switched A/B compartments between the two tissues, and these regions
PU.1
were associated with altered gene expression and H3K27ac signals
Mouse

EHF
Brain

GATA2
chr18:81,347,542-81,351,549 GATA6 (Extended Data Fig. 10c). We predicted 1,350 topologically associating
ETS1
FOXA2 domains (TADs) in the brain and 1,238 TADs in the muscle (Extended
Human
Brain

BMAL1
chr18:76,480,477-76,485,273
HNF4A Data Fig. 10d, Supplementary Table 15). Most of the TADs were shared
HNF1
CDX2 between the two tissues (Extended Data Fig. 10b, e) and TAD boundaries
phyloP

4
ERRA
-4 GATA4
GATA3
were enriched for CTCF-binding sites, SINEs and satellite elements

P-value
NUR77
MEF2D (Extended Data Fig. 10f–h).
MYOD
SIX1 20 We identified 7,708 and 5,312 chromatin loops in the adult brain and
P53
ETV1 0 muscle, respectively (Fig. 5a, Supplementary Table 16). The major-
Blood

Colon

Muscle

Blood

Muscle

ity of the loop anchors had convergent CTCF-binding motifs (72% in

Colon
Intestine

Kidney

Intestine

Kidney
Heart

Heart
Liver
Brain

Brain
Skin
Spleen

Spleen

Liver

Skin
Mouse reporter
assay by VISTA
muscle and 63% in brain). Of the predicted loops in the brain, 98.6%
f g
5 12
Enhancer-to-gene pairs
were between regions that contain either at least one promoter or one
Mean = 4.85 Mean = 2.67
95% Cl 9
To 1 gene = 34%
100
Expected enhancer, and 91.6% of enhancer–promoter loops overlapped with
Frequency (×103)

predicted enhancer–promoter or distal ATAC-seq peak–promoter

Hi–C interaction

6
3 transcription factors that may have a role in forming the chromatin
0
0 0
0 10 20 30 40 0 10 20 30
Number of enhancers Number of genes
linked per gene linked per enhancer
Fig. 4 | Conservation of zebrafish cis-regulatory elements and
transcriptional networks. a, Percentage of zebrafish exons and cis-regulatory
elements that have orthologous sequences in human. Total number for each
bar: exon, 1,000; promoter, 25,593; enhancer, 58,065; and random, 1,000. For
exons and random, we randomly sampled 1,000 elements and computed the
percentage conservation. The simulations were performed 20 times and the
mean percentage is shown. b, Percentage of human orthologous sequences of
zebrafish enhancers that were predicted as enhancers in human tissues. Total
number for each bar: brain, 1,241; blood, 748; colon, 775; heart, 839; intestine,
564; kidney, 173; liver, 402; muscle, 591; skin, 356; and spleen, 1,000. c, Fish
PhyloP score for the zebrafish enhancers with sequences that were not
conserved in human (number of enhancers in red line is 51,446, blue line is
50,000). d, An ultra-conserved noncoding element predicted as a brain
enhancer in zebrafish, mouse and human. This enhancer element has been
validated by transgenic reporter assay in mouse (hs1056 in the VISTA Enhancer
Browser). e, Heat map showing transcription factor motif enrichment in
tissue-specific enhancers in zebrafish and human. f, Linking distal enhancers to
their target genes by correlation of tissue-specific activity. Left, distribution of
the predicted number of enhancers per gene. Right, distribution of predicted
number of genes per enhancer. g, Validation of the predicted enhancer-to-gene
pairs by Hi-C interaction counts in brain.
tissues in zebrafish and human (Fig. 4e). To further probe the simi-
larities in transcription factor connections between zebrafish and
human, we performed the three-node network analysis as previously
described31 (Methods). We observed that CTCF was predicted as
the driver node in most tissues, whereas tissue-specific transcrip-
tion factors such as NEUROD and MYOD were predicted as middle
and passenger nodes in the networks of brain and muscle tissue,
respectively (Extended Data Fig. 9e, f). The overall patterns of the
three-node networks were highly similar between zebrafish and
human (Extended Data Fig. 9f), further demonstrating the value of
zebrafish as a system to study human transcription factor regula-
tory circuits.
40 0 500
Genomic distance (kb)
linkage pairs (Fig. 5b). We performed motif analysis to identify the
loops. CTCF and BORIS were enriched in shared loops (Fig. 5a), and
tissue-specific transcription factors were enriched in tissue-specific
chromatin interactions (Fig. 5a). For example, RFX and NeuroD2 were
enriched in brain-specific loops, whereas two muscle-specific master
regulators, Myf5 and Ascl1, were enriched in muscle-specific loops
(Fig. 5c).
Zebrafish genome evolution and TADs
TADs have been shown to be conserved among different species33–36.
To investigate the relationship between TADs and zebrafish genome
evolution, we first identified three sets of zebrafish evolutionary break-
points by aligning its genome against chicken, mouse and human,
respectively. We then compared the breakpoints with TAD annota-
tions in zebrafish and observed that 80.5% of breakpoints (984 of 1,223
zebrafish-to-human breakpoints) were located near TAD boundaries,
but depleted towards the centre of TADs (Fig. 5d, Extended Data
Fig. 11a). We divided TADs into two groups: TADs containing a break-
point and TADs not containing a breakpoint. TADs without breakpoints
had stronger interaction frequencies in the middle than TADs with
breakpoints (Fig. 5e, Extended Data Fig. 12d). Further, the expression
patterns of genes across different tissues in the TADs without break-
points were more correlated with those of their homologues in human
(Fig. 5f) than the other group, suggesting that there is an association
between TAD stability and conservation of the expression pattern. This
may be caused by strong chromatin interactions that may contribute to
TAD stability during evolution, or breaking of TADs with strong interac-
tions that are selected against in evolution, as these interactions and
the genes involved in the interactions may be physiologically important
for zebrafish.
Next, we divided zebrafish TAD boundaries into two classes, bounda-
ries overlapping with breakpoints and boundaries not overlapping with
breakpoints. We observed higher H3K4me3 signals at TAD boundaries,
as previously described36. We found a much higher level of H3K4me3 at
TAD boundaries with breakpoints (Fig. 5g, Extended Data Fig. 11b, c).
We also confirmed similar higher H3K4me3 enrichment in human
or mouse TAD boundaries with evolutionary breakpoints (Extended

Nature | Vol 588 | 10 December 2020 | 341

Article
a b Supported by
Hi-C loops linkage prediction
Brain Muscle Motif –log (P) Enhancer
24.5% 24.74%
n = 4,610 ATAC -seq
CTCF 1106 E–E 7%
8.3 Both

specific
39.73%

Brain
RFX2 106 32.53% None
34.37% E–P
NeuroD2 105 P–P
34.37%

n = 3,098 CTCF 568

d

Shared
BORIS 437 Zebrafish vs mouse Zebrafish vs human
200 200

breakpoints (BPs)
49
X-box

Number of
n = 2,219
specific CTCF 79
Muscle

Ascl1 29
Myf5 24 50 50
–500 kb +500 kb –500 kb +500 kb
TAD TAD

c
Shared loops Brain-specific loops Muscle-specific loop
akt3a zbtb18 traip
20 H3K27ac
Muscle Brain

5 RNA-seq
20 H3K27ac
5 RNA-seq
Brain Brain Brain

= 35 = 20 = 15
Muscle Muscle Muscle
chr13:13,000,000-17,500,000 chr13:10,000,000-11,800,000 chr11:33,750,000-35,600,000

e f h 3.0
Pol2
With CTCF
TADs without BPs With Without

ChIP-seq
BPs CTCF

signal
**

Without
BPs
0 1
Correlation 0.5
–1 Mb Center +1 Mb
GRO-seq
TADs with BPs
g 0.14
8 With BPs
Without BPs
GRO-seq
H3K4me3

signal

+1 Mb 0.02
–1 Mb Center 0
–500 kb +500 kb –200 kb Breakpoint +200 kb
0 1 Boundary

Fig. 5 | Higher-order chromatin structure and zebrafish genome evolution. breakpoints (bottom). f, Expression pattern of genes in TADs without an
a, Aggregate peak analysis plot and motif analysis of tissue-specific or shared evolutionary breakpoint is more highly conserved than genes in TADs that
chromatin loops. In each panel, n is the number of loops in that group. b, Left, contain a breakpoint. For each gene, we collected its expression profile across
annotation of cis elements in the predicted loop anchors in brain with a total of the same ten tissues in both zebrafish and human, and computed a Spearman
7,710 loops in the pie chart. Right, comparison of promoter–enhancer correlation coefficient between the profiles for each gene. Number of gene
chromatin loops with correlation-based linkage between ATAC-seq or histone pairs: without BPs, 4,625; with BPs, 3,918; P = 3.56 × 10 −26, two-sided Mann–
modification-based enhancer–gene pairs with a total of 4,996 loops in the pie Whitney U test. g, H3K4me3 signals were higher in TAD boundaries with
chart. c, Examples of shared, brain-specific and muscle-specific chromatin breakpoints than TAD boundaries without breakpoints. h, Higher
loops. d, Relative position of evolutionary breakpoints to TADs. Breakpoints transcriptional activities at TAD boundaries with breakpoints and containing
were between zebrafish and mouse (left) or between zebrafish and human CTCF binding sites in human GM12878 cells. Number of breakpoints: with CTCF,
(right). In all cases, we found that the evolutionary breakpoints were enriched 639; without CTCF, 625. K562 cell data is shown in Extended Data Fig. 12c. Results
at zebrafish TAD boundaries and depleted from the centre of TADs. e, TADs from 17 additional vertebrates are shown in Extended Data Figs. 11a, b, 12d.
without breakpoints (top) have stronger internal interactions than TADs with

Data Fig. 11d–f). As a control, we observe similar amounts of H3K27ac We observed that there were much higher transcription activities at
or ATAC-seq enrichment between the two groups of TAD boundaries breakpoints overlapping with CTCF-containing TAD boundaries, com-
(Extended Data Fig. 12a, b). Notably, an earlier report showed H3K4me3 pared with those without CTCF TAD boundaries (Fig. 5h, Extended
signal enrichment at recombination hotspots in mouse37, and our find- Data Fig. 12c).
ing in zebrafish further suggests its potential association with genome A key feature of the zebrafish genome is an extra genome-duplication
stability and evolution. event compared with other vertebrates41. There were 2,456 paralogous
Previous work has suggested a link between transcription at gene pairs annotated in Ensemble and the paralogues show similar
CTCF-containing TAD boundaries and their potential role in trans- expression patterns across all different tissues, with a median Pearson
locations38–40. Therefore, we investigated the transcriptional status correlation of 0.458 (Extended Data Fig. 12e). We analysed paralogue
at the evolutionary breakpoints at TAD boundaries using the Pol2, pairs located on the same chromosome and observed that paralogues
CTCF ChIP-seq and GRO-seq data in GM12878 and K562 human cells. located in the same TADs have a higher correlation in gene expression

342 | Nature | Vol 588 | 10 December 2020

patterns than paralogues located in different TADs (Extended Data
Fig. 12f).
In summary, we report a comprehensive annotation of the zebrafish
genome, and we described both conserved and divergent gene-
regulatory networks and 3D genome structures between zebrafish and
human. The breadth and depth of the data establish a genomic founda-
tion for conducting further human disease modelling and biological
studies in zebrafish.
Online content
Any methods, additional references, Nature Research reporting sum-
maries, source data, extended data, supplementary information,
acknowledgements, peer review information; details of author con-
tributions and competing interests; and statements of data and code
availability are available at https://doi.org/10.1038/s41586-020-2962-9.
1. Howe, K. et al. The zebrafish reference genome sequence and its relationship to the
human genome. Nature 496, 498–503 (2013).
2. Gerhard, G. S. et al. Life spans and senescent phenotypes in two strains of Zebrafish
(Danio rerio). Exp. Gerontol. 37, 1055–1068 (2002).
3. Lamason, R. L. et al. SLC24A5, a putative cation exchanger, affects pigmentation in
zebrafish and humans. Science 310, 1782–1786 (2005).
4. Vastenhouw, N. L. et al. Chromatin signature of embryonic pluripotency is established
during genome activation. Nature 464, 922–926 (2010).
5. Bogdanovic, O. et al. Dynamics of enhancer chromatin signatures mark the transition
from pluripotency to cell specification during embryogenesis. Genome Res. 22,
2043–2053 (2012).
6. Kaaij, L. J. et al. Enhancers reside in a unique epigenetic environment during early
zebrafish development. Genome Biol. 17, 146 (2016).
7. Aday, A. W., Zhu, L. J., Lakshmanan, A., Wang, J. & Lawson, N. D. Identification of cis
regulatory features in the embryonic zebrafish genome through large-scale profiling of
H3K4me1 and H3K4me3 binding sites. Dev. Biol. 357, 450–462 (2011).
8. Vesterlund, L., Jiao, H., Unneberg, P., Hovatta, O. & Kere, J. The zebrafish transcriptome
during early development. BMC Dev. Biol. 11, 30 (2011).
9. Buenrostro, J. D., Giresi, P. G., Zaba, L. C., Chang, H. Y. & Greenleaf, W. J. Transposition of
native chromatin for fast and sensitive epigenomic profiling of open chromatin,
DNA-binding proteins and nucleosome position. Nat. Methods 10, 1213–1218 (2013).
10. ENCODE Project Consortium. An integrated encyclopedia of DNA elements in the human
genome. Nature 489, 57–74 (2012).
11. Yue, F. et al. A comparative encyclopedia of DNA elements in the mouse genome. Nature
515, 355–364 (2014).
12. Anderson, J. L. et al. Multiple sex-associated regions and a putative sex chromosome in
zebrafish revealed by RAD mapping and population genomics. PLoS ONE 7, e40701
(2012).
13. Klemm, S. L., Shipony, Z. & Greenleaf, W. J. Chromatin accessibility and the regulatory
epigenome. Nat. Rev. Genet. 20, 207–220 (2019).
14. Meuleman, W. et al. Index and biological spectrum of human DNase I hypersensitive sites.
Nature 584, 244–251 (2020).
15. Quillien, A. et al. Robust identification of developmentally active endothelial enhancers
in zebrafish using FANS-assisted ATAC-seq. Cell Rep. 20, 709–720 (2017).
16. Letelier, J. et al. Evolutionary emergence of the rac3b/rfng/sgca regulatory cluster
refined mechanisms for hindbrain boundaries formation. Proc. Natl Acad. Sci. USA 115,
E3731–E3740 (2018).
17. Liu, G., Wang, W., Hu, S., Wang, X. & Zhang, Y. Inherited DNA methylation primes the
establishment of accessible chromatin during genome activation. Genome Res. 28,
998–1007 (2018).
18. Marlétaz, F. et al. Amphioxus functional genomics and the origins of vertebrate gene
regulation. Nature 564, 64–70 (2018).
19. Meier, M. et al. Cohesin facilitates zygotic genome activation in zebrafish. Development
145, dev156521 (2018).
20. Torbey, P. et al. Cooperation, cis-interactions, versatility and evolutionary plasticity of
multiple cis-acting elements underlie krox20 hindbrain regulation. PLoS Genet. 14,
e1007581 (2018).
21. Paik, E. J. et al. A Cdx4–Sall4 regulatory module controls the transition from mesoderm
formation to embryonic hematopoiesis. Stem Cell Reports 1, 425–436 (2013).
22. Kang, J. et al. Modulation of tissue repair by regeneration enhancer elements. Nature 532,
201–206 (2016).
23. Kaufman, C. K. et al. A zebrafish melanoma model reveals emergence of neural crest
identity during melanoma initiation. Science 351, aad2197 (2016).
24. Goldman, J. A. et al. Resolving heart regeneration by replacement histone profiling.
Dev. Cell 40, 392–404 (2017).
25. Pérez-Rico, Y. A. et al. Comparative analyses of super-enhancers reveal conserved
elements in vertebrate genomes. Genome Res. 27, 259–268 (2017).
26. Lister, R. et al. Global epigenomic reconfiguration during mammalian brain development.
Science 341, 1237905 (2013).
27. Visel, A. et al. Ultraconservation identifies a small subset of extremely constrained
developmental enhancers. Nat. Genet. 40, 158–160 (2008).
28. Dimitrieva, S. & Bucher, P. UCNEbase–a database of ultraconserved non-coding elements
and genomic regulatory blocks. Nucleic Acids Res. 41, D101–D109 (2013).
29. Visel, A., Minovitsky, S., Dubchak, I. & Pennacchio, L. A. VISTA Enhancer Browser—a
database of tissue-specific human enhancers. Nucleic Acids Res. 35, D88–D92 (2007).
30. Corces, M. R. et al. The chromatin accessibility landscape of primary human cancers.
Science 362, eaav1898 (2018).
31. Neph, S. et al. Circuitry and dynamics of human transcription factor regulatory networks.
Cell 150, 1274–1286 (2012).
32. Yang, T. et al. HiCRep: assessing the reproducibility of Hi-C data using a stratum-adjusted
correlation coefficient. Genome Res. 27, 1939–1949 (2017).
33. Krefting, J., Andrade-Navarro, M. A. & Ibn-Salem, J. Evolutionary stability of topologically
associating domains is associated with conserved gene regulation. BMC Biol. 16, 87
(2018).
34. Lazar, N. H. et al. Epigenetic maintenance of topological domains in the highly
rearranged gibbon genome. Genome Res. 28, 983–997 (2018).
35. Fishman, V. et al. 3D organization of chicken genome demonstrates evolutionary
conservation of topologically associated domains and highlights unique architecture of
erythrocytes’ chromatin. Nucleic Acids Res. 47, 648–665 (2019).
36. Dixon, J. R. et al. Topological domains in mammalian genomes identified by analysis of
chromatin interactions. Nature 485, 376–380 (2012).
37. Smagulova, F. et al. Genome-wide analysis reveals novel molecular features of mouse
recombination hotspots. Nature 472, 375–378 (2011).
38. Canela, A. et al. Genome organization drives chromosome fragility. Cell 170, 507–521
(2017).
39. Gothe, H. J. et al. Spatial chromosome folding and active transcription drive DNA fragility
and formation of oncogenic MLL translocations. Mol. Cell 75, 267–283 (2019).
40. Canela, A. et al. Topoisomerase II–induced chromosome breakage and translocation is
determined by chromosome architecture and transcriptional activity. Mol. Cell 75,
252–266 (2019).
41. Postlethwait, J. H. et al. Vertebrate genome evolution and the zebrafish gene map.
Nat. Genet. 18, 345–349 (1998).
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in
published maps and institutional affiliations.
© The Author(s), under exclusive licence to Springer Nature Limited 2020
Nature | Vol 588 | 10 December 2020 | 343
Article
Methods
No statistical methods were used to predetermine sample size. The
experiments were not randomized. The investigators were not blinded
to allocation during experiments and outcome assessment.
Adult tissue ChIP-seq
One-year-old adult Tübingen zebrafish (sex information shown in
Supplementary Table 17) were dissected to separate tissues that were
washed twice in 1× PBS buffer before flash freezing on dry ice. Collec-
tion of peripheral blood from adult fish was performed as described42.
All procedures on live animals have been approved by the Institutional
Animal Care and Use Committee (IACUC) at the Pennsylvania State
University (PRAMS201445659). At the beginning of ChIP-seq, all tissues
(except peripheral blood) were ground in liquid nitrogen and fixed by 1%
formaldehyde at room temperature for 15 min. 2.5 M glycine was added
at a final concentration of 0.2 M and incubated at room temperature
for 5 min to quench the fixation. Fixed tissues were then washed once
with cold 1× PBS. Tissue pellets were resuspended and incubated on ice
for 10 min in 100 μl of ChIP-seq lysis buffer (20 mM Tris-HCl, pH 8.0, 1%
SDS, 50 mM EDTA, 1× proteinase inhibitor cocktail). Next, 900 μl cold
1× TE buffer (10 mM Tris-HCl, 1 mM EDTA, pH 8.0) was added to dilute
the SDS, and the nuclei suspension was sonicated using a Covaris E220
with the following parameters: 140 W, duty factor 5, 200 per burst. To
check the chromatin fragmentation size, 20 μl of input chromatin was
reverse cross-linked in elution buffer (20 mM Tris-HCl, pH 8.0, 1% SDS,
1 mM EDTA) at 65 °C overnight, treated with RNase A and proteinase K
and purified by phenol–chloroform extraction. Input DNA was then
loaded on a Lonza flash gel to ensure that the majority of DNA was
between 100–300 bp. To prepare the antibody–beads complex, 3 μg
of histone H3K27ac antibody (Active Motif, 39133), H3K4me3 antibody
(EMD Millipore, 07-473), H3K9me2 antibody (Cell Signaling, 4658) or
H3K9me3 antibody (Abcam, ab8898) was mixed with 12 μl M-280 sheep
anti-rabbit (ThermoFisher, 11203D) or sheep anti-mouse IgG Dynabeads
(ThermoFisher, 11201D) in 150 μl of 5 mg ml−1 BSA in 1× PBS buffer, with
rotation at 4 °C for 3 h. After incubation, the antibody–beads complexes
were washed once with BSA in 1× PBS buffer. About 200 μg chromatin
was used per immunoprecipitation. An equal volume of master mix
(1× TE, 2% Triton X-100, 0.2% sodium deoxycholate, 2× proteinase inhibi-
tor cocktail) was mixed with 200 μg chromatin and then incubated with
antibody–beads complexes overnight with rotation. The next morn-
ing, the beads were washed 5 times with cold RIPA wash buffer (20 mM
Tris-HCl, pH 8.0, 1% NP-40, 0.7% sodium deoxycholate, 500 mM LiCl,
1 mM EDTA, 1× proteinase inhibitor cocktail). Then the bead-bound
chromatin was eluted using 150 μl of elution buffer at 65 °C for 30 min.
To prepare the library, eluted chromatin was reverse cross-linked and
purified by phenol-chloroform extraction. Then DNA was end-repaired
with the END-IT DNA end-repair kit (Epicentre, ER81050) according to
the kit protocol, adenylated using Klenow fragment (3′→5′ exo-) (NEB,
M0212S), ligated with Illumina TruSeq adaptor (Illumina, FC-121-3001)
and subsequently amplified by PCR (Roche, kk2601). The quality and
quantity of all the libraries were checked using a BioAnalyzer High
Sensitivity DNA Kit (Agilent).
Embryonic tissue ChIP-seq
Zebrafish embryonic tissue ChIP-seq was performed similarly to the
adult tissue with a few modifications. For embryonic trunk, 50 trunks of
1-dpf embryos were dissected and digested in 500 μl of 0.25% trypsin at
room temperature for 20 min. The reaction was then neutralized with
FBS. Trunk cells were washed once with cold 1× PBS and cross-linked
with 1% formaldehyde at room temperature for 12 min. For embry-
onic neuronal cells, Tg (HuC:Kaede) transgenic fish were crossed with
wild-type Tübingen fish. At 1 dpf, embryos were checked under the
fluorescence microscope, and green positive embryos were selected,
dechorionated, and pipetted in calcium-free Ringer’s solution to
remove the yolk. The supernatant was removed, and 500 μl of 0.25%
trypsin was added to digest embryos at room temperature for 20 min.
After neutralization with 500 μl of FBS and washed once with cold
1× PBS, digested cells were then sorted, and green fluorescent cells,
which corresponded to embryonic neuronal cells, were collected and
cross-linked with 1% formaldehyde at room temperature for 12 min.
RNA-seq
For each RNA-seq experiment, tissues from at least two Tübingen fish
were combined to use as one replicate. For embryonic trunk, ten 1-dpf
fish were dechorionated with pronase, and the trunk was separated
for RNA-seq. For embryonic neurons, green cells from Tg(Huc:Kaede)
fish were sorted by FACS, and approximately 20,000 cells were used
for one replicate. The tissue RNA was extracted from Trizol according
to the manufacturer’s protocol (Invitrogen). The cDNA libraries were
constructed using SureSelect Strand Specific RNA Library Prepara-
tion Kit (Agilent) according to the manufacturer’s protocol. In brief,
polyA RNA was purified from 1,000 ng of total RNA using oligo (dT)
beads (Invitrogen). Extracted RNA was first fragmented, then followed
by reverse transcription, end-repair, adenylation, adaptor ligation,
and subsequent PCR amplification. The final product was checked
by size distribution and concentration using a BioAnalyzer High Sen-
sitivity DNA Kit (Agilent) and Kapa Library Quantification Kit (Kapa
Biosystems).
ATAC-seq
Tübingen adult tissues were freshly dissected and processed imme-
diately for ATAC-seq. In brief, the tissues were resuspended in 1 ml of
lysis buffer (1× PBS, 0.2% NP-40, 5% BSA, 1 mM DTT, protease inhibi-
tors), followed by Dounce homogenization with a loose pestle using
20 strokes. The lysate was then filtered through a 40-μm cell strainer,
and nuclei were collected at 500g for 5 min. Tagmentation was
performed immediately according to the previously reported
ATAC-seq protocol9.
Whole-genome bisulfite sequencing
Tübingen adult tissues were dissected, and the genomic DNA was
extracted using DNeasy Blood & Tissue Kit (Qiagen, 69504). Then, 1 μg
of genomic DNA of each tissue was subjected to bisulfite conversion
using EZ DNA Methylation-Gold Kit. The final libraries were prepared
using the Accel-NGS Methyl-Seq DNA Library Kit.
BioNano optical mapping
Genomic DNA from one Tübingen female muscle was extracted for
BioNano optical mapping. DNA extraction was done according to the
BioNano Prep Animal Tissue DNA Isolation Fibrous Tissue Protocol-
30071. The homogenized genomic DNA was then directly labelled
using DLE enzyme (BioNano, 80005) according to the BioNano Prep
Direct Label and Stain (DLS)-30206 protocol. The labelled and stained
genomic DNA was then loaded onto a Saphyr chip, and 407 Gb of data
were collected for genome assembly.
scATAC-seq
scATAC-seq was performed using one female brain and one male brain
on the 10X Genomics platform. To isolate nuclei, a freshly dissected sin-
gle brain was transferred to 1 ml NbActiv1 medium (BrainBits, NbActiv1
500) with a wide-bore pipette tip to break the tissue into small pieces.
Then the tissue was further fragmented using regular-bore pipette
tips followed by filtering through a 30-μm cell strainer. The isolated
cells were spun down at 500g for 5 min at 4 °C and then lysed in 100 μl
chilled 0.1× lysis buffer (10 mM Tris HCl, pH 7.5, 10 mM NaCl, 3 mM MgCl2,
1% BSA,0.01% Tween-20, 0.01% NP40 and 0.001% digitonin) on ice for
5 min. Then, 1 ml chilled wash buffer (10 mM Tris HCl, pH 7.5, 10 mM
NaCl, 3 mM MgCl2, 1% BSA, 0.1% Tween-20) was added to the lysed cells
and cells were spun down at 500g for 5 min at 4 °C. Finally, 300 μl chilled
diluted nuclei buffer (10X Genomics, 2000153/2000207) was added to
resuspend the nuclei. The nuclei were filtered again using a 30-μm cell
strainer before cell counting. Around 12,000 nuclei were used for one
Tn5 tagmentation reaction, and the scATAC-seq library was prepared
and sequenced according to the 10X Genomics user guide.
Zebrafish tissue Hi-C
Hi-C experiments on adult zebrafish brain and muscle tissues were per-
formed according to a previously published protocol43 with a few modi-
fications. For brain replicate 1, two adult Tübingen zebrafish brains were
gently ground into small granules in liquid nitrogen and resuspended
in 1 ml cold hypotonic buffer (20 mM Tris-HCl, pH 8.0, 10 mM NaCl,
20 mM EDTA). For brain replicate 2, only one female Tübingen brain
was used. Granular brains were then Dounce homogenized with a loose
pestle with 20 strokes. The upper layer of the homogenate was carefully
transferred into a new tube and fixed with 2% formaldehyde at room
temperature for 10 min. 0.2 M of glycine was added to stop fixation.
For muscle tissue, 60 mg of Tübingen muscle was first chopped into
small pieces and digested with 0.25% trypsin at room temperature for
30 min. After neutralization with FBS, muscle cells were resuspended
in cold 1× PBS and fixed with 2% of formaldehyde at room temperature
for 10 min. Glycine (0.2 M) was then added to stop fixation. Two muscle
Hi-C experiments were performed using tissues from female and male
Tübingen fish, respectively.
Enhancer reporter assays
Selected potential tissue-specific enhancers were evaluated using a
Tol2 transposon-mediated zebrafish transgenesis approach as pre-
viously described44. Selected enhancers were PCR amplified from
Tübingen genomic DNA and subcloned to the upstream of the hsp70
promoter-eGFP cassette in the pT2HE vector (Supplementary Table 18
for primers). These enhancer reporter plasmids were subsequently
co-injected with transposase mRNA in one-cell-stage Tübingen
wild-type zebrafish embryos. eGFP expression patterns were moni-
tored by Zeiss SteREO DiscoveryV8 microscope by AxioVison Rel.4.8
software, at four different developmental time points: 24–36 hours
post-fertilization, 2 dpf, 3 dpf and 5 dpf. Expression patterns were
recorded only if eGFP signal was consistently displayed by at least
30–40% of total embryos (approximately 100 embryos were analysed
per enhancer).
Novel transcript identification
All the RNA-seq reads were mapped to the genome GRCz10
using Tophat245, with ‘–mate-inner-dis 50–mate-std-dev 1000–
read-mismatches 2–read-edit-dist 2’ parameters excluding ‘-G’, then
assembled using Cufflinks46. The cuffmerge function was applied to
combine cufflinks assemblies from each library. The cuffcompare func-
tion was then used to analyse assembled transcripts of all libraries
and to merge assembled transcripts with ENSEMBL and RefSeq refer-
ence annotations (ftp://ftp.ncbi.nlm.nih.gov/genomes/Danio_rerio/
ARCHIVE/ ANNOTATION_RELEASE.105/Gnomon/ and ftp://ftp.ensembl.
org/pub/release-91/gtf/danio_rerio/). Novel transcripts were defined as
those with a u, x, j or i class code and presenting in both biological rep-
licates. Class code u represents no overlap with any known transcript;
x represents exonic overlap with known transcripts on the opposite
strand; j represents potential novel isoforms with at least one splice
junction is shared with a known transcript; and i represents a transcript
falling entirely within a reference intron.
lncRNA identification and novel transcript annotation
To identify lncRNAs from novel transcripts, we first selected novel
transcripts with more than one exon and longer than 150 bp. The coding
potential of these transcripts was predicted using CPAT with a zebrafish
model47. The transcripts with a coding potential score less than 0.38
were identified as lncRNAs. The transcripts with a coding potential
score greater than 0.38 were considered as novel protein-coding gene
candidates. TransDecoder (https://transdecoder.github.io/) was then
used to predict open reading frames of these novel protein-coding gene
candidates. The predicted coding sequences were blasted with BLASTp
against known protein databases of other species (Xenopus, cattle, pig,
chicken, mouse and human) to identify homologues in these species
(cut-off: 10−3). The GTF files for the predicted lncRNAs and the novel
transcripts were integrated with reference GTF files for further analysis.
Quantification for RNA-seq signals and identification of
tissue-specific genes
RNA-seq reads were trimmed using TrimGalore (https://github.com/
FelixKrueger/TrimGalore). We aligned trimmed RNA-seq reads to the
zebrafish genome (GRCz10) and generated bigwig files using STAR48.
QC and expression values for the RNA-seq libraries were computed
with the RSEM package49 and the previously generated GTF files were
used for the genome annotation. To correct the batch effect across all
libraries prepared at different times, we used the limma package in R on
the log2(TPM + 1) matrix, and genes with TPM <1 in two biological repli-
cates in all tissues were further filtered out. Pearson correlation coef-
ficients were then computed between biological replicates using the
read counts of 10 kb-binned matrices. The average TPM value between
two biological replicates was used to represent the gene expression
value of each tissue to identify the tissue-specific expressed genes.
TPM values for the same gene in different tissues were transformed
into Z-scores to identify the tissue-specific genes at a threshold of Z > 2
and a minimum of threefold change.
Comparison of orthologous genes between zebrafish and
human
To compare the orthologue gene expression pattern between zebrafish
and human, we downloaded the TPM matrix from the GTEx dataset from
the same tissue (except testis and embryonic tissues), and the median
TPM value of replicates was used to represent the gene expression value
of each tissue. The one-to-one orthologous genes detected by BioMart
were selected for the comparison.
ATAC-seq data processing
ATAC-seq library adapters were detected and trimmed using detect_
adapter.py (https://github.com/ENCODE-DCC/atac-seq pipeline/blob/
master/src/detect_adapter.py) and cutadapt50. The trimmed reads were
aligned using bowtie251 with the following parameters: -X2000 –mm.
PCR duplication was removed using Picard (http://broadinstitute.
github.io/picard/). We calculated the Pearson correlation coef-
ficient between two biological replicates using the reads counts of
10 kb-binned matrices and merged paired files with high correlation.
Then we process the ATAC-seq data using parameters recom-
mended by ENCODE-DCC (https://github.com/ENCODE-DCC/
atac-seq-pipeline). Peaks were filtered to keep those with P <0.00001
and q <0.01 and peak length were fixed to 500 bp. We only selected
the peak with the most significant signal if several peaks overlapped in
each tissue. Then we separated the proximal ATAC-seq peaks (average
n = 23,233), which were defined as overlapping with any (transcription
start site) TSS regions (2.5 kb upstream to 500 bp downstream of TSS),
and distal ATAC-seq peaks (average n = 122,979).
Chromatin accessibility can reveal a full range of regulatory ele-
ments, including both active and poised enhancers and promoters.
To generate the zebrafish open chromatin landscape, we merged the
ATAC-seq peaks across 11 tissues. The peaks with the most significant
signals were selected as representative open chromatin regions, and
all the peaks overlapping with representative open chromatin regions
of at least one base-pair were removed. In total, we identified 436,035
representative open chromatin regions in the zebrafish genome. We
annotated the ATAC-seq peaks using the ChIPseeker package. Footprint
was identified by the HINT software v.0.13.0 (Hmm-based IdeNtification
Article
of Transcription factor footprints)52 based on ATAC-seq data. In brief,
ATAC-seq narrowpeaks were used as input, the footprint region were
filtered by footprint score >10, transcription factor motifs overlap with
footprints was identified using the MOODS package v.1.9.453 (https://
github.com/jhkorhonen/MOODS), with motifs from the HOCOMOCO
database54 (http://hocomoco11.autosome.ru/).
H3K27ac and H3K4me3 ChIP-seq data processing
We mapped the ChIP-seq reads to the zebrafish genome GRCz10 using
BWA aligner55. The mapped reads with MAPQ less than 30 were removed,
and PCR duplicated reads were removed by Picard. We calculated the
Pearson correlation coefficient between two biological replicates
and merged paired files with a high correlation. For the H3K27ac
and H3K4me3 histone marks, we used parameters recommended by
ENCODE-DCC (https://github.com/ENCODE-DCC/chip-seq-pipeline2).
In brief, candidate narrow peaks were first selected with −logP >5, and
−logQ >2. Reads per million (RPM) of immunoprecipitate (IP) data
(RPMIP) and input data (RPMinput) in each peak region were calculated,
and the qualified peaks should pass the threshold of twofold enrich-
ment (RPMIP ≥ 2 × RPMinput) and RPMIP − RPMInput > 1.
Reads counts of peak in IP
RPM(IP) =
Number of non−duplicated reads per million in IP
RPM(Input)
Reads counts of peak in input
= minMatch >0.1, the danRer10ToHg38/Mm10.over.chain files were
Number of non−duplicated reads per million in input
H3K9me3 and H3K9me2 ChIP-seq data processing
Since the H3K9me3 and H3K9me2 markers are broad domain, we called
the peaks using Homer with the parameter ‘-region -size 1,000’, and
peaks within 5 kb were merged together.
Reproducibility of ChIP-seq data
To check the reproducibility of biological replicates, we divided the
reference genome into 10-kb bins and computed the number of reads
within each bin. The Pearson correlation coefficients between biologi-
cal replicates for H3K4me3, H3K27ac, H3K9me3 and H3K9me2 were
calculated using the normalized 10-kb binned reads. After confirming
that all replicates were highly correlated, we pooled the .bam files of
biological replicates together with the merge function of Samtools56
for further analysis.
Identification of the zebrafish genome blacklist
Genomic experiments based on Illumina sequencing (for example,
ChIP-seq and ATAC-seq) often produce artificial high signals in certain
genomic regions, such as centromeres, telomeres and satellite repeats.
It is therefore essential to identify and remove these artificial signals
that exist ChIP-seq and ATAC-seq experiments. To flag these artifi-
cial regions, ChIP-seq input sequencing data were used as IP with the
whole genome sequence as background to call the artificial peaks by
MACS257. The peaks with −log10(q-value) less than 5 were filtered out.
Then, the ChIP sequencing data were used to call peaks with the genome
sequence as background, with the same threshold. The overlapped
peaks of these two datasets crossing all tissues were defined as the
zebrafish genome blacklist, and these regions were filtered out for all
ChIP-seq and ATAC-seq peak-calling analyses (Supplementary Table 19).
Identification of cis-regulatory elements
To systematically compare ChIP-seq data, we used a 1-kb flanking region
of summit peaks to define the promoters and enhancers across all tis-
sues. We observed, on average, 96% of H3K27ac and H3K4me3 were
overlapping with representative open chromatin regions (ATAC-seq
peaks ± 500 bp). Weak promoters were defined by H3K4me3 peaks that
overlapped with representative open chromatin region but without
H3K27ac peaks. Active promoters were defined by H3K4me3 peaks that
overlapped with representative open chromatin region and H3K27ac
peaks. Active enhancers were defined as H3K27ac peaks not overlap-
ping with H3K4me3 peaks or TSS regions (2.5 kb upstream to 500 bp
downstream of TSS) but overlapping with representative open chroma-
tin 500-bp flanking regions. As testis contains broad-spread H3K4me3
signals across the whole genome, instead of using H3K4me3 peaks,
we used the annotated TSS regions that overlapped with H3K27ac
peaks and representative open chromatin regions to define active
promoters. Those that overlapped with representative open chroma-
tin regions without H3K27ac peaks defined weak promoters, whereas
active enhancers were defined by H3K27ac peaks that were away from
the annotated TSS regions but overlapped with representative open
chromatin regions.
To generate the non-redundant union sets of each type of cis-
regulatory elements, we merged the active enhancers, active promoters
and weak promoters from different tissues, respectively, if the peaks
were within 500 bp, and used the middle points to represent the loca-
tion of merged active enhancers, active promoters and weak promoters,
then fixed the length of each element into 2 kb.
We defined heterochromatin in each tissue by merging the H3K9me3
and H3K9me2 peaks within 500 bp. We converted the cis-regulatory
elements from zebrafish genomic locations(danRer10) to human and
mouse genomic locations (hg38 and mm10, respectively), using the
liftOver tools with the centre 100 bp of each element and required
modified based on the zebrafish conversed CNE database58 (http://
zebrafish.stanford.edu).
Identification of tissue-specific cis-regulatory elements
We identified tissue-specific cis-regulatory elements based on the
union sets of each type of cis-regulatory element. Then we computed
the H3K27ac RPM change to represent the active enhancer/active pro-
moter intensity, and the resulting matrix was quantile normalized.
The normalized matrix was transformed into Z-scores to identify the
tissue-specific elements at a threshold of Z > 2 and a minimum of two-
fold change in magnitude. GO term analysis of top 1,000 tissue-specific
enhancers or active promoters (ranked by intensity) was performed
by using GREAT 3.0.058 after liftover the genomic coordinate of each
type of cis-regulatory elements into Zv9/danRer7. Motif analysis was
performed by HOMER2. For super-enhancers, only the narrow peaks
within super-enhancers were used for the GO term and motif analysis.
WGBS data processing
WGBS paired-end reads were mapped to the zebrafish genome assem-
bly GRCz10, as previously described with the minor modifications59.
To increase the mapping efficiency, the first ten low-quality base
pairs of the sequence read 1 s, and the first 15 of the sequence read
2 s were trimmed along with adaptor sequences using Trim Galore!
(The Babraham Institute) version 0.6.1 with the following param-
eters: --clip_R1 10 --clip_R2 15 --paired --retain unpaired -r1 21 -r2 21.
The trimmed reads were mapped to in silico bisulfite-converted
zebrafish genome reference by using Bismark60 v.0.18.1 with the fol-
lowing parameters: -X 2000 --un -N 1 -L 28. Unpaired or unmapped
read 1s were then mapped as single read mode by using Bismark with
the following parameters: -N 1 -L 28. Unpaired or unmapped read 2s
were also mapped as single read mode by using Bismark with the fol-
lowing parameters: --pbat -N 1 -L 28. The redundant reads from PCR
amplification were then removed by using the following Bismark com-
mand: deduplicate_bismark --bam. The methylation information for
individual cytosines was extracted from the deduplicated reads by
using Bismark with the following commands: bismark_methylation_
extractor --comprehensive --merge_non_CpG --gzip. After merging
paired-end and single-end extracted files, the Bismark commands
bismark2bedGraph --CX and coverage2cytosine --CX was used to cal-
culate total read count and methylation read count per each C. The
methylation levels and read coverage of each CpG were visualized on
the WashU Epigenome Browser using a methylC track61.
The mean CpG methylation levels, percentages of CpGs with low,
medium, and high methylation levels, and distribution of CpG methyla-
tion levels were calculated by using CpGs with at least five read coverage.
UMRs and LMRs
UMRs and LMRs were identified by using MethylSeekR v.1.22.062, fol-
lowing the tool’s recommendations with the minor modifications. The
random number generator seed of 123 was set at the beginning of the
analysis to ensure reproducibility. Partially methylated domains were
identified and masked using the smallest chromosome 25 as a training
set. UMRs and LMRs were identified using cut-offs of less than 0.5 of
methylation levels and at least 5 or 6 CpGs, ensuring FDRs below 5%.
UMRs and LMRs were assigned as active promoters, weak promoters,
or active enhancers if they overlap with these cis-regulatory elements
defined in the same tissue by using BEDTools v.2.27.1. UMRs and LMRs
were similarly intersected with proximal or distal ATAC-seq peaks.
DMRs
DMRs between two tissues were identified by using DSS v.2.14.063.
First, mean methylation levels of each CpG site was estimated with
smoothing64 in the DSS package. Then, dispersion at each CpG site
was estimated, and the Wald test on each CpG site was performed to
calculate statistical significance of methylation difference across dif-
ferent samples. Without replicates, DMRs were detected by using DSS
package’s callDMR function with the following parameters: delta = 0.2,
p.threshold = 0.05, minlen = 200, minCG = 5, dis.merge = 50, pct.
sig = 0.5. Hypomethylated DMRs in a given tissue were further filtered
by intersecting with UMRs or LMRs in the same tissue. Tissue-specific
hypoDMRs were defined as DMRs hypomethylated in at least 8 out of
10 pairwise comparisons. For the hypoDMRs with a heat map of DNA
methylation levels (Fig. 3c), union set of hypoDMRs were obtained by
using BEDtools’ merge function. DNA methylation levels of union set
of hypoDMRs were calculated using CpGs with at least 5 read coverage.
Only regions that exist uniquely in hypoDMRs of each tissue and have
DNA methylation levels available across all 11 tissues were used. Gene
ontology and wild-type expression enrichment analysis was performed
by the GREAT v.3.0.0, after converting genomic coordinate of hypoD-
MRs into Zv9/danRer7. Heat maps of DNA methylation levels, H3K27ac
ChIP-seq and ATAC-seq signals of tissue-specific hypoDMRs along with
their neighbouring regions were plotted by using deepTools65.
Repetitive elements enrichment analysis
We compared the predicted cis-regulatory elements with different
subtypes of repetitive elements, including LTR, DNA, SINE, LINE, sat-
ellite, unknown and rolling circle (RC), annotated by RepeatMasker to
investigate whether some cis-regulatory elements were enriched or
depleted of any specific types of repetitive elements. We analysed the
enrichment by calculating the number of overlapped base pairs between
each cis-regulatory element and subtype of repetitive elements. The
equation of calculated fold enrichment is shown below: j represents the
H3K27ac peaks in the active enhancer regions/ H3K9me3 peak regions/
H3K9me2 peak regions; i represents the subtype of repetitive elements.
Length of j overlap with i
Length of j
Fold enrichment = Length of i
Length of genome
Genome assembly
The Oxford nanopore sequencing reads of zebrafish Tu (50× cov-
erage) were assembled using Canu software66 (v.1.8) with default
settings for nanopore reads. To improve the accuracy of our assembly,
we performed four rounds of genome polishing. In the first two rounds,
we mapped the nanopore long reads to the contigs using minimap2
(v.2.16-r922) and polished the contigs with Nanopolish (v.0.11.1). In
rounds 3 and 4, the 10X Genomics reads (barcodes trimmed) were
aligned to the contigs with BWA-MEM (0.7.15-r1140). Duplicated reads
were marked by Picard (v.2.17.2). Pilon software67 (v.1.23) was used to
polish the contigs. 1,325 Nanopore contigs (maximum length = 40.077
Mbp) were generated with N50 of 9.418 Mbp. Next, we applied BioNano
Solve v.3.2.1 software (‘non-haplotype’ with ‘no extend split’ and ‘no
cut seg dups’) to combine the Nanopore contigs with BioNano opti-
cal mapping data to generate 114 chromosome arm-level scaffolds
(N50 = 29.725 Mbp and maximum length = 42.585 Mbp). At the last
step, we leveraged the brain Hi-C data with 3d-dna software68 (default
parameters) to generate the final chromosome level scaffolds with a
total length of 1,519 Mbp (N50 = 56.768 Mbp, maximum length = 75.632
Mbp). The structure variation is called by BioNano_Access software for
the de novo assembled and GRCz10 reference genome.
Linking distal ATAC-seq peaks and active enhancers to genes by
correlation
We strictly followed previously proposed strategies30. In brief, we
first removed ATAC-seq peaks whose normalized RPMs were less than
1 across all tissues. We then computed the Pearson correlation coef-
ficient (PCC) between all ATAC-seq peaks (log2(RPM)) and genes
(log2(TPM + 1)) whose TSSs were located within 500 kb of ATAC-seq
peaks. To determine PCC cut-off and estimate FDR, we randomly
selected 10,000 ATAC-seq peaks and computed their PCC with genes
which were located on other chromosomes. To compute the P value for
each correlation value, we calculated the z-score by compare it with the
mean and the standard deviation of the correlations of all the random
ATAC-seq-to-gene pairs. The z-score was converted into a two-tailed
P value. We then estimated FDR using the Benjamini–Hochberg proce-
dure. The distal H3K27ac peaks were processed by the same way with
FDR <0.01 and PCC cut-off = 0. 56142.
Motif comparative analysis between zebrafish and human
First, we predicted the motif enrichment in each zebrafish tissue-
specific enhancer group using Homer findMotifGenome function.
Then, we merged the enriched motifs in all tissues to generate a
tissue-motif matrix. The P value of motifs in each tissue was quantile
normalized. The human motif matrix was generated with the same
method using human Roadmap ChIP-seq datasets69. Then we ranked
the motifs by normalized P value, and the top three motifs of each tis-
sue in zebrafish were used to compare with the human motif matrix
Core transcriptional regulatory network analysis
To identify the regulatory transcription factor motifs for each tissue in
zebrafish and human, we used HOMER70 to analyse the nucleosome-free
regions (based on H3K27ac intensity) within the 10-kb flanking regions
of TSS sites for 661 transcription factor genes. We next used FIMO71 to
scan motif occupancy in the nucleosome-free regions with P < 1 × 10−5
as the threshold. Then, we performed the three-node motif network
analysis using the previously described method31.
Hi-C data processing
Mapping and matrix generation. For Hi-C data, adaptor sequences
were trimmed, and low-quality reads were removed. Paired-end reads
were mapped to the GRCz10 genome using HiC-Pro v.2.9.072 (https://
github.com/nservant/HiC-Pro). Singleton, multi-mapped, dumped,
dangling, self-circle paired-end reads, and PCR duplicates were all re-
moved by HiC-Pro after mapping. We generated raw contact matrices
at 25-kb, 40-kb, 100-kb, 500-kb and 1-Mb resolutions. Visualization of
Hi-C contact matrices was done using juicer tools v.1.8.973 and juicebox
v1.1174 (https://github.com/aidenlab/juicer/wiki/Download). HiCRep
Article
was used to compute correlations between replicates32. The cool file
was generated by cooler v.0.8.675, cooltools v.0.4.0 (https://github.
com/mirnylab/cooltools) was used to compute the expected value
and observed value in different resolutions.
A/B compartment. A/B compartment analysis was performed at 40-kb,
100-kb and 250-kb resolutions using HOMER software. The positive
eigenvalues were set to A compartments, and negative values were set
to B compartments. We identified the regions with changes in sign of
the PC1 value between muscle and brain as A/B compartment switched
regions.
Directionality index calculation. The directionality index of the 40-kb
binned raw Hi-C matrix was calculated as previously described36.
Insulation, boundary calculation and TAD calling. The TAD structure
(insulation/boundaries) was defined by the insulation score as previ-
ous studies76,77. The matrices which were used to calculate the insula-
tion score were normalized by ICE method78 for discarding the bias of
raw matrices. The insulation score of the ICE matrix was calculated
by the following parameters: -is 480,000 –ids 320,000 -im iqrMean
-ss 160,000.
Hi-C loop calling. Loops were computed from Hi-C matrices using
HiCCUPS43 with the parameter ‘(–ignore_sparsity -r 10000- k KR -f .1,.1,.1
-p 4,2,1 -i 7,5,3 -t 0.02,1.5,1.75,2 -d 20000,20000,50000’, as previously
described and part of the juicer tools package. Consistent loops were
identified using pairtopair function in BEDtools with the parameter
‘–type both –f 0.5’. Tissue-specific loops were identified using pairtopair
function in BEDtools with the parameter ‘-type notboth -slop 40000’.
Identification of evolutionary breakpoints distribution in TADs
We used progressiveMauve79 to identify the genomic rearrangement
breakpoints between zebrafish and human, gibbon, chimp, gorilla,
mouse, cat, dog, pig, sheep, cattle, chicken, zebra finch, Xenopus, platy-
fish, stickleback, tilapia, medaka and fugu, respectively. To calculate
the breakpoint density in TADs, we divided the zebrafish genome into
100-kb bins and computed the number of overlapped breakpoints.
We plotted the profile figure using the computeMatrix and plotPro-
file packages from deepTools, with a TAD length of 2,800 kb (median
size of the TADs) and boundary length of 40 kb. We defined TADs with
breakpoints using bedtools intersect function with parameter ‘-f 0.3’.
H3K4me3 pattern analysis of pile-up TADs
To integrate histone signals and cis-regulatory element intensity in
TADs, we divided the zebrafish genome into 100-kb bins. For H3K4me3,
we computed the H3K4me3 per base-pair signal in muscle within each
bin. We used computeMatrix and plotProfile packages from deepTools
to plot the density of H3K4me3. To compare the expression similarity
of the genes in TADs with breakpoints and TADs without breakpoints,
we calculated the Spearman correlation coefficients of gene expression
TPM across eleven tissues between zebrafish and human.
Interaction pattern analysis of pile-up TADs
To visualize the overall interaction patterns within TADs, we performed
a pile-up analysis similar to APA plotting43. We first normalized the
interaction matrix by dividing each entry of the raw 25-kb matrix by
the expected interaction frequency at the corresponding genomic
distance. To make the TAD boundary sharper, we only considered
the whole-genome intra-TAD interaction frequencies and excluded
inter-TAD interactions when we calculated the expected values. Then
the TAD sets with and without breakpoints were separately piled up
and averaged over the normalized matrix. Specifically, for each TAD
of interest, we extended fixed 40 bins (1 Mb) on both sides of the mid-
point and extracted the 80 × 80 matrix within the extended region.
Hundreds of such matrices were then aggregated, and the average
interaction intensity was calculated for each location of the matrix.
In the resulting plot, the centre point corresponds to the centre of
each TAD, and boundaries of the centre block represent the average
locations of piled-up TAD boundaries.
Single-cell analysis
We have generated 781,123,374 reads from 23,871 cells using 10X Genom-
ics Chromium single cell ATAC-seq solution protocol. The BCL files
generated from sequencing were used as inputs to the 10X Genomics
Cell Ranger ATAC-seq pipeline; then the FASTQ files were aligned to the
GRCz10 genome using BWA80 and the fragments with MAPQ >30 were
kept for further analysis and each fragment is associated with a single
cell barcode. To qualify scATAC-seq, we calculated the Pearson correla-
tion coefficient between the aggregated signal of scATAC-seq and bulk
ATAC-seq. The results showed that single-cell data was highly correlated
with bulk ATAC-seq (Fig. 2g). Then a cell-by-bin matrix was generated
by segmenting the genome into 5-kb windows and scoring each cell
for reads in each window. The high-quality cells were kept with the
log10(UMI) value between 3 and 5 and the fraction of reads in promoters
between 10% and 65% (Extended Data Fig. 4a). Finally, 19,955 cells passed
quality control for further analysis. We used the SnapATAC (https://
github.com/r3fang/SnapATAC) method to reduce the dimensionality
of the dataset and the deep neural network-based scAlign81 to remove
the batch effect of two replicates. We then identified 25 clusters using
the graph-based clustering. To identify cell-type-specific regulatory
elements, we called peaks on aggregated single cells from each cluster
using MACS2. Finally, we identified a total of 268,268 non-redundant
peaks, of which 195,004 peaks were also identified as peaks in bulk
ATAC-seq data. A total of 171,159 cluster-specific differential peaks (DA
peaks) were identified by Fisher’s exact test, and the threshold was set
as FDR <0.05. To determine the cell type of each cluster, we performed
motif analysis based on known transcription factor binding motifs in
DA peak regions of each cluster using HOMER. We used ChromVAR82
to estimate bias-corrected deviations of transcription factor binding
motif enrichment and the result was consistent with the top motif
enrichments for DA peaks of each cluster.
Statistics
All box plots in main and extended data figures were plotted using R
and Python. In box plots, the horizontal line shows the median, the
box encompasses the interquartile range, and whiskers extend to 5th
and 95th percentiles.
Reporting summary
Further information on research design is available in the Nature
Research Reporting Summary linked to this paper.
Data availability
All the sequencing data are deposited in the NCBI Gene Expression
Omnibus under accession code GSE134055. All the genomic data
generated in this study can be visualized in the WashU Epigenome
Browser (https://epigenome.wustl.edu/zebrafishENCODE/). The
human histone-modification ChIP-seq data were downloaded from
the ROADMAP Project. The mouse histone modification ChIP-seq data
were downloaded from the mouse ENCODE Consortium. The human tis-
sue transcriptome data were downloaded from the GTEx Consortium.
The public zebrafish ChIP-seq and ATAC-seq data used in this study are
listed in Supplementary Table 6. The human h1-ESC Hi-C data were
downloaded from GSE52457. GM12878 and K562 GRO-seq data
were downloaded from GSE60456. GM12878 and K562 CTCF ChIP-seq
were downloaded from GSE31477. GM12878 and K562 Pol2 ChIP-seq were
downloaded from GSE91426 and GSE31477. Source data are provided
with this paper.
42. Pedroso, G. L. et al. Blood collection for biochemical analysis in adult zebrafish. J. Vis.
Exp. 3865, e3865 (2012).
43. Rao, S. S. et al. A 3D map of the human genome at kilobase resolution reveals principles
of chromatin looping. Cell 159, 1665–1680 (2014).
44. Xie, W. et al. Epigenomic analysis of multilineage differentiation of human embryonic
stem cells. Cell 153, 1134–1148 (2013).
45. Kim, D. et al. TopHat2: accurate alignment of transcriptomes in the presence of insertions,
deletions and gene fusions. Genome Biol. 14, R36 (2013).
46. Trapnell, C. et al. Transcript assembly and quantification by RNA-Seq reveals
unannotated transcripts and isoform switching during cell differentiation. Nat.
Biotechnol. 28, 511–515 (2010).
47. Wang, L. et al. CPAT: Coding-Potential Assessment Tool using an alignment-free logistic
regression model. Nucleic Acids Res. 41, e74 (2013).
48. Dobin, A. et al. STAR: ultrafast universal RNA-seq aligner. Bioinformatics 29, 15–21 (2013).
49. Li, B. & Dewey, C. N. RSEM: accurate transcript quantification from RNA-Seq data with or
without a reference genome. BMC Bioinformatics 12, 323 (2011).
50. Maertin, M. Cutadapt removes adapter sequences from high-throughput sequencing
reads. EMBnet J. 17, 10–12 (2011).
51. Langmead, B. & Salzberg, S. L. Fast gapped-read alignment with Bowtie 2. Nat. Methods
9, 357–359 (2012).
52. Li, Z. et al. Identification of transcription factor binding sites using ATAC-seq. Genome
Biol. 20, 45 (2019).
53. Korhonen, J., Martinmäki, P., Pizzi, C., Rastas, P. & Ukkonen, E. MOODS: fast search for
position weight matrix matches in DNA sequences. Bioinformatics 25, 3181–3182
(2009).
54. Kulakovskiy, I. V. et al. HOCOMOCO: towards a complete collection of transcription
factor binding models for human and mouse via large-scale ChIP-seq analysis. Nucleic
Acids Res. 46 (D1), D252–D259 (2018).
55. Gerstein, M. B. et al. Integrative analysis of the Caenorhabditis elegans genome by the
modENCODE project. Science 330, 1775–1787 (2010).
56. Li, H. et al. The Sequence Alignment/Map format and SAMtools. Bioinformatics 25,
2078–2079 (2009).
57. Liu, T. Use model-based Analysis of ChIP-Seq (MACS) to analyze short reads generated by
sequencing protein-DNA interactions in embryonic stem cells. Methods Mol. Biol. 1150,
81–95 (2014).
58. Hiller, M. et al. Computational methods to detect conserved non-genic elements in
phylogenetically isolated genomes: application to zebrafish. Nucleic Acids Res. 41, e151
(2013).
59. Lee, H. J. et al. Regenerating zebrafish fin epigenome is characterized by stable
lineage-specific DNA methylation and dynamic chromatin accessibility. Genome Biol. 21,
52 (2020).
60. Krueger, F. & Andrews, S. R. Bismark: a flexible aligner and methylation caller for
Bisulfite-Seq applications. Bioinformatics 27, 1571–1572 (2011).
61. Zhou, X., Li, D., Lowdon, R. F., Costello, J. F. & Wang, T. methylC Track: visual integration of
single-base resolution DNA methylation data on the WashU EpiGenome Browser.
Bioinformatics 30, 2206–2207 (2014).
62. Burger, L., Gaidatzis, D., Schübeler, D. & Stadler, M. B. Identification of active regulatory
regions from DNA methylation data. Nucleic Acids Res. 41, e155 (2013).
63. Wu, H. et al. Detection of differentially methylated regions from whole-genome bisulfite
sequencing data without replicates. Nucleic Acids Res. 43, e141 (2015).
64. Hansen, K. D., Langmead, B. & Irizarry, R. A. BSmooth: from whole genome bisulfite
sequencing reads to differentially methylated regions. Genome Biol. 13, R83 (2012).
65. Ramírez, F. et al. deepTools2: a next generation web server for deep-sequencing data
analysis. Nucleic Acids Res. 44 (W1), W160–W165 (2016).
66. Koren, S. et al. Canu: scalable and accurate long-read assembly via adaptive k-mer
weighting and repeat separation. Genome Res. 27, 722–736 (2017).
67. Walker, B. J. et al. Pilon: an integrated tool for comprehensive microbial variant detection
and genome assembly improvement. PLoS ONE 9, e112963 (2014).
68. Dudchenko, O. et al. De novo assembly of the Aedes aegypti genome using Hi-C yields
chromosome-length scaffolds. Science 356, 92–95 (2017).
69. Kundaje, A. et al. Integrative analysis of 111 reference human epigenomes. Nature 518,
317–330 (2015).
70. Heinz, S. et al. Simple combinations of lineage-determining transcription factors prime
cis-regulatory elements required for macrophage and B cell identities. Mol. Cell 38,
576–589 (2010).
71. Grant, C. E., Bailey, T. L. & Noble, W. S. FIMO: scanning for occurrences of a given motif.
Bioinformatics 27, 1017–1018 (2011).
72. Servant, N. et al. HiC-Pro: an optimized and flexible pipeline for Hi-C data processing.
Genome Biol. 16, 259 (2015).
73. Durand, N. C. et al. Juicer provides a one-click system for analyzing loop-resolution Hi-C
experiments. Cell Syst. 3, 95–98 (2016).
74. Robinson, J. T. et al. Juicebox. js provides a cloud-based visualization system for Hi-C
data. Cell Syst. 6, 256–258 (2018).
75. Abdennur, N. & Mirny, L. A. Cooler: scalable storage for Hi-C data and other genomically
labeled arrays. Bioinformatics 36, 311–316 (2020).
76. Crane, E. et al. Condensin-driven remodelling of X chromosome topology during dosage
compensation. Nature 523, 240–244 (2015).
77. Giorgetti, L. et al. Structural organization of the inactive X chromosome in the mouse.
Nature 535, 575–579 (2016).
78. Imakaev, M. et al. Iterative correction of Hi-C data reveals hallmarks of chromosome
organization. Nat. Methods 9, 999–1003 (2012).
79. Darling, A. E., Mau, B. & Perna, N. T. progressiveMauve: multiple genome alignment with
gene gain, loss and rearrangement. PLoS ONE 5, e11147 (2010).
80. Li, H. & Durbin, R. Fast and accurate short read alignment with Burrows–Wheeler
transform. Bioinformatics 25, 1754–1760 (2009).
81. Johansen, N. & Quon, G. scAlign: a tool for alignment, integration, and rare cell
identification from scRNA-seq data. Genome Biol. 20, 166 (2019).
82. Schep, A. N., Wu, B., Buenrostro, J. D. & Greenleaf, W. J. chromVAR: inferring
transcription-factor-associated accessibility from single-cell epigenomic data.
Nat. Methods 14, 975–978 (2017).
Acknowledgements This work was supported by NIH grants R35GM124820, R01HG009906,
R24DK106766 (R.C.H. and F.Y.) and R01DK107735 (G.S.G.). F.Y. is also supported by
U01CA200060. T.W. is supported by NIH grants R01HG007175, R01HG007354, R01ES024992,
U24ES026699 and U01HG009391. We thank J. A. Stamatoyannopoulos for discussion and
suggestions; H. Lyu for proof reading and other Yue lab members for discussion; and E.
DeForest, S. Stella, P. Hubley and Penn State Zebrafish Functional Genomics Core for fish
husbandry and embryo collection.
Author contributions F.Y. conceived and supervised the project. H.Y. and T.L. collected tissue
and conducted experiments. Y.L. led the data analysis. Y.L., H.Y., H.J.L., Y.W., X.W., B.Z., L.F. and
J.W. conducted analyses. D.L. and T.W. provided the website for data presentation. K.C.A. and
K.C.C. provided animal support. Q.J., X.X., J.X., F.S., I.S., C. K., T.S., M.N.K.C., J.T., K.W., G.S.G.,
R.C.H., T.W. and K.C.C. helped with data interpretation. H.Y., Y.L., T.L. and F.Y. prepared the
manuscript with input from all authors.
Competing interests The authors declare no competing interests.
Additional information
Supplementary information is available for this paper at https://doi.org/10.1038/s41586-020-
2962-9.
Correspondence and requests for materials should be addressed to F.Y.
Peer review information Nature thanks Michael Beer, Jesse Dixon and the other, anonymous,
reviewer(s) for their contribution to the peer review of this work.
Reprints and permissions information is available at http://www.nature.com/reprints.
Article
Extended Data Fig. 1 | Tissue-specific gene expression in zebrafish.
a, Clustering analysis of transcripts from RNA-seq data in embryonic and adult
tissues (n = 31,842). b, c, Gene Ontology and KEGG pathway analysis for the
tissue-specific genes in adult brain, heart and testis (the number of
tissue-specific genes in these two figures are, Brain = 3,693, Heart = 392,
Testis = 1,605). d, Distribution of H3K4me3 signals surrounding the known and
predicted novel transcripts. e, Human orthologues of zebrafish tissue-specific
genes were more tissue-specific compared to human orthologues of
non-tissue-specific zebrafish genes (n = 14,764, 3,739, 6,043, Mann–Whitney U
Test, two-sided, ***P < 2.2 × 10 −16).
Extended Data Fig. 2 | Comparative analysis of zebrafish cis-regulatory
elements. a, Comparison of the predicted regulatory elements identified with
previous data. Enhancers were based on H3K27ac signals in the same four
tissues (brain, heart, intestine, testis) from Perez-Rico et al. 2017. The data we
generated are from Tübingen zebrafish strain and the published results were
from the AB strain. b, Number of predicted cis-regulatory elements in each
tissue. E-brain stands for 1 dpf embryonic neuron cells. E-trunk stands for 1 dpf
zebrafish whole trunk region. c, An example showing genes with active
promoters have higher expression level. Blue hollow bar indicates the known
mrpl39 promoter. Orange hollow bar indicates the potential novel promoter.
The mrpl39 promoter has H3K4me3 peaks in both muscle and brain, but only
has strong H3K27ac signals in muscle and its expression is higher (4.43-fold).
d, Gene Ontology results for the muscle-specific enhancers and skin-specific
enhancers. We used the GREAT tool for this analysis (the numbers of
tissue-specific enhancers used in this figure are muscle = 813, skin = 512).
Article
Extended Data Fig. 3 | Enhancer reporter assay for tissue-specific
enhancers. In total, 28 of 32 predicted tissue-specific enhancers showed
consistent GFP signals in the corresponding tissues. For the eight brain
enhancers tested, 63/95, 51/86, 85/119, 112/143, 27/45, 34/48, 27/41, 62/77, and
37/45 embryos, respectively, had green signals in the brain region. For the six
tested heart enhancers, 64/94, 52/85, 79/121, 20/41, 51/95, 32/55 and 20/31
embryos, respectively, had green signals in the heart region. For the six tested
muscle enhancers, 52/57, 26/30, 107/124, 53/63, 93/114, 61/67 and 66/78
embryos, respectively, had green signals in the trunk muscle. For the four
selected kidney enhancers, 47/82, 35/67, 44/62, 15/42 and 56/110 embryos,
respectively, had green signals in the kidney region.
Extended Data Fig. 4 | Single-cell ATAC-seq in zebrafish brain. a, Barcode
selection of single cell ATAC-seq. The x-axis represents the log value of the
number of unique molecular identifiers (UMI); the y axis represents the ratio of
fragments in promoter regions; the red lines represent threshold, and the grey
shadows represent that the barcode passed the filter. b, Genomic distribution
of all differentially accessible (DA) peaks. c, Overlap of all differentially
accessible peaks with enhancers predicted in bulk brain. d, Top, the cluster
distribution in the tSNE projection. Bottom left, pileups of differentially
accessible ATAC-seq signals for each cluster. Shown in the figure is the +/− 10kb
flanking region surrounding peak centres. Bottom right, most significantly
enriched transcription factor motif for each cluster. e, t-SNE projection of all
scATAC-seq cells colored by Z-score of peak enrichment. f, Motif enrichment of
known neuron-specific TFs in scATAC-seq predicted clusters (n = 19,955).
Article
Extended Data Fig. 5 | Heterochromatin annotation in adult tissues.
a, WashU Epigenome Browser screenshot of H3K9me3 and H3K9me2 histone
ChIP-seq signals in 11 zebrafish adult tissues. The values on the y-axis were
input-normalized. b, Distribution of H3K9me3 and H3K9me2 sites in the
zebrafish genome. c, Venn diagram shows the overlap between H3K9me3 and
H3K9me2 sites in zebrafish genome. d, Overlapping percentile of H3K9me3
and H3K9me2 peaks in adult tissues. e, H3K9me3 and H3K9me2 sites were
depleted of ATAC-seq, H3K4me3 and H3K27ac ChIP-seq signals (n = 68,789
H3K9me3 sites and n = 73,777 H3K9me2 sites). f, Overlap of H3K9me3 sites,
H3K9me2 sites, and ATAC-seq peaks with repetitive elements (The total
number of each bar, from left to right, 68,789, 73,777 and 436,036). g, Examples
of H3K9me3 sites in one tissue found to be active regions in other tissues.
Horizontal scale 0-20 for H3K27ac and H3K4me3, 0-10 for RNA-seq, 0-5 for
H3K9me3 and H3K9me2.
Extended Data Fig. 6 | DNA methylation level and distribution in adult
tissues. a, Fraction of total CpGs with low (<25%), medium (≥25% and <75%), and
high (≥75%) methylation levels and mean CpG methylation levels (mCG/CG) in
zebrafish adult tissues (the mCG/CG ratio, from left to right, 0.788, 0.859,
0.790, 0.777, 0.791, 0.797, 0.781, 0.777, 0.804, 0.789, 0.781). b, Distribution of
CpG methylation levels across zebrafish adult tissues. c, The distribution of
non CpG methylation in 11 adult tissues. d, Mean methylation levels of the
tissue-specific gene promoters. n represents the number of tissue-specific
gene promoter. e, Mean methylation level of CpGs overlapping different
genomic features or repetitive element classes. CDS, coding sequence.
f, Number of UMRs and LMRs in zebrafish tissues and their overlap with
enhancer and promoters (left panel) (number of UMR and LMR, from top to
bottom, 14,990, 10,569, 14,569, 14,587, 14,831, 14,289, 13,842, 13,569, 14,424,
14,374, 13,908, 30,009, 7,916, 19,038, 21,411, 22,591, 16,796, 14,961, 16,268,
17,481, 15,932, 15,665) and ATAC-seq peaks (right panel)(numbers of UMR and
LMR are the same with left panel). g, Clustering of tissue-specific hypoDMRs.
Values in the heat map are mean methylation levels of hypoDMRs (n = 17,654,
number of tissue-specific hypoDMRs).
Article
Extended Data Fig. 7 | De novo assembly of zebrafish chromosome 4 of the
Tübingen strain. a, WashU Epigenome Browser snapshot showing that
heterochromatic marks H3K9me2 and H3K9me3 signals were enriched on
chromosome 4 in zebrafish testis. The values on the y-axis were input-normalized.
b, H3K9me2, H3K9me3, and DNA methylation level on chr4 long arm are
significantly higher than other regions in all tissues (n = 11, two-sided, t-test).
c, Overall strategy of de novo assembly of the Tübingen chr4 by integrating
10X, Nanopore, Bionano, and Hi-C data. d, Bionano long molecule sequencing
data shows that there were many SVs on chr4 when mapped to the GRCz11
reference genome. e, SVs on chr4 detected by Bionano when the data were
mapped to the de novo assembled chr4.
Extended Data Fig. 8 | Conservation of cis-regulatory elements from
zebrafish to other vertebrates. a, Percentage of zebrafish enhancers whose
sequences were conserved in human (the number of each bar, from left to right,
13,307, 7,018, 11,940, 7,499, 14,783, 14,272, 8,995, 13,777, 10,757, 15,505, 1,734,
4,011, 5,247). b, c, Similar to Fig. 4a. Percentage of zebrafish exons and
cis-regulatory elements that have orthologous sequences in mouse and other
fish species. Total number of each bar, from left to right: 1,000, 25,593, 58,065,
1,000. For exons and random, we randomly sample 1000 elements and
computed their conservation percentage. The simulations were performed
20 times and the average percentage was presented. d, Another example of
ultra-conserved noncoding element (UCNE). This element (FOXP1_Finn_1) is
predicted to be a muscle enhancer in zebrafish, mouse, and human. Grey
vertical bar marks the ultra-conserved region. Red vertical bar is the enhancer
sequence in the human genome that was validated as a limb enhancer by
transgenic mouse reporter assay in the VISTA Enhancer Browser (#hs956).
Article
Extended Data Fig. 9 | Distal ATAC-seq peak-to-gene pairs, enhancer-
to-gene pairs, and transcriptional regulation network. a, b, Distance
distribution of cis-regulatory elements to their linked gene TSS. c, Correlation
of ATAC-seq peak-to-gene pairs and Enhancer-to-gene pairs (n from left to
right = 3,292, 3,827, 3,544, 3,281, 3,008, 2,795, 2,357, 2,001, 1,106). d, Validation
of predicted enhancer-to-gene pairs by Hi-C interaction counts in muscle.
e, mef2d is a regulator in both zebrafish muscle and heart, but it regulates
different downstream targets by motif prediction analysis. f, The overall
structure of the regulatory network is conserved between human and
zebrafish. FFL connection analysis was performed, in this analysis, there are
three types of nodes: A, driver node that regulates B and C; B, middle node,
regulated by A but regulating node C; C, passenger node, regulated by both A
and B.
Extended Data Fig. 10 | See next page for caption.
Article
Extended Data Fig. 10 | Compartment and TADs in zebrafish. a, Heat map of
genome-wide Hi-C interaction matrices in zebrafish brain (blue) and muscle
(red). b, Active marks (H3K4me3, H3K27ac, and ATAC-seq) were enriched in
compartment A and depleted in compartment B. Repressive marks (H3K9me2
and H3K9me3) were enriched in compartment B. Error bands represent
standard error of the mean. c, Genome browser snapshot of A/B compartment
in brain and muscle. The blue vertical shaded area marks a region that is located
in compartment B in brain but in compartment A in muscle. As expected, A
compartment which is associated with more ATAC-seq peaks, H3K27ac and
RNA-seq signals. d, Examples of shared TADs between zebrafish brain and
muscle. e, Average DI scores surrounding TAD boundaries identified in brain
(upper panel) and muscle (lower panel). f, ChIP-seq data shows that CTCF
binding sites were enriched at TAD boundaries. g, Footprint analysis of ATAC-
seq peaks in the TAD boundaries shows enrichment of CTCF binding motif
(number of each bar, from left to right, 0.213, 0.24, 0.22, 0.237, 0.251, 0.232,
0.24, 0.262, 0.271, 0.281, 0.37, 0.27, 0.253, 0.25, 0.252, 0.253, 0.26, 0.23, 0.238,
0.24, 0.22). h, Repetitive elements enriched at TAD boundaries (left panel) and
loop anchors (right panel).
Extended Data Fig. 11 | Comparing zebrafish evolutionary breakpoints
with TAD annotation. a. Similar to Fig. 5d. Enrichment of evolutionary
breakpoints at TAD boundaries. Relative positions of evolutionary breakpoints
to TADs in 15 vertebrates. In all cases, we found that the evolutionary
breakpoints were enriched at zebrafish TAD boundaries and depleted from the
centre of TADs. Grey vertical bar labels the TAD body area. b, By comparing
zebrafish with 17 vertebrates, H3K4me3 signals were found to be more
enriched at TAD boundaries with breakpoints than those without breakpoints.
Orange vertical bar labels the TAD boundaries. c, Higher H3K4me3 levels at
breakpoint-containing TAD boundaries when using TADs annotation from
zebrafish muscle were found as well, similar to Fig. 5g. d, H3K4me3 enrichment
in human ESCs (H1) TAD boundaries with or without zebrafish-to-human
breakpoints. e, H3K4me3 enrichment in mouse ESCs TAD boundaries with or
without zebrafish-to-mouse breakpoints. f, H3K4me3 enrichment in human
ESCs (H1) TAD boundaries with or without mouse-to-human breakpoints.
Article
Extended Data Fig. 12 | TADs with and without breakpoints. a, H3K27ac and
ATAC-seq signals do not show differences at TAD boundaries with breakpoints
compared to those without breakpoints. Orange vertical bar labels the TAD
boundaries. b, Sizes of TADs with and without evolutionary breakpoints were
similar (n = 573, 777, two-sided, t-test). c, Enrichment of transcription at
breakpoints (BP) that overlap with CTCF TAD boundaries in K562 cells
(the number of breakpoints in blue line is 639, red line is 625). d, In 17
vertebrates, TADs without evolutionary breakpoints (bottom panel) have
stronger interaction frequencies in the middle than TADs with evolutionary
breakpoints (upper panel). Breakpoints in these 17 vertebrates were defined by
comparing their genomes to the zebrafish genome. e, Distribution of
correlations between the expression pattern of each pair of paralogs across 11
adult zebrafish tissues. f, Correlations between pairs of paralogs located on the
same chromosome. Among them, 17 pairs were located within the same TAD,
and the rest of the 65 pairs were located in different TADs. As a control, we
randomly sampled 100 genes. Number of each bar, from left to right, 17, 65, 100.
 nature research | reporting summary
Corresponding author(s): Yue
Last updated by author(s): 07/05/2020

