EDI TO R I A L
Rooting out scientific misconduct
S
cientific misconduct is an issue rife with contro-
versy, from its forms and definitions to the poli-
cies that guide how allegations are handled. A
survey published nearly 15 years ago reported
that 2% of researchers said they had fabricated
or falsified data in their published work. This is
not just an academic issue. Fake data promote
ineffective or even dangerous treatments, for example,
and thwart the discovery of real solutions for society.
In the United States, the Office of Research Integrity
(ORI) is tasked with rooting out misconduct in research
funded by the National Institutes of Health (NIH). Last
October, ORI proposed changes to how it functions.
The agency’s recommendations—the first since 2005—
have evoked mixed reactions, but the real problem is
that ORI is underfunded and lacks
the resources and authority needed to
make a difference. Unless its charter
is revised by Congress, the ORI can
sadly do little more than tinker at the
“…ORI…lacks
edges of scientific fraud.
It is a wonder that the ORI accom-
the resources
plishes much at all. Its current budget
is $12 million per year to oversee work
and authority
funded by NIH, a $48 billion agency.
Add to that the frequent internal
needed to make
strife over ORI’s proper role and a di-
rectorship that has often been vacant, a difference.”
and one can see how its ability to be
effective does not meet the expecta-
tions for upholding the integrity of research activities.
The regulations proposed by ORI’s new director,
Sheila Garrity, include fine-tuning of definitions and
processes that is long overdue. For example, they clarify
the term “reckless,” which was used more often recently
by the ORI to prosecute fraud cases. The term empha-
sizes indifference to or disregard for the truth of the
matter being asserted, according to a recent article.
But what should happen if someone has supervised,
but not performed, the research at issue, as in the case
of former Stanford University president Marc Tessier-
Lavigne, who failed to correct problems in work by his
trainees? As the article asks, what is reasonable supervi-
sion and when is it so lacking that it becomes reckless?
The agency seems much more open to disclosing the
results of university investigations, a transparency that
has predictably been met with criticism from academic
institutions that claim it might “violate privacy laws
SCIENCE science.org
or distort the actual findings.” Still, ORI has missed an Ivan Oransky
opportunity to hold institutions accountable. Although is a cofounder of
the agency suggests that it is an institution’s responsi- Retraction Watch,
bility to foster an environment that promotes integrity, New York, NY, USA;
how should this be measured and judged? The revision distinguished
still exclusively addresses misconduct by individuals. It journalist in
would be best if an institution could be held responsible residence at New
for a toxic, unsupportive research environment. York University,
Even if its recommendations are further adjusted, New York, NY, USA;
ORI lacks the personnel and budget to address the
and editor-in-chief
potential scope of alleged misconduct. The office is
of The Transmitter,
p
largely limited to supervising university investigations
The Simons
instead of carrying them out itself, which would avoid
the obvious institutional conflict of interest. ORI also Foundation, New
lacks subpoena power to compel witness testimony. York, NY, USA. ivan@
This point may help explain why the retractionwatch.com
National Science Foundation’s Office
of the Inspector General, which has Barbara Redman
g
subpoena power, tends to make far
more findings of misconduct than
is a courtesy adjunct
professor at New
ORI each year.
There is good news, though. Some
York University’s
School of Nursing
y
publishers have become more willing
to correct the scientific record. This
and an associate of
New York University
led to more than 10,000 retractions in
2023—reflecting about 0.2% of the lit-
Langone’s Division
of Medical Ethics,
erature across all fields, as indicated
in a recent analysis. According to the
New York, NY, USA.
br68@nyu.edu
study, this is a 10-fold increase com-
pared with two decades ago. Not all
of these were because of misconduct, but studies have
consistently found that two-thirds of retractions are.
y g
However, it is not clear whether the incidence of mis-
conduct has increased over time. There is no question
that the work of today’s sleuths, who often use software
not available 20 years ago, has pushed these numbers
higher. And the fraudulent activity of research paper
y
mills that produce fake manuscripts is likely also a fac-
tor. On a larger scale, universities are starting to take
,
a harder look at the suitability of perverse “publish or
perish” incentives for faculty promotion and tenure.
Thirty years ago, ORI was created in response to a se-
ries of scandals at prominent institutions, some involv-
ing faked data, that caught the attention of Congress.
Congress should strengthen what it set out to do—
address misconduct in science by giving ORI the teeth it
needs to sink into the problem.
–Ivan Oransky and Barbara Redman
10.1126/science.adn9352
12 JANUARY 2024 • VOL 383 ISSUE 6679 131
 NEWS
We’ll never be a science superpower behind a visa paywall.
“ Former U.K. science minister George Freeman, in Times Higher Education, on rising
immigration fees and the government’s plan to raise the minimum salary needed to get a skilled
worker visa to £38,700—above the typical starting salary of a U.K. postdoctoral researcher.
”