Reporting Summary
Nature Research wishes to improve the reproducibility of the work that we publish. This form provides structure for consistency and transparency
in reporting. For further information on Nature Research policies, see our Editorial Policies and the Editorial Policy Checklist.

Statistics
For all statistical analyses, confirm that the following items are present in the figure legend, table legend, main text, or Methods section.
n/a Confirmed
The exact sample size (n) for each experimental group/condition, given as a discrete number and unit of measurement
A statement on whether measurements were taken from distinct samples or whether the same sample was measured repeatedly
The statistical test(s) used AND whether they are one- or two-sided
Only common tests should be described solely by name; describe more complex techniques in the Methods section.

A description of all covariates tested

A description of any assumptions or corrections, such as tests of normality and adjustment for multiple comparisons
A full description of the statistical parameters including central tendency (e.g. means) or other basic estimates (e.g. regression coefficient)
AND variation (e.g. standard deviation) or associated estimates of uncertainty (e.g. confidence intervals)

For null hypothesis testing, the test statistic (e.g. F, t, r) with confidence intervals, effect sizes, degrees of freedom and P value noted
Give P values as exact values whenever suitable.

For Bayesian analysis, information on the choice of priors and Markov chain Monte Carlo settings
For hierarchical and complex designs, identification of the appropriate level for tests and full reporting of outcomes
Estimates of effect sizes (e.g. Cohen's d, Pearson's r), indicating how they were calculated
Our web collection on statistics for biologists contains articles on many of the points above.

Software and code

Policy information about availability of computer code
Data collection Images of reporter assays in zebrafish embryos were collected using Zeiss SteREO Discovery.V8 microscope by AxioVison Rel.4.8 software. All
the list of public zebrafish ChIP-seq and ATAC-seq used in this study were listed in the supplementary table 6.

Data analysis The ENCODE histone ChIP-seq pipeline(https://github.com/ENCODE-DCC/chip-seq-pipeline2)

The ENCODE ATAC-seq pipeline(https://github.com/ENCODE-DCC/atac-seq-pipeline)
RepeatMasker 4.1.0
bowtie 1.2.2
Trim Galore 0.6.0 (https://github.com/FelixKrueger/TrimGalore).
macs2 2.2.4
bedtools v2.29.0
Bowtie2 2.3.4
Samtools 1.7
STAR 2.6.0
RSEM 1.3.0
April 2020

Deeptools 3.3.1
GREAT 3.0.0
Tophat2 2.1.0
Cufflinks 2.2.1
CPAT v1.2.3
TransDecoder V5.1.0 (https://transdecoder.github.io/)
blast/2.7.1
biomaRt 2.36.0
Cutadapt 2.5

1
 picard 1.126 (http://broadinstitute.github.io/picard/).
ChIPseeker 1.19.1

nature research | reporting summary

Bismark 0.18.1
MethylseekR 1.22.0
DSS v2.14.0
Homer 4.10
FIMO 5.0.5
Hi-C pro 2.11.1 (https://github.com/nservant/HiC-Pro)
cooler 0.8.6 (https://github.com/mirnylab/cooler)
cooltools v.0.4.0 (https://github.com/mirnylab/cooltools)
HiCrep 0.4.0
progressiveMauve 2.4.0
BWA 0.7.17
cell ranger 3.1.0
snapATAC 1.0.0 (https://github.com/r3fang/SnapATAC)
scAlign 1.0
ChromVAR 1.6.0
HINT v0.13.0
Juicer 1.8.9 (https://github.com/aidenlab/juicer/wiki/Download)
Juicebox 1.11.08
Bionano_Access1.5.1
Bionano_Solve Solve3.5.1_01142020
3d-dna 180419
lastz version 1.04.00
Canu Version 2.0 (https://github.com/marbl/canu)
Pilon Version 1.23
limma 3.45.9
MOODS 1.9.4 49(https://github.com/jhkorhonen/MOODS)
minimap2 2.16-r922
Nanopolish 0.11.1
BWA-MEM 0.7.15-r1140
Picard 2.17.2

For manuscripts utilizing custom algorithms or software that are central to the research but not yet described in published literature, software must be made available to editors and
reviewers. We strongly encourage code deposition in a community repository (e.g. GitHub). See the Nature Research guidelines for submitting code & software for further information.

Data
Policy information about availability of data
All manuscripts must include a data availability statement. This statement should provide the following information, where applicable:
- Accession codes, unique identifiers, or web links for publicly available datasets
- A list of figures that have associated raw data
- A description of any restrictions on data availability

Next generation sequencing data have been deposited in Gene Expression Omnibus (GEO) under the following accession numbers ：GSE134055
The genome browser link: https://epigenome.wustl.edu/zebrafishENCODE/
The human h1-ESC Hi-C data were downloaded from GSE52457.
GM12878 and K562 GRO-seq data were downloaded from GSE60456.
GM12878 and K562 CTCF ChIP-seq were downloaded from GSE31477.
GM12878 and K562 Pol2 ChIP-seq were downloaded from GSE91426 and GSE31477.

Field-specific reporting
Please select the one below that is the best fit for your research. If you are not sure, read the appropriate sections before making your selection.
Life sciences Behavioural & social sciences Ecological, evolutionary & environmental sciences
For a reference copy of the document with all sections, see nature.com/documents/nr-reporting-summary-flat.pdf

Life sciences study design

All studies must disclose on these points even when the disclosure is negative.
April 2020

Sample size The number of adult zebrafish used for each tissue were determined in order to obtain enough tissue for 4 ChIP-seq, RNA-seq, ATAC-seq,
WGBS and Hi-C experiments. In all, 160 datasets including 11 WGBS, 26 RNA-seq, 95 ChIP-seq, 22 ATAC-seq, 4 Hi-C and 2 single-cell ATAC-seq
were used in this study.

Data exclusions This is not relevant since we used all sequencing data in this study

Replication We have two replicates for each ChIP-seq (only one replicate for Kidney H3K27ac) , RNA-seq, ATAC-seq, Hi-C and single-cell ATAC-seq. We
performed two technical replicates for Whole Genome Bisulfite Sequencing (WGBS) and reached 30X coverage. We calculated the Pearson

2
 correlation coefficient between two biological replicates using the reads counts of 10 kb-binned matrices. The correlation score of all
replicates were listed in supplemental Table 1.

nature research | reporting summary

Randomization This is not relevant since we did not use different experimental groups or conditions in our study.

Blinding Blinding was not relevant to our study since we did not have experimental groups to compare.

Reporting for specific materials, systems and methods

We require information from authors about some types of materials, experimental systems and methods used in many studies. Here, indicate whether each material,
system or method listed is relevant to your study. If you are not sure if a list item applies to your research, read the appropriate section before selecting a response.

Materials & experimental systems Methods

n/a Involved in the study n/a Involved in the study
Antibodies ChIP-seq
Eukaryotic cell lines Flow cytometry
Palaeontology and archaeology MRI-based neuroimaging
Animals and other organisms
Human research participants
Clinical data
Dual use research of concern

Antibodies
Antibodies used Rabbit polyclonal histone H3K27ac antibody (Active Motif, 39133) 1:100
Rabbit polyclonal histone H3K4me3 antibody (EMD Millipore, 07-473) 1:100
Rabbit monoclonal histone H3K9me2 antibody (Cell Signaling, 4658) 1:100
Rabbit polyclonal histone H3K9me3 antibody (Abcam, ab8898) 1:100

Validation The four primary antibodies are commercial antibodies against Histone H3 modifications, validated as ChIP
grade by the manufacturer (active motif, milipore, cellsignal and Abcam):
https://www.activemotif.com/catalog/details/39133
https://www.emdmillipore.com/US/en/product/Anti-trimethyl-Histone-H3-Lys4-Antibody,MM_NF-07-473
https://www.cellsignal.com/products/primary-antibodies/di-methyl-histone-h3-lys9-d85b4-xp-rabbit-mab/4658
https://www.abcam.com/histone-h3-tri-methyl-k9-antibody-chip-grade-ab8898.html
Furthermore, these antibodies have been validated by various labs for ChIP-seq with zebrafish and also many other species of which
have the exact same amino acid sequence (PMIDs: 31794598, 24286030, 31495082, 30948728)

Animals and other organisms

Policy information about studies involving animals; ARRIVE guidelines recommended for reporting animal research
Laboratory animals All adult Tissues were collected from one-year old fish raised of Tuebingen strains ordered from ZIRC. The Gender information is
provided in supplementary table 17. Tg(huc:kaede) fish was used to cross with Tuebingen fish to obtain kaede labeled neuron cells.

Wild animals The study did not involve animals in the wild.

Field-collected samples The study did not involve animals collected from field.

Ethics oversight All procedures on live animals were approved by the Institutional Animal Care and Use Committee (IACUC) at the Pennsylvania State
University, ID: PRAMS201445659
Note that full information on the approval of the study protocol must also be provided in the manuscript.

ChIP-seq
Data deposition
April 2020

Confirm that both raw and final processed data have been deposited in a public database such as GEO.
Confirm that you have deposited or provided access to graph files (e.g. BED files) for the called peaks.

Data access links GEO_number: GSE134055

May remain private before publication.

Files in database submission GSE134055_YueLab-ChIP-seq-Brain_H3K27ac_merged.narrowPeak.gz

GSE134055_YueLab-ChIP-seq-Brain_H3K4me3_merged.narrowPeak.gz

3
 GSE134055_YueLab-ChIP-seq-Colon_H3K27ac_merged.narrowPeak.gz
GSE134055_YueLab-ChIP-seq-Colon_H3K4me3_merged.narrowPeak.gz

nature research | reporting summary

GSE134055_YueLab-ChIP-seq-Heart_H3K27ac_merged.narrowPeak.gz
GSE134055_YueLab-ChIP-seq-Heart_H3K4me3_merged.narrowPeak.gz
GSE134055_YueLab-ChIP-seq-Intestine_H3K27ac_merged.narrowPeak.gz
GSE134055_YueLab-ChIP-seq-Intestine_H3K4me3_merged.narrowPeak.gz
GSE134055_YueLab-ChIP-seq-Kidney_H3K27ac_merged.narrowPeak.gz
GSE134055_YueLab-ChIP-seq-Kidney_H3K4me3_merged.narrowPeak.gz
GSE134055_YueLab-ChIP-seq-Liver_H3K27ac_merged.narrowPeak.gz
GSE134055_YueLab-ChIP-seq-Liver_H3K4me3_merged.narrowPeak.gz
GSE134055_YueLab-ChIP-seq-Muscle_H3K27ac_merged.narrowPeak.gz
GSE134055_YueLab-ChIP-seq-Muscle_H3K4me3_merged.narrowPeak.gz
GSE134055_YueLab-ChIP-seq-Skin_H3K27ac_merged.narrowPeak.gz
GSE134055_YueLab-ChIP-seq-Skin_H3K4me3_merged.narrowPeak.gz
GSE134055_YueLab-ChIP-seq-Spleen_H3K27ac_merged.narrowPeak.gz
GSE134055_YueLab-ChIP-seq-Spleen_H3K4me3_merged.narrowPeak.gz
GSE134055_YueLab-ChIP-seq-Testis_H3K27ac_merged.narrowPeak.gz
GSE134055_YueLab-ChIP-seq-Testis_H3K4me3_merged.narrowPeak.gz
GSE134055_YueLab-ChIP-seq-embryonic-Brain_H3K27ac_merged.narrowPeak.gz
GSE134055_YueLab-ChIP-seq-embryonic-Brain_H3K4me3_merged.narrowPeak.gz
GSE134055_YueLab-ChIP-seq-embryonic-Trunk_H3K27ac_merged.narrowPeak.gz
GSE134055_YueLab-ChIP-seq-embryonic-Trunk_H3K4me3_merged.narrowPeak.gz
GSE134055_YueLab-ChIP-seq-embryonic-Blood_H3K27ac_merged.narrowPeak.gz
GSE134055_YueLab-ChIP-seq-embryonic-Blood_H3K4me3_merged.narrowPeak.gz
GSE134055_YueLab-ATAC-seq-Brain.narrowPeak.gz
GSE134055_YueLab-ATAC-seq-Liver.narrowPeak.gz
GSE134055_YueLab-ATAC-seq-Muscle.narrowPeak.gz
GSE134055_YueLab-ATAC-seq-Skin.narrowPeak.gz
GSE134055_YueLab-ATAC-seq-Spleen.narrowPeak.gz
GSE134055_YueLab-ChIP-seq-Brain_H3K9me3.peak.gz
GSE134055_YueLab-ChIP-seq-Brain_H3K9me2.peak.gz
GSE134055_YueLab-ChIP-seq-Colon_H3K9me3.peak.gz
GSE134055_YueLab-ChIP-seq-Colon_H3K9me2.peak.gz
GSE134055_YueLab-ChIP-seq-Heart_H3K9me3.peak.gz
GSE134055_YueLab-ChIP-seq-Heart_H3K9me2.peak.gz
GSE134055_YueLab-ChIP-seq-Intestine_H3K9me3.peak.gz
GSE134055_YueLab-ChIP-seq-Intestine_H3K9me2.peak.gz
GSE134055_YueLab-ChIP-seq-Kidney_H3K9me3.peak.gz
GSE134055_YueLab-ChIP-seq-Kidney_H3K9me2.peak.gz
GSE134055_YueLab-ChIP-seq-Liver_H3K9me3.peak.gz
GSE134055_YueLab-ChIP-seq-Liver_H3K9me2.peak.gz
GSE134055_YueLab-ChIP-seq-Muscle_H3K9me3.peak.gz
GSE134055_YueLab-ChIP-seq-Muscle_H3K9me2.peak.gz
GSE134055_YueLab-ChIP-seq-Skin_H3K9me3.peak.gz
GSE134055_YueLab-ChIP-seq-Skin_H3K9me2.peak.gz
GSE134055_YueLab-ChIP-seq-Spleen_H3K9me3.peak.gz
GSE134055_YueLab-ChIP-seq-Spleen_H3K9me2.peak.gz
GSE134055_YueLab-ChIP-seq-Testis_H3K9me3.peak.gz
GSE134055_YueLab-ChIP-seq-Testis_H3K9me2.peak.gz
GSE134055_YueLab-ChIP-seq-embryonic-Brain_H3K9me3.peak.gz
GSE134055_YueLab-ChIP-seq-embryonic-Brain_H3K9me2.peak.gz
GSE134055_YueLab-ChIP-seq-embryonic-Trunk_H3K9me3.peak.gz
GSE134055_YueLab-ChIP-seq-embryonic-Trunk_H3K9me2.peak.gz
GSE134055_YueLab-ChIP-seq-embryonic-Blood_H3K9me3.peak.gz
GSE134055_YueLab-ChIP-seq-embryonic-Blood_H3K9me2.peak.gz

Genome browser session https://epigenome.wustl.edu/zebrafishENCODE/

(e.g. UCSC)

Methodology
Replicates Brain_RNA-seq 2 0.888
Blood_RNA-seq 2 946
Colon_RNA-seq 2 0.908
Heart_RNA-seq 2 0.849
Intestine_RNA-seq 2 0.853
E-brain_RNA-seq 2 0.911
April 2020

E-trunk_RNA-seq 2 0.875
Kidney_RNA-seq 2 0.91
Liver_RNA-seq 2 0.871
Muscle_RNA-seq 2 0.871
Skin_RNA-seq 2 0.916
Spleen_RNA-seq 2 0.852
Testis_RNA-seq 2 0.953
Brain_h3K27ac 2 0.97
Colon_h3K27ac 2 0.93

4
 Blood_h3K27ac 2 0.91
Heart_h3K27ac 2 0.97

nature research | reporting summary

Intestine_h3K27ac 2 0.81
E-brain_h3K27ac 2 0.97
E-trunk_h3K27ac 2 0.98
Kidney_h3K27ac 1
Liver_h3K27ac 2 0.97
Muscle_h3K27ac 2 0.91
Skin_h3K27ac 2 0.96
Spleen_h3K27ac 2 0.91
Testis_h3K27ac 2 0.93
Brain_h3k4me3 2 0.97
Colon_h3k4me3 2 0.93
Blood_h3k4me3 2 0.91
Heart_h3k4me3 2 0.97
Intestine_h3k4me3 2 0.94
E-brain_h3k4me3 2 0.97
E-trunk_h3k4me3 2 0.98
Kidney_h3k4me3 2 0.94
Liver_h3k4me3 2 0.82
Muscle_h3k4me3 2 0.98
Skin_h3k4me3 2 0.97
Spleen_h3k4me3 2 0.91
Testis_h3k4me3 2 0.88
Brain_H3K9me3 2 0.942
Colon_H3K9me3 2 0.91
Blood_H3K9me3 2 0.903
Heart_H3K9me3 2 0.926
Intestine_H3K9me3 2 0.91
Kidney_H3K9me3 2 0.937
Liver_H3K9me3 2 0.936
Muscle_H3K9me3 2 0.926
Skin_H3K9me3 2 0.913
Spleen_H3K9me3 2 0.916
Testis_H3K9me3 2 0.974
Brain_H3K9me2 2 0.916
Colon_H3K9me2 2 0.901
Blood_H3K9me2 2 0.893
Heart_H3K9me2 2 0.881
Intestine_H3K9me2 2 0.925
Kidney_H3K9me2 2 0.897
Liver_H3K9me2 2 0.899
Muscle_H3K9me2 2 0.89
Skin_H3K9me2 2 0.889
Spleen_H3K9me2 2 0.933
Testis_H3K9me2 2 0.922
Brain_ATAC 2 0.913
Colon_ATAC 2 0.908
Blood_ATAC 2 0.901
Heart_ATAC 2 0.925
Intestine_ATAC 2 0.865
Kidney_ATAC 2 0.914
Liver_ATAC 2 0.968
Muscle_ATAC 2 0.905
Skin_ATAC 2 0.941
Spleen_ATAC 2 0.911
Testis_ATAC 2 0.925
Brain_WGBS 1
Colon_WGBS 1
Blood_WGBS 1
Heart_WGBS 1
Intestine_WGBS 1
Kidney_WGBS 1
Liver_WGBS 1
Muscle_WGBS 1
Skin_WGBS 1
Spleen_WGBS 1
April 2020

Testis_WGBS 1
Brain_HiC 2 0.981
Muscle_HiC 2 0.976
Brain_scATAC-seq 0.925

Sequencing depth Brain_RNA-seq_rep1 57,399,962 paired-ed 60

Blood_RNA-seq_rep1 33,709,554 paired-ed 60
Colon_RNA-seq_rep1 64,407,490 paired-ed 60
Heart_RNA-seq_rep1 57,992,414 paired-ed 60

5
Intestine_RNA-seq_rep1 62,593,691 paired-ed 60
E-brain_RNA-seq_rep1 81,655,736 paired-ed 60

nature research | reporting summary

E-trunk_RNA-seq_rep1 81,968,040 paired-ed 60
Kidney_RNA-seq_rep1 53,192,824 paired-ed 60
Liver_RNA-seq_rep1 66,886,775 paired-ed 60
Muscle_RNA-seq_rep1 61,018,951 paired-ed 60
Skin_RNA-seq_rep1 43,248,349 paired-ed 60
Spleen_RNA-seq_rep1 19,289,511 paired-ed 98
Testis_RNA-seq_rep1 80,040,638 paired-ed 60
Brain_h3K27ac_rep1 22,919,396 paired-ed 150
Colon_h3K27ac_rep1 2,913,190 paired-ed 150
Blood_h3K27ac_rep1 12,855,702 paired-ed 150
Heart_h3K27ac_rep1 25,213,922 paired-ed 150
Intestine_h3K27ac_rep1 37,480,680 paired-ed 150
E-brain_h3K27ac_rep1 81,331,004 paired-ed 150
E-trunk_h3K27ac_rep1 61,361,014 paired-ed 150
Kidney_h3K27ac_rep1 31,757,872 paired-ed 150
Liver_h3K27ac_rep1 18,410,148 paired-ed 150
Muscle_h3K27ac_rep1 2,951,696 paired-ed 150
Skin_h3K27ac_rep1 54,680,744 paired-ed 150
Spleen_h3K27ac_rep1 7,119,136 paired-ed 150
Testis_h3K27ac_rep1 12,227,784 paired-ed 150
Brain_h3k4me3_rep1 20,554,470 paired-ed 150
Colon_h3k4me3_rep1 4,982,782 paired-ed 150
Blood_h3k4me3_rep1 14,199,752 paired-ed 150
Heart_h3k4me3_rep1 25,106,098 paired-ed 150
Intestine_h3k4me3_rep1 5,716,110 paired-ed 150
E-brain_h3k4me3_rep1 22,133,422 paired-ed 150
E-trunk_h3k4me3_rep1 50,430,196 paired-ed 150
Kidney_h3k4me3_rep1 3,922,484 paired-ed 150
Liver_h3k4me3_rep1 18,711,566 paired-ed 150
Muscle_h3k4me3_rep1 29,684,894 paired-ed 150
Skin_h3k4me3_rep1 58,024,592 paired-ed 150
Spleen_h3k4me3_rep1 18,813,056 paired-ed 150
Testis_h3k4me3_rep1 22,526,244 paired-ed 150
Brain_H3K9me3_rep1 13,902,969 paired-ed 150
Colon_H3K9me3_rep1 7,721,719 paired-ed 150
Blood_H3K9me3_rep1 10,840,250 paired-ed 150
Heart_H3K9me3_rep1 7,947,696 paired-ed 150
Intestine_H3K9me3_rep1 12,958,288 paired-ed 150
Kidney_H3K9me3_rep1 7,238,470 paired-ed 150
Liver_H3K9me3_rep1 12,298,856 paired-ed 150
Muscle_H3K9me3_rep1 14,540,724 paired-ed 150
Skin_H3K9me3_rep1 9,293,298 paired-ed 150
Spleen_H3K9me3_rep1 9,013,540 paired-ed 150
Testis_H3K9me3_rep1 9,339,132 paired-ed 150
Brain_H3K9me2_rep1 17,983,130 paired-ed 150
Colon_H3K9me2_rep1 9,611,265 paired-ed 150
Blood_H3K9me2_rep1 10,932,641 paired-ed 150
Heart_H3K9me2_rep1 10,006,810 paired-ed 150
Intestine_H3K9me2_rep1 13,028,365 paired-ed 150
Kidney_H3K9me2_rep1 7,917,567 paired-ed 150
Liver_H3K9me2_rep1 16,595,355 paired-ed 150
Muscle_H3K9me2_rep1 18,683,245 paired-ed 150
Skin_H3K9me2_rep1 10,006,417 paired-ed 150
Spleen_H3K9me2_rep1 9,392,285 paired-ed 150
Testis_H3K9me2_rep1 18,683,245 paired-ed 150
Brain_ATAC_rep1 11,256,714 paired-ed 150
Colon_ATAC_rep1 64,407,490 paired-ed 150
Blood_ATAC_rep1 44,349,981 paired-ed 150
Heart_ATAC_rep1 13,318,779 paired-ed 150
Intestine_ATAC_rep1 21,043,459 paired-ed 150
Kidney_ATAC_rep1 21,561,836 paired-ed 150
Liver_ATAC_rep1 13,617,294 paired-ed 150
Muscle_ATAC_rep1 20,315,786 paired-ed 150
Skin_ATAC_rep1 14,112,784 paired-ed 150
Spleen_ATAC_rep1 7,587,349 paired-ed 150
April 2020

Testis_ATAC_rep1 10,244,549 paired-ed 150

Brain_WGBS_rep1 446,611,815 paired-ed 150
Colon_WGBS_rep1 421,379,280 paired-ed 150
Blood_WGBS_rep1 255,237,608 paired-ed 150
Heart_WGBS_rep1 403,339,595 paired-ed 150
Intestine_WGBS_rep1 409,562,435 paired-ed 150
Kidney_WGBS_rep1 429,903,275 paired-ed 150
Liver_WGBS_rep1 425,121,792 paired-ed 150
Muscle_WGBS_rep1 454,789,887 paired-ed 150

6
Skin_WGBS_rep1 429,245,446 paired-ed 150
Spleen_WGBS_rep1 409,400,141 paired-ed 150

nature research | reporting summary

Testis_WGBS_rep1 416,803,341 paired-ed 150
Brain_HiC_rep1 191,703,862 paired-ed 150
Muscle_HiC_rep1 323,085,950 paired-ed 150
Brain_RNA-seq_rep2 42,061,260 paired-ed 60
Blood_RNA-seq_rep2 32,282,328 paired-ed 60
Colon_RNA-seq_rep2 41,366,162 paired-ed 60
Heart_RNA-seq_rep2 56,548,432 paired-ed 60
Intestine_RNA-seq_rep2 63,100,907 single-end 56
E-brain_RNA-seq_rep2 42,438,841 paired-ed 99
E-trunk_RNA-seq_rep2 48,844,849 paired-ed 99
Kidney_RNA-seq_rep2 33,744,731 paired-ed 60
Liver_RNA-seq_rep2 38,926,981 paired-ed 60
Muscle_RNA-seq_rep2 40,352,658 paired-ed 51
Skin_RNA-seq_rep2 56,024,033 paired-ed 60
Spleen_RNA-seq_rep2 77,413,639 paired-ed 60
Testis_RNA-seq_rep2 41,712,778 paired-ed 60
Brain_h3K27ac_rep2 12,514,004 paired-ed 150
Colon_h3K27ac_rep2 28,798,332 paired-ed 150
Blood_h3K27ac_rep2 13,316,975 paired-ed 150
Heart_h3K27ac_rep2 1,470,040 paired-ed 150
Intestine_h3K27ac_rep2 1,043,600 paired-ed 150
E-brain_h3K27ac_rep2 35,534,782 paired-ed 150
E-trunk_h3K27ac_rep2 52,479,248 paired-ed 150
Liver_h3K27ac_rep2 19,622,870 paired-ed 150
Muscle_h3K27ac_rep2 18,465,766 paired-ed 150
Skin_h3K27ac_rep2 49,624,004 paired-ed 150
Spleen_h3K27ac_rep2 42,619,886 paired-ed 150
Testis_h3K27ac_rep2 10,459,980 paired-ed 150
Brain_h3k4me3_rep2 40,158,348 paired-ed 150
Colon_h3k4me3_rep2 2,213,328 paired-ed 150
Blood_h3k4me3_rep2 17,854,940 paired-ed 150
Heart_h3k4me3_rep2 40,874,600 paired-ed 150
Intestine_h3k4me3_rep2 3,381,112 paired-ed 150
E-brain_h3k4me3_rep2 35,017,716 paired-ed 150
E-trunk_h3k4me3_rep2 50,553,374 paired-ed 150
Kidney_h3k4me3_rep2 14,636,426 paired-ed 150
Liver_h3k4me3_rep2 25,704,652 paired-ed 150
Muscle_h3k4me3_rep2 9,398,482 paired-ed 150
Skin_h3k4me3_rep2 29,972,300 paired-ed 150
Spleen_h3k4me3_rep2 30,352,882 paired-ed 150
Testis_h3k4me3_rep2 31,154,908 paired-ed 150
Brain_H3K9me3_rep2 11,381,274 paired-ed 150
Colon_H3K9me3_rep2 10,271,857 paired-ed 150
Blood_H3K9me3_rep2 9,695,703 paired-ed 150
Heart_H3K9me3_rep2 10,271,857 paired-ed 150
Intestine_H3K9me3_rep2 9,473,110 paired-ed 150
Kidney_H3K9me3_rep2 7,702,242 paired-ed 150
Liver_H3K9me3_rep2 10,159,060 paired-ed 150
Muscle_H3K9me3_rep2 11,160,372 paired-ed 150
Skin_H3K9me3_rep2 11,160,372 paired-ed 150
Spleen_H3K9me3_rep2 11,079,087 paired-ed 150
Testis_H3K9me3_rep2 8,508,473 paired-ed 150
Brain_H3K9me2_rep2 15,194,540 paired-ed 150
Colon_H3K9me2_rep2 11,329,112 paired-ed 150
Blood_H3K9me2_rep2 12,092,485 paired-ed 150
Heart_H3K9me2_rep2 15,194,540 paired-ed 150
Intestine_H3K9me2_rep2 12,221,343 paired-ed 150
Kidney_H3K9me2_rep2 11,735,304 paired-ed 150
Liver_H3K9me2_rep2 15,746,699 paired-ed 150
Muscle_H3K9me2_rep2 12,168,092 paired-ed 150
Skin_H3K9me2_rep2 12,557,397 paired-ed 150
Spleen_H3K9me2_rep2 15,521,034 paired-ed 150
Testis_H3K9me2_rep2 12,351,120 paired-ed 150
Brain_ATAC_rep2 36,189,602 paired-ed 150
Colon_ATAC_rep2 41,366,162 paired-ed 150
April 2020

Blood_ATAC_rep2 44,609,138 paired-ed 150

Heart_ATAC_rep2 50,275,820 paired-ed 150
Intestine_ATAC_rep2 8,289,902 paired-ed 150
Kidney_ATAC_rep2 62,647,690 paired-ed 150
Liver_ATAC_rep2 46,710,702 paired-ed 150
Muscle_ATAC_rep2 10,111,831 paired-ed 150
Skin_ATAC_rep2 86,132,201 paired-ed 150
Spleen_ATAC_rep2 55,632,860 paired-ed 150
Testis_ATAC_rep2 56,153,663 paired-ed 150

7
 Brain_HiC_rep2 546,389,510 paired-ed 150
Muscle_HiC_rep2 331,791,203 paired-ed 150

nature research | reporting summary

Brain_scATAC-seq_rep1 324,127,750 paired-ed 150
Brain_scATAC-seq_rep2 330,445,166 paired-ed 150

Antibodies Rabbit polyclonal histone H3K27ac antibody (Active Motif, 39133) Lot. 31814008 Dilution ratio. 1:100
Rabbit polyclonal histone H3K4me3 antibody (EMD Millipore, 07-473) clone MC135. Lot. 2591879 Dilution ratio. 1:100
Rabbit monoclonal histone H3K9me2 antibody (Cell Signaling, 4658) Lot. GR3247768-1 Dilution ratio. 1:100
Rabbit polyclonal histone H3K9me3 antibody (Abcam, ab8898) Lot. GR3247768-1 Dilution ratio.1:100

Peak calling parameters H3K27ac and H3K4me3: macs2 callpeak -f BED -n -p 1e-2 --nomodel --shift 0 --extsize 150 --keep-dup all -B --SPMR
ATAC-seq: macs2 callpeak --shift 75 --extsize 150_window --nomodel -B --SPMR --keep-dup all --call-summits
H3K9me3 and H3K9me2 : findPeaks -style histone -region -size 1000 -minDist 5000

Data quality 10496 blood_h3k9me3

20872 brain_h3k9me3
14715 colon_h3k9me3
15671 heart_h3k9me3
19356 intestine_h3k9me3
15291 kidney_h3k9me3
14825 liver_h3k9me3
7058 muscle_h3k9me3
13345 skin_h3k9me3
10774 spleen_h3k9me3
13164 testis_h3k9me3
12985 blood_h3k9me2
19872 brain_h3k9me2
13672 colon_h3k9me2
13258 heart_h3k9me2
13330 intestine_h3k9me2
8537 kidney_h3k9me2
11715 liver_h3k9me2
8876 muscle_h3k9me2
12237 skin_h3k9me2
10858 spleen_h3k9me2
13377 testis_h3k9me2
110482 blood.atac.fixed.500bp.most_significant.peak
142651 brain.atac.fixed.500bp.most_significant.peak
66771 colon.atac.fixed.500bp.most_significant.peak
137064 heart.atac.fixed.500bp.most_significant.peak
113787 intestine.atac.fixed.500bp.most_significant.peak
145525 kidney.atac.fixed.500bp.most_significant.peak
90895 liver.atac.fixed.500bp.most_significant.peak
105594 muscle.atac.fixed.500bp.most_significant.peak
180788 skin.atac.fixed.500bp.most_significant.peak
103231 spleen.atac.fixed.500bp.most_significant.peak
143452 testis.atac.fixed.500bp.most_significant.peak
26529 blood.h3k27ac.merge_x_ctl_for_rep1.peak.rpkm.filter_FC2_change_1
29767 brain.h3k27acxbrain.input.peak_region.rpkm.filter_FC2_change_1
38384 colon.h3k27acxcolon.input.peak_region.rpkm.filter_FC2_change_1
24990 e-brain.h3k27acxe-brain.input.peak_region.rpkm.filter_FC2_change_1
56391 e-muscle.h3k27acxe-muscle.input.peak_region.rpkm.filter_FC2_change_1
47199 heart.h3k27acxheart.input.peak_region.rpkm.filter_FC2_change_1
35349 intestine.h3k27acxintestine.input.peak_region.rpkm.filter_FC2_change_1
45105 kidney.h3k27acxkidney.to.h3k27ac.input.peak_region.rpkm.filter_FC2_change_1
52000 liver.h3k27acxliver.input.peak_region.rpkm.filter_FC2_change_1
74777 muscle.h3k27acxmuscle.input.peak_region.rpkm.filter_FC2_change_1
8724 skin.h3k27acxskin.input.peak_region.rpkm.filter_FC2_change_1
22015 spleen.h3k27acxspleen.to.h3k27ac.input.peak_region.rpkm.filter_FC2_change_1
24238 testis.h3k27acxtestis.input.peak_region.rpkm.filter_FC2_change_1
15521 blood.h3k4me3.merge_x_ctl_for_rep1.peak.rpkm.filter_FC2_change_1
21845 brain.h3k4me3xbrain.input.peak_region.rpkm.filter_FC2_change_1
24878 colon.h3k4me3xcolon.input.peak_region.rpkm.filter_FC2_change_1
36789 e-brain.h3k4me3xe-brain.input.peak_region.rpkm.filter_FC2_change_1
47734 e-muscle.h3k4me3xe-muscle.input.peak_region.rpkm.filter_FC2_change_1
April 2020

20386 heart.h3k4me3xheart.input.peak_region.rpkm.filter_FC2_change_1
22039 intestine.h3k4me3xintestine.input.peak_region.rpkm.filter_FC2_change_1
26275 kidney.h3k4me3xkidney.input.peak_region.rpkm.filter_FC2_change_1
21490 liver.h3k4me3xliver.input.peak_region.rpkm.filter_FC2_change_1
23063 muscle.h3k4me3xmuscle.input.peak_region.rpkm.filter_FC2_change_1
18720 skin.h3k4me3xskin.input.peak_region.rpkm.filter_FC2_change_1
20429 spleen.h3k4me3xspleen.to.h3k4me3.input.peak_region.rpkm.filter_FC2_change_1
32333 testis.h3k4me3xtestis.input.peak_region.rpkm.filter_FC2_change_1
13171 brain_merge_scATAC.10_peaks.narrowPeak

8
 123336 brain_merge_scATAC.11_peaks.narrowPeak
127338 brain_merge_scATAC.12_peaks.narrowPeak

nature research | reporting summary

55723 brain_merge_scATAC.13_peaks.narrowPeak
121441 brain_merge_scATAC.14_peaks.narrowPeak
17349 brain_merge_scATAC.15_peaks.narrowPeak
60244 brain_merge_scATAC.16_peaks.narrowPeak
85839 brain_merge_scATAC.17_peaks.narrowPeak
70176 brain_merge_scATAC.18_peaks.narrowPeak
58501 brain_merge_scATAC.19_peaks.narrowPeak
109890 brain_merge_scATAC.1_peaks.narrowPeak
95033 brain_merge_scATAC.20_peaks.narrowPeak
86813 brain_merge_scATAC.21_peaks.narrowPeak
127874 brain_merge_scATAC.22_peaks.narrowPeak
93747 brain_merge_scATAC.23_peaks.narrowPeak
63592 brain_merge_scATAC.24_peaks.narrowPeak
135147 brain_merge_scATAC.25_peaks.narrowPeak
85722 brain_merge_scATAC.2_peaks.narrowPeak
168874 brain_merge_scATAC.3_peaks.narrowPeak
15599 brain_merge_scATAC.4_peaks.narrowPeak
59250 brain_merge_scATAC.5_peaks.narrowPeak
59074 brain_merge_scATAC.6_peaks.narrowPeak
65282 brain_merge_scATAC.7_peaks.narrowPeak
65523 brain_merge_scATAC.8_peaks.narrowPeak
30041 brain_merge_scATAC.9_peaks.narrowPeak

Software MACS2, Homer

April 2020

9
Article

Structure of LRRK2 in Parkinson’s disease

and model for microtubule interaction
­­­­
https://doi.org/10.1038/s41586-020-2673-2 C. K. Deniston1,7,11, J. Salogiannis1,2,11, S. Mathea3,11, D. M. Snead1, I. Lahiri1,8, M. Matyszewski1,
O. Donosa2, R. Watanabe4,9, J. Böhning4,10, A. K. Shiau5,6, S. Knapp3, E. Villa4,
Received: 10 November 2019
S. L. Reck-Peterson1,2,6 ✉ & A. E. Leschziner1,4 ✉
Accepted: 12 August 2020

Published online: 19 August 2020

Leucine-rich repeat kinase 2 (LRRK2) is the most commonly mutated gene in familial
Check for updates Parkinson’s disease1 and is also linked to its idiopathic form2. LRRK2 has been
proposed to function in membrane trafficking3 and colocalizes with microtubules4.
Despite the fundamental importance of LRRK2 for understanding and treating
Parkinson’s disease, structural information on the enzyme is limited. Here we
report the structure of the catalytic half of LRRK2, and an atomic model of
microtubule-associated LRRK2 built using a reported cryo-electron tomography
in situ structure5. We propose that the conformation of the LRRK2 kinase domain
regulates its interactions with microtubules, with a closed conformation favouring
oligomerization on microtubules. We show that the catalytic half of LRRK2 is
sufficient for filament formation and blocks the motility of the microtubule-based
motors kinesin 1 and cytoplasmic dynein 1 in vitro. Kinase inhibitors that stabilize an
open conformation relieve this interference and reduce the formation of LRRK2
filaments in cells, whereas inhibitors that stabilize a closed conformation do not. Our
findings suggest that LRRK2 can act as a roadblock for microtubule-based motors and
have implications for the design of therapeutic LRRK2 kinase inhibitors.

LRRK2 is a large (288 kDa) multi-domain protein. Its amino-terminal half

consists of repetitive protein interaction motifs (armadillo, ankyrin, and
leucine-rich repeats), and its carboxy-terminal catalytic half contains
a Ras-like GTPase (Ras-of-complex (ROC) domain), a kinase domain,
and two other domains (C-terminal of ROC (COR), and WD40) (Fig. 1a).
High-resolution structural data on LRRK2 is limited to bacterial homo-
logues6,7 or isolated domains8,9, whereas low-resolution full-length
protein structures have been obtained using negative-stain electron
microscopy10 and cryo-electron microscopy (cryo-EM)11. A recent study
of LRRK2 bound to microtubules in cells using cryo-electron tomog-
raphy (cryo-ET) and subtomogram analysis led to a 14 Å structure and
a proposed model of the catalytic half of LRRK25.
The interaction of LRRK2 with microtubules is linked to disease
as four of the five major mutations that cause Parkinson’s disease1
(Fig. 1a) enhance the microtubule association of LRRK212. Furthermore,
Rab GTPases, which mark membranous cargos that move along micro-
tubules, are physiological LRRK2 kinase substrates13,14. These membra-
nous cargos, and others implicated in Parkinson’s disease pathology3,
are transported by the microtubule-based motors dynein and kinesin.
Here, we set out to determine a high-resolution structure of the cata-
lytic half of LRRK2 using cryo-EM, and to understand how it interacts
with microtubules and how this affects microtubule-based motor
movement.
Structure of the catalytic half of LRRK2
High-resolution studies on human LRRK2 have been limited by the lack
of efficient expression systems that yield stable protein. We tested many
constructs (Extended Data Fig. 1a) and identified one that consisted
of the C-terminal half of wild-type LRRK2 (amino acids 1327–2527),
which resulted in robust insect cell expression and well-behaved pro-
tein (Extended Data Fig. 1b, c). This construct consists of the ROC,
COR, kinase and WD40 domains of LRRK2 (Fig. 1a), which we refer to
as LRRK2RCKW. The COR domain was previously defined as consisting
of two subdomains, COR-A and COR-B6.
We determined a 3.5 Å structure of LRRK2RCKW in the presence of
GDP using cryo-EM (Fig. 1b, c, Extended Data Figs. 1, 2). On our grids,
we observed a mixture of monomers, dimers and head-to-tail trim-
ers; we used the trimer to solve the structure (Fig. 1b, Extended Data
Figs. 1, 2). Although crucial for reaching high resolution, the trimer
is probably specific to the cryo-EM grid preparation as LRRK2RCKW is
predominantly monomeric, with a smaller percentage of dimers, in
solution (Extended Data Fig. 2n), and we only observed the trimer when
preparing grids with high concentrations of LRRK2RCKW (Methods).
Owing to flexibility, the ROC and COR-A domains were lower resolu-
tion in our structure (Fig. 1b, c). We improved this part of the map using
signal subtraction and focused 3D classification and refinement (Fig. 1d,

Department of Cellular and Molecular Medicine, University of California San Diego, La Jolla, CA, USA. 2Howard Hughes Medical Institute, Chevy Chase, MD, USA. 3Institute of Pharmaceutical
1

Chemistry, Goethe-Universität, Frankfurt, Germany. 4Division of Biological Sciences, Molecular Biology Section, University of California San Diego, La Jolla, CA, USA. 5Small Molecule Discovery
Program, Ludwig Institute for Cancer Research, La Jolla, CA, USA. 6Division of Biological Sciences, Cell and Developmental Biology Section, University of California San Diego, La Jolla, CA,
USA. 7Present address: Genomics Institute of the Novartis Research Foundation, La Jolla, CA, USA. 8Present address: Department of Biological Sciences, Indian Institute of Science Education
and Research Mohali, Mohali, India. 9Present address: La Jolla Institute for Immunology, La Jolla, CA, USA. 10Present address: Sir William Dunn School of Pathology, Oxford University, Oxford,
UK. 11These authors contributed equally: C. K. Deniston, J. Salogiannis, S. Mathea. ✉e-mail: sreckpeterson@ucsd.edu; aleschziner@ucsd.edu

344 | Nature | Vol 588 | 10 December 2020

a LRRK2RCKW f g
ANK LRR ROC COR Kinase WD40
COR-B
2527
1327
R1441C/G Y1699C G2019S I2020T N2081D
Parkinson’s Crohn’s
disease disease
b d
N-lobe
Kinase
Å
3
C-lobe
4 int actions
5
90°
6
c e
7
WD40
Fig. 1 | Cryo-EM structure of LRRK2RCKW. a, Schematic of the construct used in
this study. The N-terminal half of LRRK2, absent from our construct, is shown in
dim colours. The same colour-coding of domains is used throughout the
Article. The five major familial mutations in Parkinson’s disease and a mutation
linked to Crohn’s disease are indicated. b, c, A 3.5 Å cryo-EM map (b) and local
resolution (c) of the LRRK2RCKW trimer, with one monomer highlighted.
d, e, A 3.8 Å cryo-EM map (d) and local resolution (e) of a LRRK2RCKW monomer
e, Extended Data Fig. 2a–d). The final model was generated using the
signal-subtracted maps of the ROC and COR-A domains, and then com-
bined with the COR-B, kinase and WD40 domains from the trimer map
(Fig. 1f, Extended Data Fig. 2e–m, Supplementary Video 1). Our model
fits well into an 8.1 Å reconstruction we obtained of a LRRK2RCKW mono-
mer (Fig. 1g, Extended Data Fig. 6), which indicates that trimer forma-
tion does not cause major structural changes in the protein.
LRRK2RCKW adopts an overall J-shape, with the WD40, kinase and
COR-B domains arranged along one axis, and COR-A and ROC turning
around back towards the kinase. This brings the COR-A and the tightly
associated ROC domain into close proximity to the kinase C-lobe (Fig. 1f,
Supplementary Video 1). This arrangement probably underpins the
crosstalk between the LRRK2 kinase and GTPase15,16. Part of the FERM
domain in the FAK–FERM complex approaches the FAK C-lobe in a
similar way17 (Extended Data Fig. 3a, b). The ROC, COR-A and COR-B
domains are arranged as seen in crystal structures of LRRK2 bacte-
rial homologues6,7,18. The N-lobe of the kinase domain, in particular
its αC helix, forms an extensive interaction with the COR-B domain,
with COR-B occupying a location similar to cyclin A in CDK2–cyclin
A19 (Extended Data Fig. 3a, c).
The kinase in our LRRK2RCKW structure is in an open, inactive confor-
mation. Its activation loop contains the site of two familial mutations
found in Parkinson’s disease (G2019S and I2020T) and is disordered
beyond G2019 (Fig. 1h, Extended Data Fig. 2h, Supplementary Video 2).
R1441 and Y1699 are the sites of three other familial Parkinson’s disease
mutations and are located at the ROC and COR-B interface (Fig. 1h,
Extended Data Fig. 2j, Supplementary Video 2). Because the kinase and
GTPase interact with each other via the COR-A domain, it is possible
that these mutations, located at the interface between the GTPase
and COR-B, alter the conformational landscape of LRRK2 in response
to ligands and/or regulatory signals and therefore affect the crosstalk
between the LRRK2 catalytic domains.
A unique feature of LRRK2 is a 28-amino-acid α-helix located at its
extreme C terminus, after the WD40 domain (Fig. 1i, j, Supplementary
h Parkinson’s and
Crohn’s mutations
ROC Y1699
180°
GDP
R1441
G2019
[
[I2020]
]
N2081
i j Electrostatic
rostatic Hydrophobic
interactions interactions
C2139 H1977 P1930 H1929 H1928
COR-A
L2500 E2503 L2507 H2510 R2514 M2521
j
N2422 H2502 E2503 H2510 R2514
R1973 H1977 D1980
C-terminal
terminal
100°
00°
(WD40)
Kinase
helix
30°
C
with improved resolution for the ROC and COR-A domains. f, Ribbon diagram
of the atomic model of LRRK2RCKW. g, A 8.1 Å cryo-EM map of monomeric
LRRK2RCKW with the model in f docked in. h, Location of the Parkinson’s and
Crohn’s disease mutations listed in a. i, j, Interface between the C-terminal
helix and the kinase domain in LRRK2RCKW with residues involved in
electrostatic and hydrophobic interactions indicated.
Video 3). Although several other kinases have α-helices in the same gen-
eral location, none form interactions as extensive as those observed in
LRRK2 (Fig. 1i, j, Extended Data Fig. 3d–i). Deletion of this helix resulted
in an insoluble protein (Extended Data Fig. 1a, b). A residue near its end
(T2524) is a known phosphorylation site for LRRK220. Owing to the close
proximity between T2524 and the N-lobe of the kinase domain, as well
as the adjacent COR-B domain, we hypothesize that phosphorylation
of this residue may be involved in regulation of the kinase. Because the
last two residues of the C-terminal helix are disordered in our structure,
as is a neighbouring loop in COR-B, it is possible that conditions exist
in which these regions become ordered and turn the C-terminal helix
into a scaffolding element that connects COR-B, the kinase and the
WD40 domains.
We modelled the leucine-rich repeats (LRR) into LRRK2RCKW by using
a crystal structure of the LRR, ROC and COR domains of the Chlorobium
tepidum Roco protein7 (Extended Data Fig. 3l–p). In our model, the
LRR wraps around the N-lobe of the kinase and approaches the C-lobe,
placing the known S1292 autophosphorylation site in the LRR close to
the active site of the kinase, and the Crohn’s-disease-associated residue
N208121, located in the kinase C-lobe, next to the LRR (Extended Data
Fig. 3q), suggesting the functional relevance of this predicted interface.
Model of microtubule-bound filaments
A 14 Å structure of microtubule-associated filaments of full-length
LRRK2 (carrying the filament-promoting I2020T mutation12) was
recently determined using in situ cryo-ET and subtomogram anal-
ysis5 (Fig. 2a). The LRRK2 filaments formed on microtubules are
right-handed5. Because microtubules are left-handed and no strong
density connected the LRRK2 filament to the microtubule surface5,
it is not known whether the LRRK2 microtubule interaction is direct.
To address this, we combined purified microtubules and LRRK2RCKW,
either wild type or I2020T, and imaged them by cryo-EM. Both wild-type
and I2020T mutant LRRK2RCKW bound to microtubules, and diffraction
Nature | Vol 588 | 10 December 2020 | 345
Article
a c
b d
LRRK2RCKW LRRK2RCKW
Microtubule (WT) (I2020T)
Image
Diffraction
pattern
h i
Kinase
Open
Closed
Fig. 2 | Modelling the microtubule-associated LRRK2 filaments. a, A 14 Å
cryo-ET map of a segment of microtubule-associated LRRK2 filament
in cells. The microtubule is shown in blue and the LRRK2 filament in grey.
b, Microtubule-associated LRRK2RCKW filaments reconstituted in vitro from
purified components. Top, single cryo-EM images of a naked microtubule
(left), and wild-type (WT) (centre) and I2020T (right) LRRK2RCKW filaments.
Bottom, diffraction patterns (power spectra) calculated from the images
above. Filled and open arrowheads indicate the layer lines corresponding to the
microtubule and LRRK2RCKW, respectively. Scale bar, 20 nm. c, Fitting of the
LRRK2RCKW structure, which has its kinase in an open conformation, into the
cryo-ET map. d, Atomic model of the LRRK2RCKW filaments from c. The white
patterns calculated from the images showed layer lines consistent with
the formation of ordered filaments (Fig. 2b). Therefore, the interaction
between LRRK2 and microtubules is direct and the catalytic C-terminal
half of LRRK2 is sufficient for the formation of microtubule-associated
filaments. The layer line patterns of wild-type and I2020T mutant
LRRK2RCKW are different, with the I2020T diffraction pattern having an
additional layer line of lower frequency, which indicates longer-range
order in the filaments (Fig. 2b). This is consistent with the observation
that the I2020T mutation promotes microtubule association by LRRK2
in cells12. Understanding the structural basis for this effect will require
high-resolution structures of the filaments formed by wild-type and
I2020T mutant LRRK2.
Integrative modelling was previously used to build a model into the
in situ structure of microtubule-associated LRRK25. This modelling indi-
cated that the well-resolved cryo-ET density closest to the microtubule
consisted of the ROC, COR, kinase and WD40 domains and gave orienta-
tion ensembles for each domain5 that are in good agreement with our
high-resolution structure of LRRK2RCKW (Extended Data Fig. 4a). Here,
we built an atomic model of the microtubule-bound LRRK2 filaments
by combining our 3.5 Å structure of LRRK2RCKW with the 14 Å in situ
structure of microtubule-associated LRRK2 (Extended Data Fig. 4b–f).
This showed that the LRRK2RCKW structure is sufficient to account for
the density seen in the in situ structure (Fig. 2c), in agreement with our
ability to reconstitute microtubule-associated LRRK2RCKW filaments
in vitro (Fig. 2b), and with the earlier modelling5.
Although our LRRK2RCKW structure fits the overall shape of the cryo-ET
map, there were notable clashes at the COR domain interfaces (Fig. 2d).
Because the kinase in our LRRK2RCKW structure is in an open confor-
mation, we hypothesized that filament formation might require the
346 | Nature | Vol 588 | 10 December 2020
Open f Closed
g
WD40
WD40
WD40
WD40
COR
COR
circle highlights the filament interface mediated by interactions between COR
domains, where clashes are found. e, Superposition of the LRRK2RCKW structure
(coloured by domains) and a model of LRRK2RCKW with its kinase in a closed
conformation in blue. The dashed blue arrow indicates the closing of the
kinase. f, Fitting of the closed-kinase model of LRRK2RCKW into the cryo-ET map.
g, Atomic model of the closed-kinase LRRK2RCKW filaments from f with a white
circle highlighting the same interface as in d. h, i, Cartoon representation of the
two filament models, highlighting the clashes observed with open-kinase
LRRK2RCKW (h) and resolved with the closed-kinase model (i). In total, 82% of
clashes were resolved using the closed-kinase LRRK2RCKW model (Methods).
LRRK2 kinase to be in a closed conformation. To test this, we modelled
a kinase-closed LRRK2RCKW (Fig. 2e, Extended Data Fig. 4g–j) and used
it to rebuild the LRRK2 filament. The kinase-closed LRRK2RCKW model
resolved more than 80% of the backbone clashes we had observed
with our kinase-open LRRK2RCKW structure (Fig. 2c, d, f, g). A closed
conformation for the kinase was also proposed by the integrative
modelling5. Given these data, we hypothesize that the conformation
of LRRK2 controls its ability to oligomerize on microtubules, with a
closed kinase promoting oligomerization and an open (inactive) one
disfavouring it (Fig. 2h, i).
The LRRK2 filaments in our kinase-closed model are formed by two
homotypic interactions: one is mediated by the WD40 domain and the
other by the COR-A and COR-B domains (Fig. 3a–d). Similar interfaces
were reported on the basis of the cryo-ET structure5. We also solved
structures of LRRK2RCKW dimers, using the same grids that yielded
the 3.5 Å structure of LRRK2RCKW. We obtained structures of both the
WD40–WD40- and COR–COR-mediated dimers, which indicates that
both interfaces mediate dimerization in the absence of microtubules
(Fig. 3e, f, Extended Data Figs. 5, 6). The interface in the COR-mediated
dimer of LRRK2RCKW differs from that reported for the C. tepidum Roco
protein6,7; although the GTPase domains interact directly in the dimer of
the bacterial protein7, they are not involved in the dimerization interface
we observed for LRRK2 (Extended Data Fig. 6c).
We built an independent model of a closed-kinase LRRK2RCKW by
splitting our 3.5 Å structure in half at the junction between the N- and
C-lobes of the kinase, and fitting the fragments into our cryo-EM map
of a WD40–WD40 dimer obtained in the presence of a LRRK2-specific
type I kinase inhibitor MLi-222,23, which is predicted to stabilize the
closed conformation of the kinase (Extended Data Fig. 7a–c).
 a WD40-mediated b c COR-mediated d g
LRRK2RCKW dimer LRRK2RCKW dimer

h
e f
Apo MLi-2 Apo MLi-2 90°

90° 90°
90° 90°

Fig. 3 | LRRK2RCKW forms WD40- and COR-mediated dimers outside the
filaments. a–d, The filament model shown in Fig. 2h, i is shown here in grey,
with either a WD40-mediated (a), or COR-mediated (c) LRRK2RCKW dimer
highlighted with domain colours. The corresponding molecular models are
shown next to the cartoons (b, d). e, f, Cryo-EM reconstructions of LRRK2RCKW
dimers obtained in the absence of inhibitor (‘apo’), or in the presence of MLi-2.
For each reconstruction, two orientations of the map are shown: down the
two-fold axis at the dimerization interface (left), which matches the orientation
of the models shown in b and d, and perpendicular to it (right). The top row
shows the cryo-EM map and the bottom row a transparent version of it with a
model docked in. g, Molecular models of the WD40-mediated and
COR-mediated LRRK2RCKW dimers obtained in the presence of MLi-2 (e, f) were
aligned in alternating order. This panel shows the resulting right-handed helix.
h, The helix has dimensions that are compatible with the diameter of a
12-protofilament microtubule (EMD-5192)44, which was the species used to
obtain the cryo-ET map shown in Fig. 2a5, and has its ROC domains pointing
towards the microtubule surface.

We then docked this closed-kinase model (Extended Data Fig. 7c) into
the cryo-EM maps of WD40- and COR-mediated dimers obtained in
the presence of MLi-2 to generate molecular models of both dimers
(Extended Data Fig. 7d, e). We aligned these models to build a polymer
in silico. This resulted in a right-handed helix with the same general
geometric properties seen in the cellular LRRK2 filaments, which
indicates that those properties are largely encoded in the structure
of LRRK2RCKW itself (Fig. 3g, h, Extended Data Fig. 7f). Docking the
same two halves of LRRK2RCKW into the cryo-EM map of a monomer we
obtained in the absence of inhibitors or ATP led to a structure simi-
lar to our 3.5 Å structure obtained from trimers, further confirming
that trimer formation does not alter the conformation of LRRK2RCKW
(Extended Data Fig. 7g, h).
microtubule-associated proteins, such as MAP2 and Tau, also inhibit
kinesin, but not dynein26,27, probably owing to the ability of dynein to
side-step on the microtubule28–30. The unusual ability of LRRK2 to inhibit
dynein may be a consequence of it forming oligomers that cannot be
overcome by sidestepping.
We also tested the inhibition of kinesin by I2020T mutant LRRK2RCKW,
which promotes the formation of filaments when overexpressed in
cells12. I2020T mutant LRRK2RCKW inhibited kinesin to a similar extent
as wild-type LRRK2RCKW (Extended Data Fig. 8c, d). Because the in vitro
reconstituted filaments of I2020T mutant LRRK2RCKW show longer range
a Dynein-TMR/dynactin/ b c
100 100

(% per microtubule)
(% per microtubule)
These data, along with the apparent lack of any residue-specific Kinesin-GFP
Kines ninein-like
Kinesin motility

Dynein motility
80 80
interactions between LRRK2 and the microtubule lattice, suggest that 60 60
****
the microtubule may provide a surface for LRRK2 to oligomerize on Microtubule 40 40
using interfaces that exist in solution, therefore explaining the sym- – + 20 **** 20
****
metry mismatch between the microtubule and the LRRK2 filament. Streptavidin 0 0
Biotin 0 6.25 12.50 25.00 0 25
Consistent with this, the surface charge of the microtubule facing the LRRK2RCKW [nM] LRRK2RCKW [nM]
LRRK2RCKW filament is acidic, whereas there are basic patches on the d e f
Kinesin Kinesin Dynein
LRRK2RCKW filament that face the microtubule (Extended Data Fig. 7i–l). 1.0 1.5 1.0
Relative frequency
Relative frequency

Velocity (μm s–1)

The unstructured C-terminal tails of α- and β-tubulin, which were not 1.0
****
included in the surface charge calculations, are also acidic. 0.5 [LRRK2RCKW], tau 0.5
0 nM, 1.667 μm
6.25 nM,1.570 μm 0.5 [LRRK2RCKW], tau
12.5 nM, 1.048 μm 0 nM, 4.980 μm
25 nM, 0.813 μm 25 nM, 0.846 μm
0.0
LRRK2RCKW inhibits kinesin and dynein 0.0
0 5 10 15
0.0
0 6.25 12.50 25.00 0 20 40
Run length (μm) LRRK2RCKW [nM] Run length (μm)
To test our hypothesis that the conformation of the LRRK2 kinase
domain regulates its interaction with microtubules, we needed a sen- Fig. 4 | LRRK2RCKW inhibits the motility of kinesin and dynein. a, Schematic of
sitive assay to measure the association of LRRK2RCKW with microtubules the single-molecule motility assay. b, c, The percentage (mean ± s.d.) of motile
and a means to control the conformation of its kinase. We monitored events per microtubule as a function of LRRK2RCKW concentration for kinesin (b)
microtubule association by measuring the effect of LRRK2RCKW on and dynein (c). ****P < 0.0001, Kruskal–Wallis test with Dunn’s post hoc for
microtubule-based motor motility. We used a truncated human kine- multiple comparisons for (b) or Mann–Whitney test (c). d, Cumulative frequency
sin 1, KIF5B (‘kinesin’)24, which moves towards the microtubule plus end, distribution of kinesin run lengths as a function of LRRK2RCKW concentration.
and the activated human cytoplasmic dynein-1–dynactin–ninein-like Mean decay constants (tau) are shown. The 12.5 nM and 25 nM, but not 6.25 nM,
conditions were significantly different (P < 0.0001) than the 0 nM condition
complex (‘dynein’)25, which moves in the opposite direction. Using
(one-way analysis of variance (ANOVA) with Dunnett’s test for multiple
single-molecule in vitro motility assays (Fig. 4a), we found that low
comparisons using error generated from a bootstrapping analysis). e, Velocity of
nanomolar concentrations of LRRK2RCKW inhibited the movement
kinesin as a function of LRRK2RCKW concentration. Data are mean ± s.d.
of both kinesin and dynein, with near complete inhibition at 25 nM ****P < 0.0001, one-way ANOVA with Dunn’s post hoc for multiple comparisons.
LRRK2RCKW (Fig. 4b, c, Extended Data Fig. 8a, b). We hypothesized that f, Cumulative frequency distribution of dynein run lengths as a function of
LRRK2RCKW was acting as a roadblock for the motors. In agreement LRRK2RCKW concentration. Mean decay constants (tau) are shown. Data were
with this, the distance that single kinesins moved (run length) was resampled with bootstrapping analysis and were significant. P < 0.0001,
reduced (Fig. 4d), whereas their velocity remained relatively con- unpaired t-test with Welch’s correction using error generated from a
stant (Fig. 4e). We obtained similar results with dynein (Fig. 4f). Other bootstrapping analysis.

Nature | Vol 588 | 10 December 2020 | 347

Article
a e
DMSO Ponatinib GZD-824 MLi-2 LRRK2-IN-1 Kinase

Kinesin motility per MT (%)

100
Open
80

60 COR

40 ****
****
20
******** ******** WD40
0 WD40
0 25 50 0 25 50 0 25 50 0 25 50 0 25 50 WD40
LRRK2RCKW [nM] Kinase
b c d COR
Closed
DMSO Pon. GZD MLi-2 IN-1 COR
Dynein motility per MT (%)

100 60 50
**** COR

Cells with LRRK2

Cells with LRRK2

80 40

filaments (%)

filaments (%)
40
60 30

40 20 WD40
20

20 *** **** 10
****
0 0 0

SO
M O
0 25 0 25 0 25 0 25 0 25

-2

10
5
S
GZD (μM)

Li

DM
DM
LRRK2RCKW [nM]

Fig. 5 | Type II, but not type I, kinase inhibitors rescue kinesin and dynein of wild-type GFP–LRRK2 filaments in 293T cells. Data are mean ± s.d.
motility and reduce LRRK2 filament formation in cells. a, b, Effects of ****P = 0.0002, Mann–Whitney test. d, Treatment with GZD-824 (5 μM) for
different kinase inhibitors on LRRK2RCKW’s inhibition of kinesin (a) and dynein 30 min decreases the formation of GFP–LRRK2(I2020T) filaments in 293 cells.
(b) motility. Data shown are the percentage of motile events per microtubule Data are mean ± s.d. *P = 0.0133, **P = 0.0012, Kruskal–Wallis with Dunn’s
(MT) as a function of LRRK2RCKW concentration in the absence (DMSO) or post hoc test for multiple comparisons. e, Schematic of our hypothesis. The
presence of the indicated inhibitors (ponatinib (Pon.) and GZD-824 (GZD): LRRK2 kinase can be in an open or closed conformation. The different species
10 μM; MLi-2 and LRRK2-IN-1 (IN-1): 1 μM). Data are mean ± s.d. ***P < 0.001, we observed are represented in the rounded rectangles, but only monomers
****P < 0.0001, Kruskal–Wallis test with Dunn’s post hoc for multiple are shown on the microtubule for simplicity. Our model proposes that the
comparisons within drug only. DMSO conditions reproduced from Fig. 4c for kinase-closed form of LRRK2 favours oligomerization on microtubules.
comparison. c, Treatment with MLi-2 (500 nM) for 2 h increases the formation

order than wild-type LRRK2RCKW (Fig. 2b), it is possible that the high
sensitivity of the single-molecule motility assays does not allow us to GZD-824 reduces filaments in cells
distinguish differences in oligomer length and/or stability between In cells expressing high levels of LRRK2, LRRK2 forms filaments that
wild-type and I2020T mutant LRRK2RCKW. colocalize with a subset of microtubules and are sensitive to the
microtubule-depolymerizing drug nocodazole12. This association is
enhanced by the Parkinson’s disease-linked mutations R1441C, R1441G,
Type II inhibitors rescue the motors Y1699C and I2020T12,36 and by type I kinase inhibitors34,37. We tested
Our hypothesis predicts that the closed conformation of the LRRK2 our kinase conformation hypothesis in human 293T cells by determin-
kinase domain will favour the oligomerization of LRRK2 on micro- ing whether type I and type II kinase inhibitors had opposite effects
tubules. By contrast, it predicts that conditions that stabilize the on the formation of microtubule-associated LRRK2 filaments in cells.
kinase in an open conformation will prevent oligomerization and Consistent with previous findings37,38, the type I inhibitor MLi-2
therefore decrease microtubule binding by LRRK2RCKW, resulting in enhanced the microtubule association of LRRK2 (Fig. 5c), which
the relief of LRRK2RCKW-dependent inhibition of kinesin and dynein suggests that the closed conformation of the kinase favours bind-
motility. To test these predictions, we searched for a type II kinase ing to microtubules in cells. By contrast, the type II inhibitor
inhibitor that binds tightly to LRRK2 with structural evidence that GZD-824 reduced the filament-forming ability of overexpressed
it stabilizes an open kinase conformation. We selected ponatinib, LRRK2(I2020T) (Fig. 5d). This reduction in LRRK2 filament formation
which has an inhibition constant (Ki) for LRRK2 of 31 nM31, and crystal was not due to changes in the amount of LRRK2 protein expression or
structures show it bound to RIPK232 and IRAK4 (Protein Data Bank the overall architecture of the microtubule cytoskeleton (Extended
(PDB) code 6EG9) in open conformations (Extended Data Fig. 9a). Data Fig. 9l–n).
Ponatinib inhibited the ability of LRRK2RCKW to phosphorylate Rab8A13
(Extended Data Fig. 9f).
As predicted, ponatinib rescued kinesin motility in a dose-dependent Conclusions
manner at concentrations of LRRK2RCKW (25 nM) that had resulted in Here we reported the 3.5 Å structure of the catalytic half of LRRK2
almost complete inhibition of kinesin (Fig. 5a, Extended Data Fig. 9g–j). and used it, in combination with a 14 Å cryo-ET structure of
We observed similar effects with GZD-824, a chemically related type II microtubule-associated LRRK2 filaments5, to build an atomic model
kinase inhibitor33 (Fig. 5a, Extended Data Fig. 9i, j). Our hypothesis also of these filaments. This modelling led us to hypothesize that the confor-
predicted that kinase inhibitors that stabilize the closed form of the mation of the LRRK2 kinase controls its association with microtubules
kinase would not rescue the motors and might enhance the inhibitory (Fig. 5e). Cryo-EM structures of dimers of LRRK2RCKW obtained in the
effect of LRRK2RCKW by increasing its ability to form filaments on micro- absence of microtubules showed that the same interfaces are involved
tubules. Indeed, MLi-222,23 and another LRRK2-specific type I inhibitor, in both dimerization and filament formation, and aligning them in silico
LRRK2-IN-134, which are both expected22,35 to stabilize a closed confor- resulted in a right-handed filament with similar properties to LRRK2
mation of the kinase (Extended Data Fig. 9b–e), further enhanced the filaments observed in cells, which suggests that the ability of LRRK2
inhibitory activity of LRRK2RCKW on kinesin (Fig. 5a). Dynein motility was to form filaments is a property inherent to LRRK2, and specifically the
also rescued by ponatinib and GZD-824, but not by MLi-2 or LRRK2-IN-1 RCKW domains. We propose that both the surface charge and geometric
(Fig. 5b, Extended Data Fig. 9i, k). Similar to ponatinib, GZD-824, MLi-2 complementarity between the microtubule and LRRK2 promote the
and LRRK2-IN-1 inhibited the phosphorylation of Rab8A by LRRK2RCKW formation of LRRK2 filaments. It remains to be determined whether
(Extended Data Fig. 9f). LRRK2RCKW monomers or dimers are the minimal filament-forming unit.

348 | Nature | Vol 588 | 10 December 2020

 We tested our model that the conformation of the LRRK2 kinase
regulates microtubule association using kinase inhibitors expected
to stabilize either the open (type II) or closed (type I) conformations
of the kinase. In support of our model, type II inhibitors relieved the
LRRK2RCKW-dependent inhibition of kinesin and dynein and reduced the
formation of LRRK2 filaments in cells, whereas type I inhibitors were
unable to rescue motor motility and enhanced filament formation in
cells. Notably, our single-molecule motility assays showed that low
nanomolar concentrations of LRRK2RCKW negatively affect both kinesin
and dynein. At these low concentrations, it is likely that LRRK2 would
not form the long, highly ordered filaments and microtubule bundles
observed in cells that overexpress the protein; instead, we hypothesize
that at lower expression levels in cells LRRK2 forms short oligomers
on microtubules.
The physiological role of non-pathogenic microtubule-associated
LRRK2 is not known. Our data show that LRRK2 acts as a roadblock
for microtubule-based motors in vitro. In cells, dynein and kine-
sin bind directly or indirectly to many Rab-marked membranous
cargos39–43. Our data also show that the microtubule-associated form
of LRRK2 has its kinase in a closed (and potentially active) confor-
mation. Microtubule-associated LRRK2 stalling of kinesin or dynein
could therefore increase the likelihood that LRRK2 phosphorylates
cargo-associated Rab proteins13, modulating effector binding13 and
resulting in changes in cargo dynamics. In support of this, the four
mutations associated with Parkinson’s disease that enhance LRRK2
microtubule binding12 also show higher levels of Rab phosphorylation
in cells than the G2019S mutant13,14, which does not show enhanced
microtubule binding compared with wild-type LRRK212.
Our data have important implications for the design of LRRK2 kinase
inhibitors for therapeutic purposes. We predict that inhibitors that
favour the closed conformation of the kinase will promote LRRK2 fila-
ment formation, and therefore could block microtubule-based traf-
ficking, whereas inhibitors that favour an open conformation of the
kinase will not. These results should be taken into account to enhance
the therapeutically beneficial effects of LRRK2 kinase inhibition and
to avoid potential unintended side effects.
Online content
Any methods, additional references, Nature Research reporting sum-
maries, source data, extended data, supplementary information,
acknowledgements, peer review information; details of author con-
tributions and competing interests; and statements of data and code
availability are available at https://doi.org/10.1038/s41586-020-2673-2.
1. Monfrini, E. & Di Fonzo, A. Leucine-rich repeat kinase (LRRK2) genetics and parkinson’s
disease. Adv. Neurobiol. 14, 3–30 (2017).
2. Di Maio, R. et al. LRRK2 activation in idiopathic Parkinson’s disease. Sci. Transl. Med. 10,
eaar5429 (2018).
3. Abeliovich, A. & Gitler, A. D. Defects in trafficking bridge Parkinson’s disease pathology
and genetics. Nature 539, 207–216 (2016).
4. Gloeckner, C. J. et al. The Parkinson disease causing LRRK2 mutation I2020T is associated
with increased kinase activity. Hum. Mol. Genet. 15, 223–232 (2006).
5. Watanabe, R. et al. The in situ structure of Parkinson’s disease-linked LRRK2. Cell 182,
1508–1518.e16 (2020).
6. Gotthardt, K., Weyand, M., Kortholt, A., Van Haastert, P. J. M. & Wittinghofer, A. Structure
of the Roc-COR domain tandem of C. tepidum, a prokaryotic homologue of the human
LRRK2 Parkinson kinase. EMBO J. 27, 2239–2249 (2008).
7. Deyaert, E. et al. Structure and nucleotide-induced conformational dynamics of the
Chlorobium tepidum Roco protein. Biochem. J. 476, 51–66 (2019).
8. Deng, J. et al. Structure of the ROC domain from the Parkinson’s disease-associated
leucine-rich repeat kinase 2 reveals a dimeric GTPase. Proc. Natl Acad. Sci. USA 105,
1499–1504 (2008).
9. Zhang, P. et al. Crystal structure of the WD40 domain dimer of LRRK2. Proc. Natl Acad.
Sci. USA 116, 1579–1584 (2019).
10. Guaitoli, G. et al. Structural model of the dimeric Parkinson’s protein LRRK2 reveals a
compact architecture involving distant interdomain contacts. Proc. Natl Acad. Sci. USA
113, E4357–E4366 (2016).
11. Sejwal, K. et al. Cryo-EM analysis of homodimeric full-length LRRK2 and LRRK1 protein
complexes. Sci. Rep. 7, 8667 (2017).
12. Kett, L. R. et al. LRRK2 Parkinson disease mutations enhance its microtubule association.
Hum. Mol. Genet. 21, 890–899 (2012).
13. Steger, M. et al. Phosphoproteomics reveals that Parkinson’s disease kinase LRRK2
regulates a subset of Rab GTPases. eLife 5, e12813 (2016).
14. Steger, M. et al. Systematic proteomic analysis of LRRK2-mediated Rab GTPase
phosphorylation establishes a connection to ciliogenesis. eLife 6, e31012 (2017).
15. Ito, G. et al. GTP binding is essential to the protein kinase activity of LRRK2, a
causative gene product for familial Parkinson’s disease. Biochemistry 46, 1380–1388
(2007).
16. West, A. B. et al. Parkinson’s disease-associated mutations in LRRK2 link enhanced
GTP-binding and kinase activities to neuronal toxicity. Hum. Mol. Genet. 16, 223–232
(2007).
17. Lietha, D. et al. Structural basis for the autoinhibition of focal adhesion kinase. Cell 129,
1177–1187 (2007).
18. Terheyden, S., Ho, F. Y., Gilsbach, B. K., Wittinghofer, A. & Kortholt, A. Revisiting the Roco
G-protein cycle. Biochem. J. 465, 139–147 (2015).
19. Cheng, K.-Y. et al. The role of the phospho-CDK2/cyclin A recruitment site in substrate
recognition. J. Biol. Chem. 281, 23167–23179 (2006).
20. Pungaliya, P. P. et al. Identification and characterization of a leucine-rich repeat kinase 2
(LRRK2) consensus phosphorylation motif. PLoS ONE 5, e13672 (2010).
21. Hui, K. Y. et al. Functional variants in the LRRK2 gene confer shared effects on risk for
Crohn’s disease and Parkinson’s disease. Sci. Transl. Med. 10, eaai7795 (2018).
22. Scott, J. D. et al. Discovery of a 3-(4-pyrimidinyl) indazole (MLi-2), an orally available and
selective leucine-rich repeat kinase 2 (LRRK2) inhibitor that reduces brain kinase activity.
J. Med. Chem. 60, 2983–2992 (2017).
23. Fell, M. J. et al. MLi-2, a potent, selective, and centrally active compound for exploring the
therapeutic potential and safety of LRRK2 kinase inhibition. J. Pharmacol. Exp. Ther. 355,
397–409 (2015).
24. Case, R. B., Pierce, D. W., Hom-Booher, N., Hart, C. L. & Vale, R. D. The directional
preference of kinesin motors is specified by an element outside of the motor catalytic
domain. Cell 90, 959–966 (1997).
25. Redwine, W. B. et al. The human cytoplasmic dynein interactome reveals novel activators
of motility. eLife 6, e28257 (2017).
26. Dixit, R., Ross, J. L., Goldman, Y. E. & Holzbaur, E. L. F. Differential regulation of dynein and
kinesin motor proteins by tau. Science 319, 1086–1089 (2008).
27. Monroy, B. Y. et al. A combinatorial MAP Code dictates polarized microtubule transport.
Dev. Cell 53, 60–72.e4 (2020).
28. Reck-Peterson, S. L. et al. Single-molecule analysis of dynein processivity and stepping
behavior. Cell 126, 335–348 (2006).
29. Qiu, W. et al. Dynein achieves processive motion using both stochastic and coordinated
stepping. Nat. Struct. Mol. Biol. 19, 193–200 (2012).
30. DeWitt, M. A., Chang, A. Y., Combs, P. A. & Yildiz, A. Cytoplasmic dynein moves through
uncoordinated stepping of the AAA+ ring domains. Science 335, 221–225 (2012).
31. Liu, M. et al. Type II kinase inhibitors show an unexpected inhibition mode against
Parkinson’s disease-linked LRRK2 mutant G2019S. Biochemistry 52, 1725–1736 (2013).
32. Canning, P. et al. Inflammatory signaling by NOD-ripk2 is inhibited by clinically relevant
type II kinase inhibitors. Chem. Biol. 22, 1174–1184 (2015).
33. Ren, X. et al. Identification of GZD824 as an orally bioavailable inhibitor that targets
phosphorylated and nonphosphorylated breakpoint cluster region-Abelson (Bcr-Abl)
kinase and overcomes clinically acquired mutation-induced resistance against imatinib.
J. Med. Chem. 56, 879–894 (2013).
34. Deng, X. et al. Characterization of a selective inhibitor of the Parkinson’s disease kinase
LRRK2. Nat. Chem. Biol. 7, 203–205 (2011).
35. Gilsbach, B. K. et al. Structural characterization of LRRK2 inhibitors. J. Med. Chem. 58,
3751–3756 (2015).
36. Godena, V. K. et al. Increasing microtubule acetylation rescues axonal transport and
locomotor deficits caused by LRRK2 Roc-COR domain mutations. Nat. Commun. 5,
5245 (2014).
37. Blanca Ramírez, M. et al. GTP binding regulates cellular localization of Parkinson’s
disease-associated LRRK2. Hum. Mol. Genet. 26, 2747–2767 (2017).
38. Schmidt, S. H. et al. The dynamic switch mechanism that leads to activation of LRRK2
is embedded in the DFGψ motif in the kinase domain. Proc. Natl Acad. Sci. USA 116,
14979–14988 (2019).
39. Wang, Y. et al. CRACR2a is a calcium-activated dynein adaptor protein that regulates
endocytic traffic. J. Cell Biol. 218, 1619–1633 (2019).
40. Etoh, K. & Fukuda, M. Rab10 regulates tubular endosome formation through KIF13A and
KIF13B motors. J. Cell Sci. 132, jcs226977 (2019).
41. Horgan, C. P., Hanscom, S. R., Jolly, R. S., Futter, C. E. & McCaffrey, M. W. Rab11-FIP3 links
the Rab11 GTPase and cytoplasmic dynein to mediate transport to the
endosomal-recycling compartment. J. Cell Sci. 123, 181–191 (2010).
42. Niwa, S., Tanaka, Y. & Hirokawa, N. KIF1Bβ- and KIF1A-mediated axonal transport of
presynaptic regulator Rab3 occurs in a GTP-dependent manner through DENN/MADD.
Nat. Cell Biol. 10, 1269–1279 (2008).
43. Matanis, T. et al. Bicaudal-D regulates COPI-independent Golgi-ER transport by recruiting
the dynein-dynactin motor complex. Nat. Cell Biol. 4, 986–992 (2002).
44. Sui, H. & Downing, K. H. Structural basis of interprotofilament interaction and lateral
deformation of microtubules. Structure 18, 1022–1031 (2010).
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in
published maps and institutional affiliations.
© The Author(s), under exclusive licence to Springer Nature Limited 2020
Nature | Vol 588 | 10 December 2020 | 349
Article
Methods
Data reporting
No statistical methods were used to predetermine sample size, and
experiments were not randomized.
Cloning, plasmid construction and mutagenesis
For baculovirus expression, the DNA coding for wild-type LRRK2
residues 1327 to 2527 (taken from Mammalian Gene Collection) was
PCR-amplified using the forward primer TACTTCCAATCCATGA
AAAAGGCTGTGCCTTATAACCGA and the reverse primer TATCCACCTTT
ACTGTCACTCAACAGATGTTCGTCTCATTTTTTCA. The T4 polymerase-
treated amplicon was inserted into the expression vector pFB-6HZB
by ligation-independent cloning. According to Bac-to-Bac expression
system protocols (Invitrogen), this plasmid was used for the generation
of recombinant Baculoviruses.
For mammalian expression vectors, pDEST53-GFP-LRRK2 (WT)45
from Addgene (25044) was used. pDEST53-GFP-LRRK2 (I2020T) was
cloned using QuikChange site-directed mutagenesis (Agilent) with the
forward primer AAGATTGCTGACTACGGCACTGCTCAGTACTGCTG and
the reverse primer CAGCAGTACTGAGCAGTGCCGTAGTCAGCAATCTT.
pET17b-Kif5b(1-560)-GFP-His46 was obtained from Addgene (15219).
For pET28a-ZZ-TEV-Halo-NINL1–702, Ninein-like1–702 (NINL) was synthe-
sized as previously described25 and inserted into a pET28a expression
vector with a synthesized ZZ-TEV-Halo gBlock fragment (IDT) using
Gibson assembly.
LRRK2RCKW expression and purification
The expression construct contained an N-terminal His6-Z-tag, cleav-
able with TEV protease (Extended Data Fig. 1a–c). For LRRK2RCKW puri-
fication, the pelleted Sf9 cells were washed with PBS, resuspended in
lysis buffer (50 mM HEPES pH 7.4, 500 mM NaCl, 20 mM imidazole,
0.5 mM TCEP, 5% glycerol, 5 mM MgCl2, 20 μM GDP) and lysed by soni-
cation. The lysate was cleared by centrifugation and loaded onto a
Ni-NTA (Qiagen) column. After vigorous rinsing with lysis buffer the
His6-Z-tagged protein was eluted in lysis buffer containing 300 mM
imidazole. Immediately thereafter, the eluate was diluted with a buffer
containing no NaCl, to reduce the NaCl-concentration to 250 mM and
loaded onto an SP sepharose column. His6-Z-TEV-LRRK2RCKW was eluted
with a 250 mM to 2.5 M NaCl gradient and treated with TEV protease
overnight to cleave the His6-Z-tag. Contaminating proteins, the cleaved
tag, uncleaved protein and TEV protease were removed by another com-
bined SP sepharose Ni-NTA step. Finally, LRRK2RCKW was concentrated
and subjected to gel filtration in storage buffer (20 mM HEPES pH 7.4,
800 mM NaCl, 0.5 mM TCEP, 5% glycerol, 2.5 mM MgCl2, 20 μM GDP)
using an AKTA Xpress system combined with an S200 gel filtration
column. The final yield as calculated from UV absorbance was 1.2 mg l−1
LRRK2RCKW insect cell medium.
SEC–MALS
SEC–MALS experiments were performed using an ÄKTAmicro chroma-
tography system hooked up to a Superdex 200 Increase 3.2/300 size
exclusion chromatography column coupled in-line to a DAWN HELEOS
II multiangle light scattering detector (Wyatt Technology) and an Opti-
lab T-rEX refractive index detector (Wyatt Technology). SEC–MALS
was performed in 50 mM HEPES pH 7.4, 200 mM NaCl 0.5 mM TCEP,
5% glycerol, 5 mM MgCl2, and 20 μM GDP. For a typical sample, 50 μl of
approximately 7 μM LRRK2RCKW was injected onto the column. Molar
mass was calculated using ASTRA 6 software, with protein concentration
derived from the Optilab T-rEX. LRRK2RCKW used for SEC-MALS experi-
ments contained an extra 16 residue N-terminal Gly-Ser linker sequence.
Electron microscopy
Electron microscopy sample preparation and imaging of trimer
dataset. Purified LRRK2RCKW was dialysed into a final buffer consisting
of 20 mM HEPES pH7.4, 80 mM NaCl, 0.5mM TCEP, 5% glycerol, 2.5 mM
MgCl2 and 20 μM GDP and then diluted to a final concentration of 4
μM in the same buffer. We have only observed the trimer species when
preparing grids with LRRK2RCKW at concentrations of 4 μM or above and
for our apo samples (that is, in the absence of inhibitors). This sample
was applied to glow-discharged (20 mA for 20 s in a K100 Instrument)
UltrAuFoil Holey Gold R 1.2/1.3 grids (Quantifoil). A Vitrobot (FEI) was
then used to blot away excess sample and plunge freeze the grids in
liquid ethane. Grids were stored in liquid nitrogen until imaged.
Cryo-EM data were collected at UCLA California NanoSystems
Institute in a Titan Krios (FEI) operated at 300 kV, equipped with a K2
Summit direct electron detector (Gatan) and a Quantum energy filter
(Gatan). Automated data collection was performed using Leginon47.
We recorded a total of 3,824 movies in ‘counting mode’ at a dose rate of
6.65 electrons Å−2 s−1 with a total exposure time of 8 s sub-divided into
200-ms frames, for a total of 40 frames. The images were recorded at
a nominal magnification of 130,000×, resulting in an object pixel size
of 1.07 Å. The defocus range of the data was −1 μm to −1.8 μm.
Electron microscopy map and model generation of trimer data-
set. We aligned the movie frames using UCSF MotionCor248, using
the dose-weighted frame alignment option. We estimated the CTF
on dose-weighted images using GCTF version 1.0649 as implemented
in Appion50 with per-particle CTF generation. Images having CTF fits
worse that 5 Å (as determined by GCTF) were excluded from further pro-
cessing. Using this approach, 3,693 micrographs were kept for further
processing. We selected particles from micrographs using FindEM51
with projections of a trimeric LRRK2RCKW map, created from an initial
Cryosparc ab initio model generation, serving as a reference. Particle
picking was performed within the framework of Appion, resulting in
a dataset of 836,956 particles.
We carried out subsequent processing first in Relion 3.052 then in
CryoSparc253. A series of 2D and 3D classifications were performed
as shown in Extended Data Fig. 1f to generate the final map. The initial
reference was created in a similar manner to those used for template
picking, from an initial ab initio model generated in CryoSparc. All ref-
erences were filtered to either 60 Å (default in Relion) or 30 Å (default
in CryoSparc) before refinement processes. The final map, generated
in Cryosparc2 using non-uniform refinement and while applying C3
symmetry, reached 3.47 Å resolution. Initial 2D classifications used
binned data (4.28 Å pixel−1) while all subsequent 3D classifications and
refinement steps used unbinned images (1.07 Å pixel−1).
To improve the density corresponding to the ROC and COR-A
domains, a second map was generated following a different process-
ing scheme, which used signal subtraction and is shown in Extended
Data Fig. 2a. The final refinement led to a 3.8 Å map of LRRK2RCKW. All
the steps used unbinned images (1.07 Å pixel−1).
The resolutions of the cryo-EM maps, here and below, were esti-
mated from Fourier shell correlation (FSC) curves calculated using the
gold-standard procedure and the resolutions are reported according to
the 0.143 cut-off criterion54–56. FSC curves were corrected for the convo-
lution effects of a soft mask applied to the half maps by high-resolution
phase randomization57. For display and analysis purposes, we sharp-
ened the maps with automatically estimated negative B factors from
Relion or CryoSparc2.
We built the LRRK2RCKW models using both the 3.47 Å C3 and 3.8 Å
density-subtracted maps. We used a combination of Rosetta58 and
manual building in Coot59 to build all models. Starting models were
found via a sequence alignment search in HHpred60 and each sec-
tion was built from the following: (1) the ROC and COR domains were
built using the top two alignment results, 3DPT_B and 3DPU_B, which
scored much higher than those ranked third and below; (2) the kinase
N-lobe, kinase C-lobe, and WD40 domain were all built from the top
five sequence matches: 5JGA_A, 1OPK_A, 4BTF_A, 5GZA_A, 5K00_A for
the kinase N-lobe; 5CYZ_A, 2IZR_A, 5YKS_B, 3S95_B, 4TWC_B for the
kinase C-lobe; and 4J87_A, 3OW8_C, 5U69_A, 5O9Z_F, 5T2A_7 for the
WD40 domain. The C-terminal helix was built manually. We built the
COR-B, KIN and WD40 domains using the density of the 3.47 Å C3 map
and the ROC and COR-A domains using the 3.8 Å map, then connected
the two after fitting them into the 3.8 Å map. Finally, we performed
multiple iterations of both the CM and Relax functions in Rosetta, along
with manual manipulation in Coot, to build our final 20 models
(10 including GDP-Mg2+ in the ROC domain, and 10 excluding it). In
areas of weak density, we either removed part of the polypeptide chain
or, where only side-chain density was poor, converted the chain to
poly-alanine.
The GDP-Mg2+ was placed in our model by initially aligning a structure
of the ROC domain containing a bound GDP-Mg2+ (PDB code 2ZEJ)8 to
the ROC domain in our structure. The GDP-Mg2+ was then added to our
model in the aligned position and run through Rosetta to allow for fine
movements into our density and re-arrangement of nearby chains.
The phosphorylation of the threonine residue 1343, which we
observe in our map and is a known phosphorylation site of Roco family
GTPases61, was confirmed by phosphor-enrichment mass spectrometry
(data not shown).
Electron microscopy sample preparation, imaging, and process-
ing of apo, MLi2, ponatinib LRRK2RCKW monomer or dimers. For all
samples, purified LRRK2RCKW was dialysed into the same final buffer as
described for the trimer data, and then diluted to its final concentration
in the same buffer. Unlike with the trimer data, however, the following
datasets were collected from multiple grids prepared using slightly
different sample conditions and imaged using a range of microscope
settings.
The apo LRRK2RCKW monomer dataset had sample diluted to final
concentrations ranging between 1 μM and 6 μM. In addition, one data-
set was collected with the grid tilted to 30° to overcome preferred
orientation issues. Otherwise, grids were prepared as described for
the trimer dataset.
The apo LRRK2RCKW dimer dataset had sample diluted to final concen-
trations ranging between 4 μM and 12 μM. Two of the samples contained
0.05 mM digitonin (Sigma, D141) or 0.03% octyl glucoside (Sigma,
O8001) detergents to overcome preferred orientation issues. One data
set was collected with the grid tilted to 30°, also to overcome preferred
orientation issues. Otherwise, grids were prepared as described for
the trimer dataset.
The MLi-2 LRRK2RCKW dimer dataset had sample diluted to final con-
centrations of either 3 μM or 4 μM. MLi-2 was added post-dialysis to a
final concentration of 5 μM. The sample was incubated on ice for at least
1 h before being applied to the grid. Otherwise, grids were prepared as
described for the trimer dataset.
The ponatinib LRRK2RCKW dimer dataset had sample diluted to final
concentrations of 2 μM or 4 μM. Ponatinib was added post-dialysis at
a concentration of either 5 μM or 100 μM. The sample was incubated
on ice for at least 1 h before being applied to the grid. Otherwise, grids
were prepared as described for the trimer dataset.
The apo LRRK2RCKW monomer cryo-EM data were collected on a Talos
Arctica (FEI) operated at 200 kV, equipped with a K2 Summit direct
electron detector (Gatan). Automated data collection was performed
using Leginon47. A total of 11,354 movies were collected. We imaged
samples at dose rates between 4.2 and 10 electrons Å−2 s−1 with total
exposure times ranging from 6 s to 12 s sub-divided into 200-ms frames,
for a total of 30 or 60 frames. All the images were recorded at a nominal
magnification of 36,000× (either counting mode or super resolution
mode), resulting in object pixel sizes of either 1.16 Å or 0.58 Å, respec-
tively. The defocus range of the data was −1 μm to −2 μm.
Frame alignment, CTF estimation and image selection were per-
formed as described for the trimer dataset except that per-particle
CTF was not used, instead the CTF information of the whole image
was used. After selection, we had 7,067 micrographs. Particles were
extracted using crYOLO45. We carried out subsequent processing in
CryoSparc253 on binned images (2.32 Å pixel−1).
A series of 2D and 3D classifications were performed as shown in
Extended Data Fig. 6 to generate the final map. The initial monomer
reference, used for refinement of the monomer map, was generated
from our LRRK2RCKW model. The initial dimer references, used for par-
ticle sorting, were generated as shown in Extended Data Fig. 5a–e. All
references were filtered to 30 Å, the default value in CrypSparc2, before
refinement processes. This final map reached a resolution of 8.08 Å
using non-uniform refinement.
The apo LRRK2RCKW dimer cryo-EM data were collected as described
for the apo LRRK2RCKW monomer. A total of 5,303 movies were collected.
We imaged samples at dose rates between 4.6 and 7.8 electrons Å−2 s−1
with total exposure times ranging between 7 s and 11 s sub-divided
into 200ms frames, for a total of 35 or 55 frames. All the images were
recorded at a nominal magnification of 36,000× (either counting mode
or super resolution mode), resulting in object pixel sizes of either
1.16 Å or 0.58 Å, respectively. The defocus range of the data was
−1 μm to −2 μm.
Frame alignment, CTF estimation, image selection, and particle pick-
ing were performed as described for the apo LRRK2RCKW monomer.
After selection we had 3,100 micrographs. We carried out subsequent
processing in CryoSparc253 on binned images (2.32 Å pixel−1).
The classification and refinement scheme for the apo LRRK2RCKW
WD40- and COR-mediated dimer maps is shown in Extended Data Fig. 6.
The same initial dimer references used for the apo LRRK2RCKW mono-
mer, filtered to the same resolution, were used here for refinement of
the dimer maps. In addition, a linear trimer reference, generated by
combining two initial dimer references and filtered to the same
resolution, was used during the initial 3D classification step to sort
out species longer than dimer, which negatively affected subsequent
alignment. The final maps had resolutions of 13.39 Å (WD40-mediated
dimer) and 9.52 Å (COR-mediated dimer) with C2 symmetry applied
to both.
The MLi-2 LRRK2RCKW dimer cryo-EM data were collected as described
for the apo LRRK2RCKW monomer. We recorded a total of 4,139 movies.
We imaged all datasets in ‘counting mode’ at a dose rate of 5.5 elec-
trons Å−2 s−1, with total exposure times of either 9 s or 10 s sub-divided
into 200ms frames, for a total of 45 or 50 frames. All the images were
recorded at a nominal magnification of 36,000× (counting mode),
resulting in object pixel sizes of 1.16 Å. The defocus range of the data
was −1 μm to −2 μm.
Frame alignment, CTF estimation, image selection, and particle pick-
ing were performed as described for the apo LRRK2RCKW monomer.
After selection, 4,030 micrographs were kept for further processing.
Processing was done in Relion 3.052 then CryoSparc253 using binned
images (2.32 Å pixel−1).
The classification and refinement scheme for the MLi-2 LRRK2RCKW
WD40- and COR-mediated dimer maps is shown in Extended Data Fig. 5.
The same references used for the apo LRRK2RCKW dimers, and filtered to
the same resolution, were used here for the same purposes. The final
maps had resolutions of 9.74 Å (WD40-mediated dimer) and 9.04 Å
(COR-mediated dimer), with no symmetry applied.
The ponatinib LRRK2RCKW dimer cryo-EM data were collected as
described for the apo LRRK2RCKW monomer. We recorded a total of
1,797 movies. We imaged all datasets at dose rates of either 5.5 or 9.7
electrons Å−2 s−1 with total exposure times of 7 s or 10 s sub-divided
into 200-ms frames, for a total of 35 or 50 frames. All the images were
recorded at a nominal magnification of 36,000× (counting mode),
resulting in object pixel sizes of 1.16 Å. The defocus range of the data
was −1 μm to −2 μm.
Frame alignment, CTF estimation, image selection, and particle pick-
ing were performed as described for the apo LRRK2RCKW monomer. A
total of 1,455 micrographs were kept for further processing. Processing
was done in CryoSparc253 on binned images (2.32 Å pixel−1). Ponatinib
Article
LRRK2RCKW dimer particles were sorted via 2D classification leading to
the final 2D averages shown in Extended Data Fig. 6b.
Electron microscopy sample preparation and imaging of
microtubule-associated LRRK2RCKW. Purified LRRK2RCKW and unpo-
lymerized bovine tubulin were added together to the microtubule
polymerization buffer (1× BRB80, 1 mM DTT, 10% DMSO, 1 mM GTP, 1
mM MgCl2, 10 μM Taxol). LRRK2RCKW was added in a 2× molar excess rela-
tive to the α/β tubulin dimer during the polymerization procedure. To
promote the interaction of LRRK2RCKW with microtubules, NaCl concen-
tration was capped at 90 mM; this set an upper limit to the LRRK2RCKW
concentrations we could use. For wild-type LRRK2RCKW, we used 2.5 μM
LRRK2RCKW and 1.25 μM tubulin. For the I2020T variant, we used 4.5 μM
LRRK2RCKW(I2020T) and 2.25 μM tubulin, respectively. The mixture was
incubated for 1 h at room temperature and was then diluted threefold
into cryo-EM buffer (20 mM HEPES pH 7.4, 80 mM NaCl, 0.5 mM TCEP,
2.5 mM MgCl2, 10 μM taxol) immediately before application to the
grid and plunge-freezing in a Vitrobot (FEI). The grids (Lacey Carbon
on copper, 300 mesh; EMS) were glow-discharged (20 mA for 40 s in a
K100 Instrument) before the sample was applied.
Cryo-EM data were collected on a Talos Arctica (FEI) operated at 200
kV, equipped with a K2 Summit direct electron detector (Gatan) using
Leginon47. We imaged at a dose rate of 5.85 electrons Å−2 s−1 with a total
exposure time of 10 s sub-divided into 200 ms frames, for a total of
50 frames. All the images were recorded at a nominal magnification
of 36,000×, resulting in an object pixel size of 1.16 Å. The defocus range
of the data was −1 μm to −2 μm. Images were aligned using MotionCor248
with the dose-weighted frame alignment option.
For layer line analysis, filaments were selected and extracted into
600 Å boxes using Relion 3.052 from the dose-corrected images.
The images were binned fourfold into a pixel size of 4.64 Å and a final
box size of 164 pixels. Before calculating the Fourier transforms, the
images were padded with an additional 82 pixels on all sides.
Building the molecular model of microtubule-associated
LRRK2RCKW filaments
Given that the WD40 densities are clearly identifiable in the cryo-ET
map of microtubule-associated LRRK2, we used these as a starting point
for docking the structure of LRRK2RCKW into the cryo-ET map. First, a
synthetic dimer of the WD40 domains from the LRRK2RCKW structure was
generated by aligning them to a crystal structure of the isolated WD40
domain, which formed a dimer in the crystal (PDB code 6DLP)9. This
synthetic dimer was then docked into the cryo-ET map in Chimera62,
using the Fit in Map function with the options of filtering the structure
to the resolution of the map (14 Å) and optimizing correlation. A WD40
dimer was placed into each of the two corresponding densities present
in the map. Then, four copies of the LRRK2RCKW structure were added
by aligning their WD40 domains to those previously docked into the
cryo-ET map. The same procedure was followed to build the filament
using the ‘closed’ kinase model of LRRK2RCKW (see ‘Modelling a closed
kinase version of LRRK2RCKW’ for how that model was generated).
Backbone clashes at the COR-mediated interface in the model fila-
ment were measured in Chimera62 (with default settings) after convert-
ing the four LRRK2RCKW monomers to poly-alanine models.
Modelling a closed kinase version of LRRK2RCKW
To identify a good reference to model the closed state of the LRRK2RCKW
kinase, we ran separate structural searches (using the DALI server) with
the N- and C-lobes of the LRRK2RCKW kinase domain. We looked through
the matches for a kinase that scored highly with both lobes, and whose
structure is in a closed state. We selected interleukin-2 inducible T-cell
kinase (ITK) bound to an inhibitor as our reference (PDB code 3QGY)63.
LRRK2RCKW was split at the junction between the N- and C-lobes of its
kinase domain (L1949–A1950), resulting in one half containing the ROC,
COR, and kinase (N-lobe) domains and another containing the kinase
(C-lobe) and WD40 domains. The C-lobe of 3QGY was then aligned
(in Chimera) to the C-lobe of the LRRK2RCKW kinase domain, and subse-
quently the N-lobe of the LRRK2RCKW kinase was aligned to the N-lobe
of 3QGY. The two halves were then combined to generate the closed
kinase model of LRRK2RCKW.
Docking of LRRK2RCKW into cryo-EM maps of monomers and
dimers
To build models of WD40- and COR-mediated dimers of LRRK2RCKW in
the presence of MLi-2, we again split LRRK2RCKW at the junction between
the N- and C-lobes (L1949–A1950). The two halves were fitted into
one half of the cryo-EM map of a WD40-mediated dimer of LRRK2RCKW
obtained in the presence of MLi-2 (we chose this map as its resolution
was higher than that of the COR-mediated dimer). We also docked the
two halves of LRRK2RCKW into a cryo-EM map of a LRRK2RCKW monomer
obtained in the absence of inhibitor. The fitting was done in Chimera62
using the Fit in Map function with the options of filtering the structure
to the resolution of the map and optimizing correlation. The two halves
were then joined to generate a full model of LRRK2RCKW.
The WD40- and COR-mediated dimers of LRRK2RCKW in the presence
of MLi-2 were built by docking the models built above into the cor-
responding cryo-EM maps, using the same approach in Chimera62 as
outlined above.
The LRRK2RCKW filament shown in Extended Data Fig. 7f was generated
by aligning, in alternating order, several copies of the two dimer models
(WD40- and COR-mediated) built into the cryo-EM maps obtained in
the presence of MLi-2.
Kinase inhibitors
Stocks of the kinase inhibitors MLi-2 (10 mM; Tocris), ponatinib (10 mM;
ApexBio), GZD-824 (10 mM; Cayman Chemical) and LRRK2-IN-1 (2 mM;
Michael J. Fox Foundation) were stored in DMSO at −20 °C.
Antibodies
All antibodies used for immunocytochemistry were diluted to 1:500.
Primary antibodies used were chicken anti-GFP (Aves Labs) and rab-
bit anti-α-tubulin (ProteichTech). Secondary antibodies used were
goat anti-chicken-Alexa 488 (ThermoFisher) and goat anti-chicken
Alexa568 (ThermoFisher). DAPI was used at 1:5,000 according to
the manufacturer’s suggestions (ThermoFisher). Primary antibod-
ies used for western blots were mouse anti-GFP (Santa Cruz, 1:1,000
dilution) mouse anti-GAPDH (ProteinTech, 1:5,000 dilution) and mouse
anti-gamma-tubulin (ProteinTech, 1:5,000 dilution). Secondary anti-
bodies (1:15,000) used for western blots were IRDye goat anti-mouse
680RD and IRDye goat anti-rabbit 780RD (Li-COR).
Rab8a expression and purification
N-terminally tagged (His6-ZZ) Rab8A containing a TEV cleavage site was
cloned into a PET28a expression vector and expressed in BL21(DE3)
Escherichia coli cells. Transformed cells were grown overnight at 37 °C
in 10 ml LB medium containing kanamycin (50 μg ml−1), then diluted into
200 ml LB medium containing kanamycin (50 μg ml−1), grown to an opti-
cal density at 600 nm of approximately 1–2, diluted into 4 l LB medium
containing kanamycin (50 μg ml−1), and grown to an optical density at
600 nm of 0.4. IPTG was added (final concentration 0.5 mM) to induce
protein expression for around 18 h at 18 °C. Cells were collected by
centrifugation at 8,983g for 10 min at 4 °C, followed by resuspension
in 15 ml LB medium and centrifugation at 2,862g for 10 min at 4 °C. The
cell pellet was flash frozen in liquid nitrogen and stored at −80 °C. For a
typical protein purification, cell pellets were resuspended in lysis buffer
(50 mM HEPES pH 7.4, 200 mM NaCl, 2 mM DTT, 10% glycerol, 5 mM
MgCl2, 0.5 mM Pefabloc, and protease inhibitor cocktail tablets) and
lysed by sonication on ice. The lysate was clarified by centrifugation
at 164,700g for 40 min at 4 °C and then incubated with Ni-NTA agarose
beads (Qiagen) for 1 h at 4 °C. Beads were extensively washed with wash
buffer (50 mM HEPES pH 7.4, 150 mM NaCl, 2 mM DTT, 10% glycerol,
5 mM MgCl2); His6-ZZ-Rab8a was eluted in 40 ml elution buffer (50
mM HEPES pH 7.4, 150 mM NaCl, 300 mM imidazole, 2 mM DTT, 10%
glycerol, 5 mM MgCl2). The protein eluate was diluted twofold in wash
buffer, incubated with IgG sepharose 6 fast flow beads equilibrated
in wash buffer, incubated at 4 °C for 2.5 h, and washed extensively in
wash buffer. Protein-bound IgG beads were then transferred into TEV
buffer (50 mM HEPES pH 7.4, 200 mM NaCl, 2 mM DTT, 10% glycerol,
5 mM MgCl2), and untagged Rab8A was cleaved off of IgG sepharose
beads by incubation with TEV protease at 4 °C overnight. The next day,
cleaved Rab8A was separated from His6-TEV protease and any remaining
uncleaved protein or residual tag by incubation with Ni-NTA agarose
beads (Qiagen), followed by washing with TEV buffer containing 25
mM imidazole. Lastly, purified Rab8A was run over a Superdex 200
increase 10/300 size exclusion column equilibrated in S200 buffer (50
mM HEPES pH 7.4, 200 mM NaCl, 2 mM DTT, 1% glycerol, 5 mM MgCl2),
and concentrated and exchanged into buffer containing 10% glycerol
for storage at −80 °C.
In vitro phosphorylation of Rab8A by LRRK2RCKW
Purified Rab8A (approximately 3.8 μM) was phosphorylated by
LRRK2RCKW (38 nM) in a buffer containing 50 mM HEPES pH 7.4, 80
mM NaCl, 10 mM MgCl2, 0.5 mM TCEP, 1 mM ATP, 200 μM GDP); 34 μl
reaction mixtures containing kinase inhibitor or an equivalent volume
DMSO were incubated at 30 °C, and samples were taken at 45 min, and
90 min. An effective reaction volume of 0.75 μl was run on a 4–12%
Bis-Tris protein gel, transferred to nitrocellulose, and blotted with
a commercially available antibody to pT72-Rab8a (MJFF-pRab8) as
previously described49 and per manufacturer’s instructions, with the
exception that HRP-labelled secondary antibody was used at a dilu-
tion of 1:2,000.
Purification of molecular motors
Protein purification steps were done at 4 °C unless otherwise indicated.
Human KIF5B1–560(K560)–GFP was purified from E. coli using an adapted
protocol previously described64. pET17b-Kif5b(1-560)-GFP-His was
transformed into BL-21[DE3] RIPL cells (New England Biolabs) until
optical density at 600 nm of 0.6–0.8 and expression was induced with
0.5 mM IPTG for 16 h at 18 °C. Frozen pellets from 2 l culture were resus-
pended in 40 ml lysis buffer (50 mM Tris, 300 mM NaCl, 5 mM MgCl2,
and 0.2 M sucrose, pH 7.5) supplemented with 1 cOmplete EDTA-free
protease inhibitor cocktail tablet (Roche) per 50 ml and 1 mg ml−1
lysozyme. The resuspension was incubated on ice for 30 min and lysed
by sonication. Sonicate was supplied with 10 mM imidazole and 0.5 mM
PMSF and clarified by centrifuging at 30,000g for 30 min in Type 70
Ti rotor (Beckman). The clarified supernatant was incubated with 5 ml
Ni-NTA agarose (Qiagen) and rotated in a nutator for 1 h. The mixture
was washed with 30 ml wash buffer (50 mM Tris, 300 mM NaCl, 5 mM
MgCl2, 0.2 M sucrose, and 20 mM imidazole, pH 7.5.) by gravity flow.
Beads were resuspended in elution buffer (50 mM Tris, 300 mM NaCl,
5 mM MgCl2, 0.2 M sucrose, and 250 mM imidazole, pH 8.0), incubated
for 5 min, and eluted stepwise in 0.5 ml increments. Peak fractions
were combined and buffer exchanged on a PD-10 desalting column
(GE Healthcare) equilibrated with storage buffer (80 mM PIPES, 2 mM
MgCl2, 1 mM EGTA, and 0.2 M sucrose, pH 7.0). From this, peak fractions
of motor solution were either flash frozen at −80 °C until further use or
immediately subjected to microtubule bind and release purification.
A total of 1 ml motor solution was incubated with 1 mM AMP-PNP and
20 μM taxol on ice for 5 min and warmed to room temperature. For
microtubule bind and release, polymerized bovine brain tubulin was
centrifuged through a glycerol cushion (80 mM PIPES, 2 mM MgCl2, 1
mM EGTA, and 60% glycerol (v/v) with 20 μM taxol and 1 mM DTT) and
resuspended as previously described25 was incubated with motor solu-
tion in the dark for 15 min at room temperature. The motor-microtubule
mixture was laid on top of a glycerol centrifuged in a TLA120.2 rotor at
278,835g for 12 min at room temperature. Final pellet (kinesin-bound
microtubules) was washed with BRB80 (80 mM PIPES, 2 mM MgCl2,
and 1 mM EGTA, pH 7.0) and incubated in 100 μl of release buffer (80
mM PIPES, 2 mM MgCl2, 1 mM EGTA, and 300 mM KCl, pH 7 with 5 mM
Mg-ATP) for 5 min at room temperature. The supernatant was supplied
with 660 mM sucrose and flash frozen. A typical kinesin prep yielded
0.5 to 1 μM K560–GFP dimer.
Human dynactin was purified from stable cell lines expressing
p62-Halo-3×Flag as previously described65,66. In brief, cells were col-
lected from 160 × 15 cm plates and resuspended in 80 ml of dynactin-lysis
buffer (30 mM HEPES, pH 7.4, 50 mM potassium acetate, 2 mM magne-
sium acetate, 1 mM EGTA, 1 mM DTT, 10% (v/v) glycerol) supplemented
with 0.5 mM Mg-ATP, 0.2% Triton X-100 and 1 cOmplete EDTA-free pro-
tease inhibitor cocktail tablet (Roche) per 50 ml and rotated slowly for
15 min. The lysate was clarified by centrifuging at 66,000g for 30 min in
Type 70 Ti rotor (Beckman). The clarified supernatant was incubated
with 1.5 ml of anti-Flag M2 affinity gel (Sigma-Aldrich) overnight on a
roller. The beads were transferred to a gravity flow column, washed
with 50 ml of wash buffer (dynactin-lysis buffer supplemented with
0.1 mM Mg-ATP, 0.5 mM Pefabloc and 0.02% Triton X-100), 100 ml of
wash buffer supplemented with 250 mM potassium acetate, and again
with 100 ml of wash buffer. Dynactin was eluted from beads with 1 ml
of elution buffer (wash buffer with 2 mg ml−1 of 3×Flag peptide). The
eluate was collected, filtered by centrifuging with Ultrafree-MC VV
filter (EMD Millipore) in a tabletop centrifuge and diluted to 2 ml in
buffer A (50 mM Tris-HCl, pH 8.0, 2 mM magnesium acetate, 1 mM
EGTA and 1 mM DTT) and injected onto a MonoQ 5/50 GL column (GE
Healthcare and Life Sciences) at 1 ml min−1. The column was pre-washed
with 10 column volumes (CV) of buffer A, 10 CV of buffer B (50 mM
Tris-HCl, pH 8.0, 2 mM magnesium acetate 1 mM EGTA, 1 mM DTT and
1 M potassium acetate) and again with 10 CV of buffer A at 1 ml min−1.
To elute, a linear gradient was run over 26 CV from 35–100% buffer B.
Pure dynactin complex eluted from 75–80% buffer B. Peak fractions
containing pure dynactin complex were pooled, buffer exchanged
into a GF150 buffer supplemented with 10% glycerol, concentrated to
0.02–0.1 mg ml−1 using a 100 K MWCO concentrator (EMD Millipore)
and flash frozen in liquid nitrogen. Typical dynactin prep yields are
between 150 and 300 nM.
Human dynein was purified from stable cells lines expressing an
IC2-SNAPf-3×Flag as previously described25. Frozen pellets collected
from approximate 60–100 15-cm plates were resuspended in dynein
lysis buffer (25 mM HEPES pH 7.4, 50 mM potassium acetate, 2 mM
magnesium acetate,1 mM EGTA, 10% glycerol (v/v) and 1 mM DTT) sup-
plemented with 0.2% Triton X-100, 0.5 mM Mg-ATP, and cOmplete
EDTA-free protease inhibitor cocktail. The lysate was centrifuged at
66,000g in a Ti-70 rotor for 30 min. The clarified supernatant was
incubated with 1 ml of anti-Flag M2 affinity gel (Sigma-Aldrich) over-
night on a roller. Beads were collected by gravity flow and washed with
50 ml wash buffer (dynein lysis buffer with 0.02% Triton X-100 and
0.5 mM Mg-ATP) supplemented with protease inhibitors (cOmplete
Protease Inhibitor Cocktail, Roche). Beads were then washed with 50
ml high-salt wash buffer (25 mM HEPES, pH 7.4, 300 mM potassium
acetate, 2 mM magnesium acetate, 10% glycerol, 1 mM DTT, 0.02%
Triton X-100, 0.5 mM Mg-ATP), and then with 100 ml wash buffer. For
labelling, beads were resuspended in 1 ml wash buffer and incubated
with 5 μM SNAP-Cell TMR Star (New England BioLabs) for 10 min on
the column at room temperature. Unbound dye was removed with 100
ml wash buffer at 4 °C. Dynein was eluted with 1 ml of elution buffer
(wash buffer containing 2 mg ml−1 3×Flag peptide). The eluate was col-
lected, diluted to 2 ml in buffer A (50 mM Tris pH 8.0, 2 mM magnesium
acetate, 1 mM EGTA and 1 mM DTT) and injected onto a MonoQ 5/50 GL
column (GE Healthcare Life Sciences) at 0.5 ml min−1. The column was
washed with 20 CV of buffer A at 1 ml min−1. To elute, a linear gradient
was run over 40 CV into buffer B (50 mM Tris pH 8.0, 2 mM magnesium
acetate, 1 mM EGTA, 1 mM DTT, 1 M potassium acetate). Pure dynein
Article
complex elutes from ~60–70% buffer B. Peak fractions were pooled
and concentrated, 0.1 mM Mg-ATP and 10% glycerol were added and
the samples were snap frozen in liquid nitrogen. A typical preparation
yielded 150–300 nM dynein.
Human NINL was purified as previously described 65.
pET28a-ZZ-TEV-Halo-NINL1–702 was transformed into BL-21[DE3] cells
(New England Biolabs) until an optical density at 600 nm of 0.4–0.6
and expression was induced with 0.1 mM IPTG for 16 h at 18 °C. Frozen
cell pellets from 1 l culture were resuspended in 40 ml of activator-lysis
buffer (30 mM HEPES [pH 7.4], 50 mM potassium acetate, 2 mM mag-
nesium acetate, 1 mM EGTA, 1 mM DTT, 0.5 mM Pefabloc, 10% (v/v)
glycerol) supplemented with 1 cOmplete EDTA-free protease inhibitor
cocktail tablet (Roche) per 50 ml and 1 mg ml−1 lysozyme. The resus-
pension was incubated on ice for 30 min and lysed by sonication. The
lysate was clarified by centrifuging at 66,000g for 30 min in Type 70 Ti
rotor (Beckman). The clarified supernatant was incubated with 2 ml of
packed IgG Sepharose 6 Fast Flow beads (GE Healthcare Life Sciences)
for 2 h on a roller. The beads were transferred to a gravity flow column,
washed with 100 ml of activator-lysis buffer supplemented with 150 mM
potassium acetate and 50 ml of cleavage buffer (50 mM Tris–HCl, pH
8.0, 150 mM potassium acetate, 2 mM magnesium acetate, 1 mM EGTA, 1
mM DTT, 0.5 mM Pefabloc and 10% (v/v) glycerol). The beads were then
resuspended and incubated in 15 ml of cleavage buffer supplemented
with 0.2 mg ml−1 TEV protease overnight on a roller. The supernatant
containing cleaved proteins were concentrated using a 50-K MWCO
concentrator (EMD Millipore) to 1 ml, filtered by centrifuging with
Ultrafree-MC VV filter (EMD Millipore) in a tabletop centrifuge, diluted
to 2 ml in buffer A (30 mM HEPES, pH 7.4, 50 mM potassium acetate, 2
mM magnesium acetate, 1 mM EGTA, 10% (v/v) glycerol and 1 mM DTT)
and injected onto a MonoQ 5/50 GL column (GE Healthcare and Life Sci-
ences) at 0.5 ml min−1. The column was pre-washed with 10 CV of buffer
A, 10 CV of buffer B (30 mM HEPES, pH 7.4, 1 M potassium acetate, 2 mM
magnesium acetate, 1 mM EGTA, 10% (v/v) glycerol and 1 mM DTT) and
again with 10 CV of buffer A at 1 ml min−1. To elute, a linear gradient was
run over 26 CV from 0–100% buffer B. The peak fractions containing
unlabelled Halo-tagged NINL were collected and concentrated to using
a 50K MWCO concentrator (EMD Millipore) to 0.2 ml. A typical NINL
prep yield was approximately 5–10 μM dimer.
Single-molecule microscopy and motility assays
Single-molecule imaging was performed using total internal reflection
fluorescence (TIRF) microscopy with an inverted microscope (Nikon,
Ti-E Eclipse) equipped with a 100× 1.49 NA oil immersion objective
(Nikon, Plano Apo), and a MLC400B laser launch (Agilent), with 405 nm,
488 nm, 561 nm and 640 nm laser lines. Excitation and emission paths
were filtered using single bandpass filter cubes (Chroma), and emitted
signals were detected with an electron multiplying CCD camera (Andor
Technology, iXon Ultra 888). Illumination and image acquisition were
controlled with NIS Elements Advanced Research software (Nikon),
and the xy position of the stage was controlled with a ProScan linear
motor stage controller (Prior).
Single-molecule motility were performed in flow chambers as
previously described25 using the setup shown in the schematic in
Fig. 4a. Biotin-PEG-functionalized coverslips (Microsurfaces) were
adhered to a Superfrost Plus Microscope slide (ThermoFisher)
using double-sided scotch tape. Each slide contained four
flow-chambers. Taxol-stabilized microtubules (approximately 15
mg ml−1) with 10% biotin-tubulin and 10% Alex405-tubulin were pre-
pared as previously described25. For each motility experiment, 1
mg ml−1 Strepdavidin (in 30 mM HEPES, 2 mM magnesium acetate,
1 mM EGTA, 10% glycerol) was incubated in the flow chamber for
3 min. A 1:150 dilution of taxol-stabilized microtubules in motility
assay buffer (30 mM HEPES, 50 mM potassium acetate, 2 mM magne-
sium acetate, 1 mM EGTA, 10% glycerol, 1 mM DTT and 20 μM Taxol, pH
7.4) was added to the flow chamber for 3 min to adhere polymerized
microtubules to the coverslip. Flow chambers containing adhered
microtubules were washed twice with LRRK2 buffer (20 mM HEPES
pH 7.4, 80 mM NaCl, 0.5 mM TCEP, 5% glycerol, 2.5 mM MgCl2 and
20 μM GDP). Flow chambers were then incubated for 5 min either
with (1) LRRK2 buffer alone or with the indicated kinase inhibitors
(0 nM LRRK2RCKW condition) or (2) LRRK2 buffer containing LRRK2RCKW
alone, with DMSO, or with kinase inhibitors. DMSO or drugs were
incubated with LRRK2 buffer (± LRRK2RCKW) for 10 min at room tem-
perature before adding to the flow chambers. Before the addition
of dynein and kinesin motors, the flow chambers were washed twice
with motility assay buffer containing 1 mg ml−1 casein. To assemble
dynein-dynactin-ninein-like (NINL) complexes, purified dynein (10–15
nM), dynactin and NINL were mixed at 1:2:10 molar ratio and incubated
on ice for 10 min. The final imaging buffer for motors contained motil-
ity assay buffer supplemented with an oxygen scavenger system,
71.5 mM β-mercaptoethanol and either 1 mM ATP (kinesin) or 2.5 mM
ATP (dynein). The final concentrations of kinesin and dynein in the flow
chambers were 2.5 nM and 0.3 nM, respectively. K560–GFP was imaged
every 500 ms for 2 min with 25% laser (488) power at 150 ms exposure
time. Dynein-TMR-dynactin-NINL was imaged every 300 ms for 3 min
with 25% laser (561) power at 100 ms. Each sample was imaged no longer
than 15 min. Each technical replicate consisted of movies from at least
two fields of view containing between 5 and 10 microtubules each.
Single-molecule motility assay analysis
Kymographs were generated from motility movies and quantified for
run lengths, percent motility, and velocity using ImageJ (NIH). Specifi-
cally, maximum-intensity projections were generated from time-lapse
sequences to define the trajectory of particles on a single microtubule.
The segmented line tool was used to trace the trajectories and map them
onto the original video sequence, which was subsequently re-sliced to
generate a kymograph. Motile and immotile events (>1 s) were manually
traced. Bright aggregates, which were less than 5% of the population,
were excluded from the analysis. For dynein–dynactin–NINL, both
stationary and diffusive events were grouped as immotile. Run-length
measurements were calculated from motile events only. Error bars for
run length analysis were generated using a bootstrapping method (run
length values from each condition were resampled 200 times using an
XLSTAT program for Excel) and statistical significance was established
using a one-way ANOVA with Dunnett’s test for multiple comparisons or
an unpaired Welch’s t-test for only one comparison. For percent motility
per microtubule measurements, motile events (>1 s and >1 μm) were
divided by total events per kymograph. Velocity measurements were
calculated from the inverse slopes of the motile event traces (>1 s and
>1 μm) only. Statistical analyses were performed in Prism8 (Graphpad).
Cell line
HEK293T cells were purchased from ATCC (CRL-3216) and authenticated
by ATCC. This cell line was tested for mycoplasma before expanding
and freezing. After thawing each time, it was tested again. In addition,
all cells grown in the laboratory are routinely tested for mycoplasma
every three months. The cells used in the experiments reported here
were last tested on 10/16/19 and showed no contamination.
Western blot analysis
293T cells were maintained in DMEM (containing 10% fetal bovine serum
and 1% penicillin/streptomycin). For western blot quantification of
LRRK2 protein expression (Extended Data Fig. 9l), cells were plated on
6-well dishes (150,000 cells per well) 24 h before transfection. Cells were
transfected with 1 μg of GFP–I2020T using polyethylenimine (PEI, Poly-
sciences). After 48 h, cells were treated for 30 min with either 5 μM GZD-
824 or DMSO-matched control. Cells were lysed on ice in RIPA buffer
(50 mM Tris pH 7.5, 150 mM NaCl, 0.2% Triton X-100, 0.1% SDS, 0.5%
Na-deoxycholate, with cOmplete protease inhibitor cocktail). Lysates
were further rotated for 15 min at 4 °C and clarified by centrifuging
at 13,000g for 15 min. Clarified supernatants were boiled for 5 min in
Laemmli buffer. The experiments were performed in triplicate.
For western blots, lysates were run on 4–12% gradient SDS–PAGE (Life
Technologies) for 60 min and transferred to nitrocellulose for 3 h at
250 mA. Blots were dried at room temperature for 30 min, rinsed in 1×
TBS, followed by blocking with 5% milk in TBS. Antibodies were diluted
in 5% milk in TBS with 0.1% Tween-20 (TBS-T). Primary antibodies were
incubated overnight at 4 °C and infrared secondary antibodies were
incubated at room temperature for 45 min. For quantification of LRRK2
expression levels, blots were imaged on an Odyssey CLx controlled by
Imaging Studio software (v.5.2). DMSO and GZD-824 conditions were
quantified in triplicate and normalized to a GAPDH loading control
using Empiria Studio software (Li-COR). To ensure quantification was
in the combined linear range for antibodies detecting both GFP–LRRK2
and GAPDH, a linear dilution series of lysates from cells expressing
GFP–LRRK2 was also quantified by infrared western blot.
Immunofluorescence, confocal microscopy and image analysis
The day before transfection, 293T cells were plated on acid-treated
coverslips (Bellco Glass) pre-coated with 100 μg ml−1 poly-d-lysine
(Sigma) and 4 μg ml−1 mouse laminin (ThermoFisher) in 24-well plates
(35,000 cells per well). Cells were transfected with 500 ng plasmid of
either pDEST53-GFP-LRRK2 or pDEST53-GFP-LRRK2(I2020T) using
PEI. After 48–72 h, cells were incubated with either a kinase inhibitor
or DMSO-matched control (matched for time and concentration).
For the type I inhibitor experiment, cells were incubated with DMSO
or Mli-2 (500 nM) for 2 h. For the type II inhibitor, cells were incubated
with DMSO or GZD-824 (5 or 10 μM) for 30 min. Cells were quickly
washed once on ice with ice-cold PBS, and fixed with ice-cold 4% PFA,
90% methanol, 5 mM sodium bicarbonate for 10 min at −20 °C. After
fixation, the wells were immediately washed three times with ice-cold
PBS. Blocking buffer (1% BSA, 5% normal goat serum, 0.3% Triton X-100
in PBS) was added for 1 h at room temperature. Primary antibodies
were diluted (1:500) in antibody dilution buffer (1% BSA, 0.1% Triton
X-100 in PBS) and incubated overnight at 4 °C. After overnight incu-
bation, the wells were washed three times in PBS and incubated with
secondary antibodies (1:500) in antibody dilution buffer for 1 h at room
temperature. After secondary incubation, the wells were washed three
times in PBS, once in ddH2O and mounted using CitiFluor AF2 (EMS)
on Superfrost Plus Microscope slides (ThermoFisher). Coverslips were
sealed with nail polish and stored at 4 °C.
For the LRRK2 filament analysis in Fig. 5, experimenters were blinded
to condition for both the imaging acquisition and analysis. Cells were
imaged using a Nikon A1R HD confocal microscope with a LUN-V laser
engine (405 nm, 488 nm, 561 nm and 640 nm) and DU4 detector using
bandpass and long-pass filters for each channel (450/50, 525/50, 595/50
and 700/75). Slides were imaged on a Nikon Ti2 body using an Apo 60×
1.49 NA objective. Image stacks were acquired in resonant scanning
mode with bidirectional scanning and 4× line averaging and 1.2 airy
units. The lasers used were 405 nm, 488 nm and 561 nm. Illumination and
image acquisition were controlled by NIS Elements Advanced Research
software (Nikon Instruments). ImageJ was used to quantify the percent-
age of cells with LRRK2 filaments. Maximum-intensity projections were
generated from z-stack confocal images. Using the GFP immunofluo-
rescence signal, transfected cells were traced. Cells were scored for
the presence or absence of filaments using both the z-projection and
z-stack micrographs as a guide. The presence of filaments was scored
if the cells had either (1) a GFP filament signal greater than 5 μm or (2)
bundles of filaments with at least two identifiable crosses. To calculate
the percentage cells with filaments, the number of cells with filaments
was divided by the total number of transfected cells per technical rep-
licate (defined as one 24-well coverslip). Approximately 20 cells were
quantified per replicate for each condition in Fig. 5c (DMSO versus
Mli-2) and between 40 and 100 cells were quantified per replicate for
each condition in Fig. 5d (DMSO versus GZD-824). The quantification
of all cellular experiments comes from data collected on three separate
days except for the 10 μM GZD-824 condition in Fig. 5d, which was per-
formed on two separate days. All statistical analyses were performed
in Prism8 (Graphpad).
Reporting summary
Further information on research design is available in the Nature
Research Reporting Summary linked to this paper.
Data availability
All reagents and data will be made available upon request. Model coor-
dinates for the LRRK2RCKW structure are deposited in the PDB as follows:
(1) PDB accession code 6VP6: LRRK2RCKW with the adjacent COR-B and
WD40 domains (from the trimer) used to optimize residues at those
interfaces during refinement in Rosetta, with GDP-Mg2+ bound; (2)
PDB accession code 6VNO: the top 10 models for LRRK2RCKW without
adjacent domains, with GDP-Mg2+ bound; (3) PDB accession code
6VP8: LRRK2RCKW with the adjacent COR-B and WD40 domains (from
the trimer) used to optimize residues at those interfaces during refine-
ment in Rosetta, no GDP-Mg2+; (4) PDB accession code 6VP7: the top 10
models for LRRK2RCKW without adjacent domains, no GDP-Mg2+ bound.
Cryo-EM maps for the different LRRK2RCKW structures are deposited
at the EMDB as follows: (1) Electron Microscopy Data Bank (EMDB)
accession code EMD-21250: this deposition contains both the 3.5 Å
map of LRRK2RCKW trimer (used to build the COR-B, kinase and WD40
domains) and the 3.8 Å map of the signal-subtracted LRRK2RCKW trimer
(used to build the ROC and COR-A domains); (2) EMDB accession code
EMD-21306: 8.1 Å map of LRRK2RCKW monomer; (3) EMD accession code
21309: 9.5 Å map of COR-mediated LRRK2RCKW dimer in the absence of
kinase ligand (apo); (4) EMDB accession code EMD-21310: 13.4 Å map
of WD40-mediated LRRK2RCKW dimer in the absence of kinase ligand
(apo); (5) EMDB accession code EMD-21311: 9.0 Å map of COR-mediated
LRRK2RCKW dimer in the presence of MLi-2; (6) EMDB accession code
EMD-21312: 10.2 Å map of WD40-mediated LRRK2RCKW dimer in the pres-
ence of MLi-2. All other data that support the findings of this study are
available from the corresponding authors upon reasonable request.
45. Wagner, T. et al. SPHIRE-crYOLO is a fast and accurate fully automated particle picker for
cryo-EM. Commun. Biol. 2, 218 (2018).
46. Woehlke, G. et al. Microtubule interaction site of the kinesin motor. Cell 90, 207–216
(1997).
47. Suloway, C. et al. Automated molecular microscopy: the new Leginon system. J. Struct.
Biol. 151, 41–60 (2005).
48. Zheng, S. Q. et al. MotionCor2: anisotropic correction of beam-induced motion for
improved cryo-electron microscopy. Nat. Methods 14, 331–332 (2017).
49. Lis, P. et al. Development of phospho-specific Rab protein antibodies to monitor in vivo
activity of the LRRK2 Parkinson’s disease kinase. Biochem. J. 475, 1–22 (2018).
50. Lander, G. C. et al. Appion: an integrated, database-driven pipeline to facilitate EM image
processing. J. Struct. Biol. 166, 95–102 (2009).
51. Roseman, A. M. FindEM—a fast, efficient program for automatic selection of particles
from electron micrographs. J. Struct. Biol. 145, 91–99 (2004).
52. Zivanov, J. et al. New tools for automated high-resolution cryo-EM structure
determination in RELION-3. eLife 7, e42166 (2018).
53. Punjani, A., Rubinstein, J. L., Fleet, D. J. & Brubaker, M. A. cryoSPARC: algorithms for rapid
unsupervised cryo-EM structure determination. Nat. Methods 14, 290–296 (2017).
54. Henderson, R. et al. Outcome of the first electron microscopy validation task force
meeting. Structure 20, 205–214 (2012).
55. Scheres, S. H. W. & Chen, S. Prevention of overfitting in cryo-EM structure determination.
Nat. Methods 9, 853–854 (2012).
56. Rosenthal, P. B. & Henderson, R. Optimal determination of particle orientation, absolute
hand, and contrast loss in single-particle electron cryomicroscopy. J. Mol. Biol. 333,
721–745 (2003).
57. Chen, S. et al. High-resolution noise substitution to measure overfitting and validate
resolution in 3D structure determination by single particle electron cryomicroscopy.
Ultramicroscopy 135, 24–35 (2013).
58. Wang, R. Y.-R. et al. Automated structure refinement of macromolecular assemblies from
cryo-EM maps using Rosetta. eLife 5, 352 (2016).
59. Emsley, P. & Cowtan, K. Coot: model-building tools for molecular graphics. Acta
Crystallogr. D 60, 2126–2132 (2004).
60. Söding, J., Biegert, A. & Lupas, A. N. The HHpred interactive server for protein homology
detection and structure prediction. Nucleic Acids Res. 33, W244–W248 (2005).
61. Greggio, E. et al. The Parkinson’s disease kinase LRRK2 autophosphorylates its GTPase
domain at multiple sites. Biochem. Biophys. Res. Commun. 389, 449–454 (2009).
Article
62. Pettersen, E. F. et al. UCSF Chimera—a visualization system for exploratory research and
analysis. J. Comput. Chem. 25, 1605–1612 (2004).
63. Charrier, J.-D. et al. Discovery and structure-activity relationship of 3-aminopyrid-2-ones
as potent and selective interleukin-2 inducible T-cell kinase (Itk) inhibitors. J. Med. Chem.
54, 2341–2350 (2011).
64. Nicholas, M. P., Rao, L. & Gennerich, A. An improved optical tweezers assay for measuring
the force generation of single kinesin molecules. Methods Mol. Biol. 1136, 171–246 (2014).
65. Htet, Z. M. et al. LIS1 promotes the formation of activated cytoplasmic dynein-1
complexes. Nat. Cell Biol. 22, 518–525 (2020).
66. Kendrick, A. A. et al. Hook3 is a scaffold for the opposite-polarity microtubule-based
motors cytoplasmic dynein-1 and KIF1C. J. Cell Biol. 218, 2982–3001 (2019).
67. Deniston, C. K. Parkinson’s disease-linked LRRK2 structure and model for microtubule
interaction. PhD thesis, Univ. California, San Diego (2020).
Acknowledgements We thank S. Taylor for her role in initiating this collaborative work, which
was partially supported by multi-investigator grants from the Michael J. Fox Foundation:
grants: 11425 and 11425.02 (PI: S. Taylor) and 18321 (PIs: A.E.L. and S.L.R.-P.). We also thank the
UC San Diego Cryo-EM Facility, the Nikon Imaging Center at UC San Diego, where the confocal
microscopy was performed, the use of instruments at the Electron Imaging Center for
NanoMachines supported by NIH (1S10RR23057, 1S10OD018111, and 1U24GM116792), NSF (DBI-
1338135) and CNSI at UCLA; J. P. Gillies and A. Kendrick for technical support with protein
purifications, and A. Dickey for feedback on the manuscript. C.K.D. was initially supported by
the Molecular Biophysics Training Grant (NIH grant T32 GM008326) and subsequently by a
Predoctoral Fellowship from the Visible Molecular Cell Consortium and Center for Trans-scale
Structural Biology (UC San Diego). D.S. is supported by an A. P. Giannini Foundation
postdoctoral fellowship. A.K.S. receives salary and support from the Ludwig Institute for
Cancer Research. E.V. is supported by a NIH Director’s New Innovator Award DP2GM123494.
S.L.R.-P. is an investigator of the Howard Hughes Medical institute and is also supported by
R01GM121772. A.E.L. is supported by R01GM107214. S.K. is grateful for support from the SGC, a
registered charity that receives funds from AbbVie, Bayer Pharma AG, Boehringer Ingelheim,
Canada Foundation for Innovation, Eshelman Institute for Innovation, Genome Canada,
Innovative Medicines Initiative EUbOPEN (agreement No 875510), Janssen, Merck KGaA, MSD,
Ontario Ministry of Economic Development and Innovation, Pfizer, São Paulo Research
Foundation-FAPESP, Takeda, and the Wellcome, as well as Boehringer Ingelheim for funding
initial structural studies of this project. Most of this work is described in the thesis67 by C.K.D.
Author contributions C.K.D. collected and processed the cryo-EM data. J.S. performed the
single-molecule and cellular assays with the help of O.D. S.M. designed the LRRK2RCKW construct
and purified the protein. C.K.D. and I.L. built the molecular model of LRRK2RCKW. D.M.S.
performed the SEC–MALS and phosphorylation assays. M.M. collected and analysed the
LRRK2RCKW and microtubule cryo-EM data. R.W. and J.B. performed the cellular cryo-ET. A.K.S.
contributed to the structural analysis and provided advice on the selection of kinase inhibitors.
S.K., E.V., S.L.R.-P. and A.E.L. directed and supervised the research. C.K.D., J.S., S.L.R.-P. and
A.E.L. wrote the manuscript and S.M., D.S., M.M., O.D., A.K.S., S.K. and E.V. edited it.
Competing interests The authors declare no competing interests.
Additional information
Supplementary information is available for this paper at https://doi.org/10.1038/s41586-020-
2673-2.
Correspondence and requests for materials should be addressed to S.L.R.-P. or A.E.L.
Peer review information Nature thanks Asa Abelovich, Henning Stahlberg and the other,
anonymous, reviewer(s) for their contribution to the peer review of this work.
Reprints and permissions information is available at http://www.nature.com/reprints.
Extended Data Fig. 1 | Optimization of LRRK2 constructs and cryo-EM
analysis of a LRRK2RCKW trimer. a, We systematically scanned domain
boundaries (amino acid numbers of boundaries noted above domain names) to
generate LRRK2 constructs that expressed well in baculovirus-infected insect
cells and yielded stable and soluble protein. These attempts included
full-length LRRK2, the kinase domain alone or with the WD40 domain, and
other isolated domains. In this approach, only the GTPase domain on its own
expressed well. Next, we gradually shortened LRRK2 from its amino terminus.
Red asterisks indicate constructs that were soluble. b, After identifying domain
boundaries yielding constructs that expressed soluble protein, additional fine
tuning of boundaries was performed. A Coomassie-stained SDS–PAGE gel
shows systematic N-terminal truncations at the ROC domain resulting in the
identification of a construct with the highest expression levels: amino acids
1327–2527 (red asterisk, ‘LRRK2RCKW’ here). c, A Coomassie-stained SDS–PAGE
gel of purified LRRK2RCKW after elution from an S200 gel filtration column. As
predicted by its primary structure, LRRK2RCKW runs at approximately 140 kDa.
d, Electron micrograph of LRRK2RCKW. e, 2D class averages of the LRRK2RCKW
trimer. f, 2D/3D classification scheme used to obtain the 3.5 Å structure of the
LRRK2RCKW trimer. g, h, Fourier shell correlations (from Cryosparc) (d) and
Euler angle distribution (e) for the LRRK2RCKW trimer.
Article