IN BRIEF Edited by Jeffrey Brainard

Billionaire vows plagiarism checks

RESEARCH INTEGRITY | Hedge fund
manager and billionaire Bill Ackman says
he plans to look for signs of plagiarism
in work by all of the faculty and board
members of the Massachusetts Institute of
Technology (MIT), including its president.

p
His campaign, announced last week, comes
after Ackman was angered by two Business
Insider exposés this month reporting that
a 2010 doctoral dissertation and other
academic papers by Neri Oxman, his wife,
included multiple passages substantially
lifted from other sources without proper

g
attribution. Without offering evidence,
Ackman alleged that MIT leaders likely
prompted the articles. He has pushed
to oust several university presidents

y
including MIT’s head, cell biologist Sally
Kornbluth, claiming they did not take
antisemitism on campus seriously enough.
(Kornbluth has held on, but political
scientist Claudine Gay stepped down last
week from Harvard University’s helm
after examinations of her scholarly work
revealed instances of plagiarism, which
Ackman and others had highlighted pub-
licly.) Plagiarism and publishing experts

y g
told Science that identifying the hundreds
The ill-fated Peregrine Lunar Lander is shown nestled within a Vulcan rocket before its launch. of thousands of research papers and other
writings of nearly 1100 MIT faculty mem-
bers, obtaining the full text, and running
PLANETARY SCIENCE the documents through plagiarism-checking

y
software could take months, even with an
Failure spoils U.S. return to Moon unlimited budget.
,
Spain reforms research reviews

H
ours after the 8 January launch of the Peregrine lander from Cape
Canaveral, Florida, a catastrophic leak crippled its propulsion P O L I C Y | Spain’s new approach to evaluat-
system, ending hopes it would become the first commercial mis- ing scientists at public universities is
sion to land on the Moon. Developed by Astrobotic Technology, drawing praise for emphasizing quality
a Pittsburgh startup, Peregrine was the first mission in NASA’s over quantity of research publications.
Spanish researchers face evaluations
Commercial Lunar Payload Services program, which has sought every 6 years to gain promotions and
to spur lunar-landing capabilities by financing private missions. NASA salary increases, and in the past, evalu-
has repeatedly warned that it expects some of these low-cost probes to ations hinged on just one criterion:
fail; a second mission, led by Houston’s Intuitive Machines, will launch whether scientists published at least five
papers in high-impact journals in that
next month. The mishap raises questions for Astrobotic, which has a period. Critics complained the approach
contract to deliver NASA’s $500 million Volatiles Investigating Polar incentivized publishing mediocre papers
Exploration Rover to the Moon later this year. and unethical ways to boost authorship

132 12 JANUARY 2024 • VOL 383 ISSUE 6679 science.org SCIENCE

 ANIMAL PHYSIOLOGY

Snake teeth predict strike tactics

T
he shape and position of a snake’s teeth say a lot about
how it hunts, a study has found. Using 3D x-ray scans
of their mouths, herpetologist William Ryerson of the
Cornell University College of Veterinary Medicine identi-
fied two distinct groups, each correlated with different
methods of striking prey. “Strikers,” such as boa constrictors
(right), quickly attack from above, immediately grabbing the prey
with their tall lower teeth to enable the snake to coil around the prey
and strangle it. “Lungers,” such as king snakes, attack an animal
straight on and less quickly, piercing it with both their top and bottom
teeth at the same time. Their gape is smaller than that of strikers.
Ryerson, who presented the findings last week at the Society for
Integrative and Comparative Biology’s annual meeting, studied about
70 individuals from 13 snake species, representing four of the five main
groups. He filmed them as they attacked a dead rodent that he had heated
to attract them. He hopes a better understanding of snake teeth will
inspire practical engineering applications.

p
credits. Under the new method, which such research, but some federal and state opposed the TMT and long campaigned
took effect on 1 January, the National lawmakers have introduced new propos- against overdevelopment of the summit,
Evaluation and Accreditation Agency will als to prohibit it. which is home to 13 observatories. The

also consider a range of research outputs
besides papers—for example, patents and
exhibitions—along with the work’s social
impact. The reforms are also crafted to
Curie lab demolition postponed
| A planned demolition of part
H I S T O RY
g
CSO is the first to be taken down after a
2015 pledge by Hawaii’s then-Governor
David Ige (D) to reduce the number. From
1987 to 2015, the CSO pioneered studies

reduce plagiarism and deceitful multi-
authorship, and to reward scholars for
publishing in open-access journals. Spain
is one of the first countries to adopt such
a policy; in several others, only individual
institutions have.
Fetal tissue yields brain organoids
BIOMEDICINE | Researchers say they have
of Marie Curie’s historic laboratory in
Paris has been suspended after an outcry
from preservationists and an interven-
tion from France’s culture minister.
The targeted building, the Pavillon des
Sources, was one of three constructed
when the Radium Institute, now known
as the Curie Institute, was established in
1909. Curie used the building for some of
her pioneering work on radioactivity and
y
in the wavelength range between infrared
and radio. Its components will be shipped
to Chile and reassembled in the high
Atacama Desert, where it will be reborn as
the Leighton Chajnantor Telescope.
Radio circles linked to gas blasts
ASTRONOMY | Scientists this week pro-
posed a cause of odd radio circles (ORCs),