Extended Data Fig. 2 | See next page for caption.

Extended Data Fig. 2 | Cryo-EM analysis of a signal-subtracted LRRK2RCKW
trimer and map-to-model fit. a, Processing strategy used to obtain a 3.8 Å
structure of LRRK2RCKW generated from a signal-subtracted trimer where only
one monomer contains the ROC and COR-A domains. This structure improved
the resolution of the ROC and COR-A domains relative to the full trimer
(Extended Data Fig. 1). b–d, 2D class averages (b), Fourier shell correlations
(from Relion) (c), and Euler angle distribution (from Relion) (d) for the 3.8 Å
resolution signal-subtracted LRRK2RCKW structure. e, Close-ups (f–l) of
different parts of the final model fit into the map. f, Section of the WD40
domain. g, C-terminal helix and its interface with the kinase domain. h, Active
site of the kinase. Residues in the DYG motif are labelled. G2019, the site of a
major PD-associated mutation (G2019S) and the last residue of the activation
loop seen in our structure, is highlighted by a black rounded square. i, Interface
between COR-B and the αC helix of the N-lobe of the kinase domain. j, Interface
between the ROC and COR-B domains. R1441 and Y1699, two residues mutated
in Parkinson’s disease, are labelled. k, l, Two different views of the ROC and
COR-A domains with GDP-Mg 2+ modelled into the density. Side chains were
omitted in these two panels, corresponding to the lowest-resolution parts of
the map. m, Map-to-model FSC plots for the top-ranked LRRK2RCKW models,
with (left) or without (right) GDP-Mg 2+ (right) in the ROC domain.
The 0.143 FSC values are reported in Supplementary Table 1. n, Size exclusion
chromatography–multiple angle light scattering (SEC–MALS) analysis of
LRRK2RCKW under the conditions used for cryo-EM (Fig. 1). The table shows the
calculated molecular weights (MW) of LRRK2RCKW according to SEC standards
and MALS.
Article

Extended Data Fig. 3 | See next page for caption.

Extended Data Fig. 3 | Comparisons between LRRK2 and other kinases and
modelling of the LRR into LRRK2RCKW. a, View of the LRRK2RCKW atomic model
with COR-A, COR-B and kinase domains coloured. The N- and C-lobes of the
kinase are labelled, as is the αC helix in the N-lobe. b, c, The FAK-FERM (PDB
code 2J0J)17 (b) and CDK2-cyclin A (PDB code 2CCH)19 (c) complexes, shown in
the same orientation as the kinase in a. The αC helix of CDK2 is also labelled.
d, Same view as in a with only the kinase domain and the C-terminal helix
coloured. e, Rotated view of the LRRK2 kinase domain with the C-terminal helix
facing the viewer. f, g, CDKL3 (PDB code 3ZDU) (f) and RIPK2 (PDB code 4C8B)32
(g) shown in the same orientation as the LRRK2 kinase in e, with alpha helices
with the same general location as the LRRK2 C-terminal helix coloured in green.
h, KSR2-MEK1 complex (PDB code 2Y4I), with the kinase oriented as in e (left)
and after removing KSR2 for clarity (right). The alpha helix associated with the
kinase is shown in green. i, HCK (PDB code 2HCK) in complex with its SH2 and
SH3 domains with the kinase oriented as in e (left), and after removal of the SH2
and SH3 domains for clarity (right). A remaining alpha helix from the SH2
domain is shown in yellow. j, Front view of the LRRK2 kinase with the C-spine
and R-spine residues coloured in grey and white, respectively. k, Close-up of
the DYG motif and neighbouring R-spine residues. A putative hydrogen bond
between Y2018 and the backbone carbonyl of I1933 is shown (O–O distance:
2.7 Å). This interaction provides a structural explanation for the
hyperactivation of the kinase resulting from a Y2018F mutation38, which would
release the activation loop. l, Crystal structure of the LRR–ROC–COR(A/B)
domains from C. tepidum Roco (PDB code 6HLU)7. m, Homology model for
human LRR–ROC–COR(A/B) based on the C. tepidum Roco structure (from
SWISS-MODEL). n, Chimeric model combining LRRK2RCKW and the homology
model for the LRR domain from m obtained by aligning their ROC–COR(A/B)
domains. o, p, Two views of the hybrid LRRK2 LRCKW model. q, Close-up showing
the proximity between the active site of the kinase (with the side chains of its
DYG motif shown) and the S1292 autophosphorylation site on the LRR. The
close-up also highlights the proximity between N2081, a residue implicated in
Crohn’s disease, and the LRR.
Article

Extended Data Fig. 4 | See next page for caption.

Extended Data Fig. 4 | Comparison between LRRK2RCKW and integrative
models built into cryo-ET maps of LRRK2 filaments in cells and docking of
LRRK2RCKW into those maps. a, Root-mean-square deviation (r.m.s.d.)
between the atomic model of LRRK2RCKW and each of the 1,167 integrative
models previously generated5. r.m.s.d. values were calculated in Chimera62
using 100% residue similarity and with pruning iterations turned off. r.m.s.d.
values are grouped into 53 clusters of related models (see ref. 5 for details), with
the mean and standard deviation shown whenever the cluster contains two or
more models. Integrative models that gave the lowest, median and highest
r.m.s.d. values are shown. The models are coloured according to the per-
residue r.m.s.d. with the atomic model of LRRK2RCKW. b, The WD40s in the
crystal structure of a dimer of the LRRK2 WD40 (PDB code: 6DLP)9 were
replaced with the WD40s from our cryo-EM structure of LRRK2RCKW. c, The
resulting dimer was fitted into the 14 Å cryo-ET map of cellular microtubule-
associated LRRK2 filaments 5. d, Two views of the same fitting shown in c,
displayed with a higher threshold for the map to highlight the fitting of the
WD40 β-propellers into the density. The white arrows point towards the holes
at the centre of the β-propellers densities. e, Four copies of LRRK2RCKW were
docked into the cryo-ET map by aligning their WD40 domains to the docked
WD40 dimer. f, Model containing the four aligned LRRK2RCKW. g–j, Modelling of
the kinase-closed form of LRRK2RCKW. g, h, The structure of ITK bound to an
inhibitor (PDB code 3QGY)63, which is in a closed conformation, was aligned to
LRRK2RCKW using only the C-lobes of the two kinases. i, The N-terminal portion
of LRRK2RCKW, comprising ROC, COR-A, COR-B and the N-lobe of the kinase, was
aligned to ITK using only the N-lobes of the kinases. ROC, COR-A and COR-B
were moved as a rigid body in this alignment. j, Kinase-closed model of
LRRK2RCKW.
Article
Extended Data Fig. 5 | Ab initio models for cryo-EM of LRRK2RCKW dimers
and cryo-EM analysis of WD40- and COR-mediated dimers of LRRK2RCKW in
the presence of the inhibitor MLi-2. a, An initial dataset was collected from a
sample of LRRK2RCKW incubated in the presence of the kinase inhibitor MLi-2
and dimers were selected. b, Representative two-dimensional class averages
used for ab initio model building. c, Ab initio models with the structure of
LRRK2RCKW docked in. d, Volumes generated form the molecular models in b,
filtered to 30 Å resolution. e, Projections of the volumes in d shown in the same
order as their corresponding 2D class averages in b. f, Data processing strategy
for obtaining cryo-EM structures of WD40- and COR-mediated dimers of
LRRK2RCKW in the presence of the inhibitor MLi-2. The models used during this
processing (Methods) are those shown in d along with an additional linear
trimer (Methods) used for particle sorting.
Extended Data Fig. 6 | See next page for caption.
Article
Extended Data Fig. 6 | Cryo-EM analysis of a monomer and WD40- and
COR-mediated dimers of LRRK2RCKW in the absence of inhibitor (apo) and
dimerization of LRRK2RCKW outside the filaments. a, Data-processing
strategy for obtaining cryo-EM structures of a monomer and WD40- and COR-
mediated dimers of LRRK2RCKW in the absence of inhibitor. The models used
during the processing of the dimers (Methods) are those shown in Extended
Data Fig. 5d, along with an additional linear trimer (Methods) used for particle
sorting. The models used for processing of the monomer (Methods) were the
same dimer models as in Extended Data Fig. 5d (used for particle sorting) in
addition to a monomer model generated from our LRRK2RCKW model (used for
refinement). b, Two-dimensional (2D) class averages of WD40- and COR-
mediated LRRK2RCKW dimers obtained in the absence of inhibitors (apo) or in
the presence of either ponatinib or MLi-2. The same molecular models of the
two dimers shown in Fig. 3 are shown on the left but in orientations similar to
those represented by the 2D class averages shown here. For each class average,
a projection from the corresponding model in the best-matching orientation is
shown to its left. c, Two copies of the LRRK2RCKW structure were aligned to the
ROC–COR domains of the LRR–ROC–COR structure from the C. tepidum Roco
protein (PDB code 6HLU) to replicate the interface observed in the bacterial
homologue in the context of the human protein. This panel shows a
comparison between the dimer modelled based on the C. tepidum LRR–ROC–
COR structure and the dimer observed for LRRK2RCKW in this work. Although the
bacterial structure shows a dimerization interface that involves the GTPase
(ROC), LRRK2RCKW interacts exclusively through its COR-A and -B domains, with
the ROC domains located away from this interface. The two arrangements are
shown schematically in cartoon form below the structures.
Extended Data Fig. 7 | See next page for caption.
Article
Extended Data Fig. 7 | Properties of the microtubule-associated LRRK2RCKW
filaments. a, b, The LRRK2RCKW structure solved in this work (a) was split at the
junction between the N- and C-lobes of the kinase domain (L1949-A1950) (b).
c, Docking of the two halves of LRRK2RCKW into a cryo-EM map of a LRRK2RCKW
dimer solved in the presence of MLi-2. The dimer map is the same one shown in
Fig. 3 and Extended Data Figs. 10 and 11. d, The model obtained in c was docked
into cryo-EM maps of either WD40- or COR-mediated dimers obtained in the
presence of MLi-2. e, Molecular models resulting from the docking in d.
f, Aligning, in alternating order, copies of the dimer models generated in d and
e results in a right-handed filament with dimensions compatible with those of a
microtubule, and its ROC domains pointing inwards (see Fig. 3g, h for more
details). g, Docking of the two halves of LRRK2RCKW into a cryo-EM map of a
LRRK2RCKW monomer solved in the absence of inhibitor (apo). The map is the
one shown in Fig. 1g and Extended Data Fig. 6. h, Three-way comparison of
LRRK2RCKW (with domain colours) and the models resulting from the dockings
into the MLi-2 WD40-mediated dimer map (c) (dark blue) and apo monomer
map (g) (light blue). The three structures were aligned using the C-lobes of
their kinases and the WD40 domain. The superposition illustrates that the
docking into the apo map results in a structure very similar to that obtained
from the trimer (Fig. 1) and that the presence of MLi-2 leads to a closing of the
kinase. i, Molecular model of the microtubule-associated LRRK2RCKW filament
obtained by docking a fragment of a microtubule structure (PDB code 6O2S)
into the corresponding density in the sub-tomogram average (Fig. 2a). j, Same
view as in i with the models shown as surface representations coloured by their
Coulomb potential. k, l, ‘Peeling off’ of the structure shown in j, with the
LRRK2RCKW filament seen from the perspective of the microtubule surface (k)
and the microtubule surface seen from the perspective of the LRRK2RCKW
filament (l). Note that the acidic C-terminal tubulin tails are not ordered in the
microtubule structure and are therefore not included in the surface charge
distributions. The Coulomb potential colouring scale is shown on the right.
Extended Data Fig. 8 | Inhibition of motor motility by wild-type and I2020T
mutant LRRK2RCKW. a, Example kymographs showing that increasing
concentrations of LRRK2RCKW reduce kinesin runs. b, Example kymographs
showing that 25 nM LRRK2RCKW reduces dynein runs. c, Representative
kymographs of kinesin motility in the presence or absence of wild-type and
I2020T mutant LRRK2RCKW. d, The percentage of motile kinesin events per
microtubule in the absence of LRRK2 or in the presence of 25 nM wild-type or
I2020T mutant LRRK2RCKW. Data are mean ± s.d. (n = 12 microtubules per
condition quantified from two independent experiments). There is a
significant difference between 0 nM and both 25 nM RCKW conditions
(P < 0.0001), but no significant (ns) difference between the inhibitory effects of
wild-type LRRK2RCKW versus I2020T mutant LRRK2RCKW as calculated using the
Kruskal–Wallis test with Dunn’s posthoc for multiple comparisons (compared
to no LRRK2RCKW).
Article

Extended Data Fig. 9 | See next page for caption.

Extended Data Fig. 9 | Type II kinase inhibitors rescue kinesin and dynein
motility. a–e, Ponatinib is a type II, ‘DFG out’ inhibitor. a, Superposition of the
structures of Ponatinib-bound RIPK2 (PDB code 4C8B)32 and IRAK4 (PDB code
6EG9). Ponatinib is shown in yellow, and the DYG motif residues are shown in
white. b, c For comparison, the structures of Roco4 bound to LRRK2-IN-1 (PDB
code 4YZM)35, a LRRK2-specific type I, ‘DFG in’ inhibitor (b), and a model of
MAPK1 bound to MLi-2 (PDB code 5U6I)22, another LRRK2-specific type I, ‘DFG
in’ inhibitor (c) are shown. The inhibitor and DFG residues are coloured as
in a. d, The structures in a–c, as well as the kinase from LRRK2RCKW are shown
superimposed. The colour arrowheads point to the N-lobe β-sheet to highlight
the difference in conformation between kinases bound to the two different
types of inhibitors. Note that the LRRK2RCKW kinase is even more open than the
two ponatinib-bound kinases. e, Rotated view of d, now highlighting the
position of the N-lobe αC helix. An additional alpha helix in the N-lobe of MAPK1
was removed from this view for clarity. f, The kinase inhibitors MLi-2 (1 μM),
LRRK2-IN-1 (1 μM), ponatinib (10 μM) and GZD-824 (10 μM) all inhibit the
LRRK2RCKW kinase activity in vitro compared to a DMSO control. A western blot
using a phospho-specific antibody to Rab8A at the indicated time points is
shown. g, A dose–response curve showing the percentage of motile kinesin
events per microtubule as a function of ponatinib concentration with
LRRK2RCKW (25 nM) or without LRRK2RCKW. Data are mean ± s.d. (from left to
right: n = 12, 18, 16, 14 and 9 microtubules quantified from one experiment).
****P < 0.0001, Kruskal–Wallis test with Dunn’s posthoc for multiple
comparisons, compared to DMSO without LRRK2RCKW. h, Dose–response curve
of run lengths from data in g represented as a cumulative frequency
distribution. From top to bottom: n = 654, 173, 584, 293 and 129 motile kinesin
events. Mean decay constants (tau) ± confidence interval are (from top to
bottom) 2.736 ± 0.113, 1.291 ± 0.181, 2.542 ± 0.124, 2.285 ± 0.134, and
1.653 ± 0.17. i, Representative kymographs of kinesin and dynein with DMSO or
type II inhibitors with or without LRRK2RCKW. j, The type II kinase inhibitors
ponatinib and GZD-824 rescue kinesin run length, represented as a cumulative
frequency distribution of run lengths with LRRK2RCKW (25 nM) or without
LRRK2RCKW. From top to bottom: n = 893, 355, 507, 499, 524 and 529 runs from
two independent experiments. Mean decay constants (tau) ± 95% confidence
intervals are (from top to bottom) 2.070 ± 0.058, 0.8466 ± 0.091, 1.938 ± 0.065,
2.075 ± 0.07, 1.898 ± 0.065 and 1.718 ± 0.064. Data were resampled with
bootstrapping analysis and statistical significance was established using a one-
way ANOVA with Dunnett’s test for multiple comparisons. DMSO run lengths
were significantly different (P < 0.0001) between conditions (0 vs 25 nM
RCKW). Ponatinib (0 vs 25 nM RCKW) and GZD-824 (0 vs 25 nM LRRK2) were not
significant. k, As in j but with dynein. From top to bottom: n = 659, 28, 289, 306,
254 and 339 runs from two independent experiments. Mean decay constants
(tau) ± 95% confidence intervals; micrometres are 4.980 ± 0.147, 0.846 ± 0.415,
4.686 ± 0.142, 4.445 ± 0.172, 3.156 ± 0.09, 3.432 ± 0.188 (from top to bottom).
Statistical significance as in j and run lengths were significantly different
(P < 0.0001) between DMSO conditions (0 vs 25 nM RCKW), and not significant
for ponatinib or GZD0824 conditions. The DMSO conditions are reproduced
from Fig. 4f for comparison. l, Expression levels of GFP-LRRK2 (I2020T) in 293T
cells treated with either DMSO or GZD-824 (5 μM). An immunoblot with anti-
GFP (LRRK2) and anti-GADPH (loading control), which is a representative
image from three replicates, is shown. m, Quantification of GFP–LRRK2
(I2020T) expression levels from western blots similar to l. Data are mean ± s.d.
(n = 3 per condition). GZD-824 is not significantly different from the DMSO-
treated control (Mann–Whitney test). n, 293T cells immunostained for tubulin
showing that the microtubule architecture is not affected by GZD-824 or
MLi-2 compared to DMSO treatment. See Supplementary Table 1 for all source
data and replicate information.
Article
Extended Data Table 1 | Cryo-EM data collection and model refinement statistics

The model refinement statistics are reported for four different types of model, two including GDP-Mg2+ in the ROC domain and two excluding it. In each case, we report statistics for two types
of model: ‘Monomer w/interfaces’ consists of an LRRK2RCKW monomer plus fragments from the neighbouring monomers in the C3 trimer that were used during model building and refinement;
‘Top 10 monomers’ are the top-10 results from Rosetta Relax with the neighbouring fragments removed after processing in Rosetta. PDB accession numbers for the models and the EMD code
of the maps used for model-building and refinement are indicated. EMD-21250 contains both the C3 map of the LRRK2RCKW trimer used to build the COR-B, kinase and WD40 domains and the
signal-subtracted monomer used to build the ROC and COR-A domains. The final models reported here were refined into the signal-subtracted monomer map (Methods) *C3 reconstruction.
#
Signal-subtracted monomer.
a
WD40-mediated dimer.
b
COR-mediated dimer.
**Numbers represent the average of the values for all 10 models.
 nature research | reporting summary
Samara L Reck-Peterson and Andres E
Corresponding author(s): Leschziner
Last updated by author(s): Aug 6, 2020

Reporting Summary
Nature Research wishes to improve the reproducibility of the work that we publish. This form provides structure for consistency and transparency
in reporting. For further information on Nature Research policies, see Authors & Referees and the Editorial Policy Checklist.

Statistics
For all statistical analyses, confirm that the following items are present in the figure legend, table legend, main text, or Methods section.
n/a Confirmed
The exact sample size (n) for each experimental group/condition, given as a discrete number and unit of measurement
A statement on whether measurements were taken from distinct samples or whether the same sample was measured repeatedly
The statistical test(s) used AND whether they are one- or two-sided
Only common tests should be described solely by name; describe more complex techniques in the Methods section.

A description of all covariates tested

A description of any assumptions or corrections, such as tests of normality and adjustment for multiple comparisons
A full description of the statistical parameters including central tendency (e.g. means) or other basic estimates (e.g. regression coefficient)
AND variation (e.g. standard deviation) or associated estimates of uncertainty (e.g. confidence intervals)

For null hypothesis testing, the test statistic (e.g. F, t, r) with confidence intervals, effect sizes, degrees of freedom and P value noted
Give P values as exact values whenever suitable.

For Bayesian analysis, information on the choice of priors and Markov chain Monte Carlo settings
For hierarchical and complex designs, identification of the appropriate level for tests and full reporting of outcomes
Estimates of effect sizes (e.g. Cohen's d, Pearson's r), indicating how they were calculated
Our web collection on statistics for biologists contains articles on many of the points above.

Software and code

Policy information about availability of computer code
Data collection For electron microscopy experiments, data was collected with Leginon. All the electron microscopy data collection software sources are
referenced in the methods section. For light microscopy experiments, data was collected with Nikon Elements Software (commercially
available). For Western blot, data was collected using Image Studio v5.2 (Li-COR).

Data analysis For electron microscopy experiments, data was processed with Appion, GCTTF, MotionCor2, FindEM, Cryolo, Relion3, and Cryosparc2. All
the electron microscopy data processing software sources are referenced in the methods section. For light microscopy experiments, data
was analyzed with ImageJ and used to make image z-maximum projections and kymographs. Graphpad Prism8 were used for all
statistical analysis of light microscopy data. For Western blots, data was quantified using EmpiriaStudio software (Li-COR).
For manuscripts utilizing custom algorithms or software that are central to the research but not yet described in published literature, software must be made available to editors/reviewers.
We strongly encourage code deposition in a community repository (e.g. GitHub). See the Nature Research guidelines for submitting code & software for further information.

Data
Policy information about availability of data
All manuscripts must include a data availability statement. This statement should provide the following information, where applicable:
- Accession codes, unique identifiers, or web links for publicly available datasets
October 2018

- A list of figures that have associated raw data

- A description of any restrictions on data availability

For electron microscopy experiments, maps and coordinates are deposited on the PDB and EMDB. No image sets or particle stacks will be made available. All the
raw data that went into the biochemical and cell biological analyses for Figures 4 and 5 (and associated Extended Data Figures) were deposited in a spreadsheet
with the manuscript.

1
 nature research | reporting summary
Field-specific reporting
Please select the one below that is the best fit for your research. If you are not sure, read the appropriate sections before making your selection.
Life sciences Behavioural & social sciences Ecological, evolutionary & environmental sciences
For a reference copy of the document with all sections, see nature.com/documents/nr-reporting-summary-flat.pdf

Life sciences study design

All studies must disclose on these points even when the disclosure is negative.
Sample size For all experiments, we determined the sample size by following conventions in the field.

Data exclusions For single molecule kinesin experiments, clear bright aggregates (less than 5% of runs) were excluded from the analysis, as these runs display
longer run lengths than typical single-molecule kinesin runs (Brouhard, 2010, Methods Cell Biol). No conclusions change with the addition or
exclusion of these aggegrates, and we would be happy to provide the data without exclusion of aggregates if deemed necessary.

Replication All single molecule experiments in Figure 4 and 5 (including dynein and kinesin data with or without drugs) were performed with at least two,
but up to four technical replicates on different days (except EDF10g,h that was only performed with one technical replicate). Major findings
with kinesin and dynein single molecule data have been confirmed by two different protein preps. All cellular data from Figure 5 was
quantified from at least four, but up to ten technical replicates (defined in the Methods as at least 20 cells per coverslip) and independent
experiments were performed on multiple days as outlined in the Methods section.

Randomization This is not relevant. We have no data involving organisms or subjects that would require randomization.

Blinding For the cell biology data in Fig 5c-f, experimenter was blinded to conditions for both the imaging acquisition and analysis of LRRK2 filaments.

Reporting for specific materials, systems and methods

We require information from authors about some types of materials, experimental systems and methods used in many studies. Here, indicate whether each material,
system or method listed is relevant to your study. If you are not sure if a list item applies to your research, read the appropriate section before selecting a response.

Materials & experimental systems Methods

n/a Involved in the study n/a Involved in the study
Antibodies ChIP-seq
Eukaryotic cell lines Flow cytometry
Palaeontology MRI-based neuroimaging
Animals and other organisms
Human research participants
Clinical data

Antibodies
Antibodies used mouse anti-GFP (Santa Cruz, clone: B-2, Cat: sc-9996, Lot: ); chicken anti-GFP (AvesLabs, Cat: GFP-1010, Lot: GFP879484); rabbit
anti-alpha-tubulin (ProteinTech, Cat: 11224-1-AP); mouse anti-GAPDH (ProteinTech, Cat: 60004-1-Ig)

Validation All antibodies used are well-validated and highly-specific commercially available antibodies. For LiCOR quantification, linear

Eukaryotic cell lines

Policy information about cell lines
Cell line source(s) HEK293T used were from ATCC (CRL-3216)
October 2018

Authentication ATCC authenticated

Mycoplasma contamination Every new cell line we receive is tested for mycoplasma before expanding and freezing. After thawing, each cell line is tested
again. Once every three months, our lab tests all growing cells for mycoplasma as well. The cells we used in our experiments
were last test on 10/16/19 and did not contain contamination.

Commonly misidentified lines None

(See ICLAC register)

2
Article

Structures and pH-sensing mechanism of

the proton-activated chloride channel

https://doi.org/10.1038/s41586-020-2875-7 Zheng Ruan1,4, James Osei-Owusu2,4, Juan Du1, Zhaozhu Qiu2,3 ✉ & Wei Lü1 ✉

Received: 4 April 2020

Accepted: 14 August 2020 The proton-activated chloride channel (PAC) is active across a wide range of
Published online: 4 November 2020 mammalian cells and is involved in acid-induced cell death and tissue injury1–3. PAC
has recently been shown to represent a novel and evolutionarily conserved protein
Check for updates
family4,5. Here we present two cryo-electron microscopy structures of human PAC in
a high-pH resting closed state and a low-pH proton-bound non-conducting state.
PAC is a trimer in which each subunit consists of a transmembrane domain (TMD),
which is formed of two helices (TM1 and TM2), and an extracellular domain (ECD).
Upon a decrease of pH from 8 to 4, we observed marked conformational changes in
the ECD–TMD interface and the TMD. The rearrangement of the ECD–TMD interface
is characterized by the movement of the histidine 98 residue, which is, after
acidification, decoupled from the resting position and inserted into an acidic pocket
that is about 5 Å away. Within the TMD, TM1 undergoes a rotational movement,
switching its interaction partner from its cognate TM2 to the adjacent TM2. The
anion selectivity of PAC is determined by the positively charged lysine 319 residue
on TM2, and replacing lysine 319 with a glutamate residue converts PAC to a
cation-selective channel. Our data provide a glimpse of the molecular assembly of
PAC, and a basis for understanding the mechanism of proton-dependent activation.

Acidic pH is crucial for the function of intracellular organelles in the
secretory and endocytic pathways. It is also one of the pathological
hallmarks of many diseases, including cerebral and cardiac ischaemia,
cancer, infection and inflammation. The activity of PAC is stimulated by
the lowering of extracellular pH and has been recorded in a wide range
of mammalian cells1. By mediating the influx of Cl− and subsequent cell
swelling, PAC is implicated in acid-induced cell death2,3. We and others
recently used unbiased RNA interference screens4,5 to identify a novel
gene, PACC1 (also known as TMEM206), that encodes the PAC channel.
Our functional studies revealed that PAC has a key role in acid-induced
neuronal cell death in vitro and in ischaemic brain injury in mice4,6.
With no obvious sequence homology to other membrane proteins,
PAC represents a completely new family of ion channels4,5. PAC is highly
conserved across vertebrates and is predicted to have two transmem-
brane helices4,5, similar to the acid-sensing ion channel (ASIC) and the
epithelial sodium channel (ENaC)7,8. Although the structure and func-
tion of ASIC have been extensively studied7,9–12, the architecture of PAC
and the mechanisms that underlie its pH sensing and anion selectivity
are unknown. To address these questions, we determined structures of
human PAC using single-particle cryo-electron microscopy (cryo-EM)
combined with patch-clamp electrophysiological studies.
Structural determination
PAC is activated at a pH below 5.5 at room temperature, and is maxi-
mally stimulated by protons at around pH 4.6–41. We determined
cryo-EM structures of PAC reconstituted in lipid nanodiscs at pH 8
(pH8-PAC) and pH 4 (pH4-PAC) with estimated resolutions of 3.60
and 3.73 Å, respectively (Extended Data Figs. 1a–c, 2, 3, Extended Data
Table 1). The maps were of sufficient quality to carry out de novo model
building of the protein (Fig. 1, Extended Data Fig. 4a–d, Extended Data
Table 1). The cytoplasmic N and C termini (residues 1–60 and 339–350
in pH8-PAC and 1–52 and 340–350 in pH4-PAC) are disordered in our
cryo-EM maps.
Overall architecture
PAC is a trimer. It has a small, ball-shaped ECD sitting on top of a
slim and elongated TMD that contains two transmembrane heli-
ces (TM1 and TM2) in each subunit (Fig. 1a, b, e, f). This trimeric
two-transmembrane-helix architecture is reminiscent of ASIC
(Extended Data Fig. 5a–e) and ENaC7,8. The ECD of PAC is heavily gly-
cosylated, with four N-glycosylation sites in each subunit (Fig. 1c, g)—
consistent with a previous report5 and a deglycosylation assay
(Extended Data Fig. 1d).
Alkaline and acidic pH yielded two PAC structures with distinct
shapes—pH4-PAC is shorter and bulkier than pH8-PAC, and they differ
mainly at the TMD and the ECD–TMD interface. At pH 8, the TM1 helix
runs nearly parallel to and forms interactions only with its cognate TM2
(Fig. 1b, d). When the pH drops to 4, TM1 switches its interaction from
its cognate TM2 to the adjacent TM2 (Fig. 1f, h). This domain-swapped
movement of TM1 has not, to our knowledge, been observed in any

1
Department of Structural Biology, Van Andel Institute, Grand Rapids, MI, USA. 2Department of Physiology, Johns Hopkins University School of Medicine, Baltimore, MD, USA.
3
Solomon H. Snyder Department of Neuroscience, Johns Hopkins University School of Medicine, Baltimore, MD, USA. 4These authors contributed equally: Zheng Ruan, James Osei-Owusu.
✉e-mail: zhaozhu@jhmi.edu; wei.lu@vai.org

350 | Nature | Vol 588 | 10 December 2020

 a b 90° c 180° d
Disulfide
bond TM1 TM2
M2
2
ECD

N TM1
TM2 N
TM1 TM2
18.4 Å ECD–TMD 57.7 Å
interface NAG C
TM1–β1
linker H98 (N148) NAG NAG
(N190) (N162)
β14–TM2 TM1 10.4 Å
NAG
TMD linker TM2 (N155)
27.8 Å ECD 30.2 Å TM2 TM1
C N

e f 90° g 180° h
Disulfide
bond TM2 TM1
N
TM1
TM2 C

TM2
28.9 Å
H98 50.6 Å TM1
TM1
TM1–β1
Pre-TM2
linker TM2
TM2
TM1
25.9 Å ECD 8.9 Å
30.8 Å
N

Fig. 1 | Overall architecture of PAC. a, e, Cryo-EM maps of pH8-PAC and
pH4-PAC viewed parallel to the membrane. The map refined without using a
mask is shown as a transparent envelope. The horizontal dimension of the ECD–
TMD interface is represented by the distance between the Cα atoms of adjacent
His98 residues. The density for His98 is coloured in yellow in the pH8-PAC (a)
and pH4-PAC (e) maps. b, f, Atomic models of pH8-PAC and pH4-PAC. The green
subunit is shown as a cartoon and the other two subunits are shown in surface
representation. The distances between the centre of mass of the ECD and of the
TMD in pH8-PAC (b) and pH4-PAC (f) are shown on the right. c, g, The ECDs of
pH8-PAC and pH4-PAC viewed from the extracellular side. Four putative
glycosylation sites (Asn148, Asn155, Asn162 and Asn190) are labelled in c. The
distances between the centre of mass of the ECDs of each subunit in pH8-PAC
(c) and pH4-PAC (g) are shown by the triangles. NAG, N-acetyl-d-glucosamine.
d, h, The TMDs of pH8-PAC and pH4-PAC viewed from the intracellular side. A
light salmon arrow in d indicates the rotation of TM1 of PAC after acidification
to pH 4. The relative position and distance of TM1 and TM2 in pH8-PAC (d) and
pH4-PAC (h), which are represented by the Cα atoms of Iso73 and Lys319,
respectively, are shown at the bottom. The double-headed arrow indicates the
interaction between TM1 and TM2.

other two-transmembrane-helix channels,12,13 implying a novel gating
mechanism.
The ECD–TMD interface consists of part of the TM1 helix on the extra-
cellular side and two linkers that connect the TMD and the ECD—the
TM1–β1 linker and the β14–TM2 linker. This interface differs substan-
tially between the two PAC structures (Fig. 1b, f). At pH 8, TM1 and the
short TM1–β1 linker hold the adjacent ECD through an ‘anchor’ residue,
His98, while the β14–TM2 linker is extended as a loop close to the pore
axis (Fig. 1b). At pH 4, the β14–TM2 linker is remodelled into a short
pre-TM2 helix, and the TM1–β1 linker moves outward, causing a verti-
cal compression and an expansion of the ECD–TMD interface (Fig. 1a,
b, e, f, Supplementary Video 1). In tandem with the rearrangement of
the ECD–TMD interface, the ECD at pH 4 shows a vertical movement
towards the TMD and a contraction towards the pore axis, resulting in
a shorter overall structure and a more compact ECD in comparison to
that at pH 8 (Fig. 1b, c, f, g).
Structure of single subunits
At pH 8, each PAC protomer adopts an arm-like structure, with the
ECD as the hand, the ECD–TMD interface as the wrist and the TMD as
the forearm (Extended Data Fig. 4e). The hand-like ECD is composed
of a palm, a finger, a thumb and a β-ball domain, all of which consist
of β-strands except for the thumb domain, which contains two short
α-helices (Extended Data Fig. 4e–g). The finger and the β-ball domains
are connected by a disulfide bond (Cys128–Cys149), forming a rigid
structure that occupies the peripheral region of the ECD. The con-
nection between the ECD and the TMD is achieved by the TM1–β1 and
β14–TM2 linkers, which together form the wrist domain.
The TMD consists of two transmembrane helices, TM1 and TM2, at the
N terminus and C terminus of the protein, respectively. TM1 contains
mostly hydrophobic residues and makes direct contacts with the lipid
bilayer. TM2 contains both hydrophilic and hydrophobic residues and
lines the ion-conducting pore. Although the ECD mostly maintains
its conformation in both pH states, the ECD–TMD interface and TMD
differ substantially at pH 4 and pH 8, characterized by the distinct con-
formations of His98 and TM1 (Extended Data Fig. 4e, f). At pH 8, TM1
is approximately parallel to TM2, whereas at pH 4, the two transmem-
brane helices form an angle of 64°.
The TM2 helix of PAC is a continuous α-helix and differs from the
ASIC TM2, which has a characteristic two-segment structure and a
Gly-Ala-Ser belt10 (Extended Data Fig. 5b, e). The ECD of PAC shows
strong similarities to the β-sheet core of the ECD in ASIC, despite sharing
limited protein-sequence identity (Extended Data Figs. 5e, 6). Notably,
the PAC ECD lacks the large exterior helical structures of the ASIC ECD,
which are involved in pH sensing14 (Extended Data Figs. 5a–d, 6), so PAC
must have a different pH-sensing mechanism.
Channel assembly
The major interactions between PAC subunits occur at the ECD, the
ECD–TMD interface and in the upper part of the TMD. The lower part

Nature | Vol 588 | 10 December 2020 | 351

Article
a b c d e
Upper Palm
β11 D91 N302 TM1
ECD β14 β10 N305
β1 R87
Lower β1 Q296 W304
ECD β1 β14 K248
Finger S102 F83 L309 A308
ECD–TMD
interface I298 G312 C311
H98
M101 A316 L315
TM1
TMD β12 F318
F196 β14–TM2 K319
β8 αA αB TM1
linker TM1 TM2 TM2
M101
β-ball β1 V103
TM1–TM2 TM2–TM2
interface interface
f Upper
g h i j
Acidic β1–β2 linker
ECD pocket
D109
Lower β10
ECD β1 β11
β14 TM1 TM2
E107
ECD–TMD
L75
interface E250 β1 TM2
I71 TM1
V103 H98 Y74
β14 β1 A320
TMD β12 L70 A321
F196 M101
αA αB
β8 V103 Pre-TM2 A324
β1 TM1
TM1 TM2 S327

Fig. 2 | Intersubunit interfaces. Each row represents the same view of
pH8-PAC (top row; a–e) and pH4-PAC (bottom row; f–j). a, f, The overall
structure of PAC shown in cartoon and surface representation. The ECD is
divided into the upper ECD and lower ECD for discussion. b, g, The upper ECD
viewed from the extracellular side. Phe196, which mediates the intersubunit
interaction in the upper ECD, is shown as spheres. c, h, The lower ECD viewed
from the extracellular side. At pH 8, Met101 at the beginning of the β1 strand is
in the centre of the lower ECD (c, bottom right). At pH 4, the lower ECD
undergoes a clockwise inward rotation so that Val103 in the middle of the β1
strand moves to the centre of the lower ECD (h, bottom right). d, i, The ECD–
TMD interface viewed parallel to the membrane. e, j, The interaction interfaces
at the TMD.

of the TMD lacks extensive interactions and is thus flexible. Analysis
of the conformational changes at the ECD revealed a rigid-body con-
traction of the entire ECD and an iris-like rotation of the lower ECD
(Supplementary Video 1). We thus looked at both the upper ECD (finger
and β-ball domains) and the lower ECD (palm and thumb domains) to
study their intersubunit interfaces (Fig. 2a, f).
At pH 8, both the upper and the lower ECD have loose intersubu-
nit interfaces with an obvious gap between subunits (Fig. 2b, c). The
upper ECD has two major intersubunit interactions. The first is formed
between the adjacent β8 and β12 strands that run approximately paral-
lel to each other. The second is formed between the β6–β7 linker and
the adjacent finger domain, where the Phe196 residue on the β6–β7
linker is inserted into the finger domain, forming both hydrophobic
and cation–π interactions (Fig. 2b, g, Extended Data Fig. 1e). Substitu-
tion of Phe196 with an alanine residue (F196A) yielded a misassembled
mutant protein that exhibited a markedly decreased channel activity
compared to the wild type (Extended Data Fig. 1f, g). This suggests that
the upper ECD has an important role in channel assembly. The lower
ECD has a single major interface at the centre, where the three Met101
residues on the N terminus of the β1 strand tightly interact with each
other. This interface disconnects the central pore from the ECD to the
TMD (Fig. 2c). At pH 4, the gaps in both the upper and the lower ECDs are
mostly filled, creating extensive interactions between subunits (Fig. 2g,
h). Moreover, in the centre of the lower ECD, owing to the iris-like rota-
tion of the ECD from pH 8 to pH 4, the Val103 residue in the middle of
the β1 strand now mediates the contact (Fig. 2h).
The intersubunit contact at the ECD–TMD interface is mediated
through His98 at the TM1–β1 linker. At pH 8, His98 is surrounded by
hydrophilic and hydrophobic residues of the β1 and β14 strands of the
adjacent ECD, constituting a resting ECD–TMD interface (Fig. 2d). At
pH 4, this interface is remodelled; His98 interacts with a pocket formed
by residues in the β10–β11 linker of its cognate ECD and in the β1–β2
linker of the adjacent ECD (Fig. 2i). Because this pocket is constructed
solely of negatively charged residues, we call it an ‘acidic pocket’.
At the TMD, PAC has two major intersubunit interfaces (Fig. 2e, j): the
TM1–TM2 interface and the TM2–TM2 interface. At pH 8, both inter-
faces are near the extracellular part of the TMD. At pH 4, the TM2–TM2
interface remains mostly unchanged, but the TM1–TM2 interface slips
towards the intracellular side as a result of the domain-swapped move-
ment of TM1.
Ion-conducting pathway and selectivity
The PAC channel has a central pore along the symmetry axis with wide
openings at the extracellular and intracellular ends in both pH states
(Fig. 3a–d). Within the TMD, the ion-conducting pore is lined by TM2
and the β14–TM2 linker (Fig. 3b, d). At pH 8, the ion-conducting pore is
occluded at several positions (Fig. 3e, g), thus representing a high-pH
resting closed state. At pH 4 (Fig. 3f, g), the intracellular part of the pore
has an enlarged radius of 0.82 Å, but is still not wide enough to allow
the permeation of Cl− ions, thus representing a low-pH protonated
non-conducting state. We found that PAC exhibits a strong outward
rectification such that either the open probability or the single-channel
conductance is low at 0 mV, the voltage at which the cryo-EM structures
were determined (Fig. 3h). Moreover, PAC showed a marked desensiti-
zation after prolonged treatment with a pH 4 solution (Extended Data
Fig. 7a–f). Therefore, we suggest that a closed pore in the pH4-PAC
structure represents either a pre-open state or a desensitized state.
To reveal the molecular determinants that are responsible for the
anion selectivity of PAC, we examined the positively charged residues
within the ion-conducting pore, all of which are located in the intracel-
lular half of the TM2 helix. The Lys319 residue appears to be an ideal
candidate, because it is immediately below the intracellular restric-
tion site (Leu315) and forms a positively charged ‘triad’ around the
intracellular entry point (Fig. 3e, f). In line with this hypothesis, we
found that a charge-reversing mutation (K319E) converted PAC from
an anion-selective to a cation-selective channel with a pronounced
inwardly rectifying current (Fig. 3h–j, Extended Data Fig. 8a, b). By con-
trast, mutation of two other lysine residues, K325E and K329E, resulted
in mutant proteins that behaved similarly to wild-type PAC (Extended
Data Fig. 8a–e). The crucial role of Lys319 is further supported by its
conservation across species (Extended Data Fig. 6), and by the fact that
PAC(K319C) is not functional5. Together, our data provide evidence that
Lys319 is the determinant of anion selectivity for PAC.

352 | Nature | Vol 588 | 10 December 2020

a b c d a 3Å b c
E107′
E107’
D109′
D109’

β12 E250′
E250’ 4.6 Å
β12 Palm Q296
Vestibule H98’
H98′ TM1′
TM1’
Vestibule
β14 H98′ H98 TM2′
TM2’
Fenestration β14 Acidic
Fenestration H98 TM1
β1 β1 pocket Resting
β14–TM2 β14–TM2 interface TM2
TM1’
TM1′ TM2
linker linker TM1’
TM1′ TM1 22.8°
K319 TM2′ TM1
TM2’ 47.4°
TM2 TM2

d e 6.5 2.0 × 10–16

1.0 WT
7.9 × 10–4
H98R

Normalized current
e f g H98A 1.2 × 10–7
6.0
M101 ECD–TMD ECD–TMD
ECD–TMD ECD–TMD 6.6 × 10–8
V103 Q296A
seal
seal seal

Distance along pore axis (Å)

seal

pH50
V100 0.5 E107R 5.5
Fenestration
T300 Fenestration N302
Upper
Upper 5.0
N305 gate
gate
N305 0 4.5
A308 7.0 6.0 5.0 4.0 WT H98R H98A Q296A E107R
L309 G312 pH
G312 Lower
L315 gate Fig. 4 | Mechanisms of pH sensing and channel activation. a, Superposition
L315 pH 8
K319 of a single subunit of pH8-PAC (blue) and pH4-PAC (red) aligned using the ECD
10 Å pH 4
K319
palm domain. The 3 Å centre-of-mass distance indicates the rigid-body
1 2 3 4
Pore radius (Å) movement of the ECD. b, Close-up view of the conformational change in the
WT K319E
j ECD–TMD interface in a. Structural elements and residues in the pH4-PAC
h 2 1 i WT K319E 6.1 0.05
I (nA)

8 structure are labelled with a prime symbol. Residues from adjacent subunits
–100 30
I (nA)

1 are coloured in bright and light colours, respectively. At pH 8, His98 interacts

PCl/PNa
Vrev (mV)

100
0 with Gln296. At pH 4, the side chain of His98 interacts with an acidic pocket.
–100 V (mV)
–1
100 –30
V (mV) 150 mM
–1 –2 15 mM –60
Fig. 3 | Ion-conducting pathways and anion selectivity. a, c, pH8-PAC (a) and
pH4-PAC (c) in surface representation, coloured according to the electrostatic
surface potential from −3 to 3 kT/e (red to blue). Titratable residues are
assigned to their predominant protonation state at pH 8 (a) or pH 4 (c) based on
PROPKA. b, d, The pore profiles of pH8-PAC (b) and pH4-PAC (d) models along
the symmetry axis. Pore-lining residues are shown. e, f, Enlargements of the
boxed areas in b (e) and d (f), respectively. The positions of the ECD–TMD seal
and fenestration site are labelled. g, Pore radius plots of the profiles in e, f.
h, The representative current (I)–voltage (V) relationship for wild-type (WT)
PAC and PAC(K319E). The pipette solution contains 150 mM NaCl; the bath
solution contains 150 mM (black) or 15 mM (red) NaCl. i, The reversal potential
(Vrev) of wild-type PAC and PAC(K319E) from recordings in h (n = 16 (wild type)
and n = 7 (K319E)). Data are mean ± s.e.m. Individual data points are shown as
dots. j, The relative Cl−/Na+ permeability (PCl/PNa) for wild-type PAC (n = 15) and
PAC(K319E) (n = 11) calculated from current induced at pH 5 and 100 mV. Data
are mean ± s.e.m. of the permeability ratio. Individual data points are shown as
solid dots. The average PCl/PNa permeability values are indicated at the top for
each construct.
4
c, Comparison of the TMD viewed from the intracellular side. The structures of
pH8-PAC (blue) and pH4-PAC (red) are aligned using the ECD. d, pH dose–
WT K319E
response curve of wild-type PAC and various PAC mutants. Data are
mean ± s.e.m. of the current at 100 mV, normalized to pH-4.6-induced current
(n = 10 (wild type, H98R, E107R), n = 9 (H98A) and n = 11 (Q296A)). The Hill
coefficients for wild-type PAC and PAC(E107R) are 2.44 ± 0.18 and 1.18 ± 0.19
(mean ± s.e.m.), respectively. e, pH50 value estimated from the pH dose–
response curve. The centre and error bar represent the estimated pH50 value
and s.e.m. from the nonlinear fitting in d. A one-way analysis of variance
(ANOVA) with Bonferroni post-hoc test was used to determine the significance
(P values are indicated).
the pH4-PAC structure are hydrated, whereas those in the pH8-PAC
structure are less accessible to solvent (Extended Data Fig. 8h, k). Our
data suggest that the lateral fenestrations could be an extracellular
ion-entry point that is common to two-transmembrane-helix chan-
nels9,13,15. This agrees with a previous report in which treatment with
a thiol-reactive reagent, MTSES, partially inhibited the ion-channel
activity of PAC when Thr306—which is part of the fenestration—was
replaced by a cysteine residue5.

In the ECD, the pore along the symmetry axis has a large vestibule in
the middle (Fig. 3b, d). This vestibule is constricted at the ECD–TMD Mechanisms of pH sensing and channel activation
interface by an ECD–TMD seal in both the pH8-PAC and the pH4-PAC To elucidate the pH-sensing mechanism of PAC, we compared the
structures (Fig. 3e–g, Extended Data Fig. 8f, i). This leads to the question structures of pH8-PAC and pH4-PAC. We focused on the ECD and the
of how ions might enter the ion-conducting pore from the extracellular ECD–TMD interface, because PAC is activated by extracellular acid.
side. Just below the seal, we observed three lateral fenestrations that Superimposing a single subunit revealed that, upon protonation, the
connect to the central pore. Fenestrations at similar locations have been major motion of the extracellular region occurred at the ECD–TMD
defined as an ion-entry point in both the ASIC and P2X channels9,13,15. At interface, whereas the ECD showed minor rigid-body movement
pH 8, the fenestration in PAC is formed by the extracellular portion of (Fig. 4a). This suggests that the ECD–TMD interface probably par-
the TM1 helix and the β14–TM2 linker of the adjacent subunit (Extended ticipates in pH sensing. We hypothesized that the His98 residue in the
Data Fig. 8f). The entrance is surrounded by several negatively charged TM1–β1 linker is one of the key pH sensors, because it showed a large
residues, making it unfavourable for conducting anions (Extended Data movement from the high-pH resting state to the low-pH proton-bound
Fig. 8g). At pH 4, a different fenestration is established by the β1 strand state and because its side-chain pKa value is close to the pH50 (pH of
and the pre-TM2 helix in the adjacent subunit (Extended Data Fig. 8i). half-maximal activation) value of PAC16 (Extended Data Fig. 9a).
The fenestration at pH 4 is wider than that at pH 8, and has several posi- At pH 8, His98 is in close contact with the Gln296, Iso298 and Ser102
tively charged residues lining the entry point, rendering it favourable residues of the adjacent ECD. We speculated that the side chain of His98
for anions (Extended Data Fig. 8i, j). To provide evidence that these fen- forms a hydrogen bond with the side-chain amine group of Gln296,
estrations could be extracellular ion-entry points in PAC, we performed which locks the TM1 helix in a conformation parallel to its cognate
molecular dynamics simulations and found that the fenestrations in TM2 helix (Fig. 4a, b). To investigate whether the interaction between

Nature | Vol 588 | 10 December 2020 | 353

Article
His98 and Gln296 is critical for stabilizing the channel in a resting
closed state, we engineered a disulfide bond connecting these two
residues. Indeed, double mutation of both His98 and Gln296 to cysteine
(H98C/Q296C) fixed His98 in the resting position and thus rendered
PAC insensitive to pH, whereas the control serine double mutant (H98S/
Q296S) showed increased pH sensitivity (Extended Data Fig. 9b–f). At
pH 4, the protonated His98 is decoupled from Gln296 and flipped into
the acidic pocket, which is 4.6 Å away (Fig. 4b). The acidic pocket is
formed by Glu107, Asp109 and Glu250, and interacts favourably with
the protonated His98 residue because Asp109 is predicted to remain
unprotonated at pH 4 (Extended Data Fig. 9a). The flipping of His98
pulls TM1 away from its cognate TM2 and creates a new interface with
the TM2 of the adjacent subunit (Fig. 4b). Concurrent with a 47.4° swing
of TM1, the pore-lining TM2 undergoes a counterclockwise rotation of
22.8° when viewed from the intracellular side (Fig. 4b, c).
We hypothesized that the flipping of His98 from the resting position
to the acidic pocket is a critical element for the proton-induced acti-
vation of the PAC channel. To test this hypothesis, we first generated
mutants of His98 and its interacting partner in the pH8-PAC structure
(Gln296) and examined their pH sensitivity (Fig. 4d, e). H98R, H98A and
Q296A mutants all resulted in an increased pH sensitivity by disengag-
ing the hydrogen bond between His98 and Gln296, which supports
the idea that the decoupling of protonated His98 from the resting
interface has a role in channel activation. However, the H98A and H98R
mutants showed a similar pH50 value, which is unexpected because an
alanine residue would be less attracted by the acidic pocket than an
arginine residue. This might be because an arginine residue at position
98 caused additional conformational changes rather than a simple
side-chain substitution. Next we studied Glu107, which is close to His98
in the pH4-PAC structure. Mutation of Glu107 to arginine (E107R) not
only markedly increased the pH sensitivity but also decreased the Hill
coefficient of the pH dose–response curve (Fig. 4d, e). As Glu107 has
a predicted pKa value of around 6 in the pH4-PAC structure (Extended
Data Fig. 9a), it probably also participates in PAC pH sensing. The E107R
mutation might cause a rearrangement of the acidic pocket, which
could lead to an altered interaction with His98. Consequently, the
activation of the channel might require fewer protons, resulting in an
increased pH sensitivity.
Discussion
Our work on PAC provides a glimpse of the molecular structures
and pH-sensing mechanism of a proton-activated chloride chan-
nel (Extended Data Fig. 9g). Similarly to ASIC12,14,17–19, the pH-sensing
mechanism of PAC is almost certainly determined by multiple resi-
dues, because several tested mutations altered the pH sensitivity but
none of them abolished it, and because titratable residues are dis-
tributed throughout the ECD (Extended Data Fig. 9a). Our structural
and functional data suggest that the pH4-PAC structure represents a
proton-bound pre-open state or a proton-bound desensitized state. We
acknowledge the limits of using a proton-bound non-conducting con-
formation to discuss the activation mechanism, because the TMD may
354 | Nature | Vol 588 | 10 December 2020
differ from that in an open state. Indeed, cysteine mutants of multiple
pore-lining residues in TM2 (for example, Ala316, Leu315, Gly312) can
still be accessed by the thiol-reactive reagent MTSES from the extracel-
lular side5, indicating that the ion-conducting pore and lateral fenestra-
tions in an open state are probably substantially larger than in either
of the present structures. Further studies are required to develop a
thorough understanding of this family of proton-sensitive ion channels.
Online content
Any methods, additional references, Nature Research reporting sum-
maries, source data, extended data, supplementary information,
acknowledgements, peer review information; details of author con-
tributions and competing interests; and statements of data and code
availability are available at https://doi.org/10.1038/s41586-020-2875-7.
1. Capurro, V. et al. Functional analysis of acid-activated Cl− channels: properties and
mechanisms of regulation. Biochim. Biophys. Acta 1848, 105–114 (2015).
2. Wang, H. Y., Shimizu, T., Numata, T. & Okada, Y. Role of acid-sensitive outwardly rectifying
anion channels in acidosis-induced cell death in human epithelial cells. Pflugers Arch.
454, 223–233 (2007).
3. Sato-Numata, K., Numata, T. & Okada, Y. Temperature sensitivity of acid-sensitive
outwardly rectifying (ASOR) anion channels in cortical neurons is involved in
hypothermic neuroprotection against acidotoxic necrosis. Channels 8, 278–283 (2014).
4. Yang, J. et al. PAC, an evolutionarily conserved membrane protein, is a proton-activated
chloride channel. Science 364, 395–399 (2019).
5. Ullrich, F. et al. Identification of TMEM206 proteins as pore of PAORAC/ASOR
acid-sensitive chloride channels. eLife 8, e49187 (2019).
6. Osei-Owusu, J., Yang, J., Del Carmen Vitery, M., Tian, M. & Qiu, Z. PAC proton-activated
chloride channel contributes to acid-induced cell death in primary rat cortical neurons.
Channels 14, 53–58 (2020).
7. Jasti, J., Furukawa, H., Gonzales, E. B. & Gouaux, E. Structure of acid-sensing ion channel 1
at 1.9 A resolution and low pH. Nature 449, 316–323 (2007).
8. Noreng, S., Bharadwaj, A., Posert, R., Yoshioka, C. & Baconguis, I. Structure of the human
epithelial sodium channel by cryo-electron microscopy. eLife 7, e39340 (2018).
9. Gonzales, E. B., Kawate, T. & Gouaux, E. Pore architecture and ion sites in acid-sensing ion
channels and P2X receptors. Nature 460, 599–604 (2009).
10. Baconguis, I., Bohlen, C. J., Goehring, A., Julius, D. & Gouaux, E. X-ray structure of
acid-sensing ion channel 1-snake toxin complex reveals open state of a Na+-selective
channel. Cell 156, 717–729 (2014).
11. Yoder, N. & Gouaux, E. Divalent cation and chloride ion sites of chicken acid sensing ion
channel 1a elucidated by X-ray crystallography. PLoS ONE 13, e0202134 (2018).
12. Yoder, N., Yoshioka, C. & Gouaux, E. Gating mechanisms of acid-sensing ion channels.
Nature 555, 397–401 (2018).
13. Mansoor, S. E. et al. X-ray structures define human P2X3 receptor gating cycle and
antagonist action. Nature 538, 66–71 (2016).
14. Vullo, S. et al. Conformational dynamics and role of the acidic pocket in ASIC
pH-dependent gating. Proc. Natl Acad. Sci. USA 114, 3768–3773 (2017).
15. Gao, C. et al. Roles of the lateral fenestration residues of the P2X4 receptor that contribute
to the channel function and the deactivation effect of ivermectin. Purinergic Signal. 11,
229–238 (2015).
16. Lambert, S. & Oberwinkler, J. Characterization of a proton-activated, outwardly rectifying
anion channel. J. Physiol. 567, 191–213 (2005).
17. Liechti, L. A. et al. A combined computational and functional approach identifies new
residues involved in pH-dependent gating of ASIC1a. J. Biol. Chem. 285, 16315–16329 (2010).
18. Smith, E. S. J., Zhang, X., Cadiou, H. & McNaughton, P. A. Proton binding sites involved in
the activation of acid-sensing ion channel ASIC2a. Neurosci. Lett. 426, 12–17 (2007).
19. Paukert, M., Chen, X., Polleichtner, G., Schindelin, H. & Gründer, S. Candidate amino acids
involved in H+ gating of acid-sensing ion channel 1a. J. Biol. Chem. 283, 572–581 (2008).
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in
published maps and institutional affiliations.
© The Author(s), under exclusive licence to Springer Nature Limited 2020
Methods
Data reporting
No statistical methods were used to predetermine sample size. The
experiments were not randomized and the investigators were not
blinded to allocation during experiments and outcome assessment.
Constructs and mutagenesis
For protein expression and purification, the human PAC gene (PACC1;
Uniprot ID Q9H813) from our previous study4 was subcloned into a pEG
BacMam vector20 with a thrombin cutting site, enhanced green fluores-
cent protein (eGFP) and 8×His tag in the C terminus. The pIRES2-eGFP
vector containing PACC1 was used for whole-cell patch-clamp record-
ings4. Site-directed mutagenesis was performed using the QuikChange
site-directed mutagenesis protocol (Agilent) and confirmed by Sanger
sequencing.
Mammalian cell culture, protein expression, purification and
nanodisc reconstitution
For small-scale protein expression, adherent tsA201 cells were main-
tained in Dulbecco’s modified Eagle medium (DMEM) supplemented
with 10% fetal bovine serum (FBS) at 37 °C. When the cell density
reached approximately 80% confluence, transient transfection was
performed by incubating the plasmid DNA and Lipofectamine-2000
reagent (Thermo Fisher Scientific) in Opti-MEM medium (Thermo
Fisher Scientific) using manufacturer-provided protocols. Sodium
butyrate (10 mM) was added to the adherent cells 24 h after transfec-
tion. The cells were then maintained at 30 °C to boost protein expres-
sion. The next day, the adherent cells were washed with 20 mM Tris,
150 mM NaCl, pH 8.0 (TBS) buffer, collected and stored at −80 °C.
For large-scale protein expression, we used the Bac-to-Bac baculo-
virus expression system21. Specifically, plasmid expressing wild-type
PAC was transfected into DH10α cells to produce the bacmid. Purified
bacmid was transfected into adherent Sf9 cells using Cellfectin II rea-
gent (Thermo Fisher Scientific) to produce P1 virus. P2 virus was then
generated by infecting suspension Sf9 cells with the P1 virus at a 1:5,000
(v/v) ratio. The expression of PAC protein was induced by infecting
tsA201 suspension cells in FreeStyle 293 medium (Gibco) with 7.5% P2
virus. After 8–12 h, sodium butyrate (5 mM) was added to the infected
suspension cells and the temperature was adjusted to 30 °C. Suspension
tsA201 cells were collected 70 h after infection and stored at −80 °C.
Mammalian cells infected with PAC were suspended in TBS buffer
(150 mM NaCl, 20 mM Tris HCl, pH 8.0) supplemented with a protease
inhibitor cocktail (1 mM phenylmethylsulfonyl fluoride, 2 mM pepsta-
tin, 0.8 μM aprotinin and 2 μg ml–1 leupeptin) and lysed by sonication.
Cell debris was removed by centrifugation at 3,000g for 10 min. The
membrane fraction was pooled by ultracentrifugation of the superna-
tant for 1 h at 186,000g. The membrane was then Dounce-homogenized
and solubilized in TBS buffer with 1% glyco-diosgenin (GDN) and the pro-
tease inhibitor cocktail. After 1 h, the sample was ultracentrifuged for 1 h
at 186,000g. The supernatant was applied to 2 ml talon resin preequili-
brated with TBS buffer with 0.02% GDN. The resin was washed with
20 ml TBS buffer with 0.02% GDN and 20 mM imidazole. The protein
was eluted with 8 ml TBS buffer with 0.02% GDN and 250 mM imidazole.
The eluent was concentrated to 500 μl using a 100-kDa concentrator
(MilliporeSigma). MSP3D1 protein and soybean lipid (SBL) extract
were mixed with the PAC protein sample using a molar ratio of 3:400:1
(MSP3D1:SBL:PAC). Three rounds of biobead incubation were carried
out to facilitate nanodisc reconstitution. The volume of the mixture
was then expanded to 12.5 ml so that the imidazole concentration was
10 mM. Empty nanodiscs were removed by passing the mixture through
talon resin a few times. MSP3D1–PAC was eluted using TBS buffer con-
taining 250 mM imidazole and concentrated to 500 μl. Thrombin (0.01
mg ml–1) was then added to the eluent to cleave eGFP at 4 °C overnight.
The mixture was further purified by size-exclusion chromatography
(SEC) using TBS as the running buffer. Peak fractions were combined
and concentrated to 4 mg ml–1 before grid freezing.
Cryo-grid preparation
The purified PAC nanodisc protein is in TBS buffer at pH 8. The pH 4
condition was made by adding 1 M acetic acid (pH 3.5) buffer to the
purified PAC nanodisc sample at a 1:20 v/v ratio. Fluorinated octyl
maltoside (0.5 mM) was added to the sample to help reduce protein
unfolding in the air–water interface. Quantifoil grids (Au 1.2/1.3 or Au
2/1, 300 mesh) were glow-discharged for 30 s. A cryo-grid was made
using the VitrobotMark III kept at 18 °C and 100% humidity. A volume
of 2.5 μl of the PAC nanodisc protein sample was loaded to the grid,
blotted for 1.5 s, plunge-frozen into liquid ethane and transferred into
liquid nitrogen for storage.
Cryo-EM data collection
Cryo-EM data were collected using a FEI Titan Krios transmission elec-
tron microscope equipped with a Gatan K2 Summit direct electron
detector. Automated data acquisition was facilitated by SerialEM
software in super-resolution counting mode22. Each raw movie stack
consists of 40 frames with a total dose of 49.6 e–/Å2 for 8 s. Nominal
defocus values were set to range from −1.2 to −1.9 μm.
Single-particle data analysis
For both the pH8-PAC and the pH4-PAC dataset, raw movies were first
motion-corrected using MotionCor v.1.2.1 (ref. 23). The contrast transfer
function (CTF) of each micrograph was estimated using Gctf v.1.06
(ref. 24) or CTFFIND25. Template-based particle picking was conducted
using Gautomatch v.0.53 (https://www2.mrc-lmb.cam.ac.uk/down-
load/gautomatch-053/). Junk particles were sorted by two rounds of
two-dimensional (2D) classification in RELION v.3.0 (ref. 26).
For the pH8-PAC dataset, particles belonging to the 2D class aver-
ages with features were selected for ab initio three-dimensional (3D)
reconstruction in cryoSPARC v.0.6.5 (ref. 27). The resulting map was then
used as the template for 3D classification using RELION v.3.0 with C1
symmetry. Class averages with high-resolution features were combined
and refined by imposing C3 symmetry. A solvent mask was generated
and was used for all subsequent refinement steps. Bayesian polishing
was conducted to refine the beam-induced motion of the particle set,
resulting in a map at 4.0-Å resolution28.
We noticed that the size of the nanodiscs was not homogeneous,
which could lead to inaccuracy of particle alignment. In addition, the
cytosolic side of the TMD is flexible, which could also influence par-
ticle alignment. To address these potential problems, we subtracted
the nanodisc signal and further classified the particles without image
alignment. A subsequent 3D refinement allowed us to obtain a map at
3.60 Å for PAC. Likewise, we also attempted to only include signals from
the ECD and part of the TMD close to the ECD. This strategy allowed us
to obtain a reconstruction at 3.36 Å. The pH 8 data-processing workflow
is summarized in Extended Data Fig. 2a.
For the pH4-PAC dataset, the initial 3D classification was performed
using the pH8-PAC map (low-pass-filtered to 50 Å) as the reference
without imposing symmetry. Particles belonging to 3D classes with
high-resolution features were pooled and refined using C3 symmetry.
This yielded a reconstruction at 5.8 Å. We then reclassified the particles
using this map (low-pass-filtered to 50 Å) and imposed C3 symmetry.
Subsequent refinement on reasonable 3D classes allowed us to obtain
a reconstruction at 4.6-Å resolution. Finally, a third 3D classification
was initiated by only low-pass-filtering the reference to 7 Å and with C3
symmetry. This classification helped obtain a homogeneous particle set
that gave a reconstruction at 4.2 Å after refinement. We then performed
signal subtraction and 3D classification without image alignment by
focusing on the ECD and part of the TMD proximal to the ECD. This step
allowed us to further push the overall map resolution to 3.73 Å for PAC
at pH 4. We also attempted to only refine the ECD and part of the TMD
Article
for the pH 4 data, which resulted in a map at 3.66-Å resolution. The pH 4
data-processing workflow is summarized in Extended Data Fig. 3a.
Model building
The pH8-PAC model was built de novo using Coot29. Registers were iden-
tified by secondary structure prediction from the JPred web server and
bulky residues in the density30. Both the full map and the ECD-focused
map were used during model building. We were able to model residues
61–338 into the map. Extra density observed on Asn148, Asn155, Asn162
and Asn190 in the ECD was modelled as N-acetyl-d-glucosamine (NAG)
to represent N-linked glycosylation. Real-space refinement was per-
formed in PHENIX to produce the final model31.
The pH4-PAC model was first generated by the RosettaEM flexible
fitting tools, with the pH8-PAC map as the starting point32. The model
was then manually adjusted in Coot and subjected to PHENIX real-space
refinement31. The final model contains residues 53–339 of PAC. Models
and maps are visualized using UCSF Chimera, UCSF ChimeraX and
PyMOL33–35.
Deglycosylation assay
Adherent tsA201 cells transiently transfected with wild-type PAC–eGFP
were solubilized using TBS buffer with 1% GDN for 1 h at 4 °C. The sample
was centrifuged at 20,000g for 30 min. Deglycosylation was facilitated
by mixing the PNGase F with the supernatant and incubating at room
temperature overnight. For the control reaction, the same amount of
water was added instead of PNGase F enzyme. The next day, the sample
was mixed with 2× SDS sample-loading buffer (Sigma) and resolved
by SDS–PAGE electrophoresis. The gel was imaged in the ChemiDoc
system by probing the far-red and GFP signal (mUV 680 and 488 nm).
Molecular dynamics simulations
The structure of PAC in the pH 8 or pH 4 state was used as the starting
model. Missing side-chain atoms were fixed using the PDB2PQR util-
ity36. Titratable residues were assigned as the predominant protona-
tion state based on the predicted pKa value from PROPKA3 at pH 8 or
pH 4 (ref. 37). The membrane orientation of the protein was calculated
using the OPM server38. Subsequent system preparation was conducted
in CHARMM-GUI39. POPC lipids were selected to construct the lipid
bilayer. The rest of the protein was solvated and neutralized in 150 mM
NaCl. The resulting simulation box had dimensions of approximately
87 × 87 × 163 Å.
All-atom molecular dynamics (MD) simulation was carried out using
Gromacs v. 2019.2 (ref. 40). CHARMM36m force field was used to param-
eterize the MD system41. The steepest-descent algorithm was used to
minimize the energy of the system so that the Fmax was below 1,000 kJ
mol–1 nm–1. The NVT ensemble was then started to keep the temperature
of the system at 310 K. Subsequently, the NPT ensemble was enabled by
maintaining the system pressure at 1 bar. Protein non-hydrogen atoms
and phosphorus groups of POPC were restrained during NVT and NPT
equilibration. Production simulation continued from the NPT equili-
brated system with the restraints disabled. A Nosé–Hoover thermostat
and a Parrinello–Rahman barostat were used to maintain system tem-
perature and pressure, respectively. Hydrogen atoms were constrained
using the LINCS algorithms42. For efficient GPU acceleration, a Verlet
cut-off scheme (12 Å) was enabled to maintain the particle neighbour
list. We performed 100-ns simulation for both PAC pH 8 and PAC pH 4
conditions using a time step of 2 fs. Analysis of the MD trajectory was
conducted using the utilities inside Gromacs. Specifically, the slice of
water molecules in each snapshot was extracted using the gmx select
command. The coordinates of oxygen atoms in the water molecules
were then projected to the x/y plan for visualization.
pKa prediction
We noticed that the pKa prediction was very sensitive to the side-chain
orientations of the input structure model. To partially account for this
issue, we generated an ensemble of PAC models based on the pH 8
and pH 4 structures using Rosetta. Specifically, the fixed backbone
design protocol was used to sample the side-chain rotamers43. A total
of 1,000 atomic models were built and subjected to pKa prediction
using PROPKA37. The mean and standard deviation of the pKa for his-
tidine, glutamate and aspartate residues are provided in Extended
Data Fig. 9a.
Electrophysiology
PAC-knockout HEK293 cells were seeded on coverslips and transfected
with wild-type or mutant PAC plasmids using Lipofectamine 2000
(Thermo Fisher Scientific). The cells were recorded around 1 day after
transfection. Whole-cell patch-clamp recordings were performed as
described previously4. The extracellular recording solution contained
(in mM): 145 NaCl, 2 KCl, 2 MgCl2, 1.5 CaCl2, 10 HEPES, 10 glucose (300–
310 mOsm/kg; pH 7.3, titrated with NaOH). Acidic extracellular solu-
tions were made of the same ionic composition with 5 mM sodium
citrate as the buffer instead of HEPES, and pH was adjusted using citric
acid. Solutions were applied locally using a gravity perfusion system
with a small tip 100–200 μm away from the recording cell. The intracel-
lular recording solution contained (in mM): 135 CsCl, 1 MgCl2, 2 CaCl2,
10 HEPES, 5 EGTA, 4 MgATP (280–290 mOsm/kg; pH 7.2, titrated with
CsOH). Pipette solution used to observe PAC current at 0 mV contained
(in mM): 50 NaCl, 100 sodium gluconate, 10 HEPES (280–290 mOsm/kg;
pH 7.2, adjusted with NaOH). Patch pipettes (2–4 MΩ) were pulled with
a Model P-1000 multi-step puller (Sutter Instruments).
For selectivity experiments, the extracellular solution used contained
(in mM): 15 or 150 NaCl, 10 MES, 10 glucose (osmolality adjusted with
mannitol to 300–310 mOsm/kg; pH adjusted with methanesulfonic
acid to 5.0). The pipette solution contained (in mM) 150 NaCl, 10 HEPES
(280–290 mOsm/kg; pH 7.2, adjusted with NaOH). Voltage ramp pulses
were applied every 3 s from −100 to +100 mV at a speed of 1 mV/ms,
and a holding potential of 0 mV. The recorded currents were used to
generate I–V curves for reversal potential determination. The per-
meability ratios were calculated from shifts in the reversal potential
using the Goldman–Hodgkin–Katz equation44. For the measurement
of pH sensitivity, currents were normalized to the maximal current at
pH 4.6. The normalized data were then fitted to a pH dose–response
curve equation
top − bottom
Y = bottom +
1 + 10(X −pH50)×HillSlope
to estimate the pH50 and Hill’s slope (Hill coefficient). Recordings were
done at room temperature with a MultiClamp 700B amplifier and 1550B
digitizer (Molecular Devices). Current signals were filtered at 2 kHz and
digitized at 10 kHz. Series resistance was compensated for at least 80%.
Clampfit v.10.6 and GraphPad Prism v.6 or 7 were used for data analyses.
Reporting summary
Further information on research design is available in the Nature
Research Reporting Summary linked to this paper.
Data availability
The cryo-EM density maps and coordinates of pH8-PAC and pH4-PAC
have been deposited in the Electron Microscopy Data Bank (EMDB)
under accession numbers EMD-22403 and EMD-22404 and in the RCSB
Protein Data Bank (PDB) under accession codes 7JNA and 7JNC.
20. Goehring, A. et al. Screening and large-scale expression of membrane proteins in
mammalian cells for structural studies. Nat. Protoc. 9, 2574–2585 (2014).
21. Haley, E. et al. Expression and purification of the human lipid-sensitive cation channel
TRPC3 for structural determination by single-particle cryo-electron microscopy. J. Vis.
Exp. 143, e58754 (2019).
22. Mastronarde, D. N. Automated electron microscope tomography using robust prediction
of specimen movements. J. Struct. Biol. 152, 36–51 (2005).
23. Zheng, S. Q. et al. MotionCor2: anisotropic correction of beam-induced motion for
improved cryo-electron microscopy. Nat. Methods 14, 331–332 (2017).
24. Zhang, K. Gctf: real-time CTF determination and correction. J. Struct. Biol. 193, 1–12
(2016).
25. Rohou, A. & Grigorieff, N. CTFFIND4: fast and accurate defocus estimation from electron
micrographs. J. Struct. Biol. 192, 216–221 (2015).
26. Scheres, S. H. W. RELION: implementation of a Bayesian approach to cryo-EM structure
determination. J. Struct. Biol. 180, 519–530 (2012).
27. Punjani, A., Rubinstein, J. L., Fleet, D. J. & Brubaker, M. A. cryoSPARC: algorithms for rapid
unsupervised cryo-EM structure determination. Nat. Methods 14, 290–296 (2017).
28. Zivanov, J., Nakane, T. & Scheres, S. H. W. A Bayesian approach to beam-induced motion
correction in cryo-EM single-particle analysis. IUCrJ 6, 5–17 (2019).
29. Emsley, P. & Cowtan, K. Coot: model-building tools for molecular graphics. Acta
Crystallogr. D 60, 2126–2132 (2004).
30. Drozdetskiy, A., Cole, C., Procter, J. & Barton, G. J. JPred4: a protein secondary structure
prediction server. Nucleic Acids Res. 43, W389–W394 (2015).
31. Afonine, P. V. et al. New tools for the analysis and validation of cryo-EM maps and atomic
models. Acta Crystallogr. D 74, 814–840 (2018).
32. Wang, R. Y. R. et al. Automated structure refinement of macromolecular assemblies from
cryo-EM maps using Rosetta. eLife 5, e17219 (2016).
33. Pettersen, E. F. et al. UCSF Chimera—a visualization system for exploratory research and
analysis. J. Comput. Chem. 25, 1605–1612 (2004).
34. Goddard, T. D. et al. UCSF ChimeraX: meeting modern challenges in visualization and
analysis. Protein Sci. 27, 14–25 (2018).
35. The PyMOL Molecular Graphics System, v.2.1. (Schrödinger, LLC, 2020).
36. Dolinsky, T. J., Nielsen, J. E., McCammon, J. A. & Baker, N. A. PDB2PQR: an automated
pipeline for the setup of Poisson–Boltzmann electrostatics calculations. Nucleic Acids
Res. 32, W665–W667 (2004).
37. Olsson, M. H. M. Søndergaard, C. R., Rostkowski, M. & Jensen, J. H. PROPKA3: consistent
treatment of internal and surface residues in empirical pKa predictions. J. Chem. Theory
Comput. 7, 525–537 (2011).
38. Lomize, M. A., Pogozheva, I. D., Joo, H., Mosberg, H. I. & Lomize, A. L. OPM database and
PPM web server: resources for positioning of proteins in membranes. Nucleic Acids Res.
40, D370–D376 (2012).
39. Jo, S., Kim, T., Iyer, V. G. & Im, W. CHARMM-GUI: a web-based graphical user interface for
CHARMM. J. Comput. Chem. 29, 1859–1865 (2008).
40. Abraham, M. J. et al. Gromacs: high performance molecular simulations through
multi-level parallelism from laptops to supercomputers. SoftwareX 1–2, 19–25 (2015).
41. Huang, J. et al. CHARMM36m: an improved force field for folded and intrinsically
disordered proteins. Nat. Methods 14, 71–73 (2017).
42. Hess, B. P-LINCS: a parallel linear constraint solver for molecular simulation. J. Chem.
Theory Comput. 4, 116–122 (2008).
43. Leaver-Fay, A., Kuhlman, B. & Snoeyink, J. An adaptive dynamic programming algorithm
for the side chain placement problem. Pac. Symp. Biocomput. 10, 16–27 (2005).
44. Yang, H. et al. TMEM16F forms a Ca2+-activated cation channel required for lipid
scrambling in platelets during blood coagulation. Cell 151, 111–122 (2012).
45. Zhang, Y. & Skolnick, J. TM-align: a protein structure alignment algorithm based on the
TM-score. Nucleic Acids Res. 33, 2302–2309 (2005).
Acknowledgements We thank G. Zhao and X. Meng for support with data collection at the
David Van Andel Advanced Cryo-Electron Microscopy Suite; the HPC team of VARI for
computational support; and D. Nadziejka for technical editing. W.L. is supported by National
Institutes of Health (NIH) grants R56HL144929 and R01NS112363; Z.Q. is supported by a
McKnight Scholar Award, a Klingenstein-Simon Scholar Award, a Sloan Research Fellowship in
Neuroscience and NIH grants R35GM124824 and R01NS118014; Z.R. is supported by an
American Heart Association (AHA) postdoctoral fellowship (grant 20POST35120556); J.O.-O. is
supported by an AHA predoctoral fellowship (grant 18PRE34060025); and J.D. is supported by
a McKnight Scholar Award, a Klingenstein-Simon Scholar Award, a Sloan Research Fellowship
in Neuroscience, a Pew Scholar in the Biomedical Sciences and NIH grant R01NS111031.
Author contributions W.L. and Z.Q. supervised the project. Z.R. purified PAC, prepared and
screened cryo-EM samples, performed cryo-EM data collection and processing and
performed computational simulation. J.O.-O. cloned the PAC constructs and performed
electrophysiological studies. Z.R., J.O.-O., J.D., Z.Q. and W.L. contributed to data analysis and
manuscript preparation.
Competing interests The authors declare no competing interests.
Additional information
Supplementary information is available for this paper at https://doi.org/10.1038/s41586-020-
2875-7.
Correspondence and requests for materials should be addressed to Z.Q. or W.L.
Peer review information Nature thanks Lily Jan, Stephan Kellenberger and Kenton Swartz for
their contribution to the peer review of this work.
Reprints and permissions information is available at http://www.nature.com/reprints.
Article
Extended Data Fig. 1 | Purification of PAC and biochemical and biophysical
analysis. a, Fluorescence size-exclusion chromatography (FSEC) of PAC–GFP
solubilized in GDN detergent. b, SDS–PAGE gel of purified PAC–GFP protein
after metal affinity chromatography. The uncropped source gel of the image
can be found in Supplementary Fig. 1a. The gel was repeated three times from
different batches of purification and similar results were obtained. c, SEC
profile of PAC in MSP3D1 nanodiscs. d, A deglycosylation assay of PAC–GFP
with or without PNGase F treatment. The GFP and far-red signal (Alexa 488 and
Alexa 680) of the gel was detected and merged using ChemiDoc imaging
system (BioRad). The uncropped source gel of the image can be found in
Supplementary Fig. 1b. The deglycosylation assay was repeated twice with
similar results. e, F196 mediates intersubunit interactions by forming a
cation-π interaction with Arg237′ and hydrophobic interactions with Tyr267′
and Phe282′ from the adjacent subunit. The two subunits are in green and blue.
f, FSEC traces of GFP-tagged wild-type PAC and the F196 mutant solubilized
using GDN detergent. The peak position of F196A is shifted and is broader
compared to the wild type, suggesting that F196A interferes with the proper
assembly of PAC. g, The whole-cell current density of wild-type PAC and
PAC(F196A) recorded at pH 4.6 with a holding potential of 100 mV. The centre
error bar represents mean and s.e.m. Two-tailed unpaired t-test was used to
determine the difference in current density between F196A and the wild type
(P = 3.09 ×10 −6). D’Agostino & Pearson omnibus test was performed to check the
normality of the data (P values are 0.846 and 0.349 for wild type (n = 10) and
F196A (n = 11), respectively). ***denotes P < 0.001.
Extended Data Fig. 2 | Workflow for cryo-EM data-processing of pH8-PAC
and data statistics. a, A total of 16,733 raw movies stacks were collected and
processed with motion correction, CTF estimation and particle picking.
Particles were subjected to two rounds of 2D classification and a 3D
classification run to obtain a homogeneous particle set. To further sort out
conformational heterogeneity, we attempted to subtract and classify (1)
particles without nanodiscs and (2) the ECD of PAC (residues 72–317) by using a
mask. Subsequent refinement allowed us to obtain a map at 3.60-Å resolution
for the entire PAC protein and 3.36-Å resolution for the ECD. b, Representative
micrograph, 2D class averages, Fourier shell correlation (FSC) curves and
angular distribution of particles used for 3D reconstruction for the pH8-PAC
dataset. The gold-standard 0.143 threshold was used to determine map
resolution based on the FSC curve. The threshold for model versus map
correlation was 0.5 to determine the resolution.
Article

Extended Data Fig. 3 | See next page for caption.

Extended Data Fig. 3 | Workflow for cryo-EM data-processing of pH4-PAC
and data statistics. a, A total of 26,689 raw movie stacks were collected and
processed with motion correction, CTF estimation and particle picking. Two
rounds of 2D classification were performed to clean up junk particles.
Subsequently, particles belonging to the 2D class averages with clear features
were subjected to three rounds of 3D classification. The initial 3D classification
was conducted by using the pH8-PAC map low-pass filter set to 50 Å as the
reference. No symmetry operator was imposed in this step. After refinement
with C3 symmetry, a 5.8-Å-resolution map for pH4-PAC was obtained.
Subsequently, the second 3D classification job was conducted by using the 5.8-
Å map as the reference and the low-pass filter set to 50 Å. We imposed C3
symmetry at this step to increase the classification efficiency. This allowed us
to obtain a map at 4.6 Å after refinement. Finally, a third 3D classification job
was launched by using the 4.6-Å pH4-PAC map as the reference and the low-pass
filter set to 7 Å. The C3 symmetry was also imposed. This classification pushed
the resolution of the pH4-PAC map to 4.2 Å. In an effort to obtain a more
homogeneous particle set, we subtracted the ECD of the pH4-PAC map
(residues 72–317) and classified the refined particles without image alignment.
In the end, we obtained a reconstruction of the pH4-PAC map at 3.73-Å
resolution and a pH4-PAC ECD map at 3.66-Å resolution. b, Representative
micrograph, 2D class averages, Fourier shell correlation (FSC) curves and
angular distribution of particles used for 3D reconstruction for the pH4-PAC
dataset. The gold-standard 0.143 threshold was used to determine map
resolution based on the FSC curve. The threshold for model versus map
correlation was 0.5 to determine the resolution.
Article
Extended Data Fig. 4 | Local-resolution cryo-EM maps, representative
densities of cryo-EM maps and domain organization of human PAC. a, The
local resolution of the pH8-PAC map. A non-sliced (left) and a sliced (right) view
of the map viewed parallel to the membrane are shown. The unit for the colour
key is Å. b, Representative densities of several secondary structural elements
of pH8-PAC. The atomic model is overlaid with the density to show the side
chain information. c, The local resolution of the pH4-PAC map. A non-sliced
(left) and a sliced (right) view of the map viewed parallel to the membrane are
shown. The unit for the colour key is Å. d, Representative densities of several
secondary structural elements of pH4-PAC. The atomic model is overlaid with
the density to show the side chain information. e, The pH8-PAC single subunit
viewed parallel to the membrane. The wrist, palm, thumb, finger and β-ball
domains are highlighted. f, The pH4-PAC single subunit viewed in the same
orientation as the right image of panel e. g, Domain organization of PAC.
Clusters of secondary structure that form the palm, finger, thumb and β-ball
domains are labelled.
Extended Data Fig. 5 | Comparison of the structures of PAC and ASIC.
a–d, Structural comparison of human PAC (a, c) with chicken ASIC1a (b, d)
viewed parallel to the membrane (a, b) and from the extracellular side (c, d).
The acidic pocket of human PAC and chicken ASIC1a are in different locations.
e, Overlay of the pH8-PAC (blue) and pH4-PAC (red) single subunit with the
chicken ASIC1a (green) subunit. The ECD of ASICa is composed of a β-sheet
core and the exterior helical structure. Although the β-sheet core shares high
similarity with the human PAC structure, the chicken ASIC1a TMD is organized
differently from that of the human PAC.
Article
Extended Data Fig. 6 | Sequence alignment of PAC homologues and ASIC.
Sequence alignment of PAC homologues (from human, frog (XENLA) and
zebrafish (DANRE)) and chicken ASIC1. The ASIC1 sequence is aligned with PAC
based on the structural alignment using TMalign45. Secondary structural (SS)
elements of PAC are labelled at the top, whereas the SS elements of ASIC1 are
indicated at the bottom. Cysteine residues mediating disulfide bonds in the
extracellular domain of PAC are marked with yellow dots. Putative N-linked
glycosylation sites of PAC are highlighted with green dots. Lys319 of PAC is
marked with red dots. The pre-TM2 helix observed in the pH4-PAC structure is
indicated with a red frame. PAC lacks the α1, α2, α3, α4 and α5 helices that form
the ECD exterior helical structure in chicken ASIC1a, whereas the αA and αB
helices are unique to PAC.
Extended Data Fig. 7 | PAC channel desensitization. a, A representative
whole-cell current trace of PAC in wild-type HEK293 cells upon extracellular
acidification at pH 4.6 and pH 4.0 with a holding potential at 100 mV.
Substantial desensitization was observed during the prolonged exposure to
the pH 4.0 solution (position 4 versus position 3), but not to the pH 4.6 solution
(position 2 versus position 1). b, Quantification of PAC desensitization (pH 4.6
(n = 12) and pH 4.0 (n = 11) as shown in a. Activation and desensitization currents
are normalized to the initial PAC currents. The x axis numbers correspond to
the red marker location in a. Each data point is represented by a solid dot. The
mean and s.e.m. are represented by the bar graph. c, Representative whole-cell
current-voltage traces of PAC at the beginning (position 3 in a) and the end
(position 4 in a) of pH 4.0 treatment. d, Reversal potential of PAC at the
beginning and the end of pH 4.6 and pH 4.0 treatment, respectively (n = 9).
Two-tailed paired t-test was used to determine significance (P values are 0.361
and 0.077 for pH 4.6 and pH 4.0, respectively). D’Agostino & Pearson omnibus
test was performed to check the normality of the data (P values are 0.673 and
0.335 for pH 4.6 and pH 4.0 conditions, respectively). NS indicates P > 0.05.
e, Whole-cell patch-clamp recording configuration with 50 mM NaCl pipette
solution and 150 mM bath solutions (scheme depicted on the left). This creates
the concentration gradient necessary to observe any potential PAC current at 0
mV. Owing to the small amplitude of endogenous PAC current at 0 mV, we
transfected PAC cDNA in PAC knockout HEK293 cells. The representative
whole-cell current trace of PAC upon acidification at 0 mV is shown on the right.
Location 1 and 3 represent initial activation of PAC immediately after acidic
buffer treatment. Location 2 and 4 represent desensitized PAC after prolonged
acidic buffer treatment. f, The desensitized currents (position 2 and 4 in e) are
normalized to the initial PAC currents (position 1 and 3 in e). The desensitized
data currents are represented by the normalized average ± s.e.m.
Article
Extended Data Fig. 8 | Lateral fenestration and ion selectivity of PAC. a, The
reversal potential (Vrev) of wild-type PAC, PAC(K325E) and PAC(K329E) at
150 mM NaCl (black) or 15 mM NaCl (red) in the bath solution (internal solution
contains 150 mM NaCl). The bar graph represents the mean and s.e.m. (n = 16
(wild type), n = 8 (K325E) and n = 6 (K329E)). Individual data points are shown as
dots. The same data points for the wild type were also used in Fig. 3i for
comparison with K319E. b, The relative Cl−/Na+ permeability for wild-type PAC
(n = 16), and K325E (n = 8) and K329E (n = 6) mutants calculated from the pH-
5-induced current at 100 mV. The centre and error bar represent the mean and
s.e.m of the permeability ratio. Individual data points are shown as solid dots.
The same data points for the wild type were also used in Fig. 3j for comparison
with K319E. The average PCl/PNa permeability values are indicated for each
construct. c, The current density of wild-type PAC (n = 10), and K325E (n = 10)
and K329E mutants (n = 10) at pH 4.6 with a holding potential of 100 mV. The bar
graph shows the average normalized current density ± s.e.m. One-way ANOVA
with Bonferroni post-hoc test was used to determine the significance (P values
are 0.832 and 0.416 for K325E and K329E, respectively). D’Agostino & Pearson
omnibus test was performed to check the normality of the data (P values are
0.255, 0.153 and 0.293 for the wild type and K325E and K329E mutants,
respectively). NS indicates P > 0.05. d, The pH dose–response curve of
wild-type PAC, PAC(K325E) and PAC(K329E). The currents are normalized to
those at pH 4.6 (n = 8 (wild-type PAC), n = 6 PAC(K325E) and n = 7 (PAC(K329E)).
The currents at different pH are represented by the average normalized
currents ± s.e.m. A nonlinear fitting to a sigmoidal dose–response curve is
generated for each construct. e, Representative whole-cell patch-clamp
recording at pH 5.0 with 150 mM NaCl pipette solution and 150 mM (black) or
15 mM NaCl (red) bath solutions. The current–voltage relationship of wild-type
(left), K325E (middle) and K329E (right) PAC in two different bath solutions are
plotted. The same wild-type traces were also shown in Fig. 3j (left) for
comparison with K319E. f, i, The pH8-PAC and pH4-PAC extracellular
fenestration viewed from the extracellular side (left) and parallel to the
membrane (right), respectively. Residues forming the fenestration are shown
in sticks, including three negatively charged residues (Asp91, Glu94 and
Glu250) for pH8-PAC and two positively charged residues (Arg93 and Lys294)
for pH4-PAC. g, j, Radius of the fenestration tunnel, estimated by CAVER v.3.0,
for pH8-PAC (g) and pH4-PAC ( j). The horizontal line marks the smallest radius
along the tunnel. The residues lining the fenestration tunnel are marked.
h, k, Fenestration water-density plot for pH8-PAC (h) and pH4-PAC (k) from a
100-ns MD simulation. Water molecules in the Z range of the side fenestration
site are projected to the X/Y plane and are shown as a 2D histogram.
Extended Data Fig. 9 | See next page for caption.
Article
Extended Data Fig. 9 | His98 is involved in PAC pH sensing. a, pKa prediction
of titratable residues for the pH8 and pH4 structures of human PAC. The mean
and error bar (standard deviation) are calculated based on 1,000 fixed-
backbone rotamer ensembles generated from each structure (see Methods). b,
SDS gel of GFP-tagged wild-type PAC, PAC(H98C/Q296C) and PAC(H98S/
Q296S). A dimeric band is observed for the H98C/Q296C mutant, but not for
the wild type and the H98S/Q296S mutant. The unedited source gel of the
image can be found in Supplementary Fig. 1c. The gel was independently
repeated twice with similar results. c, The FSEC profile of GFP-tagged wild-type
PAC, PAC(H98C/Q296C) and PAC(H98S/Q296S) solubilized using GDN
detergent. d, The whole-cell current density of wild-type PAC, PAC(H98C/
Q296C) and PAC(H98S/Q296S) recorded at pH 5.0 at 100 mV. The bar graph
shows the average current density (nA/pF) ± s.e.m. Each individual data point
represents a cell (n = 8 (wild type), n = 10 (H98C/Q296C) and n = 12 (H98S/
Q296S)). Two-tailed unpaired t-test was used to determine the difference in
current density compared to the wild type (P values are 1.08 × 10 −6 for H98C/
Q296C and 0.321 for H98S/Q296S). D’Agostino & Pearson omnibus test was
performed to check the normality of the data (P values are 0.328, 0.154 and
0.727 for the wild type and the H98C/Q296C and H98S/Q296S mutants,
respectively). e, The pH dose–response curve of wild-type PAC and PAC(H98S/
Q296S). The currents are normalized to those at pH 4.6 (n = 5 (wild-type PAC);
n = 6 (PAC(H98S/Q296S)). A nonlinear fitting to a sigmoidal dose–response
curve is generated for each construct. Bar plot shows the mean ± s.e.m. f, The
pH50 of wild-type PAC and PAC(H98S/Q296S) estimated from the pH dose–
response curve. The centre and bar represent the estimated pH50 and s.e.m.
from the nonlinear fitting in e. Two-tailed Mann–Whitney test was used to
determine the significance (P = 0.0087). g, The proposed pH-sensing
mechanism for PAC. At high pH, the deprotonated His98 residue is surrounded
by Gln296, Ser102 and Iso298, and TM1 pairs with TM2 from the same subunit.
At low pH, the protonated His98 residue undergoes a conformational change
and moves into an acidic pocket. As a result, TM1 dissociates from the resting
interface and rotates to interact with TM2 of the adjacent subunit. For all
panels, NS indicates P > 0.05, ** denotes a P value between 0.01 and 0.001 and
*** denotes P < 0.001; n represents measurements from biologically
independent cells.
Extended Data Table 1 | Cryo-EM data collection, refinement and validation statistics
 nature research | reporting summary
Corresponding author(s): Wei Lü
Last updated by author(s): Aug 5, 2020

Reporting Summary
Nature Research wishes to improve the reproducibility of the work that we publish. This form provides structure for consistency and transparency
in reporting. For further information on Nature Research policies, see Authors & Referees and the Editorial Policy Checklist.

Statistics
For all statistical analyses, confirm that the following items are present in the figure legend, table legend, main text, or Methods section.
n/a Confirmed
The exact sample size (n) for each experimental group/condition, given as a discrete number and unit of measurement
A statement on whether measurements were taken from distinct samples or whether the same sample was measured repeatedly
The statistical test(s) used AND whether they are one- or two-sided
Only common tests should be described solely by name; describe more complex techniques in the Methods section.

A description of all covariates tested

A description of any assumptions or corrections, such as tests of normality and adjustment for multiple comparisons
A full description of the statistical parameters including central tendency (e.g. means) or other basic estimates (e.g. regression coefficient)
AND variation (e.g. standard deviation) or associated estimates of uncertainty (e.g. confidence intervals)

For null hypothesis testing, the test statistic (e.g. F, t, r) with confidence intervals, effect sizes, degrees of freedom and P value noted
Give P values as exact values whenever suitable.

For Bayesian analysis, information on the choice of priors and Markov chain Monte Carlo settings
For hierarchical and complex designs, identification of the appropriate level for tests and full reporting of outcomes
Estimates of effect sizes (e.g. Cohen's d, Pearson's r), indicating how they were calculated
Our web collection on statistics for biologists contains articles on many of the points above.

Software and code

Policy information about availability of computer code
Data collection SerialEM 3.7, ClampFit 10.6

Data analysis Gctf-1.06, ctffind-4.1.10, Gautomatch-0.56, Relion-3.0, CryoSparc-v0.6.5, coot-0.8.9.2, pymol-2.3.2, Motioncor2-1.2.1,
phenix.real_space_refine_dev_3500, phenix.molprobity_dev_3500, UCSF chimera_1.13.1, UCSF chimeraX_0.91, GraphPad Prism 6 and 7,
ClampFit_10.6, GROMACS version 2019.2, OPM server (https://opm.phar.umich.edu/ppm_server), CHARMM-GUI (http://www.charmm-
gui.org/), Rosetta 2020.08.61146, propka3.1, TMalign v20190822
For manuscripts utilizing custom algorithms or software that are central to the research but not yet described in published literature, software must be made available to editors/reviewers.
We strongly encourage code deposition in a community repository (e.g. GitHub). See the Nature Research guidelines for submitting code & software for further information.

Data
Policy information about availability of data
All manuscripts must include a data availability statement. This statement should provide the following information, where applicable:
- Accession codes, unique identifiers, or web links for publicly available datasets
- A list of figures that have associated raw data
October 2018

- A description of any restrictions on data availability

The cryo-EM density map and coordinates of pH8-hsPAC and pH4-hsPAC have been deposited in the Electron Microscopy Data Bank (EMDB) under accession
numbers EMD-22403 and EMD-22404 and in the Research Collaboratory for Structural Bioinformatics Protein Data Bank under accession codes 7JNA and 7JNC.

1
 nature research | reporting summary
Field-specific reporting
Please select the one below that is the best fit for your research. If you are not sure, read the appropriate sections before making your selection.
Life sciences Behavioural & social sciences Ecological, evolutionary & environmental sciences
For a reference copy of the document with all sections, see nature.com/documents/nr-reporting-summary-flat.pdf

Life sciences study design

All studies must disclose on these points even when the disclosure is negative.
Sample size Sample size was not pre-determined for the study. All the electrophysiology experiments were repeated at least five times using different
cells. The sample size was determined based on the consistence/variability of the recordings. Overall, the samples sizes were deemed
sufficient based on the clearly visible effects of the mutations on the overall distribution of data points within and between each group (The
sample standard deviation is usually much smaller than the effect we are aiming to test). However, we acknowledge that our current sample
size may not be sufficient to detect small effect that may be present in some of the comparisons, leading us to accept the null hypothesis.

Data exclusions No data was excluded from the analysis.

Replication We have done each group of experiment with several batches of cells and different transfection, to ensure reproducibility within the lab. The
number biologically independent experimental replications is indicated in the figure legend. For electrophysiology recordings, the number is
at least 5. For SDS PAGE gel experiments, they were repeated at least twice.

Randomization Our experiment is not randomized. For electrophysiology experiments, cells with GFP fluorescence (proteins were GFP-tagged) were
randomly selected. Other experiments including protein expression, solubilization test, protein purification, deglycosylation assay, and cryo-
EM grids preparation and data collection were repeated multiple time; each time, proteins from different random batches were used.

Blinding The investigators were not blinded; it was not technically or practically feasible to do so for cryo-EM or patch-clamp studies.

Reporting for specific materials, systems and methods

We require information from authors about some types of materials, experimental systems and methods used in many studies. Here, indicate whether each material,
system or method listed is relevant to your study. If you are not sure if a list item applies to your research, read the appropriate section before selecting a response.

Materials & experimental systems Methods

n/a Involved in the study n/a Involved in the study
Antibodies ChIP-seq
Eukaryotic cell lines Flow cytometry
Palaeontology MRI-based neuroimaging
Animals and other organisms
Human research participants
Clinical data

Eukaryotic cell lines

Policy information about cell lines
Cell line source(s) Sf9 cells, tsA201 cells and HEK293 cells were purchased from ATCC

Authentication The cells were purchased and routinely maintained in our lab. They were not authenticated experimentally for these studies.

Mycoplasma contamination Sf9 cells, tsA201 cells and HEK293 cells were tested negative form Mycoplasma contamination

Commonly misidentified lines No commonly misidentified lines were used

October 2018

(See ICLAC register)

2
Matters arising
Crop asynchrony stabilizes food production
https://doi.org/10.1038/s41586-020-2965-6 Lukas Egli1,2 ✉, Matthias Schröter1, Christoph Scherber3,4, Teja Tscharntke5,6 & Ralf Seppelt1,7
Received: 12 February 2020
arising from D. Renard & D. Tilman Nature https://doi.org/10.1038/s41586-019-1316-y (2019)
Accepted: 30 September 2020
Published online: 9 December 2020
Check for updates
Stable agricultural systems are fundamental for the reliability of agri-
cultural production and food security. Recently, Renard and Tilman1
reported that crop diversity, calculated as the exponential value of
the Shannon diversity index of harvested areas of 176 crops, stabilizes
national food production. Here we show that crop asynchrony—that is,
asynchronous production trends between different crops2—is an even
better predictor of agricultural production stability than is crop diver-
sity. Our finding suggests that asynchrony is one important property
that can explain why a higher crop diversity supports the stability of
national food production, and that it should be considered in strate-
gies to stabilize agricultural production through crop diversification.
We suggest that, as well as yield stability and crop diversity, two addi-
tional aspects should be considered in the discussion of the diversity–
stability nexus. First, as well as yield stability, the stability of overall
production is another relevant aspect of food security. Second, the
actual benefits of crop diversity are not related to harvested areas as
such, but to the temporal production patterns of the cultivated crops2.
We suggest that planting multiple crops stabilizes agricultural pro-
duction only if they experience asynchronous production trends—for
example, due to distinct responses of the individual crops to climatic,
economic and political shocks3. Here we use statistical models to test
whether crop asynchrony is a better predictor of agricultural produc-
tion stability than is crop diversity.
We largely used the same datasets as Renard and Tilman1 (Extended
Data Table 1) and derived the same explanatory variables used in their
analysis, including effective crop species diversity4, irrigation4, nitrogen
use intensity4, warfare5, temperature and precipitation instability6–8 for
five ten-year intervals between 1961 and 2010 (see Supplementary Meth-
ods for details) to predict the stability of total caloric production4,9,10.
We additionally calculated synchrony between crop-specific caloric
production2,11,12, an index bounded between 0 and 1, where 1 indicates
full synchrony. Asynchrony was then calculated by subtracting syn-
chrony from 1, so that higher values indicate higher asynchrony. We
used total production instead of yield stability as the response variable,
because this offers additional insights into food security and because
it can be directly related to asynchrony (see Supplementary Methods
for details). Moreover, total production incorporates the effects of
changes in cropland area as a result of planning decisions by farmers
and of changes in global market dynamics.
First, we investigated the relationship between effective crop species
diversity and crop asynchrony and tested if this relationship changed
over time, as crop homogenization has occurred during recent dec-
ades13. To predict crop asynchrony, we used a linear mixed-effects
model with random slopes for diversity and random intercepts for
time intervals14. Second, we investigated how either crop diversity,
crop asynchrony or both affect caloric production stability. For this,
we constructed the main linear regression model used in Renard and
UFZ - Helmholtz Centre for Environmental Research, Leipzig, Germany. 2University of Potsdam, Institute of Biochemistry and Biology, Potsdam, Germany. 3University of Münster, Institute of
1
Landscape Ecology, Münster, Germany. 4Centre for Biodiversity Monitoring, Zoological Research Museum Alexander Koenig, Bonn, Germany. 5University of Göttingen, Agroecology,
Department of Crop Sciences, Göttingen, Germany. 6University of Göttingen, Centre of Biodiversity and Sustainable Land Use (CBL), Göttingen, Germany. 7Institute of Geoscience and
Geography, Martin-Luther-University Halle-Wittenberg, Halle (Saale), Germany. ✉e-mail: lukas.egli@ufz.de
Tilman1 using production stability as the response variable (see Sup-
plementary Methods for details). We then ran two additional regression
models, one in which we replaced crop diversity with crop asynchrony
and one in which we added crop asynchrony.
Crop diversity and crop asynchrony were found to be correlated
(Spearman’s ρ = 0.49, P < 0.05; Fig. 1a). However, the positive effect of
crop diversity on asynchrony decreased over time (Fig. 1a), as indicated
by a better performance of a linear mixed-effects model including time
interval (Akaike information criterion (AIC) = −340.22) compared to a
linear model including crop diversity only (AIC = −336.88). The positive
effect of crop asynchrony on caloric production stability was more
than three times the effect of crop diversity (Fig. 1b, Extended Data
Table 2). Other predictors showed similar trends, although the effect
of nitrogen use intensity, time and temperature instability was stronger
in the diversity model, whereas the effect of irrigation was lower and
insignificant. Moreover, the explanatory power of the model increased
from R2 = 0.28 in the crop-diversity model to R2 = 0.60 in the asynchrony
model (Extended Data Table 2). In the model that includes both predic-
tors, the stabilizing effect of crop asynchrony was even stronger and
the effect of crop diversity was negative (Fig. 1b, Extended Data Fig. 1);
however, explanatory power increased by only 0.01 (Extended Data
Table 2). Although crop diversity and asynchrony were correlated,
multicollinearity was not an issue in the combined model (the variance
inflation factors were less than 2). Given that crop asynchrony was a
strong predictor of caloric production stability, we further explored
their relationship in the most recent time interval (2001–2010). The
highest national crop asynchronies were mainly observed in South
and Southeast Asia, China, Central America and parts of Africa (Fig. 2).
Countries within these regions typically showed high production stabil-
ity, and all countries with high asynchrony achieved at least medium
stability. Countries with high production stability and low-to-medium
asynchrony were mainly found in North and South America (Fig. 2). The
29 countries that had low asynchrony and stability—including Russia,
Argentina and Australia—contributed more than 11% of the total crop
caloric production.
Our analysis provides an important extension to the results presented
by Renard and Tilman1. We found that the relationship between crop
diversity and crop asynchrony decreased over time, which is a potential
consequence of the increasing homogeneity of global food supplies13.
Most importantly, we identified asynchrony as one important crop
property (or trait) that can explain why a higher crop diversity sup-
ports the stability of national food production. Crop diversity as such
provides only limited insights into the mechanism that underlies stabil-
ity. The benefits of crop diversity depend on the production patterns
of the cultivated crops. Therefore, strategies to stabilize agricultural
production through crop diversification also need to account for the
asynchrony of the crops considered.
Nature | Vol 588 | 10 December 2020 | E7
Matters arising
a b
1.00
0.6
***
0.5 ***
Model
0.4 Diversity
Asynchrony
Combined
0.75 0.3

Standardized regression coefficient

0.2 *** ***
***
0.1 NS** * *
Asynchrony 0.0
0.50 NS
NS NSNS NS
–0.1
*** *** ***
–0.2
*** ***
*** ***
–0.3
Time interval ***
0.25 1961–1970 –0.4
1971–1980
1981–1990
1991–2000 –0.5
2001–2010
–0.6

5 10 15 20 25

e
ty

re
y

y
n)

y
ity

m
on

ilit
ilit
si

fa
io
Diversity

Ti
ns
r

ab
hr

ab
ar
at
ve

te
c

W
ig

st
st
Di

yn

in
Irr

in
in
As

√(

n
re
us

tio
tu
N

ita
ra
√(

pe

ip
ec
m

Fig. 1 | Crop asynchrony as a function of crop diversity and determinants of
national caloric production stability. a, Crop asynchrony as a function of
crop diversity using a linear mixed-effects model with random slopes for
diversity, and random intercepts for time intervals. Dots show national data
coloured by time interval (n = 590). b, Regression coefficients for all variables
in the linear regression models, including crop diversity (green), crop
Pr
Te
asynchrony (blue) and both (orange) (n = 590). Caloric production stability was
log-transformed, irrigation and nitrogen use intensity were square-root-transformed.
Each predictor variable was standardized to 0 mean and 1 s.d. across all nations
and time intervals. Data are mean ± s.e.m. *P < 0.05; **P < 0.01; ***P < 0.001; NS,
not significant. This figure was created with the statistical software package
R 3.6.110.

The results from the crop-diversity model are largely similar to the temporal variance2,15 at farm level, at which management decisions
findings of Renard and Tilman1, because the different response vari- are made. Moreover, growing crops in different seasons is additionally
ables (caloric yield versus production stability) were highly correlated expected to increase asynchrony, which should be further investigated.
(Spearman’s ρ = 0.84, P < 0.05). However, the effect of irrigation was Likewise, we need to better understand the conditions under which
less stabilizing for production compared to yield stability, and the asynchrony is needed for and beneficial to stability. Spain, for exam-
opposite was true for nitrogen use intensity. Moreover, overall pro- ple, experienced medium asynchrony but low stability in 2001–2010,
duction stability significantly decreased over time, which has serious whereas the opposite was true for Germany. In countries that have low
implications for food security. crop asynchrony and stability, planting additional crops with different
Asynchrony emerges from the distinct responses of individual crops responses to climatic and market disturbances might be a viable option
to climatic, economic and political shocks3. Although there is increasing to increase stability and therefore food security15, in particular in light
knowledge about the underlying drivers of overall production losses3, of climate change and increasing perturbations in global markets.
little is known about the effects on individual crops in various environ- On the national level, this is especially relevant for countries that are
mental and socioeconomic contexts—in particular regarding their facing severe food insecurity, such as Malawi. For countries such as
High
Stability
Low

Low High
Asynchrony

Fig. 2 | National crop asynchrony and caloric production stability excluded from the analysis are shown in white. The figure was created with the
worldwide. Crop asynchrony and caloric production stability are shown for statistical software package R 3.6.110.
the 2001–2010 interval and are grouped by tertiles (n = 136). Countries

E8 | Nature | Vol 588 | 10 December 2020

Russia and Argentina, which recently experienced low asynchrony
and stability but contributed more than 5% of the global crop calories,
increasing crop asynchrony is an important aspect to consider from a
global perspective. Growing trade and changing diets might further
lead to crop homogenization2,13 and therefore pose risks to the stability
of national and global food production.
Reporting summary
Further information on research design is available in the Nature
Research Reporting Summary linked to this paper.
Data availability
All datasets used and generated during this study are provided in a
public repository: https://github.com/legli/AgriculturalStability.
Code availability
The codes used for data preparation and analyses are provided in a
public repository: https://github.com/legli/AgriculturalStability.
1. Renard, D. & Tilman, D. National food production stabilized by crop diversity. Nature 571,
257–260 (2019).
2. Mehrabi, Z. & Ramankutty, N. Synchronized failure of global crop production. Nat. Ecol.
Evol. 3, 780–786 (2019).
3. Cottrell, R. S. et al. Food production shocks across land and sea. Nat. Sustain. 2, 130–137
(2019).
4. The Food and Agriculture Organization of the United Nations Statistics (FAO, accessed
22 November 2019); https://www.fao.org/faostat/en/#data/.
5. Marshall, M. G. Codebook: Major Episodes of Political Violence (MEPV) and Conflict
Regions, 1946–2015. http://www.systemicpeace.org/inscr/MEPVcodebook2016.pdf
(2016).
6. Klein Goldewijk, K., Beusen, A., Doelman, J. & Stehfest, E. New anthropogenic land use
estimates for the Holocene – HYDE 3.2. Earth Syst. Sci. Data 9, 927–953 (2017).
7. Sacks, W. J., Deryng, D. & Foley, J. A. Crop planting dates: an analysis of global patterns.
Glob. Ecol. Biogeogr. 19, 607–620 (2010).
8. Willmott, C. J. & Matsuura, K. Terrestrial Air Temperature and Precipitation: Monthly and
Annual Time Series (1950–1999). http://climate.geog.udel.edu/~climate/html_pages/
README.ghcn_ts2.html (2001).
9. Food balance sheets: A handbook (FAO, 2001).
10. R Core Team. R: A Language and Environment for Statistical Computing (R Foundation for
Statistical Computing, 2019).
11. Loreau, M. & de Mazancourt, C. Species synchrony and its drivers: neutral and nonneutral
community dynamics in fluctuating environments. Am. Nat. 172, E48–E66 (2008).
12. Hallett, L. M. et al. codyn : An R package of community dynamics metrics. Methods Ecol.
Evol. 7, 1146–1151 (2016).
13. Khoury, C. K. et al. Increasing homogeneity in global food supplies and the implications
for food security. Proc. Natl Acad. Sci. USA 111, 4001–4006 (2014).
14. Bates, D., Mächler, M., Bolker, B. M. & Walker, S. C. Fitting linear mixed-effects models
using lme4. J. Stat. Softw. 67, 1–48 (2015).
15. Knapp, S. & van der Heijden, M. G. A. A global meta-analysis of yield stability in organic
and conservation agriculture. Nat. Commun. 9, 3632 (2018).
Acknowledgements L.E. acknowledges funding from the Helmholtz Association (Research
School ESCALATE, VH-KO-613). We thank V. Grimm for discussions; M. Wu for statistical
support and D. Renard for discussions and the exchange of code to make our analysis clearer
and more consistent. The FAOSTAT database is maintained and regularly updated by FAO with
regular support from its Member States.
Author contributions L.E., M.S., T.T. and R.S. designed the study. L.E. and C.S. performed the
analysis. All authors wrote the manuscript.
Competing interests The authors declare no competing interests.
Additional information
Supplementary information is available for this paper at https://doi.org/10.1038/s41586-020-
2965-6.
Correspondence and requests for materials should be addressed to L.E.
Reprints and permissions information is available at http://www.nature.com/reprints.
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in
published maps and institutional affiliations.
© The Author(s), under exclusive licence to Springer Nature Limited 2020
Nature | Vol 588 | 10 December 2020 | E9
Matters arising
Extended Data Fig. 1 | Main determinants of national caloric production
stability. a–h, Effects of crop diversity (a), crop asynchrony (b), irrigation (c),
nitrogen use intensity (d), temperature instability (e), precipitation instability (f),
warfare (g) and time (h) on caloric production stability. Results are shown for
the linear regression models including crop diversity (green), crop asynchrony
(blue) and both (orange) (n = 590). Irrigation and nitrogen use intensity were
E10 | Nature | Vol 588 | 10 December 2020
back-transformed from square-root-transformation, predicted values were
back-transformed from log-transformation. Predictions were calculated using
the observed range of the focal predictor, while keeping all the other predictors
at their mean values. Shaded areas represent 95% confidence intervals. The
figure was created with the statistical software package R 3.6.110.
Extended Data Table 1 | Data sources underlying the analyses

See refs. 4–9.

Nature | Vol 588 | 10 December 2020 | E11

Matters arising
Extended Data Table 2 | Determinants of national caloric production stability

Linear regression models include crop diversity (diversity model), crop asynchrony (asynchrony model) or both (combined model) (n = 590). Caloric production stability was log-transformed,
irrigation and nitrogen use intensity were square-root-transformed. Predictor variables were standardized to 0 mean and 1 s.d. across all nations and time intervals.

E12 | Nature | Vol 588 | 10 December 2020

Matters arising
Reply to: Crop asynchrony stabilizes food
production
https://doi.org/10.1038/s41586-020-2966-5 Delphine Renard1 ✉ & David Tilman2,3
Published online: 9 December 2020
replying to L. Egli et al. Nature https://doi.org/10.1038/s41586-020-2965-6 (2020)
Check for updates
In the accompanying Comment, Egli et al.1 report findings related to our
Article2 on the stabilization of food production by crop diversity. In our
Article, we reported that crop diversity stabilized the combined caloric
yield of all crops in a nation2. Our analyses showed that the portfolio
effect3 was the probable cause of this greater stability, and we found no
support for the asynchrony hypothesis4. Egli et al.1 report somewhat
different findings.
The results in our Article2 differ from those of Egli et al.1 because
we analysed the year-to-year temporal stability of national caloric
yield for all crops combined and analysed for asynchrony among
individual crops using the national yield of each crop. By contrast,
Egli et al.1 analysed the stability of total national caloric production
and measured asynchrony on the basis of the annual national caloric
production of each individual crop. Although yield and production
are related to each other, they are not identical. Yield is the crop pro-
duction per unit of land, whereas production is the yield multiplied
by the cropland area. Even using their different metric, Egli et al.1
found—as did we—that greater crop diversity led to greater national
temporal crop stability.
We suggest that national yield stability is the more informative and
insightful of these two stability metrics because it directly measures
the year-to-year reliability of food production from a typical hectare of
cropland in a nation. If the total area planted and harvested had been
constant from 1961 until now in each nation, the two measures of stabil-
ity would be identical. However, harvested area has been increasing in
many lower-income nations for the past 60 years (an increase of 69%
in the least-developed nations since the 1960s5), and it first increased
and then declined in many high-income nations (a decrease of 29% in
Europe since the 1980s5). These year-to-year changes in total national
cropland area and year-to-year changes in yields both affect the sta-
bility metric used by Egli et al.1. Because the database of the Food and
Agriculture Organization of the United Nations (FAOSTAT) reports
the area harvested for each crop, not the area planted, the potential
to determine the effects of changes in area planted on national food
supply stability is limited. Finally, because yield stability is independ-
ent of year-to-year changes in national cropland area, we feel that yield
stability is more informative of underlying biological mechanisms than
is production stability. However, when more detailed data are available,
assessing changes and drivers of stability of yield and cropland area
would be informative.
1
CEFE, CNRS, Univ. Montpellier, University Paul Valéry Montpellier 3, EPHE, IRD, Montpellier, France. 2Bren School of Environmental Science and Management, University of California Santa
Barbara, Santa Barbara, CA, USA. 3Department of Ecology, Evolution and Behavior, University of Minnesota, St Paul, MN, USA. ✉e-mail: delphinerenard@hotmail.fr
We have similar concerns about the measure of asynchrony used
by Egli et al.1. To calculate asynchrony, they used data on the annual
country-level production of each crop, rather than the annual
country-level yields that we used. Their metric of asynchrony there-
fore confounds fluctuations in the performance of individual crops—
as measured by yields—with fluctuations in the area planted to, and
harvested for, each of these crops, just as does their stability metric.
Because both yields and harvested area affect the total food supply
of a country, both have important implications for the agricultural
policy of a country. We suggest that viewing each of these separately
might be better for providing insights into the best way to maximize
the year-to-year reliability of the food supply of a country. The link
found by Egli et al.1 between production asynchrony and total crop
production stability suggests the possibility that asynchronous vari-
ation in area planted with and harvested for various crops might also
contribute to national food stability. This interesting possibility merits
further exploration.
1. Egli, L., Schröter, M., Scherber, C., Tscharntke, T. & Seppelt, R. Crop asynchrony stabilizes
food production. Nature https://doi.org/10.1038/s41586-020-2965-6 (2020).
2. Renard, D. & Tilman, D. National food production stabilized by crop diversity. Nature 571,
257–260 (2019).
3. Doak, D. F. et al. The statistical inevitability of stability–diversity relationships in
community ecology. Am. Nat. 151, 264–276 (1998).
4. Loreau, M. & De Mazancourt, C. Species asynchrony and its drivers: neutral and nonneutral
community dynamics in fluctuating environments. Am. Nat. 172, E48–E66 (2008).
5. The Food and Agriculture Organization of the United Nations Statistics (FAO, accessed
January 2019); https://www.fao.org/faostat/en/#data/.
Acknowledgements We thank the Bren School of Environment Science and Management of
the University of California Santa Barbara for support leading to the initial publication. This
work was also supported by a grant overseen by the French ‘Programme Investissement
d’Avenir’ as part of the ‘Make Our Planet Great Again’ programme (reference: 17-MPGA-0004)
and by a National Science Foundation grant (LTER-1831944).
Author contributions D.R. and D.T. wrote the paper.
Competing interests The authors declare no competing interests.
Additional information
Correspondence and requests for materials should be addressed to D.R.
Reprints and permissions information is available at http://www.nature.com/reprints.
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in
published maps and institutional affiliations.
© The Author(s), under exclusive licence to Springer Nature Limited 2020
Nature | Vol 588 | 10 December 2020 | E13
Corrections & amendments
Author Correction:
Area-based conservation
in the twenty-first
century

https://doi.org/10.1038/s41586-020-2952-y

Correction to: Nature https://doi.org/10.1038/s41586-020-2773-z

Published online 07 October 2020

Check for updates

Sean L. Maxwell, Victor Cazalis, Nigel Dudley, Michael Hoffmann,

Ana S. L. Rodrigues, Sue Stolton, Piero Visconti, Stephen Woodley,
Naomi Kingston, Edward Lewis, Martine Maron,
Bernardo B. N. Strassburg, Amelia Wenger, Harry D. Jonas,
Oscar Venter & James E. M. Watson

In this Review, the affiliation to which authors Victor Cazalis and Ana
S. L. Rodrigues are attributed (affiliation 2) should be corrected from
‘Centre d’Ecologie Fonctionnelle et Evolutive CEFE UMR 5175, CNRS,
Univ. de Montpellier, Univ. Paul-Valéry Montpellier, EPHE, Montpellier,
France’ to ‘CEFE, Univ. Montpellier, CNRS, EPHE, IRD, Univ. Paul Valéry
Montpellier 3, Montpellier, France’. This error has been corrected
online.

E14 | Nature | Vol 588 | 10 December 2020

Corrections & amendments
Author Correction:
Elephant shark genome
provides unique insights
into gnathostome
evolution

https://doi.org/10.1038/s41586-020-2967-4

Correction to: Nature https://doi.org/10.1038/nature12826

Published online 08 January 2014

Check for updates

Byrappa Venkatesh, Alison P. Lee, Vydianathan Ravi,

Ashish K. Maurya, Michelle M. Lian, Jeremy B. Swann, Yuko Ohta,
Martin F. Flajnik, Yoichi Sutoh, Masanori Kasahara, Shawn Hoon,
Vamshidhar Gangu, Scott W. Roy, Manuel Irimia, Vladimir Korzh,
Igor Kondrychyn, Zhi Wei Lim, Boon-Hui Tay, Sumanty Tohari,
Kiat Whye Kong, Shufen Ho, Belen Lorente-Galdos, Javier Quilez,
Tomas Marques-Bonet, Brian J. Raney, Philip W. Ingham, Alice Tay,
LaDeana W. Hillier, Patrick Minx, Thomas Boehm, Richard K. Wilson,
Sydney Brenner & Wesley C. Warren

In this Article, the Author Information section should have stated that
the miRNA reads for 17 tissues have been deposited at the Sequence
Read Archive (SRA) under accession numbers: SRR12545166–
SRR12545182. The original Article has not been corrected.

Nature | Vol 588 | 10 December 2020 | E15

Corrections & amendments
Author Correction:
Determination of RNA
structural diversity and
its role in HIV-1 RNA
splicing

https://doi.org/10.1038/s41586-020-2949-6

Correction to: Nature https://doi.org/10.1038/s41586-020-2253-5

Published online 6 May 2020

Check for updates

Phillip J. Tomezsko, Vincent D. A. Corbin, Paromita Gupta,

Harish Swaminathan, Margalit Glasgow, Sitara Persad,
Matthew D. Edwards, Lachlan Mcintosh, Anthony T. Papenfuss,
Ann Emery, Ronald Swanstrom, Trinity Zang, Tammy C. T. Lan,
Paul Bieniasz, Daniel R. Kuritzkes, Athe Tsibris & Silvi Rouskin

In the ‘Code availability’ section of this Article, the URL from which
the DREEM clustering algorithm can be accessed was incorrect. The
correct URL is https://codeocean.com/capsule/6175523/tree/v1. The
Article has been corrected online.

E16 | Nature | Vol 588 | 10 December 2020

Corrections & amendments
Author Correction:
Innovations present in
the primate interneuron
repertoire

https://doi.org/10.1038/s41586-020-2874-8

Correction to: Nature https://doi.org/10.1038/s41586-020-2781-z

Published online 30 September 2020

Check for updates

Fenna M. Krienen, Melissa Goldman, Qiangge Zhang,

Ricardo C. H. del Rosario, Marta Florio, Robert Machold,
Arpiar Saunders, Kirsten Levandowski, Heather Zaniewski,
Benjamin Schuman, Carolyn Wu, Alyssa Lutservitz,
Christopher D. Mullally, Nora Reed, Elizabeth Bien, Laura Bortolin,
Marian Fernandez-Otero, Jessica D. Lin, Alec Wysoker,
James Nemesh, David Kulp, Monika Burns, Victor Tkachev,
Richard Smith, Christopher A. Walsh, Jordane Dimidschstein,
Bernardo Rudy, Leslie S. Kean, Sabina Berretta, Gord Fishell,
Guoping Feng & Steven A. McCarroll

In Figure 2b of this Article, two labels were inadvertently swapped:

the brown dot should be labelled ‘PVALB+’, and the yellow dot should
be labelled ‘SST+’. In addition, the y-axis label of Fig. 4g should read
“Percentage of striatal interneurons” not “Striatal interneurons (%)”
and the numbers on the axis should be 0, 10, 20, 30, 40 (rather than
0, 0.1, 0.2, 0.3, 0.4). In the sentence “The TAC3+ interneuron population
appeared to be shared between marmosets and humans (Fig. 4g), and
constituted 30% and 38% of the interneurons sampled in marmoset and
human striatum, respectively”, ‘38%’ should read ‘34%’. These errors
have been corrected online.

Nature | Vol 588 | 10 December 2020 | E17

Corrections & amendments
Publisher Correction:
Room-temperature
superconductivity in a
carbonaceous sulfur
hydride

https://doi.org/10.1038/s41586-020-2955-8

Correction to: Nature https://doi.org/10.1038/s41586-020-2801-z

Published online 14 October 2020

Check for updates

Elliot Snider, Nathan Dasenbrock-Gammon, Raymond McBride,

Mathew Debessai, Hiranya Vindana, Kevin Vencatasamy,
Keith V. Lawler, Ashkan Salamat & Ranga P. Dias

In this Article, owing to an error in the production process, the received

date was incorrectly stated as 31 August 2020 instead of 21 July 2020.
This error has been corrected online.

E18 | Nature | Vol 588 | 10 December 2020

Corrections & amendments
Publisher Correction:
Large Chinese land
carbon sink estimated
from atmospheric carbon
dioxide data

https://doi.org/10.1038/s41586-020-2986-1

Correction to: Nature https://doi.org/10.1038/s41586-020-2849-9

Published online 28 October 2020

Check for updates

Jing Wang, Liang Feng, Paul I. Palmer, Yi Liu, Shuangxi Fang,

Hartmut Bösch, Christopher W. O’Dell, Xiaoping Tang, Dongxu Yang,
Lixin Liu & ChaoZong Xia

In the legend of Fig. 2 of this Article, owing to an error during the pro-
duction process, panels a and b were inadvertently labelled ‘northeast
China (within 38–54° N, 120–135° E)’ and panels c and d were inadvert-
ently labelled ‘southwest China (within 18–30° N, 95–110° E)’, rather
than the other way around. The figure was correct. This error has been
corrected online.

Nature | Vol 588 | 10 December 2020 | E19

 Advice, technology and tools

Work Your
story
Send your careers story
to: naturecareerseditor
@nature.com
GETTY

The pandemic is taking a toll on everyone — but the burden is larger for disadvantaged groups.

’YOU ARE ALWAYS LIVING

UNDER UNCERTAINTY’
Junior scientists who are members of minority ethnic groups or
are financially disadvantaged describe the support they need.

T
he coronavirus pandemic has affected
the entire scientific world, but in
­unequal ways: although some scien-
tists have been able to carry on with
their lives and careers, many are strug-
gling with family obligations, financial strain
and tenuous employment.
The pandemic has already dimmed job
prospects in academia, and the full impacts are
probably yet to be seen. Global gross domestic
product — the total value of goods produced
and services provided — is forecast to shrink
this year by 4.9%, according to a study by the
Pew Research Center in Washington DC. That
decline is expected to hit low-income commu-
nities and nations especially hard.
For students and postdocs from less-­
privileged backgrounds — first-generation
students, members of minority ethnic groups
or those with financial stress — the pressures
of the pandemic are, and will continue to be,
particularly intense. “There are a lot of con-
cerns about very talented individuals falling
out of the pipeline,” says Barbara Natalizio,
chair of the US National Postdoctoral Asso-
ciation, based in Rockville, Maryland, which
represents more than 40,000 postdocs.
The current crisis must be a call to action,
says Bea Maas, an ecologist at the Univer-
sity of Vienna. Maas was the lead author of a
June report on the precarity of early-career
researchers during the pandemic’s first wave.
“There must be a collective effort by the entire
scientific community, especially those in lead-
ership positions, to respond to the short- and
long-term challenges of this crisis and to

Nature | Vol 588 | 10 December 2020 | 355

©
2
0
2
0
S
p
r
i
n
g
e
r
N
a
t
u
r
e
L
i
m
i
t
e
d
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
Work / Careers
protect decades of efforts to build an inclusive
scientific community,” she says.
Nature asked five early-career research-
ers from under-represented groups to share
what they’ve experienced in the pandemic and
their thoughts about surviving in the research
enterprise.
CHRYSTAL STARBIRD
BEARING A HEAVY
EMOTIONAL BURDEN
I’m the co-founder and co-chair of the Yale
Black Postdoctoral Association at Yale Uni-
versity in New Haven, Connecticut. Balanc-
ing everything during the pandemic has
been one of the biggest challenges of my life,
which is saying something because I’ve been
through a lot. On a typical day, I get up at 4:30
or 5:00 a.m. and go to the laboratory for a few
hours. Then, I come home and help my three
kids (ages 16, 14 and 7) with their school work. I
go back to the lab in the evening and work late.
My husband is a full-time student, so he can
spend a lot of time at home. I’m grateful for
that, but we still have disadvantages. Some of
my peers have hired tutors to help with school-
ing during the pandemic. We can’t afford that.
For minorities, day-to-day struggles are
compounded by what we see on the national
level. I’m not a very emotional person. But
there have been days after a shooting or killing
when I’ve had trouble focusing. It’s extremely
heavy. We’re carrying a lot of extra weight that
makes it difficult to be scientifically fruitful.
I know people from all walks of life and of
many nationalities, and it’s obvious to me that
Black and Hispanic communities are being hit
especially hard by the pandemic. You hear
about someone’s aunt, sister, grandmother
or cousin dying or going to hospital. People
are coming to the lab and collecting data while
their families are falling apart around them.
I’m amazed by their strength.
University administrators, funders and
stakeholders must think about how to level the
playing field going forward. Yale has accom-
plished positive things, including grants to
help postdocs who have lost funds because
of the pandemic. They’ve also expanded the
number of available slots at their daycare pro-
grammes, but those cost more than US$2,000
per month. I don’t know how helpful that’s sup-
posed to be for a postdoc parent. There’s a real
sense that universities are out of touch with
what disadvantaged students are enduring.
Researchers from minority ethnic groups
are seeing job opportunities that maybe
didn’t exist before the launch of diversity ini-
tiatives this year. There’s a lot of enthusiasm
at Yale and elsewhere in the United States for
hiring scholars of colour — but those schol-
ars are extremely overwhelmed. They might
356 | Nature | Vol 588 | 10 December 2020
not have the confidence to take advantage of don’t know whether they can go back to their
those opportunities. They’re still not sure that own country, but they also don’t know whether
academia really wants them to be there. they can stay where they’re at. Either way, they
I encourage minority scientists to seize wonder whetherthey will have enough results
opportunities in academia and industry. Many to defend their thesis.
companies are realizing there’s a problem with I was not working in the lab for three months,
the make-up of their board, and they’re ready and I am at a career stage in which I need to
to embrace diversity in their hiring. Now is the produce results. I am researching the human
time for us to take advantage of momentum. microbiome, and it’s one of the most competi-
Chrystal Starbird is a structural biologist
tive fields in science right now. Analysing data
is something that you can do at home, but it’s
postdoc at the Yale School of Medicine, New not the same as producing results in the lab.
Haven, Connecticut, and co-chair of the Yale On the positive side, I’ve been able to rethink
Black Postdoctoral Association. my routine. Sitting at a computer for nine
hours per day isn’t necessary for computa-
tional work. Flexibility is important for mental
EMMA HERNANDEZ-SANABRIA health, and universities should support that. I
FOREIGN STUDENTS ARE
hope that funding agencies embrace flexibility
and change how they measure productivity.
UNCERTAIN OF THEIR PLACE
Emma Hernandez-Sanabria is a senior
Since the pandemic broke out, I think many microbiology postdoc at KU Leuven in
people have been feeling really stressed, Belgium.
including myself. You are always living under
uncertainty. At European institutions, you
don’t know how long you are going to be able
ZEMMY ANG
THERE ARE PROS AND CONS
to stay if you are from elsewhere (I grew up
in Mexico and got my PhD in Canada before
moving to Belgium for my postdoc in 2016). TO VIRTUAL LEARNING
If you don’t have a permanent position (this
is my fourth postdoc), life is very precarious. More funding would be a huge help. Students
Like many foreign postdocs in Europe, I’m from disadvantaged countries who want to
on a short-term contract. Unlike some Euro- become academics or researchers need help.
pean researchers who participate in tax-free I have friends who would be very happy to go
funding schemes, I have to pay my own income overseas and then to graduate school, but they
taxes, and I don’t have any job security. Many can’t because they have to get jobs to support
international students feel stuck because they their families. The pandemic will make things
EMMA HERNANDEZ-SANABRIA
Emma Hernandez-Sanabria working with a simulator of the human microbial ecosystem.
©
2
0
2
0
S
p
r
i
n
g
e
r
N
a
t
u
r
e
L
i
m
i
t
e
d
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
 DANIEL GONZALES
IT’S TIME TO MAKE
ACADEMIA MORE INCLUSIVE
When everything shut down in March, it took
me a couple weeks to settle into working from
home. But I don’t think anyone quite knows the
full repercussions of the shutdown. Even now,
research is still slow in many places and will
probably continue to be slow for six or eight
more months, or maybe a year. For the three
months that I couldn’t enter the lab, I tried to
remain positive.
I had time to think about some theoreti-
cal aspects of the work that I had previously
never had a chance to consider. The pan-
demic gave us an opportunity to sit back
and develop some really cool theories for the
technologies that we’re working on, including
ZEMMY ANG

nanoelectric devices that would allow for

PhD student Zemmy Ang.
that much harder. They’ll have to put their
plans on hold, or just not pursue them at all.
Even with things going virtual now, classes
are challenging — describing your equations
over Zoom is much harder than writing out
what you think on the whiteboard in class.
I’m trying to imagine what my friends in
the Philippines, where I’m from, are facing.
Everything is shut down, and it might be
impossible for them to do their work. When I
lived there, I once had to give a Skype presenta-
tion from a cafe because it was the only place
nearby that had a reliable Internet connection.
Before the pandemic, my Filipino passport
limited me to conferences in Singapore and
Hong Kong. I didn’t have the visas that I would
need to attend conferences in the United
States, United Kingdom or Europe. That’s a
big challenge for early-career researchers
from the developing world. Not being able to
go to conferences puts students from disad-
vantaged backgrounds even further behind.
So in a way, going virtual is great, although it’s
not ideal to network while staring at screens.
Zemmy Ang is a PhD student in elecro-optical
engineering at Ben-Gurion University of the
Negev, Israel.
ROSA FERRERIA
older people. Now, I’m doing remote work in
physics with Arizona State University in Tempe
so that I can apply to a PhD programme.
I live in New Jersey, and there are no PhD pro-
grammes near me with courses I want to take.
I can’t just leave, because I’m a single parent.
I have to take care of everything on my own.
I am doing as much research as I can. People
used to look down on remote academic work,
but that’s changing. It’s useful for people like
me, who don’t have those programmes nearby.
I’m afraid to move to Arizona, not only because
of all the pandemic uncertainty but also
“Stories like mine are
a real wake-up call for
people who aren’t
used to struggle.”
because people have told me that minorities
don’t always feel welcome in that state. Aside
from the money problems, this programme
could create many mental and emotional
issues for me, so there’s a lot to consider.
I’ve been working for years to start a PhD
programme, but I don’t know whether I can get
the funding for one. Even before the pandemic,
none of the institutions I contacted walked me
through the ins and outs of financing a PhD.
both electrical and optical recordings of
neural activity.
Because my job has been stable during the
pandemic, I’ve been thinking more about sys-
temic racism and what I can do to support the
racial-justice movement Black Lives Matter.
I’m thinking about what kind of culture we
want in academia and how to foster an inclu-
sive culture that’s empowering to people from
all different types of backgrounds.
We need to let more people know about
opportunities and make academia more
inviting. For example, I grew up in a rural
part of West Texas, and I was able to go to my
hometown university thanks to a wonderful
scholarship for first-generation students
that paid for my schooling expenses for four
years.
There were so many things I didn’t know as
a first-generation student. I didn’t realize that
most graduate programmes in science, tech-
nology, engineering and maths were paid for. I
didn’t realize you got a stipend. I didn’t realize
there were health-care benefits.
I chose the applied-physics programme at
Rice University in Houston, Texas, for graduate
school because it didn’t require the physics
graduate-admissions examination, which
would have cost money. I knew I wanted to
run a research lab, but just didn’t think that I
could personally achieve that. But a series of
mentors empowered me to do so.
Researchers and administrators need to
ask themselves why they’re not having more

CONCERNS OVER AN
UNCERTAIN FUTURE
My path to science is really unconventional. I
was born and raised in the Dominican Repub-
lic. My family came to the United States when
I was 16, and I taught myself English when I
was 17. I was living on the street before I put
myself through college, while working with
I want to be an astronomer, but now I don’t
know whether that will work out. I’m really
scared about it. I have no idea how things will
change as the pandemic worsens, but it doesn’t
look good. Stories like mine are a real wake-up
call for people who aren’t used to struggle.
Rosa Ferreria is a remote undergraduate
student in physics at Arizona State University
in Tempe.
of the conversations that will make science
more inclusive.
Daniel Gonzales is a physics postdoc at
Purdue University, West Lafayette, Indiana.
Interviews by Carrie Arnold and Chris
Woolston
These interviews have been edited for length
and clarity.

Nature | Vol 588 | 10 December 2020 | 357

©
2
0
2
0
S
p
r
i
n
g
e
r
N
a
t
u
r
e
L
i
m
i
t
e
d
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
Work / Technology & tools

ADVANCED IMAGING CENTER, HHMI JANELIA RESEARCH CAMPUS

The Advanced Imaging Center in Ashburn, Virginia, provides technical training for core-facility managers.

CORE CURRICULUM: LEARNING

TO MANAGE A SHARED
MICROSCOPY FACILITY
High-tech tools are increasingly being consolidated in
specialized centres. Running these technological wonderlands
takes a unique blend of skills. By Sandeep Ravindran

I
t’s the long wait for equipment that
H. Krishnamurthy remembers from his
master’s studies at Bangalore University
in India. “I used to stand in line to get my
turn to use a rusted hammer to nail [down]
a frog for dissection,” he says.
Today, Krishnamurthy directs a facility
so that other researchers in Bengaluru and
throughout India need never experience that
lack of access. The Central Imaging and Flow
Cytometry Lab at the National Center for
Biological Sciences in Bengaluru “has helped
scientists to take their research to next level”,
he says. “Before I started this facility, there
was no paper published in Cell from India.”
Since then, users have published more than
half a dozen papers in Cell, as well as others
in journals such as the Proceedings of the
National Academy of Sciences.
Over the past 20 years, as instrument costs
have risen and funding levels fallen, institutions
have increasingly consolidated microscopes,
mass spectrometers, flow cytometers and
other high-tech equipment in specialized core
facilities, where dedicated staff can cost-ef-
fectively provide a breadth of expertise and
access to equipment beyond what any single
laboratory could manage. Numbers are hard to
come by, but Peter O’Toole, director of the Bio-
science Technology Facility at the University of
York, UK, has seen meetings for UK core-facility
managers grow from a dozen participants in
2006 to around 200 today. And in Germany,
the number of imaging core facilities doubled
between 2011 and 2015, from 30 to 60.
The people tasked with running these
facilities have a rare collection of skills:
in-depth knowledge of the hardware they
oversee, managerial and financial acumen to
run what is effectively a business, and scien-
tific know-how to guide researchers through
a range of experimental systems and designs.
The management aspects alone would usually
fill three jobs — financial manager, project
manager and people manager — says Graham
Wright, acting director of the Research Sup-
port Centre at the Agency for Science, Technol-
ogy and Research in Singapore. Krishnamurthy
was once asked to list his responsibilities, and
says he was shocked at how many he had. “This
is not a 9-to-5 job, it’s a 24-hour job,” he says.

358 | Nature | Vol 588 | 10 December 2020

©
2
0
2
0
S
p
r
i
n
g
e
r
N
a
t
u
r
e
L
i
m
i
t
e
d
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
 Until a few years ago, however, there was no
clear career track, and few specialized training
opportunities. “All the people in my genera-
tion figured it out as we went along,” says Jen-
nifer Waters, director of the Nikon Imaging
Center at Harvard Medical School in Boston,
Massachusetts.
But things are changing. Waters has
launched a programme at Harvard that
provides technical training for core-facility
managers ( J. C. Waters Trends Cell Biol. 30,
669–672; 2020), and other institutions have
created similar programmes, including the
Advanced Imaging Center (AIC) at the Howard
Hughes Medical Institute’s Janelia Research
Campus in Ashburn, Virginia, and the Euro-
pean Molecular Biology Laboratory (EMBL)
in Heidelberg, Germany. And fresh funding
opportunities for managers and staff are mak-
ing the career track easier to navigate.
“Probably the most important thing I’m
going to do in my career is to help train this
next generation,” says Waters.
Combining breadth and depth
At some core facilities, users drop off samples
and receive data in return. At others, includ-
ing Krishnamurthy’s, staff train users, but
are not involved in the actual experiments.
Many microscopy facilities lie between these
extremes, with staff advising users on what
imaging techniques are best suited to their
projects and working with them as they oper-
ate the equipment.
Whatever the operational model, technical
expertise is paramount. “The most important
prerequisite to work in a core facility is that you
really have to be very good at the techniques
that you are operating, be it microscopy,
flow cytometry or mass spectrometry,” says
JENNIFER C. WATERS
Teng-Leong Chew, who directs the AIC. That
means having not just technical proficiency in
that equipment, but also a deep understand-
ing of its theoretical underpinnings, as well
as the engineering skills to be able to install,
maintain and repair it.
Also required are the skills and experience
to weigh in on a diverse array of projects and
experimental models. “One minute you’re
working on yeast and bacteria and the next
minute on brain slices, and that means not only
do we see a whole variety of scientific spec-
imens, but also a huge range of microscopy
technology,” says Alison North, senior direc-
tor of the Rockefeller University Bio-Imaging
Resource Center in New York City.
Chew, for instance, helps researchers to
work out what imaging approach to use, how
to design experiments, how to handle the
equipment and how to analyse and interpret
the resulting images — a process that can take
anywhere from hours to days of one-on-one
time. “You have to provide very good train-
ing, not just in how to use the instrument, but
[in] how to make sure that your experiment is
accurate, ethical, quantitative and reproduc-
ible,” he says.
Staying up to date with technology is cru-
cial, be it through the literature, conferences
or word-of-mouth. “That’s part of the job, to
scout for new technology and make sure you’re
using it earlier,” says Stefanie Reichelt, who
runs the light-microscopy core facility at the
Cancer Research UK Cambridge Institute at the
University of Cambridge. Popular meetings
for microscopy core-facility staff include the
European Light Microscopy Initiative; Focus
on Microscopy; the UK Royal Microscopical
Society’s Microscience Microscopy Congress;
and the Seeing is Believing conference.
Yet despite their crucial role in the conduct
of research, core facilities are often as much
businesses as laboratories, and staff rarely
receive authorship unless they also provide
significant scientific input. Acknowledge-
ments are more common, although not guar-
anteed. “Sometimes you do hard work, and
you will not see that reward directly coming
in terms of an acknowledgement,” says Jan
Peychl, head of the Light Microscopy Facility
(LMF) at the Max Planck Institute of Molec-
ular Cell Biology and Genetics in Dresden,
Germany. To help address this, the Royal
Microscopical Society and the US Association
of Biomolecular Resource Facilities (ABRF)
have developed authorship guidelines that
core facilities can provide to their users.
Service providers
As microscopy core facilities become increas-
ingly prevalent, they offer an intriguing career
option for PhD and postdoctoral researchers.
“I think it is a great career,” says North. But it’s
not a position for someone who’s just applying
“because they’re worried that they won’t get a
job” as a principal investigator, she says.
North looks for applicants who have expe-
rience with different types of specimen and in
training others in microscopy, and who have
been responsible for troubleshooting. The
exact scientific training can vary, but a “lack
of ego” is a must, she says. “If it’s all about get-
ting credit for the work you’ve done, then this
is not the right job for you.”
Indeed, core-facility staff have very differ-
ent jobs from conventional scientists, and
organizations might need to implement per-
formance metrics for core-facility managers
to match, says Wright. These metrics “are
likely not focused on the number of publi-
cations or grants awarded, but more on the
users trained, level of user satisfaction, cost
recovery, acknowledgements in publications
et cetera”, he says.
Facility-director salaries can vary widely,
from around £35,000 (US$47,000) per year for
postdoc-like positions to £90,000 for senior
managers, says O’Toole. “These can be very
different roles with a very similar job title,”
he says. Those on the lower end of the scale
are likely to have many fewer responsibilities
and staff, will not write grant applications and
might not fully control their budgets. But they
can progress by expanding their core facility
or moving to a more senior position at a larger
facility.
Before he was a core-facility director, Peychl
was a medical doctor, a stage actor and an
assistant professor. The common denomina-

tor, he says, is people skills: “I deal with hun-

dreds of people, and I need to understand their
motivations, and their needs and emotions as
well,” he says.
Harvard Advanced Microscopy Fellows Rylie Walsh and Federico Gasparoli. As a junior assistant professor at the
Charles University Faculty of Medicine in

Nature | Vol 588 | 10 December 2020 | 359

©
2
0
2
0
S
p
r
i
n
g
e
r
N
a
t
u
r
e
L
i
m
i
t
e
d
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
Work / Technology & tools
Hradec Králové in the Czech Republic, Peychl
says, he was always the one taking care of his
department’s common microscopes. “The
person who first goes and fixes it probably
is highly qualified to become core-facility
manager,” he says. “It’s important that you
have this inner drive to support people with
cool technologies.”
In 2001, Peychl’s interest in microscopy
inspired a career change, and he applied for
an imaging-specialist position at the LMF. “Of
course, they were surprised that a medical doc-
tor would apply for a technician job, but they
immediately offered me a postdoctoral fellow-
ship to support the work of the core imaging
facility,” he says.
The business end
Peychl’s experience is unusual in that training
opportunities for core-facility management
are relatively rare. But they are increasing in
number.
The German Society for Microscopy and
Image Analysis, known as German BioImaging,
has offered a course on core-facility manage-
ment and leadership in Germany since 2013,
and Global BioImaging, an international net-
work of imaging infrastructures and communi-
ties coordinated from EMBL Heidelberg, offers
similar courses around the world. The ABRF
and Core Technologies for Life Sciences, a Par-
is-based non-profit association of core-facility
scientists and staff, have workshops that teach
business skills such as accounting and budget-
ing. Staff can sometimes take business-school
classes offered by the universities where their
facilities are based. And some managers,
including Wright, even complete a master’s
of business administration, which he says has
been “incredibly useful”.
German BioImaging and Global BioImaging
also organize job-shadowing programmes,
which allow staff to pick up skills from their
peers. “I see real value in visiting other core
facilities to understand how they operate and
bring home the relevant bits to help improve
our own operations,” says Wright.
Grant-writing skills are essential. “The
success of a core facility, to some extent, is
reflected by the number of shared instrumen-
tation grants that it secures,” says Chew. In the
United States, the National Science Founda-
tion and Department of Defense both fund
core facilities, and the National Institutes of
Health’s S10 shared instrumentation grant is
“almost tailor-made for core-facility direc-
tors”, Chew says. Facilities in other parts of
the world, including most European coun-
tries, are also funded mainly by government
grants.
Most of these grants cover only instru-
mentation, but some, including the German
Research Foundation’s core facilities funding
programme and the imaging scientists pro-
gramme from the philanthropic organization
the Chan Zuckerberg Initiative in Redwood
City, California cover staff, too.
Facility staff positions are often not perma-
nent, which can lead to regular loss of institu-
tional knowledge. “Usually you need minimum
of a year to get a person up to speed,” says
Peychl. Such turnover presents challenges,
he says, but it also encourages flexibility.
For instance, facility managers had to react
quickly when faced with the COVID-19 pan-
demic, both to keep their staff and users safe
and to minimize disruptions.
“In the end, it’s all
about hands-on
experience.”
Peychl, working with Elisa Ferrando-May at
the University of Konstanz Bioimaging Center
in Germany and others, published COVID-19
guidelines for core facilities, such as avoiding
face-to-face training and providing remote
support to microscope users (S. Dietzel et al.
Cytometry A 97, 882–886; 2020). Roland
Nitschke, head of the Life Imaging Center at
the University of Freiburg, Germany, set things
up at his facility so that he could remotely con-
trol microscopes to show users what to do. The
set-up required ultra-high-resolution cameras,
and scientists started using the same cameras
to check in on long-running experiments from
home. “Maybe it’s the only good thing to come
out of the coronavirus,” says Nitschke.
A core crash course
To bring new recruits up to speed, some man-
agers turn to external courses and workshops.
Waters teaches yearly microscopy courses at
Cold Spring Harbor Laboratory in New York,
and North does the same at the Woods Hole
Marine Biological Laboratory in Massachu-
setts. And Krishnamurthy provides a training
programme on confocal microscopy and flow
cytometry in Bengaluru. But, “in the end”, says
Peychl, “it’s all about hands-on experience”.
Most researchers use just one or two types
of microscopy during their PhD or postdoc.
“We have everything from single molecules to
organoids and embryos coming into the core,
and we have 15 different imaging modalities,”
says Waters. “The idea of anybody walking out
of a postdoc and into a core facility, and being
able to immediately know what they’re doing,
is ridiculous,” she says.
That’s what motivated Waters to start Har-
vard’s Advanced Microscopy Fellowship in
2013. Waters originally funded two fellows by
reallocating salary for a staff member, but has
since been supported by microscope manufac-
turer Nikon and three departments at Harvard
Medical School. The programme costs about
US$80,000 a year for each fellow, and partic-
ipants are guaranteed two years of funding.
Fellows spend half of their time on core
service work, such as training users and main-
taining microscopes, and another half on an
independent project, such as writing code to
assess the quality of microscopy images.
Fellows also learn management and budget-
ing skills, and work on grants with Waters. And
they learn the fine art of dealing with clients.
“I want them to have the confidence to sit in a
room with a group of faculty and say, ‘I know
that technology looks interesting, but you
don’t have the need for it,’” says Waters. “If I
have to write an uncomfortable e-mail to a fac-
ulty member, I will write it and then forward it
to my postdoc so they can see how I worded it,”
she says. “They get to see how those situations
are handled, and they leave with a little bit of a
toolkit of how to manage that.”
Waters has two current fellows and four
alumni who have gone on to work at or run
other core facilities. “I want them all to walk out
of here being technical experts who deserve
respect and are given autonomy,” she explains.
Staff might also go into industry, especially to
microscopy companies.
Other institutes have created similar
programmes. For instance, Janelia’s Advanced
Microscopy Fellows spend half of their time
getting hands-on experience at the AIC and the
other half on self-directed training, whether
that’s learning how to code or how to run an
international microscopy workshop. “That
50% protected time allows them to hone their
skill to become a successful core director in
the future,” says Chew.
In Europe, EMBL’s ARISE fellowship
programme is launching training for 62 fellows
over the next five years, supported by some
€12.7 million (US$15 million) from the Euro-
pean Commission and EMBL. Calls for applica-
tions went out in November, and fellows will be
hosted at one of six EMBL sites across Europe.
“We want to build through the programme
the future heads or senior staff of research
infrastructure facilities,” says Tanja Ninkovic,
ARISE programme manager at EMBL Heidel-
berg. As well as learning the ins and outs of
core-facility management, fellows will pursue
an independent research project and study
technology transfer, entrepreneurship and
science policy. “These are the skills that are
needed by research infrastructure scientists
regardless of the technology that the facility
is offering,” says Ninkovic.
Peychl, who attended a German BioImaging
course for facility managers, sees the increase
in core-facility management courses as a sign
that these positions are increasingly valued.
And he encourages others to consider the role.
“For those who like to support others, who are
technologically inclined and who have people
skills, this could be a rewarding career.”
Sandeep Ravindran is a science writer based
in Washington DC.

360 | Nature | Vol 588 | 10 December 2020

©
2
0
2
0
S
p
r
i
n
g
e
r
N
a
t
u
r
e
L
i
m
i
t
e
d
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
The back page

I
Where I work have just spent six months in the Great
Barrier Reef on the research vessel
John Fulmer Falkor. Every cruise tackles cutting-edge
science and, as lead technician, I’ve
been able to witness so many firsts: the
RV Falkor is the first ship to map much of this
area at high resolution, and to put a remotely
operated vehicle (ROV) down in the Great
Barrier Reef at depths of up to 2,000 metres.
So everything we’re seeing is new. With the
ROV’s two arms, we gather samples of rocks,
corals, sediments, jellyfish — everything we
come across down there — for collaborators
to analyse later.
Recently, we discovered a peculiar knoll
rising from the sea floor 2,000 metres down,
peaking 300 metres below the surface.
The odd thing is that it shouldn’t be there.
There’s little to no volcanic activity in the
area to create a mound, and it should have
been eroded by water like its surroundings.
Analysis of the rock samples we took from
the knoll might explain things.
I’m the main liaison between the ship’s
Photograph by the crew and the scientists travelling with us.
Schmidt Ocean Institute. Three technicians maintain all the shipboard
systems, including an instrument to measure
ocean conductivity, temperature and
density; a multibeam scanner to map the
ocean floor; and the ROV SuBastian, which I’m
working on in the photo.
We’ve been lucky that our institution,
the Schmidt Ocean Institute in Palo Alto,
California, has kept the work going during the
pandemic. Everyone undergoes a two-week
hotel quarantine and COVID‑19 testing before
getting on board. Other researchers haven’t
been able to get to sea, so we livestream our
ROV dives. Collaborators can log in and tell
us what looks worth sampling. Anyone can
watch and ask the scientists questions, and
receive live answers.
Science never sleeps. I work 12-hour shifts
every day for 6-month stretches, with no
days off, and I’m on call all the time. I’m just
starting six months on leave, but sometimes
I’ll tune in to the dives. It’s exciting stuff.
John Fulmer, lead technician for the
Schmidt Ocean Institute’s RV Falkor, is
currently on leave in Halifax, Canada.
Interview by Amber Dance.

362 | Nature | Vol 588 | 10 December 2020

©
2
0
2
0
S
p
r
i
n
g
e
r
N
a
t
u
r
e
L
i
m
i
t
e
d
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
outlook
Sustainable nutrition
For more on
sustainable nutrition
visit nature.com/
collections/
sustainable-nutrition-
outlook

I
Editorial f it was easy to change the way we eat, malnutrition in all its forms Contents
Catherine Armitage, Herb Brody, — undernutrition, overnutrition and micronutrient deficiency —
Richard Hodson, Jenny Rooke S54 GLOBAL DIET
could have been eliminated long ago. Everyone would have access
Healthy people, healthy planet
Art & Design to affordable food and choose to eat the quantity and variety that
Eating habits must change if we
Mohamed Ashour, Annthea Lewis, keeps them in optimal health. are to feed 2050’s population
Denis Mallet That’s the dream. The reality is that humans have a food problem. It
is a complex and multidimensional issue, but in broad brush strokes: S57 OPINION
Production
Cooperate to prevent food-
Nick Bruni, Kay Lewis, Ian Pope, Karl huge numbers of people go hungry; large nutritional imbalances per-
system failure
Smart sist between high- and low-income nations and regions; and the food Jessica Fanzo explains why
system, from production to supply and consumption, is both failing governments need to work
Sponsorship
Yuki Fujiwara, Natasha Boyd, society and damaging the planet. together
Takeaki Ishihama The COVID-19 pandemic has highlighted the problems, and exacer-
S58 AGRICULTURE
bated them (see page S57). Political, economic and cultural obstacles Natural solutions for
Marketing
are prolific on the pathway to achieving a sustainable global diet (S54). agricultural productivity
Gavin Buffett
Even if a suitable global menu could be agreed on, changing eating How to farm intensively and
Project Manager behaviours at scale is a formidable and understudied challenge (S70). sustainably
Rebecca Jones The mainly plant-based diet that nutritional scientists recommend S60 AQUACULTURE
Creative Director for physical and, more recently, mental health (S63) is better for the Cultivating a sea change
Wojtek Urbanek environment than diets that are heavy in meat and highly processed Growing the seafood industry
foods. To reduce our reliance on farmed meat, scientists around the sustainably
Publisher
world are developing affordable protein alternatives. Researchers are S63 HEALTH
Richard Hughes
racing to transform lab-grown meat from a headline-grabbing novelty Eating for better mental health
VP, Editorial into a viable industry supplying supermarkets (S64). And according Mood could be linked to the
Stephen Pincock to projections, aquaculture is ramping up to overtake wild fish stocks microorganisms in our gut

Managing Editor as the main source of aquatic protein in diets by 2050 (S60). Farming S64 CELLULAR AGRICULTURE
David Payne methods that intensify agricultural production while rebuilding and Cell-based meat with a side of
sustaining natural systems are also becoming more widespread (S58). science
Magazine Editor Growing meat at scale is still a
Helen Pearson
Diversity is key. There is no single solution that will guarantee sus-
tainable nutrition for everyone. In the same way that the pandemic challenge
Editor-in-Chief demands an integrated, cooperative and global response, in which S68 SUSTAINABLE NUTRITION
Magdalena Skipper science plays its part, so does feeding the global population. Research round-up
We are pleased to acknowledge the financial support of The latest studies
Ajinomoto Co., Inc., Yakult Honsha Co., Ltd and NTT Corporation in
S70 BEHAVIOUR
producing this Outlook. As always, Nature retains sole responsibility Changing diets at scale
for all editorial content. Different communities and
cultures require different
Catherine Armitage approaches
Chief editor, Nature Index

About Nature Outlooks available free online at go.nature.

On the cover
Eating well for the health
of the planet. Credit:
Sophie Casson for Nature
Nature Outlooks are supplements
to Nature supported by external
funding. They aim to stimulate
interest and debate around a subject
of particularly strong current
interest to the scientific community,
in a form that is also accessible to
policymakers and the broader public.
Nature has sole responsibility for
all editorial content — sponsoring
organizations are consulted on the
topic of the supplement, but have
no influence on reporting thereafter
(see go.nature.com/33m79fz). All
Nature Outlook supplements are
com/outlook
How to cite our supplements
Articles should be cited as part of a
supplement to Nature. For example:
Nature Vol. XXX, No. XXXX Suppl.,
Sxx–Sxx (2020).
Contact us
feedback@nature.com
For information about supporting a
future Nature Outlook supplement,
visit go.nature.com/partner
Copyright © 2020 Springer Nature
Ltd. All rights reserved.

Nature | Vol 588 | 10 December 2020 | S53

©
2
0
2
0
S
p
r
i
n
g
e
r
N
a
t
u
r
e
L
i
m
i
t
e
d
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
Sustainable nutrition
outlook

SOPHIE CASSON FOR NATURE

Healthy people, healthy planet
To provide 2050’s estimated 10 billion people with a healthy diet, global eating
habits need to become more sustainable. By Chris Woolston

E
very morsel of food from every plate,
bowl and cooking pot around the
world takes a small bite from Earth’s
resources. The human diet places
a strain on the environment, water
resources, biodiversity and just about every
other measure of planetary health. With so
much at stake, researchers have turned their
attention to a pressing question: what sort of
diet can the planet realistically support?
The answer requires insights from fields
such as nutrition, agriculture and climate
research. “We need to produce food groups
that are good for health in ways that are restor-
ative to the planet, rather than extractive,”
says Corinna Hawkes, director of the Centre
for Food Policy at City, University of London.
The particular foods on the plate will vary from
one place to another, she says, but those meals
need to add up to something more sustainable
than society’s current fare.
“When you look carefully at the big sys-
tems that regulate the stability of our planet,
food is a dominant player in essentially all of
them,” says Johan Rockström, an environmen-
tal scientist at Stockholm University. In 2019,
Rockström, Hawkes and other members of an
international group of scientists proposed the
EAT-Lancet diet1, a global meal plan that could,
in theory, feed 2050’s estimated population of
10 billion people (see ‘Planetary-health diet’).
That plan called for drastic cuts in meat con-
sumption and a much higher intake of fruits
and vegetables. But it proved controversial
with meat-industry proponents and econ-
omists, and the quest for a planetary diet

S54 | Nature | Vol 588 | 10 December 2020

©
2
0
2
0
S
p
r
i
n
g
e
r
N
a
t
u
r
e
L
i
m
i
t
e
d
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
APHRC

Elizabeth Kimani-Murage addresses community members at a meeting about food insecurity in Nairobi, Kenya.

continues. When researchers and policy-
makers convene at the United Nations Food
Systems Summit in late 2021, a healthy-planet
diet will be near the top of the agenda.
The goal will be a basic framework, not an
item-by-item menu, says Agnes Kalibata, a
food-policy specialist in Kigali, Rwanda, who
will be leading the summit as the UN special
envoy. “Diets are influenced by cultures and
custom,” she says. “We can come up with the
principles of what a good diet will look like. We
need to find a balance.”
Sustainability on a plate
Most researchers agree that the current diet
is not sustainable. A 2018 analysis2 estimated
that food production releases the equivalent
of 13.7 gigatonnes of carbon dioxide in green-
house gases into the air each year — more than
one-quarter of all human-caused greenhouse
gases. The same report estimated that agricul-
tural irrigation accounts for about two-thirds
of all fresh water used by humans. And about
37% of the planet’s land area, excluding deserts
and ice sheets, is already dedicated to food
production. That footprint is likely to grow
as the population increases.
Some foods take up many more resources
than others. At the upper end, just 100 grams
of beef protein can result in the release of
the equivalent of 105 kilograms of CO2. The
same amount of protein from a well-managed
field of peas, by contrast, typically releases
the equivalent of only about 0.2 kg of CO2.
These orders-of-magnitude differences
mean that any vision of a more sustainable
diet has to include marked reductions in the
meat consumption of high-income countries,
Hawkes says. She notes that consuming a lot
of red meat can raise the risk of cancer and
heart disease. “It’s not great for our health,
and it’s not great for our planet,” she says.
“There’s a strong alignment between health
and sustainability.”
“We have to rethink our
diets based on who are
the most vulnerable
among us”.
This convergence of nutrition and conser-
vation is a central message of the EAT-Lancet
diet. The authors started by reviewing the
best evidence for constructing a diet that
would optimize human health and reduce
the global toll of food-related health condi-
tions, such as diabetes, heart disease, cancer
and obesity. The researchers didn’t even con-
sider the impacts on climate or sustainability
until the nutritional framework had been set,
Rockström says.
The EAT-Lancet commission ultimately pro-
posed a ‘flexitarian’ diet that spans a spectrum
of food groups. It also suggested vegan and
vegetarian options. Plants form the foun-
dation of the commission’s flexitarian diet,
which recommends the daily consumption
of 300 g of vegetables, 200 g of fruit, around
230 g of whole grains and 125 g of plant-based
protein-rich foods, such as lentils, nuts and dry
beans. The diet calls for a mere five servings of
animal protein per person per week, including
about 200 g of fish and 200 g of white meat.
Controversially, the diet allows for just 100 g or
so of red meat, around one and a half servings,
per week. Rockström notes that’s a significant
reduction from the roughly 700 g of red meat
consumed each week by people in places such
as North America and Europe, but it’s much
more than the amount typically eaten by peo-
ple in low-income countries.
Despite its dire environmental impacts, meat
still has an important place in the global diet.
On a nutritional level, the proteins and minerals
from animal products could be a real boost for
malnourished populations around the world,
Hawkes says. “For infants who otherwise eat
rice or starchy cassava, meat is an incredibly
efficient way of boosting micronutrient sta-
tus,” she says. What’s more, she says, “meat has
tremendous cultural significance in people’s
lives — it’s associated with high status”.
Expensive gains
After investigating the potential environ-
mental impacts of the EAT-Lancet diet, the
authors concluded that a nutritious diet for
people could also be good for the planet.
“We found that a healthy diet combined
with sustainable agricultural practices would
have positive impacts on biodiversity, land,
water, nutrients and climate,” Rockström
says. The most significant improvements
tied to a change in diet would come from a
reduction in phosphorus and nitrogen pol-
lution in waterways and greenhouse-gas
emissions. The commission estimated that
the new meal plan could cut related green-
house-gas emissions by about half — down to

Nature | Vol 588 | 10 December 2020 | S55

©
2
0
2
0
S
p
r
i
n
g
e
r
N
a
t
u
r
e
L
i
m
i
t
e
d
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
Sustainable nutrition
outlook

PLANETARY-HEALTH DIET would look much like the EAT-Lancet diet, but
If every person had a daily food allocation to sustain not only their health but also that of the planet, what would it with a decidedly East African flavour.
look like? The answer to this question in a study¹ from the EAT-Lancet commission, places an emphasis on plant-based
The proposal from her institute is one of ten
foods, and recommends an amount of animal-derived protein much lower than that eaten in high-income countries
but much higher than the amount consumed in low-income countries. to make the finals of the Food System Vision
Macronutrient intake (grams per day) Prize, a global contest sponsored by the
Tubers or starchy Rockefeller Foundation in New York City. The
Vegetables Whole grains Protein sources Fruit vegetables
209 200
winners are due to be announced in Decem-
300 232 50
ber. Kimani-Murage says she wants Kenya to
move away the sort of large-scale industrial
Dairy foods Added fats farms that currently feed cities around the
250 52 world. “Food has been so commercialized or
Added sugars commodified,” she says. “It’s produced for
Legumes Nuts Poultry Fish Eggs 31
75 50 29 28 13 money and not for feeding people. We want
to continue this local production of food
even as the world urbanizes.” She notes that
Beef, lamb and pork local food production could also significantly
14
reduce the costs of production and shipping,
potentially increasing the affordability of an
IMPACT OF UNUSED FOOD EAT-Lancet-style diet.
Food loss (from post-harvest through the supply chain and up to, but not including, retail) and waste (at retail Feeding people at a local level is the key
and consumption level) from the main food groups have negative environmental impacts. They all have a focus of the 2021 UN Food Systems Summit.
blue-water (cubic metres of water wasted), carbon (tonnes of carbon dioxide equivalent omitted) and land
(hectares of land used) footprint per tonne of food lost or wasted.
The current global diet, Kalibata says, is unbal-
anced, largely because of gaps in wealth and
Cereals and pulses Fruits and vegetables Roots, tubers and oil-bearing crops Meat and animal products opportunity. In poor areas around the world,
9% people tend to fill their stomachs with starchy,
35% 44% Blue-water footprint carbohydrate-heavy food because they can’t

SOURCE: REF. 1 (TOP); FAO (BOTTOM)

Carbon footprint
Food loss
and waste
Land footprint
0 20
13%
about 5 gigatonnes of CO2 equivalent.
If the EAT-Lancet diet was adopted, it would
undoubtedly be a healthy step forward for
people and the environment. But it has faced
fierce opposition over its potential to devas-
tate the animal-husbandry industry, and has
been criticized as being too expensive for
many consumers. One analysis3 calculated that
nearly 1.6 billion people would be too poor to
buy the recommended mix of foods, especially
the meat, fruits and vegetables. “No amount
of nutritional knowledge is going to get them
there because they can’t afford it,” says William
Masters, one of the study’s authors and a nutri-
tional economist at Tufts University in Boston,
Massachusetts. “They may have US$1 a day to
spend on food but they would need $1.50,”
he says.
The analysis found that the EAT-Lancet
model was about 60% more expensive than the
cheapest alternative diet that could provide
all 20 essential nutrients people need to sur-
vive. That bargain diet — which consists almost
entirely of starchy staples, such as rice, cassava
and flour, with very little fish, meat, fruits or
vegetables — would also be more environmen-
tally friendly, but Masters cautions that it is not
40 60 80
Proportion (%)
healthy. It lacks, among other things, the fibre
needed for optimal digestion, the phytochem-
icals that can protect against cardiovascular
disease and cancer, and the healthy fats that
support the brain.
Poverty is a crucial barrier to improved
global nutrition, but, Masters says, it is impor-
tant not to lose sight of the greater proportion
of the world’s population that could afford to
eat better, but doesn’t. “A vastly larger num-
ber of people could walk into a grocery store
tomorrow and buy a healthier diet that’s more
environmentally sustainable than the one they
eat now,” he says.
Local eating, global impacts
Any modification to the global diet will have
to start with changes at a local level. Elizabeth
Kimani-Murage, a nutrition specialist with
the African Population and Health Research
Center in Nairobi, Kenya, predicts a future in
which residents of her city feed themselves
with locally grown foods from kitchen gardens
and urban fruit trees. Small animals, such as
chickens, rabbits, and even termites and crick-
ets, would add much-needed protein. The bal-
ance of fruits, vegetables, grains and protein
afford other, more nutritious alternatives. “We
have to rethink our diets based on who are the
most vulnerable among us,” Kalibata says. She
says that high-quality proteins, whether meat-
or plant-based, need to replace many of the
100 carbohydrates eaten in poorer areas.
Kalibata thinks it will be possible to meet
nutritional needs in the future, but it will take
concerted effort to reduce waste (see ‘Impact
of unused food’), localize production and
expand the food options of the global poor. All
these issues will be on the agenda at the sum-
mit, and Kalibata hopes they’ll inspire a global
plan of action in the years to come. “We’ve had
food summits before,” Kalibata says. “We have
to make this one different. We have to deliver
on the goals.”
Rockström hopes that the EAT-Lancet plan,
despite its shortcomings, will still serve as
inspiration for efforts to put humanity’s nutri-
tional needs on a more sustainable footing. “It
is not the final answer in any way,” he says, “But
it was the first time people got together from
disciplines in health, agriculture and sustaina-
bility to try to answer the big questions. If we’re
going to take seriously that we’ll have 10 billion
citizens who all need to eat, we’ll need to live
in a certain balance.”
Chris Woolston is a freelance writer in Billings,
Montana.
1. Willett, W. et al. Lancet 393, 447–492 (2019).
2. Poore, J. & Nemecek, T. Science 360, 987–992 (2018).
3. Hirvonen, K., Bai, Y., Headey, D. & Masters, W. A. Lancet
Glob. Health 8, e59–e66 (2020).

S56 | Nature | Vol 588 | 10 December 2020

©
2
0
2
0
S
p
r
i
n
g
e
r
N
a
t
u
r
e
L
i
m
i
t
e
d
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
 Sustainable nutrition
outlook

Cooperate to
activities involved in producing, processing, distribut-
ing, preparing and consuming food, and the people who
influence those activities — in multiple ways. It is reducing

prevent food- food-production capacity, slowing distribution and lim-

iting access to both markets and financial or nutritional

system failure safety nets. Farmers are economically vulnerable owing

to the tight profit margins associated with their industry.
Government restrictions on the movement of people have
hindered farmers’ access to necessary goods, labour and
To keep people nourished equipment, slowed the planting and harvesting of crops
and affected feeding of livestock. Restrictions have also
during a global pandemic, our impeded the ability to move food to markets, ports and
food systems must evolve, across borders, leading to increases in food loss — particu-
larly of perishables, such as meat and dairy. Food waste has
and governments must work also increased as a result of reductions in the available work-
together, says Jessica Fanzo. force at meat-processing plants. Together with higher levels

A
s journalist Joan Didion wrote in her 1967 essay
‘Goodbye to All That’, “It is easy to see the begin-
nings of things, and harder to see the ends.” She was
writing about her love affair with the city of New
York, but the same can be said of COVID-19.
How the pandemic began is reasonably well understood —
the virus, SARS-CoV-2, probably made its way from wild bats
CHRIS HARTLOVE
of unemployment and loss of income, this has resulted in
an increase in the number of people struggling to access a
healthy diet. Many people are already opting instead for
staple grains and unhealthier, highly-processed foods that
are cheaper and have longer shelf lives4. This pandemic,
and the need to stave off the next one, bolsters the already
compelling case for ensuring the global food supply is safe,
nutritious and equitably distributed.
Governments and businesses should prioritize ensuring

to humans through a food market. COVID-19 is the latest in a
long line of diseases that have crossed from animals to peo-
ple, including HIV/AIDS, severe acute respiratory syndrome
and Ebola. In fact, 60% of emerging infectious diseases are
zoonotic, and of the pathogens that cause these, at least 71%
originate in wildlife1. The reshaping of habitats around the
world, often initiated by the need to grow more food, puts
people in ever closer contact with wild animals and makes
the transmission of infections more likely.
How the pandemic will end, and what damage it will cause,
is less clear. So far, there is no end in sight. Many people will
be affected forever — economically, physically, socially and
psychologically. The World Bank estimates that up to 115 mil-
lion extra people will fall into extreme poverty (living on
less than US$1.90 per day) in 2020 owing to the economic
shocks of the pandemic. This, in turn, will have significant
impacts on food security, nutrition and health. It is projected
that 130 million more people will face acute food insecurity
by the end of 2020, in addition to the estimated 135 million
who faced it in 2019.
The health of those who are already undernourished could
decline further — particularly older, vulnerable and margin-
alized people. Disruptions to health care in many low- and
middle-income countries owing to COVID-19 could lead
to around 193,000 additional deaths among children per
month2. Obesity and non-communicable diseases are signif-
icant risk factors for hospitalization with COVID-19, and they
can result in medical complications for both young and older
people. Obesity and metabolic disorders are also factors in
the disproportionate risks of hospitalization and death in
low-income and ethnic minority populations in high-income
countries. In Chicago, Illinois, for example, nearly 70% of the
people who have died from COVID-19 were Black, although
Black people make up only 30% of the population3.
Early evidence suggests that the pandemic is trounc-
ing the functionality and efficiency of food systems — the
that producers are making healthy food and that consumers
have access to it. They should support and invest in food-as-
sistance programmes during and after the pandemic. Gov-
“This ernments must support the United Nations’ $10-billion
COVID-19 Global Humanitarian Response Plan, set up so its
pandemic agencies can provide the most marginalized and vulnerable
bolsters populations with basic services, such as COVID-19 testing
the already materials, medical equipment, food, water and basic health
coverage, such as vaccines. As of September, the programme
compelling had received less than 30% of its target.
case for The integrated One Health approach (addressing risks
ensuring at the intersection of human, animal and environmental
health) is crucial in responding to COVID-19, recovering
the global
from it and preparing for the next zoonotic pandemic. To
food supply minimize viral reservoirs and contact between virus-carry-
is equitably ing animals and people, wildlife habitats must be protected
distributed.” against urbanization and deforestation. Governments need
to police the illegal sales of wildlife in food markets and
the global food trade, and to complement this with pub-
lic-health disease-prevention programmes and messaging.
Stronger surveillance tools to track potential zoonotic and
food-borne illnesses across food systems are also needed.
These recommendations to ensure food systems func-
tion effectively during the pandemic and long after cannot
work without a united global effort. Instead of the splin-
tered responses to the COVID-19 crisis seen so far, involving
political polarization and geopolitical competition, pol-
iticians must embrace global cooperation and inclusion.
Jessica Fanzo Governments should not face inward. They should double
is a food-system down on opportunities to re-engage and collaborate on the
researcher and interlinked challenges of climate change, malnutrition and
nutritionist at Johns environmental collapse.
Hopkins University in
Baltimore, Maryland. 1. Cutler, S. J. et al. Emerg. Infect. Dis. 16, 1–7 (2020).
2. Roberton, T. et al. Lancet Global Health 8, e901–e908 (2020).
e-mail: jfanzo1@jhu. 3. Yancy, C. W. J. Am. Med. Assoc. 323, 1891–1892 (2020).
edu 4. Belén Ruiz-Roso, M. et al. Nutrients 12, 1807 (2020).

Nature | Vol 588 | 10 December 2020 | S57

©
2
0
2
0
S
p
r
i
n
g
e
r
N
a
t
u
r
e
L
i
m
i
t
e
d
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
Sustainable nutrition
outlook
Many agricultural researchers are now look-
ing to a set of practices known as sustainable

THE ‘PUSH–PULL’ FARMING SYSTEM: CLIMATE-SMART, SUSTAINABLE AGRICULTURE FOR AFRICA/ICIPE/GREEN INK LTD. UK
intensification. The specifics vary depending
on the setting, but a growing number of exam-
ples from around the world highlight the pos-
sibility of a second green revolution — one that
might better live up to its name.

Many roads to sustainability

The concept of sustainable intensification was
popularized2 in 1997 by Jules Pretty, an envi-
ronmental scientist at the University of Essex
in Colchester, UK. His goal was to challenge the
idea that increasing yield is inherently incom-
patible with environmental health, with an
agricultural philosophy that encompasses
parameters such as biodiversity and water
quality as well as the social and economic
welfare of farmers. Researchers have defined
the scope of sustainable intensification in dif-
ferent ways, but the big picture, says Pretty,
entails recognizing that agriculture is inexo-
rably connected with the environment and
designing cultivation strategies accordingly.
“Components of sustainable systems tend to be
multifunctional,” he says. “You want a diverse
system that provides support to pollinators,
A farmer inspects her maize crop, grown using a ‘push–pull’ approach. fixes nitrogen and provides a break against
insects.” Advocates of sustainable intensifica-

Natural solutions for

tion recognize that global agriculture can’t be
reinvented in one fell swoop and that progress
will come from incremental steps that improve

agricultural productivity
efficiency, as well as more-dramatic measures
that redesign the farming landscape.
Lucas Garibaldi, an agroecologist at the
National University of Río Negro in Bariloche,
Argentina, has focused on pollinators as a
Scientists are pursuing sustainability strategies for crucial component of what he calls ecologi-
cal intensification. “Crop yield depends not
intensifying production to tackle food security and only on the count of pollinators, but also on
environmental crises. By Michael Eisenstein the biodiversity of pollinators,” says Garibaldi.

O
n paper, the global agriculture sector
has done an admirable job of keep-
ing pace with a growing population.
According to the United Nations’ Food
and Agriculture Organization, agricul-
tural output per person has increased by 50%
since 1960 — impressive, considering the num-
ber of mouths to feed has more than doubled.
But the reality is messier. Many people,
including those in high-income nations, lack
reliable access to nutritious food. And food
security is an ongoing struggle for people in
poorer regions. Even transient disruptions can
have far-reaching consequences. One article1
described the global food supply as being “on a
razor’s edge” — weather events or natural disas-
ters in one part of the world can cause the price
of grain everywhere to spike by more than 50%.
“Globally, we have to increase food production
by 60%, and in some areas we have to increase
by 100%,” says P. V. Vara Prasad, a crop ecophys-
iologist at Kansas State University, Manhattan.
Over the past 50 years, producers increased
agricultural output in much of the world
through the ‘green revolution’. But this revolu-
tion has been environmentally harmful, relying
heavily on chemical pesticides and fertilizers
that have inflicted lasting damage on the soil
and water supply. Natural biodiversity has
been sacrificed to create vast monoculture
fields. And in many low-income nations, sur-
vival depends on coaxing greater productivity
from existing plots as more and more people
scramble for limited resources, says Bernard
Vanlauwe, a soil scientist based in Nairobi at the
International Institute of Tropical Agriculture.
“Millions of honeybees alone will not replace
the function of diverse species of wild bees
and butterflies and birds.” He notes that dif-
ferent bees pollinate different crops, but also
allow more efficient pollination for some plant
species. To create a haven for these airborne
assistants, Garibaldi advocates minimizing
pesticide use and including non-agricultural
zones in farmland. These could be wild-plant
borders that surround fields or just hedge-
row-like strips of flowers that are appealing to
the bees that traverse them.
Growing a mix of crops can have many bene-
fits, including attracting pollinators. Conven-
tional monoculture leaves soil exposed for
much of the year, Garibaldi says. This creates
opportunities for weeds to grow — necessi-
tating herbicides — or leaves soil susceptible
to erosion. With multiple crops or rotation
throughout the year, more durable root

S58 | Nature | Vol 588 | 10 December 2020

©
2
0
2
0
S
p
r
i
n
g
e
r
N
a
t
u
r
e
L
i
m
i
t
e
d
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
 systems that densely and extensively perme- are concerned that the approach is untested
ate the ground can be established, reinforcing and unproven. Last year, Panjab Singh, presi-
the soil and preventing the nutrient depletion dent of the National Academy of Agricultural
associated with long-term monoculture. Sciences in Delhi, told the newspaper The
Diversity can also eliminate the need for Hindu, “We are worried about the impact on
pesticides. Pretty says around 180,000 farm- farmers’ income, as well as food security.”
CHRIS GOMERSALL/2020VISION/NATUREPL.COM

ers in Kenya, Uganda and Tanzania now use
push–pull cropping practices when growing
maize. They plant grasses around the edges of
maize plots that produce chemicals that ‘pull’ a
common pest, the maize stalk borer (Busseola
fusca), away from crops, while the maize itself
attracts parasitic wasps that prey on the stalk
borer. The farmers also intersperse legumes
of the genus Desmodium with the maize that
enrich the soil with nitrogen, and produce
compounds that ‘push’ away pests and kill off
a genus of invasive weed known as Striga.
Sustainable soil management is a thorny
issue, particularly in resource-limited settings.
Vanlauwe notes that nutrient depletion is one
of the greatest threats to yield for African farm-
ers, making a hard-line approach to sustainabil-
ity unrealistic. “People who say you can trigger
agricultural development in Africa without
fertilizer do not have on the ground experi-
ence,” he says. But there are environmentally
friendly ways to feed the soil. Jo Smith, a soil
scientist at the University of Aberdeen, UK, has
been equipping farmers in Africa and Asia with
anaerobic digesters — simple systems that use
microbes to convert animal manure into biogas
for fuel and leave a nutrient-rich bioslurry. “It’s
like giving them a little fertilizer factory — it
gives you available ammonium that the crop
can take up quickly,” she says. The biogas is also
less harmful than conventional fuels, reducing
household air pollution and improving quality
of life, Smith adds.
Much of the world’s farming takes place
on smallholder plots. One study3 estimated
that one-third of the global food supply is
produced on farms of less than two hectares.
This fragmentation can make it challenging
to introduce sustainable intensification prac-
tices. “Smallholder production systems are
absolutely risk-averse,” says Vanlauwe. “Falling
from earning US$100 to $50 a month can be
the difference between being not-hungry and
being hungry.”
Close collaboration with individual farmers
is needed, but this is difficult to achieve at scale.
Fortunately, smallholders are increasingly par-
ticipating in collectives that can accelerate
information sharing and reduce the risk associ-
ated with adopting new cultivation strategies.
In August4, Pretty and his colleagues reported
that, worldwide, around 8 million such groups
have formed over the past two decades. “That’s
about 240 million people working in collec-
tive-action efforts around areas like irrigation,
Crops rely on pollinators such as bees.
forest management, pest management and
water,” says Pretty. By partnering with these
groups, researchers can design programmes
that are more likely to be compatible with
social, cultural and environmental conditions,
and establish local networks of collaborators
to facilitate the dissemination of information.
Some governments are also taking a more
active role. Ethiopia, for example, has focused
on aspects of ecological repair by establishing
‘exclosure’ areas for depleted soils. “Areas are
fenced off, and after about ten years the land
starts to recover,” Smith says.
In China, Fusuo Zhang, a plant-nutrition
specialist at the China Agricultural University
in Beijing, and his colleagues are working with
government officials to mobilize an effort to
help smallholder farmers across the nation
transition to more evidence-based, sustaina-
ble cultivation. This includes selecting seed
varieties that are suited to a given plot, using
modelling techniques to guide planting based
on levels of sunlight, water and nutrients, and
optimizing the timing and density of seed
planting. “We sent faculty members and groups
of students to live among the farmers in the vil-
lages, and work with them to try to change their
management,” says Zhengxia Dou, an agricul-
tural scientist at the University of Pennsylvania
in Philadelphia, who collaborated with Zhang’s
team. By 2015, the effort had grown to include
nearly 21 million farmers across China, who,
on average, achieved a more than 10% boost
in yield while using around 15% less fertilizer
and reducing their greenhouse-gas output5.
Many farmers in India are embracing a
national programme known as zero-budget
natural farming (ZBNF). This cultivation strat-
egy involves using soil microbes and mulch
rather than synthetic fertilizers to enrich lands.
Farmers in several Indian states are pursuing
the approach, including around half a million
farmers in Andhra Pradesh. But some scientists
Smith concurs. “It was a political move, not a
scientific move,” she says, adding that the nat-
ural farming approach has “not been properly
trialled”. To assess the technique, she and her
colleagues modelled the long-term impact
of ZBNF on soil health. They found that the
approach could meaningfully and sustaina-
bly improve nitrogen levels for low-yield lands,
but that it would offer little benefit to farms
already achieving high yields6. They concluded
that a more targeted implementation of ZBNF
is needed to protect overall national food secu-
rity. Smith remains largely positive about ZBNF,
which has been gaining momentum among
farmers. “There’s a lot of good things about it,
but it needs more science,” she says.
Outside national initiatives, smallholder sus-
tainable intensive farming requires targeted
investment and efforts to support social and
economic stability. Vanlauwe contends that, in
many parts of sub-Saharan Africa, environmen-
tal and political conditions mean that many
farmers will continue to struggle at the margins
for the foreseeable future. Still, he sees a path
towards economic mobility. “Give them access
to credit they pay back over time, and invest in
integration and value-chains so they can get rid
of or sell excess produce,” he says. “It’s about
creating incentives and access systems.”
But durable change also requires building
local expertise in crop and soil research, and in
ecosystems. Many specialists in these areas are
also involved with international education and
training. For example, as director of the Feed
the Future Innovation Lab for Collaborative
Research on Sustainable Intensification, Prasad
has helped to coordinate undergraduate- and
graduate-level agriculture programmes in
places such as Senegal, Cambodia and Bang-
ladesh. Normally, these programmes take on
a few dozen students at a time, but the shift to
online training as a result of the coronavirus
pandemic could prove to be a long-term gain
for capacity building. “We are now talking to
about 500 or even 1,000 students,” he says.
Michael Eisenstein is a science journalist in
Philadelphia, Pennsylvania.
1. Cassman, K. G. & Grassini, P. Nature Sustain. 3, 262–268
(2020).
2. Pretty, J. M. Natural Res. Forum 21, 247–256 (1997).
3. Ricciardi, V. Glob. Food Security 17, 64–72 (2018).
4. Pretty, J. et al. Glob. Sustain. 3, e23 (2020).
5. Cui, Z. et al. Nature 555, 363–366 (2018).
6. Smith, J., Yeluripati, J., Smith, P. & Nayak, D. R. Nature
Sustain. 3, 247–252 (2020).

Nature | Vol 588 | 10 December 2020 | S59

©
2
0
2
0
S
p
r
i
n
g
e
r
N
a
t
u
r
e
L
i
m
i
t
e
d
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
Sustainable nutrition
outlook

TOMMY TRENCHARD/PANOS PICTURES

Seaweed farmers in Tanzania tend to their crops. Not only is seaweed a nutritious food, but cultivating it can help to ease ocean acidification.

Cultivating a sea change

Can aquaculture overcome its sustainability challenges to
feed a growing global population? By Sarah DeWeerdt

O
n a summer morning in 2019, Andy
Suhrbier pilots a small aluminium
boat out to a mussel raft in a quiet
cove on the eastern shore of Puget
Sound in Washington State. As the
boat approaches, a mother seal and her pup
resting on the raft slip into the water. Suhrbier
climbs from his boat onto the raft; the only
sign of life is a vague smell.
Suhrbier tugs on a couple of ropes attached
to one of the raft’s beams. Soon, a mesh-
lined plastic cage emerges with water and
silt pouring out of it. He picks off several sea
stars and tosses them back into the water,
then flips open the lid like a pirate opening
a treasure chest.
Inside is more dark sediment — mostly
waste from the mussels, the source of the
smell. Suhrbier sifts through it. He is looking
for something.
“Look at this monster!” he says, holding up
a sea cucumber nearly a foot long. Its deep
red body covered in orange bumps stands out
from the muck like a gold doubloon. “That’s
definitely market size.”
Suhrbier is a biologist with the Pacific
Shellfish Institute in Olympia, Washington, a
non-profit research organization that works
to promote healthy wild shellfish populations
and sustainable shellfish aquaculture along
the US west coast. Two years earlier, he had put
sea cucumbers in cages and suspended them
beneath the mussel raft, as part of an effort to
develop aquaculture in Puget Sound. The hefty
size of the cucumbers is a promising sign.
Suhrbier and his colleagues think that
sea-cucumber farming could have two ben-
efits. First, the animals could help to prevent
excess waste from building up underneath
aquaculture installations, such as mussel
rafts or net pens used to hold bony fish such
as salmon. (Sea cucumbers, soft-bodied ani-
mals related to sea urchins, move slowly over
the sea floor eating detritus — the vacuum
cleaners of the ocean.) Second, a ready source
of farmed sea cucumbers could reduce the
poaching of wild stocks to feed the growing
market in east and southeast Asia.

S60 | Nature | Vol 588 | 10 December 2020

©
2
0
2
0
S
p
r
i
n
g
e
r
N
a
t
u
r
e
L
i
m
i
t
e
d
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
 Globally, aquaculture produced 82.1 million
tonnes of aquatic animals in 2018, and wild
fisheries produced 97 million tonnes, accord-
ing to the United Nations’ Food and Agricul-
ture Organization (FAO). But the value of
farmed fish was higher, around US$250 billion
compared with $151 billion for wild-caught
fish. Aquaculture production of animals is
projected to increase by one-third by 2030,
reaching 109 million tonnes, and will supply
the majority of aquatic protein in people’s
diets by 2050.
“We need to grow the amount of seafood
available, as world populations grow, to pro-
vide enough protein for everybody,” says
Monica Jain, founder of Fish2.0 in Carmel,
California, an organization that promotes
investment in sustainable seafood businesses.
With the catches from wild fisheries remain-
ing largely flat and some stocks already over-
exploited, “aquaculture is really the only way
to do that”. But as the industry grows, Jain and
other aquaculture advocates want to make
sure that it does so sustainably.
Double alchemy
Aquaculture is a relatively small proportion
of the global food system — terrestrial meat
production (both livestock and wild game)
totalled around 342 million tonnes in 2018, and
production of grains and cereals was 2.7 billion
tonnes. However, aquaculture is more diverse,
particularly in terms of the animals farmed.
These range widely across taxonomic groups,
including bony fish (carp, tilapia and salmon,
for example), crustaceans (shrimp, prawns and
crayfish), molluscs (clams, oysters and mus-
sels) and echinoderms (sea cucumbers). Vari-
ous species of seaweed are also gathered. There
are freshwater, saltwater, brackish water and
self-contained terrestrial aquaculture systems.
And each has its own sustainability benefits
and challenges.
One subsector that offers huge environmen-
tal advantages and has no equivalent in terres-
trial agriculture is non-fed aquaculture. Marine
bivalves, such as clams, mussels and oysters,
get their nourishment by filtering microscopic
plants, detritus and nutrients from the water
that surrounds them. They require minimal
inputs and can even improve the water qual-
ity. In this sense, Suhrbier’s sea cucumbers
represent a kind of double alchemy: non-fed
aquaculture species grown on the wastes of
other non-fed aquaculture species.
Similarly, cultivated seaweeds can remove
excess nutrients, such as nitrogen, that
contribute to the formation of areas of oxy-
gen-poor water where marine life has dif-
ficulty surviving, known as dead zones. By
taking up carbon they can also help to alleviate
ocean acidification at a local scale. More-
over, “seaweeds are incredibly nutritious,”
says Alecia Bellgrove, head of the Deakin-
Seaweed Research Group at Deakin Univer-
sity in Melbourne, Australia. “They are, for
example, fantastic sources of trace minerals,
which are often lacking in our diets based on
terrestrial foods.”
Aquatic animals that require feed — mainly
prawns and bony fish — also have an environ-
mental advantage over animals raised in
terrestrial agriculture. Because most are cold-
blooded, they convert food into body mass
more efficiently than birds and mammals,
which need energy to help regulate their body
temperature. So it takes less feed to produce a
kilogram of salmon, for example, than it does
to produce a kilogram of, say, beef or pork.
However, some of the most lucrative aqua-
culture species are carnivorous, and therefore
sit higher in the food chain than any terres-
trial species raised in agriculture. Take the
Atlantic salmon (Salmo salar), for example.
In the mid-2000s, salmon aquaculture, now
a $15.4-billion industry, was growing rapidly.
Feeding the salmon demanded an increasing
share of the world’s fish meal and fish oil, which
“Our farming practices
resulted in the net
production of fish on
the planet.”
was sourced from small forage fish, such as
anchovies, sardines and capelin. But while
demand from the salmon farms grew, fishing
yield for the forage species remained rela-
tively flat. It took at least 4 kilograms of wild-
caught forage fish to produce just 1 kilogram
of salmon.
From an environmental point of view, “It
made no sense,” says Scott Nichols, founder
of Food’s Future, a consultancy in West
Chester, Pennsylvannia, which promotes
the development of sustainable aquaculture
businesses. As a biochemist working at US
chemical company DuPont in the mid 2000s,
Nichols helped to develop a way to produce
omega-3 fatty acids from yeast. The fatty acids
were then incorporated into salmon feed to
replace some of the wild-fish component. The
new feed was tested through a partnership
between DuPont and the aquaculture firm
AquaChile based in Puerto Montt, Chile, in
the form of the salmon producer Verlasso in
Miami, Florida.
“We were able, after a couple of years of pro-
duction, to get to the point where for every kilo
of salmon that was produced, we were using
less than a kilo of wild-caught fish,” Nichols
says. “So our farming practices resulted in the
net production of fish on the planet.”
Other companies soon joined in, producing
omega-3s in genetically engineered canola oil
or single-celled algae. Meanwhile, fish-oil and
fish-meal producers are increasingly making
use of fish trimmings and other by-products
that previously went to waste. Fish meal and
fish oil, which are still used in a variety of aqua-
culture feeds, as well as in products such as
food supplements, accounted for around 10%
of the world’s total fish production in 2018,
according to the FAO. But nonetheless, Nich-
ols takes heart from developments. “What
looked on the face of it to be dismal in 2006
now looks to be very promising,” he says.
Disease detectives
An increasingly important threat to aquacul-
ture sustainability is disease, which affects all
subsectors of aquaculture and causes an esti-
mated $6 billion worth of aquatic animal losses
every year. Diseases include parasites called
sea lice in salmon; white spot syndrome virus
in prawns, which emerged in the early 1990s
and devastated prawn farming throughout
Asia before spreading to the Americas; and
tilapia lake virus, which threatens the eco-
nomic and nutritional gains that freshwater
aquaculture has made possible in many low-
and middle-income countries.
As aquaculture is scaled up, the problem
of disease will also become greater. “As you
expand the volume of production, you are
going to get significant losses,” says Grant
Stentiford, a pathologist and head of aquatic
animal health at the Centre for Environment
Fisheries and Aquaculture Science, Weymouth,
UK. “You’ve used up potentially large amounts
of resource to get absolutely nowhere.”
To deal with such threats, some large pro-
ducers who supply the export market are
moving to self-contained, land-based systems.
Others are moving away from the coast into
deeper waters that might dilute the threat of
disease. Vaccines have also made a difference,
reducing not only the threat of many fish dis-
eases, but also antibiotic use — another major
environmental concern about the industry.
And high-throughput sequencing of the
microbial DNA in aquaculture systems could
provide early warning of disease outbreaks.
But many of these solutions are expensive
and, therefore, out of reach for the small and
medium-sized producers who make up the
majority of the global aquaculture industry,
producing food for subsistence or local mar-
kets in low- and middle-income countries.
Moreover, diseases that threaten aquacul-
ture are emerging every three to five years

Nature | Vol 588 | 10 December 2020 | S61

©
2
0
2
0
S
p
r
i
n
g
e
r
N
a
t
u
r
e
L
i
m
i
t
e
d
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
Sustainable nutrition
outlook
This type of rice–fish system has been
practised for hundreds of years in China and
has been designated a Globally Important
Agricultural Heritage System by the FAO — a
designation that aims to preserve agricul-
tural knowledge that can contribute to a
more sustainable and resilient food system.
Large-scale aquaculture operations, such as
Cooke Aquaculture based in Blacks Harbour,
Canada, have also been experimenting with
multi-species systems. The company keeps
salmon in net pens near both mussel and kelp
rafts in the Bay of Fundy, Canada.
In theory, integrated aquaculture can help
to increase yields, decrease risk by diver-
sifying operations, and is generally a more
environmentally sound form of aquaculture.

Sea cucumbers are retrieved from the mesh-lined cages at Puget Sound in Washington state.
on average. The dearth of knowledge about
aquatic pathogens makes diseases hard to
predict and spot.
It can also be a challenge to deduce their
cause. For example, ice-ice disease results in
bleaching of Kappaphycus seaweed, which is
grown in large amounts in southeast Asia and
Tanzania for the production of food additives,
such as the thickening agent carrageenan. The
disease has caused yields to plummet over
the past decade, but “the causative agent is
still not known”, says Valéria Montalescot,
senior project manager for GlobalSeaweed-
STAR, a four-year research project based at
the Scottish Association for Marine Science
in Oban, UK, which aims to boost knowledge
about seaweed cultivation in low- and mid-
dle-income countries. Kappaphycus is usually
grown from cuttings, so the whole crop across
multiple countries might be the result of just a
few clones, possibly making it more vulnerable
to disease, Montalescot adds.
Diverse yields
Climate change is complicating efforts to fight
disease. Higher water temperatures can alter
the microbial community of a body of water,
encouraging the growth of pathogens, as well
as stressing organisms and making them more
vulnerable to disease. One suggested cause
of ice-ice disease is that temperature-stressed
seaweeds release compounds that attract
bacteria, for example.
And temperature is not the only issue.
Both increased rainfall and salinity intrusion
from sea-level rise can alter water chemistry
in ways that are detrimental to aquaculture
organisms. Storms can destroy aquaculture
crops or infrastructure in the water and on
land. “From an environmental point of view I
think climate change is the greatest challenge”
for the sustainability of aquaculture systems,
says Nesar Ahmed, who studies global seafood
sustainability at Deakin University.
Climate change also intersects with aquacul-
ture’s pressure on water and land resources.
Inland aquaculture demands 429 cubic kilo-
metres of fresh water each year — much less
than the demand from terrestrial agriculture,
but still enough to pose a strain on increasingly
drought-prone areas.
In south and southeast Asia, prawn cultiva-
tion has contributed to the destruction of 38%
of the world’s mangrove habitats, which have
a variety of important ecological functions,
including sequestering carbon and buffering
coastlines from storms and sea-level rise. The
loss of mangroves has also resulted in saltwater
intrusion rendering inland areas unsuitable for
terrestrial agriculture.
Some farmers are now producing prawns
among intact mangrove stands. Although there
are concerns that this practice might also dam-
age the health of the mangroves, it is part of a
larger trend to create aquaculture systems that
include multiple species and involve interrela-
tionships more like the ones that keep natural
ecosystems in balance.
Some examples of this integrated aquacul-
ture are long-established, such as stocking rice
paddy fields with fish or prawns. The animals
eat pests and fertilize the rice crop, increasing
rice yields and providing an extra source of pro-
tein or income for small-scale farmers, Ahmed
says. Growing two species in a single body of
water also reduces overall water use.
SARAH DEWEERDT
But in practice, it can be difficult to quantify
these benefits. For example, because nitro-
gen moves freely through water, it is difficult
to track uptake of excess nitrogen produced
by bony fish by seaweed growing nearby. And
then there are the complexities of managing
an operation with multiple species — not just
producing them but also harvesting, process-
ing and marketing them.
Suhrbier knows such difficulties well. The
sea cucumbers he and his team harvested from
under the mussel raft were the right size, weight
and colour for the export market, but the mus-
sel producer he was working with was unable to
renew its permit at that location. The raft was
lost, and with it Suhrbier’s chance of follow-up
experiments to develop sea-cucumber aqua-
culture techniques. “I was really shocked and
saddened to see that go because it was one of
those places where it just makes a lot of sense
for sea cucumbers to be,” Suhrbier says. The
new location of the producer’s rafts isn’t a good
habitat for sea cucumbers.
Suhrbier is still experimenting growing sea
cucumbers alongside other types of aqua-
culture operation around the Puget Sound
area. But, like an increasing number of aqua-
culture researchers, he is beginning to think
that producing the animals needs to move in
a simpler and more radical direction. Grow-
ing sea cucumbers in cages is labour intensive.
What if the animals are placed in the vicinity of
aquaculture operations and left to roam freely
— like a marine equivalent of a ranch or even a
permaculture system?
“If we could mainly enhance the wild popu-
lation around these areas, I think that would be
a great benefit for everybody,” Suhrbier says.
“I’m trying to have something that fits in: easy,
cost effective and as passive as it can be.”
Sarah DeWeerdt is a freelance writer in
Seattle covering biology, medicine and the
environment.

S62 | Nature | Vol 588 | 10 December 2020

©
2
0
2
0
S
p
r
i
n
g
e
r
N
a
t
u
r
e
L
i
m
i
t
e
d
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
 Sustainable nutrition
outlook

Eating for better mental health

Food, mood and brain function are thought to be connected through the
community of microorganisms that live in the gut. By Clare Watson

I
f you want to do right by the planet, the
general advice is to eat an abundance of
fruit and vegetables, as well as whole grains
and nuts, and to consume less meat, dairy
and processed foods. Add some high-fibre
fare, fermented food and fish a few times a
week, and you could be eating your way to
better mental health, too.
Those are the recommendations from nutri-
tional psychiatry research. The field is built
around growing evidence from population
studies and clinical trials in the past decade
suggesting that dietary improvements might
not only improve mood but also treat com-
mon and severe mental illnesses. Traditional
diets consumed by people in places such as the
Mediterranean, Norway and Japan are associ-
ated with a lower risk of depression — one of
the most common mood disorders — and, to
a lesser extent, anxiety. A change in diet can
alleviate symptoms of depression, even among
people with severe forms of the condition.
Researchers have also been investigating
how the trillions of microorganisms in the
human gut communicate with the brain to
influence the processes that take place there.
Imbalances in the gut microbiome have been
linked to a range of neurological disorders,
including Alzheimer’s disease, autism spec-
trum disorder, multiple sclerosis, Parkinson’s
SCIEPRO/SPL
adequate quantities, confer health benefits on
the host — have had compelling results, such as
reduced symptoms of depression in women
who have recently given birth. But others have
shown no more effect than a placebo. Compar-
isons are difficult owing to the varying doses
and strains of bacteria used in trials. Similarly,
although promising, the evidence from human
trials testing specific prebiotic foods — those
that are rich in high-starch dietary fibres that
stimulate beneficial bacterial colonies in the
gut — is insufficient to draw clear conclusions.
At the Food and Mood Centre, researchers
focus on the entire diet, rather than individ-
ual ingredients or certain strains of bacteria
delivered in probiotic supplements. “There
them all, people need to eat a well-rounded diet.
On the question of what mix of microbes
makes for a healthy microbiome, diversity
is thought to be important. People with a
greater variety of bacteria in their gut seem
to be healthier. A Western diet of processed
foods that is high in fat and sugar and low in
dietary fibre and micronutrients seems to
be detrimental to both the gut and the mind,
reducing the diversity of the gut microbiome,
increasing inflammation and elevating the risk
of depression.
The long-running SUN (Seguimiento Univer-
sity of Navarra) cohort study has been recruit-
ing university graduates in Spain since 2000
to analyse associations between dietary pat-
terns and health, including depression. One
of its findings is that the more ultra-processed
foods (usually, energy-dense foods that are
significantly altered from their original state)
you eat, the greater your risk of depression2.
Over the past decade, observational studies
of this kind have consistently shown that
well-rounded diets that are low in ultra-pro-
cessed foods confer some protection against
depression. But public-health researcher
Almudena Sánchez Villegas at the University
of Las Palmas of Gran Canaria, Spain, says more
randomized controlled trials testing specific
dietary interventions are needed. This will

disease and Huntington’s disease. allow researchers to refine the International

Brain function, mood and mental health
seem to be intricately linked to what peo-
ple eat through the gut–brain axis, which is
modulated by the gut microbiome.
That there is a connection between the
food people eat and the brain is “something
that ancient wisdom has been telling us for a
long time”, says Amy Loughman, a gut-micro-
biome researcher at Deakin University’s Food
and Mood Centre in Geelong, Australia. “But
it’s really exciting that science is beginning to
prove how and why.”
The gut microbiome is proving to be
extremely complex and sensitive. Changes
in diet can alter its composition in days. But
the data from human trials on dietary inter-
ventions designed to shift gut microbiota to
improve mental health are mixed, according
to one meta-analysis1. Some studies of probi-
otics — live microbes that, when introduced in
Illustration of bacteria in the intestine.
are no superfoods that are going to guarantee
positive mental health,” says Loughman. “It’s
both simpler and more complex than that.”
To support mental and brain health,
high-fibre foods such as fruit, vegetables and
whole grains are recommended, Loughman
says. In these foods, the fibre and starches that
are resistant to digestion in the small intestine
promote the growth of beneficial bacteria,
which guard against inflammation by protect-
ing the gut wall. The bacteria also produce the
short-chain fatty acids that are thought to have
a key role in crosstalk between the gut and the
brain. Chemicals such as polyphenols in plant-
based foods and omega-3 fatty acids in fish have
reported mental-health benefits, too. To get
Society for Nutritional Psychiatry’s existing
dietary guidelines3 for preventing depres-
sion, of which Sánchez Villegas is a co-author.
Because individual responses to food seem to
be mediated by a person’s gut microbiome,
researchers will need larger trials to work out
who is likely to benefit from dietary interven-
tions before these can be prescribed in the
clinic. Until then, the public-health message
is simple: everyone stands to benefit from a
balanced diet with plenty of fruit, vegetables
and fibre, and few processed foods. “That’s
universally good for people,” says Loughman.
Clare Watson is a freelance science journalist
in Wollongong, Australia.
1. Vaghef-Mehrabanya, E. et al. Clin. Nutr. 39, 1395–1410
(2020).
2. Gómez-Donoso, C. et al. Eur. J. Nutr. 59, 1093–1103 (2020).
3. Opie, R. S. et al. Nutr. Neurosci. 20, 161–171 (2017).

Nature | Vol 588 | 10 December 2020 | S63

©
2
0
2
0
S
p
r
i
n
g
e
r
N
a
t
u
r
e
L
i
m
i
t
e
d
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
Sustainable nutrition
outlook

SOPHIE CASSON FOR NATURE

Cell-based meat with a side of science
Growing a burger in a laboratory is one challenge, growing an
industry to do it at scale is quite another. By Elie Dolgin

W
hen Laura Domigan started her
research group at the University
of Auckland, New Zealand, in 2015,
she hoped to continue her work
developing protocols for growing
cell-based meat in the laboratory. But with
funding for cultivated-meat research prac-
tically non-existent in academia at the time,
Domigan pivoted to working on biomedical
materials for use in tissue engineering. A pro-
tein biochemist by training, she focused her
efforts on creating artificial corneas for eye
surgery — a far cry from anything resembling
a lab-grown steak.
Still, she never gave up on her dream of stud-
ying in vitro meat. “I had to be super patient
and keep trying,” Domigan says. And although
it took several years, Domigan’s strategy even-
tually paid off.
Initially, she secured funding for a PhD
student to begin developing formulations of
nutrient media to grow cell-based meat. Then,
in October 2020, a team led by Domigan won
a multi-million dollar grant from the New
Zealand and Singaporean governments to
explore questions such as which cells are the
best starting material for cultured meat, and
is the nutritional profile of meat grown in a lab
equivalent to the real thing. “There is so much
research that needs to be done,” Domigan says.
And much of it is only beginning to happen, at
least in any sort of transparent way.
Investors have poured hundreds of millions

S64 | Nature | Vol 588 | 10 December 2020

©
2
0
2
0
S
p
r
i
n
g
e
r
N
a
t
u
r
e
L
i
m
i
t
e
d
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
of dollars into cultured-meat research in the
past few years, bringing hype and breathless
news coverage about an agricultural revolution
that could bypass the environmental and ani-
mal-welfare issues of conventional meat pro-
duction. One estimate by the consulting firm
Kearney in Chicago, Illinois, suggests that 35%
of all meat consumed globally by 2040 will be
cultured — a change that is projected to reduce
greenhouse-gas emissions and antibiotic use.
And thanks to the COVID-19 pandemic, which
revealed crucial weaknesses in global food-sup-
ply chains, some people now expect the transi-
tion to cell-based meat to happen even faster.
Earlier this month, a US start-up called Eat
Just announced that its chicken bites — which
are 70% cultured chicken cells, with plant pro-
tein added for structure and flavour — had won
regulatory approval for sale to consumers in
Singapore, a global first for the cultivated-meat
industry.
Scientists worry that the commercial push
to bring palatable products to market means
that fundamental studies are either not hap-
pening or remain cloaked in trade secrecy.
Start-ups have made splashy demonstrations
of their lab-grown chicken nuggets, pork sau-
sages, steak strips and seafood dumplings. But
these show only that companies “can do this
on a small level”, says Abhi Kumar, a venture
partner at Lever VC, a New York-based venture
capital fund that focuses on alternative-pro-
tein start-ups. The challenge now, he says, is
making it work at scale.
Improvements in cell source material and
the nutrient media required to fuel cell growth
are needed, as are scaffolds to support 3D tis-
sue structures. Next-generation bioreactor
platforms that can grow huge numbers of
cells at high densities are also a must. These
are costly undertakings — so much so that
many in the field are dubious that private
financing can support them, and still yield an
affordable product.
That’s why leading thinkers in cellular agri-
culture, such as Erin Rees Clayton at the Good
Food Institute (GFI), argue in favour of more
open science and public investment. “There’s a
need for public-sector research in these areas,”
says Rees Clayton, who is associate director of
science and technology at the GFI, a non-profit
think tank in Washington DC. “There’s a lot of
room in the pre-competitive space for more
work to be done in an open way, so we can all
benefit from it and move forward more quickly.”
To fill the funding gap, the GFI created a
research-grant programme that has given out
close to US$3 million over the past 2 years to
16 research teams working on cultivated-meat
projects. Academic institutions are beginning
to hire with the nascent discipline in mind — the
Technical University of Munich in Germany, for
example, is accepting applications for a profes-
sorship in cellular agriculture. And, as evidenced
by Domigan’s funding success, governments too
are heeding the call for financial support.
The field of cellular agriculture is beginning
to take on some of its biggest scientific and
engineering challenges, and scientists from
a range of backgrounds are entering the fray.
“This cellular-agriculture research is
the stuff that gets me up in the morning,”
says Glenn Gaudette, a biomedical engi-
neer at Worcester Polytechnic Institute in
Massachusetts. For almost 20 years he has
studied scaffold technologies for heart-re-
generation therapies; now he is applying his
expertise to the problem of how to grow meat.
“Does it pay the bills? No, not yet — hopefully,
one day — but it’s really exciting.”
The fight for funding
In the early 2000s, the US space agency NASA
briefly supported efforts to grow goldfish
muscle in the lab as a potential source of pro-
tein for astronauts on long missions. A few
years later, the Dutch government sponsored
a €2-million ($2.3-million) project to cultivate
“This cellular-agriculture
research is the stuff that gets
me up in the morning.”
pork meat from stem cells — research that,
with an extra €250,000 infusion from Google
co-founder Sergey Brin, eventually led to the
field’s highest-profile moment so far, when
vascular biologist Mark Post at Maastricht
University in the Netherlands, unveiled the
world’s first cultured burger, in 2013.
But aside from intermittent funding oppor-
tunities to explore the social ramifications
of producing meat from cell cultures, there
have been few other public grants for culti-
vated-meat research. Government bodies
stayed away from the field in large part, says
Kate Krueger, former research director at the
non-profit organization New Harvest in Cam-
bridge, Massachusetts, because the science
was unproven and crossed disciplines in ways
that defied the conventional dividing lines of
funding-agency bureaucracy.
“Cellular agriculture falls into this funding
no-man’s land between biomedical research
and agricultural research,” says Krueger, who
now runs a consulting firm, also in Cambridge,
centred around cellular agriculture.
Money from the GFI and New Harvest has
plugged the funding gap to some extent. But
the situation is changing. As scientific interest
in the subject grows and grant applications
increase, governments have begun to inject
more cash into the field. Several large grants
have been issued in the past few months alone.
In November, for example, the government
agency Flanders Innovation and Entrepre-
neurship began funding a €2.1-million, 4-year
project called CUSTOMEAT, run by scientists
at Ghent University and KU Leuven, both in
Belgium. In the United States, the National Sci-
ence Foundation (NSF) awarded around $3.5
million in September to back a cultivated-meat
consortium at the University of California,
Davis, for the next five years.
“Our hope is that we can provide basic
knowledge and build a trained workforce,”
says chemical engineer David Block, who is
leading the Davis effort. “Those are the kinds
of things that you need to grow an industry.”
Experts say that many cultivated-meat
companies will probably over-promise and
under-deliver. But academic science can help
“keep credibility alive,” says Johannes le Cou-
tre, who led a research group at the Swiss food
giant Nestlé before joining the University of
New South Wales in Sydney, Australia, in 2019
to run a lab dedicated to cellular agriculture.
Amy Rowat, a biophysicist at the Univer-
sity of California, Los Angeles, notes that aca-
demia also offers the intellectual freedom for
researchers to work on exploratory projects,
using expertise in basic science to come up with
innovative ways of thinking, or to tackle ques-
tions not directly related to product develop-
ment but still significant to the overall domain.
And according to David Kaplan, a bioengineer at
Tufts University in Medford, Massachusetts, the
next generation of scientists entering the field
are “totally motivated to make a difference”.
“I have never seen such driven, passionate
students in all my decades of doing this,” he
says. Andrew Stout is a case in point. A PhD
student in Kaplan’s lab, Stout is rethinking the
entire process of cellular agriculture, start-
ing with the most basic ingredient: the muscle
cells themselves.
Most companies growing lab meat either
use cells taken directly from animal biopsies
or cell lines that have spontaneously become
immortal through natural mutations that
allow indefinite proliferation in the lab. Few
firms will consider genetically manipulating
the cells for optimal performance because of
a fear of consumer backlash. But Stout real-
ized that genetic engineering offered a path
to achieving the nutritional promise of cul-
tivated meat.
Starting material
He inserted three genes into cow muscle
cells1. Each encoded an enzyme involved in

Nature | Vol 588 | 10 December 2020 | S65

©
2
0
2
0
S
p
r
i
n
g
e
r
N
a
t
u
r
e
L
i
m
i
t
e
d
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
Sustainable nutrition
outlook
in vitro culture media, creating the slaugh-
ter-free products that the industry demands.
But this serum-free media is too expensive
for cultivated meat to be affordable on the
supermarket shelves. “It’s difficult to find a
cost-efficient option,” says Ka Yi Ling, chief
scientific officer at Shiok Meats, a cell-based
seafood company in Singapore.

Media matters
According to an analysis by the GFI4, growth
media currently make up the bulk of total
production costs for cultivated meat, and
proteins known as growth factors are the
most expensive ingredient. Costs are com-
ing down, as start-ups dedicated to serving
the cellular-agriculture industry devise ways

DAVID PARRY/PA IMAGES/ALAMY

Producing products such as the lab-grown meat burger at scale is still a long way from reality.
the synthesis of an antioxidant that mitigates
diseases associated with consuming red and
processed meats, such as colon cancer. The
enzymes might also help with the manufac-
ture of cultured meat, because the unstable
molecules that antioxidants attack reduce
the proliferation of some lab-grown cells. If
consumers are willing to accept these types
of DNA enhancement, “genetic and metabolic
engineering can offer a lot of impact and ben-
efit to cultured meat”, Stout says, “and could
even allow us to create novel foods that we
couldn’t get any other way”.
Other researchers are reconsidering
whether the cells that go into cultivated-meat
products need to come from species that are
already commonly consumed in Western cul-
tures. For example, Natalie Rubio, another
Tufts graduate student, has explored grow-
ing meat from insect cells to create products
that can be designed to taste like crab, prawns
and other seafood. Using muscle cells from
fruit flies (Drosophila melanogaster)2 and
the caterpillar of the moth Manduca sexta,
Rubio showed that insect cells are easier and
cheaper to grow than cells from conventional
livestock species, and might also have nutri-
tional advantages.
Meanwhile, other scientists hope to rally
the research community around the idea of
cell-based zebrafish fillets, at least as a vehi-
cle for accelerating advances in the field. The
zebrafish (Danio rerio) is an established model
organism for studying the genetic, neuronal
and behavioural basis of disease. Alain Rostain,
executive director of Clean Research, a non-
profit organization in New York, now wants
to make it the go-to species for basic-research
projects in cellular agriculture as well. “There’s
a lot of fundamental understanding that’s not
there yet,” he says. “We need the participation
of a lot of people to just think freely through
the science together.”
And, as Rostain and his colleagues have
described3, researchers can benefit from the
extensive molecular toolkit already estab-
lished for zebrafish. Plus, as a lean fish with
little fat content in the muscle, zebrafish fillets
should be easier to produce than compara-
ble lab-grown cuts of fat-laced salmon, tuna,
“I have never seen such
driven, passionate
students in all my
decades of doing this.”
beef or pork. Cultivated zebrafish will proba-
bly taste similar to white fish, such as cod or
haddock, Rostain says, and discoveries made
with zebrafish should translate to any other
edible species.
Regardless of the starting material, all cells
will require an optimized growth medium
— the rich broth of chemicals and proteins
needed to support proliferation and differ-
entiation. Companies have already devised
ways to eliminate the nutrient-rich fetal
bovine blood that is the cornerstone of most
to manufacture these products. But as Matt
Anderson-Baron, co-founder and chief sci-
entific officer at one such company, Future
Fields in Edmonton, Canada, concedes,
“There’s still so much to be done on the opti-
mization and discovery front.”
Assuming researchers find the right cell line
and growth medium combination, they then
have to grow those cells on a scaffold — ideally
one that is edible so that it doesn’t have to be
removed from the final product. For ground-
meat products, such as burgers and sausages,
small beads known as microcarriers can pro-
vide the surface properties needed by most
muscle and fat cells for growth. But for any-
thing with a more complex meat structure — a
steak or an Iberian ham, for example — a more
sophisticated tissue-engineering approach
is required.
One option comes from a team at Harvard
University in Cambridge, Massachusetts. Bio-
engineer Kevin Kit Parker and his colleagues
have developed a spinning technique that
works like a candy-floss machine to extrude
long, thin fibres from gelatin5. The researchers
put the gelatin, a protein product derived from
collagen, into their machine and produced tiny
threads — narrower than the width of a hair —
that closely matched the architecture of fibres
found in muscle tissue.
Last year, Parker and his colleagues showed
that rabbit and cow muscle cells grown on the
fibrous gelatin line up with the proper ori-
entation6. The cells were still not as densely
packed as real muscle, but Parker, together
with three of his postdocs and students, has
since created a company called Boston Meats
to improve the technology further. “Now, with
our scaffolds,” he says, “you can move from
hamburger to fillet.”
Elsewhere, researchers have made scaffolds
out of foods such as textured soya protein and
various vegetables stripped of their cellular
material so that only supporting structural

S66 | Nature | Vol 588 | 10 December 2020

©
2
0
2
0
S
p
r
i
n
g
e
r
N
a
t
u
r
e
L
i
m
i
t
e
d
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
sugar molecules and proteins remain. At the
University of Ottawa in Canada, for example,
biophysicist Andrew Pelling and his students
have taken decellularized stalks of celery and
shown that the grooves created by its natural
structure help to promote the patterning and
alignment of muscle cells7. And at Worcester
Polytechnic, Gaudette’s team has grown fat
and muscle cells on cell-free spinach leaves
— the plant’s branching network of delicate
veins provide ideal conduits for the nutrient
medium to reach every cultivated meat cell.
Because muscle and fat cells require dif-
ferent growth media, however, researchers
typically culture the two cell types separately,
each on its own scaffold in a different nutrient
bath. Some researchers, including Rowat, have
devised strategies to then interlace the muscle
and fat to achieve the flavour of a well-marbled
steak. “The cells actually fuse together with
the other partnering scaffold type on the time
scale of hours to form these composite struc-
tures,” Rowat explains. In unpublished work,
she has created miniature marbled steaks out
of mouse and rabbit cells, and has begun to
work with cells from pigs and cows, as well.
But even with the latest scaffolding strate-
gies, some muscle biologists worry that key
aspects of tissue physiology are still being dis-
counted. “This incredible focus that we have as
an industry on cell division ignores fusion and
maturation,” says James Ryall, chief scientific
officer at Vow, a start-up in Sydney working
on cell-based meat from animals such as kan-
garoo and alpaca.
To form muscle tissue, thousands of individ-
ual precursor cells must first fuse together to
become long myotubes. These cells require
physical stimuli to mature into myofibres.
Only then will muscle grown in the laboratory
have the texture and nutritional properties of
real meat, says Lieven Thorrez, a muscle-tissue
engineer at the Kortrijk campus of KU Leuven.
“And that is a process that takes time. You can-
not just differentiate cells over a period of a few
days and say the myofibres will be the same as
those found in an adult animal,” he says. “That
is largely overlooked.”
Scaling up
With so many scientific issues to resolve, cell-
based meat research, whether in academic
or private labs, remains at the experimental
stage. For commercial viability, the industry
will need to find ways to produce tissue at a
massive, and unprecedented, scale.
Tissue engineer Che Connon at Newcas-
tle University, UK, estimates that feeding
the world’s population with lab-grown meat
would necessitate building systems for grow-
ing on the order of a septillion (1024) cells
Ka Yi Ling and Sandhya Sriram are co-founders of Shiok Meats in Singapore.
annually, something that is not possible with
the types of batch bioprocessing techniques
currently used in mammalian-cell-based
manufacturing. Global capacity could fulfil
about one-billionth of that requirement. “It’s
a massive limiting factor,” says Connon, who
has developed a type of continuous cell-bio-
reactor platform that he plans to commer-
cialize through a spin-off company called
“This incredible focus that
we have as an industry on
cell division ignores fusion
and maturation.”
CellulaREvolution.
Meanwhile, computer scientist Simon
Kahan, president of life-sciences software
company Biocellion SPC in Seattle, is lead-
ing a team called the Cultivated Meat Mod-
eling Consortium that formed in 2019 with
the aim of optimizing bioprocessing tech-
niques through modelling techniques. With
funding from the German technology com-
pany Merck, the consortium has developed
a proof-of-concept model of a stirred-tank
bioreactor, involving little more than a spin-
ning rotor and free-floating minute beads for
growing muscle cells.
The bioreactor might be fairly rudimentary,
but the consortium’s computer modelling is
anything but. The simulations of fluid dynam-
ics and cellular biomechanics have revealed
a central challenge to growing muscle or fat
cells at scale. “You have these two competing
interests,” says Kahan. To support nutrient
and gas exchange, “you’re trying to keep
things well mixed while at same time trying to
subject the cells to very little mechanical fluid
stress”. With input from industry partners, the
consortium plans to build more complexity
into its models to inform the design of biore-
actors in the real world.
Block’s group, with its large NSF grant,
isn’t even attempting to work on bioreactor
designs; there is plenty to keep his team busy
tackling the issues of cell lines, media and scaf-
folds, as well as conducting feasibility assess-
ments of the cell-based meat industry. “To me,”
says Block, “it’s not clear yet that this is going
to be a viable alternative” — whether from a
technical, economic or sustainable stand-
point. But with each new grant or research
team entering the field, the goal of a perfectly
grilled, medium-rare steak grown from cells
gets a little closer to becoming a reality.
Elie Dolgin is a science journalist in
Somerville, Massachusetts.
1. Stout, A. J., Mirliani, A. B., Soule-Albridge, E. L., Cohen, J.
M. & Kaplan, D. L. Metab. Eng. 62, 126–137 (2020).
2. Rubio, N. R., Fish, K. D., Trimmer, B. A. & Kaplan, D. L. ACS
Biomater. Sci. Eng. 5, 1071–1082 (2019).
3. Potter, G. et al. One Earth 3, 54–64 (2020).
4. Specht, L. An Analysis of Culture Medium Costs and
Production Volumes for Cultivated Meat (The Good Food
Institute, 2020).
5. Reza Badrossamay, M., McIlwee, H. A., Goss, J. A. &
Parker, K. K. Nano Lett. 10, 2257–2261 (2010).
6. MacQueen, L. A. et al. npj Sci. Food 3, 20 (2019).
7. Campuzano, S., Mogilever, N. B. & Pelling, A. E. Preprint at
bioRxiv https://doi.org/10.1101/2020.02.23.958686 (2020).

Nature | Vol 588 | 10 December 2020 | S67

©
2
0
2
0
S
p
r
i
n
g
e
r
N
a
t
u
r
e
L
i
m
i
t
e
d
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
Sustainable nutrition
outlook

Research round-up
Highlights from
sustainable-
nutrition research.
By Dyani Lewis

Farming trends
deplete pollinators
Most cultivated crops depend
on insect pollinators, such as
bees, but global crop trends are
leaving pollinators worse off.
Using data from the United
Nations’ Food and Agriculture
Organization, an international
team, led by Marcelo Aizen
at the National University
of Comahue in Rio Negro,

Argentina, assessed changes
in the amount of land used for
agriculture and the types of
crops cultivated between 1961
and 2016. During that time, the
area of land used to grow crops
increased by around 40%, and
pollinator-dependent cropland
more than doubled. Soya bean,
rapeseed and oil palm — crops
associated with deforestation
and diversity loss — account
for much of the expansion and
for the increase in pollinator
dependence.
But although the land used
has increased, crop diversity
has remain largely the same
since 2000. Producers have
opted for large-scale cultivation
of one crop. That’s a problem
because monocultures don’t
provide pollinators with a
stable, year-round supply of
food. This ultimately leads
to a fall in insect numbers,
lower yields and increased
deforestation as demand for
land surges.
Greater reliance on crops
that are dependent on single-
species pollinators, coupled
with declining pollinator
populations, could cause
Rapeseed crops depend on pollinators such as bees.
problems for food security.
Poorer regions will be the
hardest hit by crop failures, but
higher-income countries that
rely on imported food will also
be affected.
Rotating a diverse range of
crops on a single piece of land
could help to stem the decline in
pollinator populations. Planting
native flowers and hedgerows
on agricultural land and
restoring neighbouring natural
environments could also
preserve pollinator habitats.
Glob. Change Biol. 25, 3516–3527
(2019)
US household food
waste calculated
Working out how much food
goes uneaten in an individual
household is notoriously
difficult. Comprehensive data
on how much food ends up
in the bin does not exist. But
Yang Yu and Edward Jaenicke at
Pennsylvania State University
in University Park used a new
method to overcome the lack
of data.
Instead of trying to
measure food waste directly,
Yu and Jaenicke calculated
a household’s ability to
efficiently convert food
brought into the household
into the energy required to
maintain the body weight of its
residents. First, they obtained
data on food purchases from
around 4,000 households
that took part in the 2012 US
Department of Agriculture’s
National Household Food
Acquisition and Purchase
Survey. The authors then
calculated the metabolic
energy requirements of
FOTOKOSTIC/ISTOCK/GETTY
the people living in each
household from attributes
such as height, weight, age
and gender. The amount of
food waste was estimated
according to the difference
between the household’s food
inputs and its members’ energy
requirements, not accounting
for overeating.
The study showed that the
average household wasted
close to one-third of the food
that it bought, which means
that the United States wastes an
estimated US$240 billion worth
of food per year. The most
efficient household in the study
wasted about 9% of its food.
Healthier diets created more
waste than unhealthier diets,
owing to the greater proportion
of fruit and vegetables. Higher-
income households wasted
about 50% more food than
lower-income households,

S68 | Nature | Vol 588 | 10 December 2020

©
2
0
2
0
S
p
r
i
n
g
e
r
N
a
t
u
r
e
L
i
m
i
t
e
d
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
and small households wasted
more per person than large
households.
Am. J. Agric. Econ. 102, 525–547
(2020)
Hidden hunger a
global problem
There is more than enough food
to feed the global population.
But local patterns of
production still leave 10% of the
world’s people with insufficient
calories, and more than half
with inadequate quantities and
variety of micronutrients —
known as hidden hunger.
These are findings of a
detailed analysis of food
production by Ozge Geyik and
colleagues at Deakin University
in Burwood, Australia. The team
gathered data on the nutrient
content of 174 individual foods
produced across 177 countries
between 1995 and 2015. The
researchers analysed whether
individual countries and
regions could meet the energy
needs of their populations,
as well as supply them with
protein, iron, zinc, vitamin A,
vitamin B12 and folate.
The study is one of the
first to take such a detailed
look at global patterns of
nutrient production using
disaggregated food data
over time. Previous work has
typically grouped foods into
broad categories, such as
cereals, dairy and vegetable
oils, which can lead to under-
or overestimates of specific
nutrients.
Global food production
increased steadily over the
two decades, and outpaced
increases in food requirements.
However, on a regional level, the
analysis found that more than
half of the countries in Africa
and Asia were not producing
enough calories for their
populations.
In 2015, more than 20% of
the global population lived
in countries with inadequate
iron, vitamin A, vitamin B12
and folate production. Food
production often fell short in
multiple nutrients. More than
70% of countries with nutrient
shortfalls produced inadequate
amounts of iron, vitamin A and
folate. And more than one-fifth
of those not producing enough
nutrients, fell short by more
than half of what was necessary
for their population.
The authors suggest that
countries with nutrient
deficiencies could prioritize
the production of foods
that contain the nutrients
that their population needs.
For example, in places
where protein production is
adequate, shifting production
to protein sources that are
higher in vitamin A and
iron could alleviate these
nutrient shortfalls. Adding
micronutrients directly to soils
and the leaves of crop plants is
another possible solution.
Glob. Food Sec. 24, 100355
(2020)
Nutrient recycling
possibilities mapped
The age-old practice of
fertilizing crops with livestock
manure has been reimagined
in a study led by Sheri Spiegal
from the US Department of
Agriculture in Las Cruces, New
Mexico. In the study, the team
introduces the concept of a
manureshed — land around
livestock farms that could
benefit from the nutrient-
rich manure that those farms
produce.
Spiegal and her colleagues
mapped a patchwork of more
than 3,000 counties across the
United States. They classified
counties as manure sources if
they could supply nutrients in
manure from livestock, or sinks
if the crops grown could use the
nutrients from manure.
The work reveals a surfeit
of opportunity to recycle
nutrients. The researchers
identified counties that
could recycle nitrogen and
phosphorous nutrients at the
local county level, as well as
four regional manuresheds —
in the northwest, southwest,
central and southeast United
States — where clusters of
source counties could join
together to develop sustainable
redistribution programmes
over longer distances. The work
suggests a pathway towards
removing manure from areas
where it can pollute the local
environment and delivering it
to nutrient-poor agricultural
lands, easing the reliance on
commercial fertilizers that
pollute the environment and
deplete finite natural resources.
But the authors note that
further research — on how
best to recover and transport
manure, for instance — will be
needed to turn the vision into a
reality.
Agric. Syst. 182, 102813 (2020)
Intervention trade-
offs assessed
Transforming the way land
is managed and food is
produced could shore up
food supplies and address the
challenges of climate change
and biodiversity loss. But
an assessment of proposed
interventions reveals that few
are up to the task of protecting
both livelihoods and the
environment.
Pamela McElwee from
Rutgers University in New
Brunswick, New Jersey, and
her colleagues assessed the
benefits and trade-offs of
40 proposed changes to land
management, food-production
chains and the management
of environmental risks. The
potential interventions are
outlined in the 2019 report
from the Intergovernmental
Panel on Climate Change, and
include improving management
of livestock, reforestation,
reducing consumer and retail
food waste and management of
urban sprawl.
The authors assessed each
of the actions against the
United Nations’ 17 Sustainable
Development Goals (SDGs),
as well as 18 measures from
the Nature’s Contributions
to People (NCP) framework,
which was drawn up by
scientists associated with the
Intergovernmental Science-
Policy Platform on Biodiversity
and Ecosystem Services in 2017.
This framework is intended
to recognize nature’s social,
cultural, spiritual and religious
significance, as well as its role
in providing food, clean water
and healthy air.
The analysis revealed
that several interventions
carried unintended negative
consequences. The production
of bioenergy, either with
or without carbon capture,
planting forests and
commercial crop insurance
all had potentially negative
consequences for both SDGs
and NCPs. For example,
bioenergy had large negative
impacts on maintaining land
biodiversity, freshwater quality
and food production, despite
providing affordable clean
energy. About one-third of
the interventions proposed
had no substantial trade-offs.
These included improving
water management, increasing
soil organic carbon content,
reducing pollution, reducing
post-harvest losses and fire
management.
The analysis could
help decision-makers to
assess environmental or
developmental policies to
avoid unintended trade-offs,
the authors say.
Glob. Change Biol. 26, 4691–4721
(2020)
For the latest
research published
online by Nature visit:
http://go.nature.
com/2k85xvs

Nature | Vol 588 | 10 December 2020 | S69

©
2
0
2
0
S
p
r
i
n
g
e
r
N
a
t
u
r
e
L
i
m
i
t
e
d
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
Sustainable nutrition
outlook

IAN WALDIE/BLOOMBERG VIA GETTY

Food labelling can help consumers to make healthier food choices.

Changing diets at scale

Researchers are working out how to achieve a widespread
change in eating behaviour. By Benjamin Plackett

I
f everyone ate a balanced diet featuring
more plant-based and sustainable ani-
mal-sourced food, up to eight billion
tonnes of carbon dioxide emissions might
be avoided globally each year by 2050,
according to the 2019 special report on climate
change and land by the Intergovernmental
Panel on Climate Change (IPCC).
Modifying diet on a global scale is a major
opportunity to combat climate change, argues
the report.
Naoko Ishii, an economist at the University of
Tokyo’s Institute for Future Initiatives, agrees.
“One of the biggest risk factors for the planet’s
health is our food system,” she says. “The way
we eat needs to change.”
That opinion might be gaining widespread
acceptance, but scientists don’t know how to
bring about the reforms needed, on the scale
required.
Much of the world’s population, even in
relatively rich nations, cannot afford the kind
of sustainable plant-based diet that scientists
favour. As the IPCC special report notes, mit-
igating climate change through dietary mod-
ification relies on consumers altering their
choices and preferences. These, in turn, are
guided by “social, cultural, environmental and
traditional factors, as well as income growth”,
the report says, all of which are hard to shift.
Studies on which levers for changing food
behaviours work best are surprisingly scant.
Most research concentrates on richer and
Western countries, which is where the majority
of behavioural changes are needed. By compar-
ison, data on what needs to happen in poorer
and subsistence-farming communities are
almost non-existent. Because the food behav-
iours of these communities are thought to be
much more sustainable than those of industri-
alized economies, the focus for these societies
is less on pushing urgent changes and more on
managing social changes to ensure unsustain-
able behaviours aren’t introduced.
The IPCC report lists school food pro-
curement, health-insurance initiatives and
public-awareness campaigns as examples of
policies that can potentially change demand.
But research to quantify the effects of vari-
ous interventions, such as taxes, labelling or
changing in-store food displays, suggests that

S70 | Nature | Vol 588 | 10 December 2020

©
2
0
2
0
S
p
r
i
n
g
e
r
N
a
t
u
r
e
L
i
m
i
t
e
d
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
achieving behavioural change is not straight-
forward. The interactions between the factors
that determine food production and consump-
tion are complex, and the risk of unintended
consequences from interventions can’t be
ignored.
According to the World Health Organization,
a sustainable diet means a large proportion of
the global population, particularly in wealthy
countries, will need to eat fewer processed
foods and reduce waste (see ‘Waste not, want
not’). Food producers around the world will
also need to cut down on plastic packaging
and use fewer antibiotics and hormones in
livestock.
One 2019 review1 concluded that a sustain-
able diet would budget for just 14 grams of red
meat per day, which is roughly one steak per
person, per month. Data from the Organisation
for Economic Co-operation and Development
(OECD) show that achieving such goals will be
a more daunting task in some countries than
in others. For example, Argentinians eat an
average of 106.7g of red meat per day, whereas
Nigerians consume just 8.3 g.
“It’s not about being anti-meat, it’s just
that the ratio of meat to plant-based food is
off-kilter for a lot of us,” says Mark Lawrence, a
public-health economist at Deakin University
in Burwood, Australia. “It’s also about being
better with what we have. Up to a third of food
is wasted and that’s terrible given the environ-
mental cost of making it.”
Cash is king
The data consistently show that one of the best
ways to influence food behaviour is through
price. If sustainable food were reliably cheaper
than environmentally damaging products,
market forces could often take care of the
problem. But achieving that is no small task —
sustainable food is often considerably more
expensive than its conventional rival products.
One study2 estimated that the cost of a sus-
tainable weekly food basket in Australia is up
to 30% more than that of a standard one. That’s
partly because sustainable practices often
carry additional expenses. For example, reduc-
ing antibiotic use in animal husbandry means
that welfare standards need to improve to keep
infections low. That doesn’t come cheap, and
the cost is passed down the supply chain.
Once a customer becomes accustomed to a
price point, it can be tough to convince them
to pay more. A survey of 600 city dwellers in
Poland3 found that higher prices were the main
barrier to them making more sustainable food
choices, a pattern that held true even among
respondents who were already interested in
sustainability.
Although fatty foods might be cheap at the
Waste not, want not
Between 2010 and 2016, food waste was
responsible for 8–10% of human-caused
greenhouse-gas emissions, according to
the 2019 special report on climate change
and land by the Intergovernmental Panel on
Climate Change (IPCC). In a 2017 literature
review of food-waste research9, which
pooled the findings of 202 studies, the
authors complained that researchers are
often forced to fall back on old data because
there is no up-to-date alternative.
The review found that the majority of
research papers focused on Western
countries. Switzerland, for example, wastes
about one-third of the food it produces;
Finnish consumers throw away around
30% of the food they buy; and the Danish
discard 23%. The United Kingdom, with
52 mentions, was the most studied country,
followed by the United States with 51. By
comparison, lower-income countries or
point of purchase, their true cost is reflected
in lost productivity and the disease burden
associated with obesity. The OECD estimates
that in the future the gross domestic product
(GDP) of a country will be 3.3% lower, on aver-
age, as a result of obesity. Similarly, the full
cost of unsustainable food is not reflected in
its price; the hit to a country’s GDP could be
at least as big as that of obesity, Lawrence
speculates. “There are a lot of distortions in
the market where the true cost, environmental
and economic, is not felt at the cashier,” he says.
“It can’t be just about
making polite nudges
here and there on price .”
So, what needs to be done to help sustaina-
ble products compete with their conventional
or cheaper counterparts? Tax is the obvious
answer. In the past ten years, tariffs have been
levied on sugary drinks in many countries,
including Barbados, Peru and the United
Kingdom. A systematic review4 evaluating
their effectiveness collected the findings of
15 studies, and concluded that, on average, for
every 1% bump in price, there’s a corresponding
1% fall in consumption.
“Most existing sugar-sweetened beverage
taxes are between 10% and 20%, so the effect on
consumption is not trivial,” says Franco Sassi, a
countries experiencing rapid development
are rarely investigated. India, for example,
was mentioned in just 12 (6%) studies,
despite it being home to almost one-fifth of
the world’s population.
A lack of data is hampering the search
for a solution, but, as the IPCC report
states, there is no panacea. Approaches are
likely to differ depending on the country.
In middle- and low-income countries,
improving food-supply-chain logistics to
ensure consistent access to refrigeration
would go a long way to decreasing waste.
But in high-income countries, more
inventive solutions will be needed, the
report says. For example, replacing 10–19%
of animal feed with protein produced by
recycling microbial protein from sewage,
for instance, would reduce greenhouse-gas
emissions associated with pastoral farming
by 6–7%.
health economist at Imperial College London.
“It actually makes me optimistic because back
in 2010 we couldn’t have imagined that govern-
ments would ever tax sugar.”
And taxes might go beyond just sugar. In a
2017 study5, Lawrence and his colleagues asked
944 people who bought household groceries
to choose between sustainable products and
more conventional foods. In one of the scenar-
ios, participants were told that brown rice had
a lower carbon footprint than white rice. They
were then asked to pick between the two. Under
normal market conditions, in which white rice
is cheaper, 61% opted for white rice. But when
brown rice was presented as 9% cheaper than
white, 57% instead chose the more sustaina-
ble option. That’s encouraging, says Lawrence,
because it shows that a small price change can
nudge enough consumers to give sustainable
products the majority market share.
This pattern does not necessarily hold true
for all products, however. The same experi-
ment was conducted for beef steak and sustain-
able alternative, kangaroo steak. Under normal
market conditions, people preferred beef. And
although some people drifted towards kanga-
roo meat when it was the cheaper option, beef
remained first choice by a significant margin
— even with a price difference of 33%. Price,
therefore, is only one of a number of factors
influencing whether consumers buy sustaina-
ble alternatives. “It can’t be just about making
polite nudges here and there on price — that

Nature | Vol 588 | 10 December 2020 | S71

©
2
0
2
0
S
p
r
i
n
g
e
r
N
a
t
u
r
e
L
i
m
i
t
e
d
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
Sustainable nutrition
outlook
middle-income countries that they can’t have
access to convenience foods,” says Lawrence.
The answer, he explains, is often to fix the
macroeconomics. For example, in some
Pacific island nations, tinned and imported
foods are now cheaper than fresh fish from
local waters. “This is often because interna-
tional trade deals have effectively subsidized
processed food,” he says. “It requires political
will to correct this for an entire economy, but
it doesn’t mean banning these products. It’s
just about making sure the economics of the
system isn’t skewed.”

Unintended consequences

Kangaroo meat is a sustainable alternative to beef, but consumers can be reluctant to switch.
won’t be enough,” says Lawrence.
Most of the evidence on changing food
behaviours comes from work on tackling obe-
sity. Findings from dietary studies with a focus
on health are being examined for their applica-
bility to the younger field of food sustainability.
One of the methods routinely deployed to
encourage healthy diets uses labels designed to
inform consumers about the nutritional value
of food. The traffic-light system in the United
Kingdom, for example, gives shoppers an idea
of how healthy a product is or isn’t at a glance.
The evidence of the effectiveness of this sort
of intervention is encouraging.
The OECD estimates that between 50% and
60% of shoppers check nutritional labels at
least some of the time. Research established
that labels indicating a product’s health cre-
dentials — or lack thereof — are linked to an
18% increase in people buying healthier food6.
Labelling on health grounds influences food
behaviour, says one of the authors of the study,
Michele Cecchini, a health-policy analyst at the
OECD’s health division in Paris. “I don’t see why
the same wouldn’t also apply to other issues
that consumers care about, like sustainability,”
he says.
Ishii says that only a proportion of consum-
ers need to change their behaviour for labelling
information to have an impact. “A relatively
small number can influence the brand to
change, and therefore they can influence the
wider supply chain,” she says.
Cultural matters
A 2020 survey7 of close to 1,200 people across
12 European countries and Uganda, high-
lighted the influence that culture can have on
food behaviours. For example, the majority
of people surveyed in each of the European
countries disagreed with statements such as
“a particular food is chosen because it makes
me look good in front of others”. In Uganda,
however, more participants agreed with such
statements.
“I don’t think we’ll be able to address food
behaviour on a global level in a uniform way,”
says Suzanne Kapelari, an educational scientist
at the University of Innsbruck, Austria, and an
author of the study. “The more we know about
the cultural attitudes to food and behaviours,
the better, but there’s quite a bit of work to be
done on that.”
“I don’t think we’ll be able
to address food behaviour
on a global level in a
uniform way.”
Food behaviours in higher-income coun-
tries such as many OECD member states are
different from those in middle- and low-in-
come countries. Consumers in wealthy
countries buy more meat, and packaged and
processed foods. “It’s been like this for decades
in high-income countries,” says Lawrence.
People in low-income countries, by compari-
son, often eat less meat and opt for locally pro-
duced products with less packaging.
The emphasis in high-income countries is,
therefore, on correcting unsustainable behav-
iours, whereas in low- and middle-income
countries, it’s on preventing unsustainable
behaviours becoming the norm.
“We have to be careful here because we
don’t want to be sitting in ivory towers telling
JULIEN VIRY/ISTOCK/GETTY
Although eliminating or significantly
reducing meat consumption would help
the environment, evidence suggests this is
unlikely to happen at scale because many
meat-eaters are reluctant to change their eat-
ing habits. A better strategy, some research-
ers argue, is to shift consumer preferences
from high carbon-producing meats, such as
lamb and beef, towards meats with a lower
environmental impact, such as chicken and
pork.
In a 2019 study8, marketing experts in
Belgium reorganized a butcher’s counter,
increasing the space given to poultry and
decreasing the space for red meat. This led
to a 13% increase in chicken sales in 4 weeks.
The only trouble is that sales of red meat didn’t
fall in tandem, so the net result was a greater
amount of meat sold, albeit not significantly.
Although this was one small study, it demon-
strates a broader point: there is no single
solution to the problem of how to change con-
sumers’ behaviour. “The common feature in
all these areas is their limited effectiveness,”
says Sassi.
The hope is that applying a range of meth-
ods in a coordinated way will have a cumula-
tive effect. But that hope lacks a solid evidence
base. Researchers are even unsure whether
different groups respond to different meth-
ods. “The truth is that we don’t really know,”
says Sassi. “It’s a gap in our evidence.”
Benjamin Plackett is a freelance science
writer based in London.
1. Willett, W. et al. Lancet 393, 447–492 (2019).
2. Barosh, L., Friel, S., Engelhardt, K. & Chan, L. Aust. N. Z. J.
Public Health 38, 7–12 (2014).
3. Rejman, K., Kaczorowska, J., Halicka, E. & Laskowski, W.
Public Health Nutr. 22, 1330–1339 (2019).
4. Teng, A. M. et al. Obes. Rev. 20, 1187–1204 (2019).
5. Hoek, A. C., Pearson, D., James, S. W., Lawrence, M. A. &
Friel, S. Food Qual. Pref. 58, 94–106 (2017).
6. Cecchini, M. & Warin, L. Obes. Rev. 17, 201–210 (2016).
7. Kapelari, S. et al. Sustainability 12, 1509 (2020).
8. Coucke, N., Vermeir, I., Slabbinck, H. & Van Kerckhove, A.
Foods 8, 186 (2019).
9. Xue, L. et al. Environ. Sci. Technol. 51, 6618–6633 (2017).

S72 | Nature | Vol 588 | 10 December 2020

©
2
0
2
0
S
p
r
i
n
g
e
r
N
a
t
u
r
e
L
i
m
i
t
e
d
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
 S P O N S O R FE AT U R E
MINT IMAGES/MINT
IMAGES RF/GETTY IMAGES

BUILDING BETTER FOOD SYSTEMS

FOR NUTRITION AND HEALTH
In advance of THE 2021 NUTRITION FOR GROWTH SUMMIT IN TOKYO, a group of leading
international academics, policy analysts and food industry representatives attended a workshop to
discuss global nutrition challenges. To address the obstacles to better health, participants discussed
the stakeholders and levers that can effect change so that nutritious — and sustainable — diets are
available for all.

A
lthough hunger was
steadily declining for
decades, progress
stalled in 20151 and since then,
the number of people who
suffer from hunger has slowly
increased. In 2018, there were
more than 820 million people
hungry2. Around 2 billion people
worldwide have micronutrient
deficiencies and 149 million
children are stunted3. And yet
in 2018, 40 million children
under five were overweight,
and in 2016, 2 billion adults
were overweight, with a third
of these obese. Many countries
are now challenged by what’s
known as the double burden of
malnutrition (DBM), in which
people are simultaneously
overweight and malnourished.
The double burden
of malnutrition is closely
associated with nutrition
transition4 which is tracking both
demographic transition (where
countries shift from patterns
of high to lower fertility and
mortality) and epidemiological
transition (where the prevalent
disease burden shifts from
infectious to chronic and
degenerative disease). With
urbanization, economic growth
and technological change,
diets shift from starchy, low
variety, low fat and high fibre
towards the ‘Western’ pattern
of increased fat, sugar and
processed foods. The effects
of malnutrition have also been
shown to be intergenerational,
where maternal nutrition can set
a trajectory for life-long health of
her offspring5.
In the 1990s, DBM mostly
affected the highest income
countries of the group of low-
and middle-income countries
(LMICs), but in the last decade
it has been seen in even the
poorest of countries, including
in rural areas and in people with
low incomes. DBM is seen in
around 14 million children in
Asia and 9.6 million children
in Africa.
Syndemics
The global synergistic
epidemic — the ‘syndemic’
— of overnutrition,
undernutrition and climate
change was described as
the greatest challenge for
human and planetary health
in the 21st century6 and
has been compounded by
unprecedented challenges to
food and nutrition security in
2020. In the biggest upsurge
seen in decades, desert locust
outbreaks in East Africa and
South Asia have had disastrous
effects on local food supplies.
COVID-19 — a pandemic
inextricably linked in cause
and effect to food systems —
could push up to 100 million
people into extreme poverty7.
Lockdowns to deal with the
spread of COVID-19 disrupted
global supply chains and
national, local and household
economies. With the increase
in poverty, and hindrance of
essential interventions and
food security, COVID-19 could
reverse many of the hard-won
gains in maternal and child
nutrition of recent decades.
Furthermore, overweight
and obesity are associated
with increased likelihood of
hospitalization, ICU admission
and worse outcomes
in COVID-19.
Building solutions
It is now evident that human
and planetary health cannot
be disentangled, nor can
nutrition be considered in

isolation from the food system.
The food system concept, as
put forward by Parsons and
Hawkes8 describes the interplay
between the food supply chain
and social, political, economic
and environmental outcomes.
It puts the individual at the core
and yet, power concentrations
within the food system exist and
markedly skew the attainment
of health and sustainable
and affordable diets for all.
Malnutrition is a consequence.
Meanwhile, nutrition
remains everyone’s business.
Food systems are a major
source of livelihoods globally.
It is estimated that one in three
workers are employed in the
agricultural sector alone9. A
Chatham House report10,11
showed that businesses in
19 countries lose up to an
estimated US$38 billion a year
from lost worker productivity
because of undernutrition, and
up to an estimated US$27 billion
a year because of obesity. The
authors conclude there is an
incentive for the food industry
to invest in better nutrition,
because it could drive a stronger
and more productive economy
generally, as well as helping
its own productivity. Around
80% of food is produced by the
private sector (mostly small-
and medium-sized enterprises)
meaning communication of the
benefits of transformation will
need to be multi-faceted, multi-
sectoral, long-term campaigns.
Within that transformation,
food systems geared towards
Ajinomoto
Group Global
Event Organizers
Brand
Guidelines
Springer Nature, Nature Food, RESULTS Japan, GGG+
Supported
メッセージ付AGBby 基本表示規定
Ministry of Foreign Affairs of Japan,
表示色
Ministry of Health,
メッセージ付AGBのカラーは、
「味の素レッド」を基本とします。特色
Labour and Welfare
を使用できない場合は、
of Japan,
プロセスカラーを使用することも可能です。
Ministry
クリアスペース
of Agriculture,
メッセージ付AGBの独自性を保ち、常に明瞭に表示するために、その
Forestry and Fisheries of Japan
周囲に最小限のクリアスペースを設定しています。
クリアスペースは、AGマークの
「白ふち」
の幅（X）を基準にし、周囲
に設けています。
クリアスペース内には、文字やロゴ、グラフィック要素を表示する
ことはできません。
クリアスペースを白く塗りつぶしたものを「白窓」と呼びます。
（「白ふち」
「白窓」については No. 2-06 参照）
最小使用サイズ
メッセージ付AGBが明瞭に表示できる最小サイズとして、最小使用
Economic, social, political and other factors determine what people eat.
nutrition must also be geared
towards sustainability. The
food system contributes about
one-third of anthropogenic
greenhouse gas emissions,
diminishes biodiversity, pollutes
water and degrades soil.
And the damage caused
by the climate emergency will
exacerbate changing diets.
Around two-fifths of the
world’s population is unable to
afford a healthy diet. This has
a huge impact on the health of
the population, as poor diet,
whether undernutrition or over-
consumption, plays a significant
role in health and health risks.
Dietary changes are driven by
a country’s wealth and level
Event Sponsors
メッセージ付AGB
クリアスペース
X X
X
X
of urbanization, and dictated
by increasing inequalities in
society. The disruption to food
supply as a result of climate
change will only serve to
throw up further challenges to
providing nutrition for all.
At the Food Systems for
Nutrition and Health workshop
held in advance of the 2021
Nutrition for Growth Summit in
Tokyo, three breakout sessions
considered the role of policy,
the role of the private sector
and the role of the nutrition
research community in bringing
about necessary change to
global food systems in order
to provide healthy, nutritious
and sustainable diets for all.
No. 2-03
カラー
プリント表示の場合
特色：味の素レッド
近似色：PANTONE 199C
プロセス：M100%+Y100%
デジタル表示の場合
RGB: 238, 28, 38
HEX: #EE1C26
最小使用サイズ
S P O N S O R FE AT U R E
While each session came at
(FROM TOP LEFT TO BOTTOM RIGHT) MONKEYBUSINESSIMAGES/ISTOCK / GETTY IMAGES PLUS;
SCIENCE PHOTO LIBRARY/GETTY IMAGES; SZEFEI/SHUTTERSTCOK; PEETERV/ISTOCK / GETTY IMAGES PLUS
the challenge from different
angles, a common theme was
the need for collaborative,
multi-stakeholder approaches
to change. The reports continue
on following pages. n
REFERENCES
1. FAO, IFAD, UNICEF, WFP and
WHO. The State of Food Security
and Nutrition in the World 2019.
Safeguarding against economic
slowdowns and downturns. (2019).
2. FAO, IFAD, UNICEF, WFP and
WHO. The State of Food Security and
Nutrition in the World 2018. Building
climate resilience for food security
and nutrition. (2018)
3. UNICEF. The State of the World’s
Children 2019: Children, food and
nutrition. (2019)
4. Popkin, B. & Gordon-Larsen, P.
The nutrition transition: worldwide
obesity dynamics and their
determinants. Int J Obes 28,
S2–S9 (2004).
5. Black, R. et al. Maternal and child
undernutrition and overweight in
low-income and middle-income
countries. Lancet 382, 427 - 451
(2013)
6. Swinburn, A. et al.
The Global Syndemic of Obesity,
Undernutrition, and Climate
Change: The Lancet Commission
report. The Lancet Commissions 393,
791–846 (2019).
7. We all stand together. Nat Food 1,
583 (2020).
8. Parsons, K. and Hawkes, C. WHO
Policy Brief 31: Connecting food
systems for co-benefits: How can food
systems combine diet-related health
with environmental and economic
policy goals? (2018)
9. FAO. FAO Statistical Yearbook 2012.
World Food and Agriculture. (2012)
10. Wellesley, L. et al. Chatham House
report: The Business Case for
Investment in Nutrition. (2020).
11. Pringle, S. Tackling Malnutrition:
Harnessing the Power of
Business. Chatham House Expert
Comment (2020).
LEARN MORE
ABOUT
PANELLISTS
 S P O N S O R FE AT U R E
BREAKOUT SESSION: NUTRITION FOR HEALTH
AHIRAO_PHOTO/ISTOCK / GETTY IMAGES PLUS
EFFECTING CHANGE IN
A FRACTURED LANDSCAPE
The double burden of malnutri-
tion (DBM) — the combination
problem of both too many
calories and simultaneous mal-
PERSPECTIVES FROM JAPAN : TOWARDS nutrition — is increasing in low-
NUTRITION FOR GROWTH, TOKYO 2021 and middle-income countries.
Inadequate nutrition is still a leading cause of global deaths. This, along with the climate
“Figures show that around 38%, or 5.3 million children, emergency and COVID-19
are dying before their fifth birthday. According to a 2013 pandemic, is having a signifi-
Lancet paper, 45% of those deaths are due to malnutrition. cant impact on global health
In other words, it is estimated that if those infants had been and increasing the risk of
able to get adequate nutrition, they would not have died non-communicable diseases.
from infectious diseases such as diarrhoea, measles, acute The Nutrition for Health
respiratory infections, malaria, tuberculosis and AIDS,” said
breakout session of the Food
SANTO Akiko, President of the House of Councillors, the
Systems for Nutrition and
National Diet of Japan.
Nutrition has an important role to play in prevention and Health workshop, comprised
control of disease, including COVID-19. of experts from industry,
“Japan achieved a dramatic reduction in the morbidity government and academia,
and mortality of communicable diseases…through nutrition discussed existing policies
improvement after the Second World War. Recent studies… to tackle DBM. Examples
have shown a close relationship between COVID-19 and include Ghana’s ‘Planting
malnutrition. Therefore, taking measures to improve for food and jobs’, which
nutrition is critical,” said SHOBAYASHI Tokuaki, Director has helped communities to
General, Health Service Bureau, Ministry of Health, Labour
expand their land use, and
and Welfare.
This requires a shift towards better food and healthier diets. community dams that support
“COVID-19 [has] brought a significant demand shift fish farms and improve land
from eating out to home,” said OSAWA Makoto, Vice- irrigation. Japan worked
Minister for International Affairs, Ministry of Agriculture, to improve undernutrition
Forestry and Fisheries, Japan. “We take this opportunity to following the Second World
propose healthier and more sustainable dietary habits by War, and overweight and
promoting traditional local diets based on local production obesity following economic
for local consumption.” growth, through education and
Achieving an improvement in diet and sustainable eating
improving access to healthy and
requires a policy shift.
nutritious food, including for
“Investment in health system strengthening is a
prerequisite for sustainable development and economic low-income populations as part
growth. Good nutrition is a foundation for healthy lives and of universal health coverage.
sustainable health systems,” said MIMURA Atsushi, Deputy Tackling DBM needs
Vice Minister of Finance for International Affairs, Ministry behaviour change of all
of Finance. food system actors and
The World Bank’s Human Capital Project is an important consumers, backed by a deep
part of the process. understanding of what drives
“The Human Capital Project considers nutrition as people’s decisions in nutrition
an essential element in unlocking human potential and
and health. All stakeholders,
economic growth,” added Mimura.
including governments,
The current global situation and the forthcoming summit
provides an opportunity to transform the way the world companies, the World Health
tackles the global malnutrition challenge. Organization, the UN Food
“We expect countries to review their existing nutrition and Agriculture Organization,
policy, consult with other nutrition stakeholders and the International Union of
announce ‘SMART’ commitments at the Tokyo N4G Summit Nutritional Sciences, non-
2021,” said ONO Keiichi, Ambassador, Director-General for government organizations and
Global Issues, Ministry of Foreign Affairs of Japan. policy makers, will need to be
involved to facilitate behavioural
change and provide food
environments that incentivize
healthy eating. Government
approaches could include
taxing highly-processed foods
and sugary drinks, improving
transport infrastructure for
affordable delivery of food,
providing safety nets for
consumers who cannot afford a
healthy diet, investing in public
awareness campaigns, and
making nutrition a focus for
health both in schools and for
the general public. Companies
need to invest in innovation
to improve the nutritional
value and safety of their food,
but not at the expense of
palatability. Investment in R&D
is expensive and needs support
to develop a sustainable
business case that does not
compromise affordability.
Change is even more
important in light of the
COVID-19 pandemic, which
demonstrated the impact of
overweight and nutrition on
the progress of infectious
disease. The strategies to keep
the spread of COVID-19 under
control have left some people
isolated, and many have lost
their jobs and experienced
serious economic hardship. The
experience of countries such
as Japan, which has survived
natural disasters, has shown
the importance of nutritional
preparedness. While staples
that can be stored long-term
are a defence against hunger,
food support should be safe
and nutritionally balanced,
with the right levels of protein,
lipids, vitamins, minerals and
dietary fibre. The store of
staples can be buttressed by
local production of nutrient-
dense food that is accessible
and affordable.

The panel heard that
the One Health Approach,
which looks at the interaction
between human and animal
health and the environment,
will be critical, particularly with
the threat of future zoonotic
diseases. Approaches aimed
at improving human health
also need to consider animal
health and the environment,
and this requires collaboration
and cooperation between re-
searchers involved in all these
fields. Education at all levels
BREAKOUT SESSION: BUSINESS FOR NUTRITION
ARE HEALTHY PROFITS
HOLDING BACK
HEALTHY DIETS?
There are many barriers to the
wide adoption of sustainably
produced food for healthy
diets. Everything from
nutrition education, through
to accessibility, convenience,
trends and marketing have
a hand in consumer choice.
With a large percentage of the
world’s population reliant on
cheap calorie-rich staples and
unable to afford more nutrient-
rich food, cost is perhaps the
largest obstacle.
A panel of leading food
industry, academic and
government representatives
formed a break-out from the
Food Systems for Nutrition and
Health workshop to discuss the
relationship between profit and
public good.
Food manufacturers often
produce foods with high levels
of calories, salt, sugar and other
unhealthy attributes. For lack
of a better term, such foods
are sometimes labelled as
‘ultra-processed’, though the
concern is not with the degree of
processing, but rather with the
nutritional content considered
as unhealthy.
(including medical schools) is
required for precision plane-
tary, population and personal
nutrition and health.
Data and evidence was also
discussed as a global challenge
related to nutrition. Researchers
need data collected at a local
level to track DBM, but there
isn’t enough available. Better
and more granular data,
collected more frequently, will
help to track diet, weight, blood
sugar and/or health. This can
be improved by working with
A report from Chatham
House entitled The Business
Case for Investment in Nutrition,
suggests that making
businesses aware of the adverse
impacts of poor nutrition on
their profits and market growth
could be a persuasive approach.
Catering for more nutritious
diets, not only through product
lines but also food in the
workplace, could benefit both
the health of food company
employees and that of the
Better product labelling is one approach to changing consumer behaviour.
local academic institutions,
food companies and schools,
and by making the most of
locally available and low-cost
point-of-care technologies
and innovations.
To really make an impact, the
nutrition community needs to be
involved in the political decision-
making process. As well as being
able to provide support and
evidence, nutrition professionals
should assess the trade-offs and
challenges that policymakers
face and feed this into education
population as a whole. It could
also contribute to increased
profits for companies across
all sectors. The panel heard
about well-known global brands
that are inspiring consumers to
use more nutritious products
to create nutrient-rich and
affordable meals.
The panel discussed how
farm subsidies could be directed
to incentivize the production
of healthy foods such as
fresh fruits and vegetables in
a more sustainable manner.
Governments can also introduce
minimum standards in public
procurement policies.
Aside from awareness-
raising about healthy diets,
governments can further have
S P O N S O R FE AT U R E
and research. Countries like
Sri Lanka and Japan have
nutritionists embedded within
their ministries of health. Japan
also has dedicated nutritionists
working within the Ministry
of Agriculture, Forestry and
Fisheries, the Ministry of
Education, and the consumer
affairs agency. In order to beat
DBM and other nutritional
challenges, nutritionists need
to reach out to, and work
with, other sectors to find
common solutions.n
a role in influencing dietary
preferences through educational
campaigns, food labelling, public
regulations, taxing unhealthy
foods and other measures.
The panel heard that some
regulatory flexibility may be
needed to allow for product
reformulation. Companies
indicated that such flexible
frameworks help attract more
investment in healthier options.
By increasing the demand for
healthier food, this could have
a knock-on effect of cutting
production costs through
innovation and economies
of scale.
Markets for ready-to-eat
meals, snacks and other comfort
foods are growing and proven to
be highly profitable. During the
COVID-19 pandemic, demand
TETRA IMAGES/GETTY IMAGES
for comfort food has increased.
However, many of these (ultra-)
processed foods can be calorie-
dense, high in added sugars and
salt, and low in nutrients. The
panel put forward examples
of governments that have
taken different approaches
to improving consumer food
selection. Chile, Mexico, the
UK and South Africa have
introduced taxes on sugary
drinks, accompanied by
marketing regulations, which
have driven companies to
change business strategies.
Indonesia has focused on
 S P O N S O R FE AT U R E
improving nutritional literacy
to promote behaviour change.
Japan has nutrition training in
schools. Israel is taking a more
positive approach by labelling
the healthy foods.
The panel acknowledged that
the COVID-19 pandemic has
impacted national economies,
and constrained budgets,
BREAKOUT SESSION: POLICY
POLICY LANDSCAPE SUPPORTING
BUSINESS AND NUTRITION
Ending hunger and malnutrition
SUBMAN/E+/GETTY IMAGES
requires political will. In a
breakout session held two weeks
after the main Food Systems for
Nutrition and Health workshop,
a panel of expert representatives
from business, academia and
government discussed how
food policy needs to shift to
drive nutrition for health and
well-being and the sustainable
production and consumption of
nutrient-dense foods. However,
as the Business for Nutrition
breakout session also discussed,
many of the food policies in
place around the world focus
on producing staples, and on
maintaining the agriculture
industry and tend to support the
consumption of energy-dense,
nutrient-poor foods.
The panel heard how
Thailand has been successful
in using policy to improve
the nation’s nutrition. In the
1970s, evidence from research
and surveys convinced the
government that nutrition was
essential for development,
and that undernutrition in the
country was linked with poverty.
The resulting poverty alleviation
plan led to a multisectoral
approach and unlocked funding
that supported community-level
planning and implementation,
which was important to ensure
ownership. The integrated
which will make setting aside
resources for improving
nutrition more difficult for both
domestic governments and
overseas donor governments.
Innovative funding sources
and private sector finance will
need to step up. Investment in
research and development will
need to come from industry,
Access to low-cost, nutrient-dense, sustainably produced food is the goal.
actions, including health,
agriculture and education,
the widespread mobilization
of village volunteers, and the
support from officials and
experts resulted in a remarkable
decline in child undernutrition
within a decade. As Thailand has
become richer, the population
began to experience DBM, with
increasing levels of obesity and
simultaneous malnutrition.
The government introduced
taxes in 2001 to tackle levels of
tobacco and alcohol use, with
a percentage of the proceeds
going to health promotion. In
2017, building on a programme
targeting children called ‘Sweet
Enough’, the government
deployed an evidence-based
sugar tax, using a tiered scale
not-for-profit funding groups
and governments.
Consumer behaviour and
choice is shaped by advertising
and marketing messages. The
panel agreed that changing
behaviours and shaping
markets, for both industry
and consumers, will need a
multi-stakeholder approach,
and giving companies two
years’ notice to reformulate
their products. The success of
these programmes has inspired
further campaigns and activities;
most recently the government
considered the merits of a salt tax.
Beyond public health, the
panel heard that nutrition
policy must also focus on
planetary health, and this will
require policies that create links
between food, health, education,
biodiversity, sustainability, the
environment, and the climate
emergency. Successful policies
will rely on all the players
working together, including
communities and small to
medium enterprises. Taxes
can play an important role,
especially if they are reinvested
and transparent and evidence-
based discussions with
clear accountability.
The first United Nations
Food Systems Summit in 2021
along with the Nutrition for
Growth Summit due to be
held in Tokyo in 2021 provide
good opportunities to advance
this agenda.n
in health, nutrition and society.
There should also be recognition
and incentives to encourage
stakeholder involvement. This
‘planetary health’ message
needs to be built into nutrition
research, and into training for
nutrition professionals.
Redirecting current
subsidies and introducing new
subsidies could play a role in
supporting both public and
planetary health, because
at the moment subsidies do
not necessarily support the
production of healthy food in
a way that is either efficient
or sustainable.
One of the biggest challenges
in nutrition policy is effectively
engaging and communicating
with the general population.
The most effective messages
are tangible, harmonized across
stakeholders and come from a
credible source. Countries can
learn from one another, but the
messaging needs to be tailored
to the social, religious and
cultural climate of the location.
The COVID-19 pandemic has
shown the impact that global
emergencies can have on food
systems. The panel discussed
the importance of learning from
the crisis, and ‘building back
better’. Food and nutrition needs
to be part of any emergency
and recovery plans, and part of
plans for the health of the nation
and the planet. This will require
stakeholders from academia,
industry, nutrition science and
the community to be invited
to the table. n