grown organoids from fetal brain tissue,
providing a new means of studying brain
cancers and developmental disorders.
In previous studies, researchers created
brainlike organoids using stem cells that
later became the first woman to win a
Nobel Prize, for discovering polonium and
radium. The institute, which still manages
the site, proposed to build a new build-
ing there to expand its cancer research,
y g
the mysterious, giant rings that pop up
in sky surveys at radio wavelengths: They
form when distant galaxies expel gas
after intense star formation. Fewer than
a dozen ORCs have been spotted since

can turn into many cell types. In this
week’s issue of Cell, a research team in
Europe reports using fetal brain tissue to
grow 3D structures that resemble those
that naturally occur in the cortex, fore-
brain, or spinal cord. The new organoids
PHOTO: LEONARDO MERCON/VWPICS/SCIENCE SOURCE
saying Curie used the existing one mainly
to store radioactive materials. Last week,
the institute said it would now consider
alternatives to demolishing the building,
which remains contaminated.
y
the first, in 2019, and astronomers did
not even know whether they originated
,
in or near our own galaxy or in distant
ones. New clues emerged after the team
discovered that a distant galaxy, which
appeared to be enveloped in an ORC,

may offer a realistic new model of how
individual brain regions grow and could
also enable better testing of drugs for
brain cancer. Pregnant people donated
the tissue from fetuses aborted between
12 and 15 weeks postconception. The team
worked with bioethicists to ensure that
the rice grain–size organoids could not
feel pain or become conscious. But in the
United States, the study of fetal tissue is
politically fraught. In 2021, President Joe
Biden’s administration reversed a ban on
Mauna Kea slim-down begins
ASTRONOMY | Workers last month
completed dismantling and removing the
Caltech Submillimeter Observatory (CSO)
atop Hawaii’s Mauna Kea, a first step
toward fulfilling a deal that was hoped to
soften opposition to construction there
of a proposed observatory, the Thirty
Meter Telescope (TMT)—the largest in the
Northern Hemisphere. Native Hawaiians,
who consider Mauna Kea sacred, have
produced a bright fluorescent glow from
excited oxygen atoms. The astronomers
developed a computer model suggesting
that during an intense period of star for-
mation there 1 billion years earlier, many
stars quickly went supernova, blasting
gas out of the galaxy. That outflow hit
cooler surrounding gas, generating the
radio ring; the fluorescent glow resulted
as some gas fell back into the galaxy. The
team describes its findings, linking the
ORC to that galaxy, this week in Nature.

SCIENCE science.org 12 JANUARY 2024 • VOL 383 ISSUE 6679 133

NE WS
IN DEP TH
CLIMATE CHANGE
The hottest year was even hotter than expected
Greenhouse gases, El Niño, and cleaner air fueled record heat in 2023
By Paul Voosen
I
t comes as no surprise to anyone who
sweated through it: 2023 was the hot-
test year in human history. Average
surface temperatures rose nearly 0.2°C
above the previous record, set in 2016,
to 1.48°C over preindustrial levels, the
European Union’s Copernicus Climate
Change Service reported this week. Only
Australia was spared record-setting heat.
The extreme conditions are a “dramatic tes-
timony of how far we now are from the cli-
mate in which our civilization developed,”
said Carlo Buontempo, Copernicus’s climate
director, in a statement.
Yet 2023’s record temperatures—
confirmed days later by analyses from
NASA, the U.S. National Oceanic and At-
mospheric Administration, the United
Kingdom’s Met Office, and Berkeley Earth—
come with a mystery. Humanity’s unabated
burning of fossil fuels is the dominant
driver of the long-term trend, but it is in-
sufficient to explain 2023’s sudden spike,
says Michael Diamond, an atmospheric sci-
entist at Florida State University.
One exacerbating factor was the end of a
La Niña climate pattern, which from 2020
to 2022 stirred up an increased amount of
deep cold water in the eastern Pacific Ocean
that absorbed heat and suppressed global
temperatures. In 2023, the pattern flipped
into an El Niño event, which blanketed the
134 12 JANUARY 2024 • VOL 383 ISSUE 6679
equatorial Pacific with warm waters and be-
gan to boost global temperatures.
But the flip is not enough to explain
2023’s record, Gavin Schmidt, director of
NASA’s Goddard Institute for Space Studies,
wrote in a blog post last week. Typically, El
Niño plays a larger role in global tempera-
tures the year after it starts—in this case,
this year. And in 2023 heat surged far from
El Niño’s influence, above the northern At-
lantic and Pacific oceans, Schmidt noted.
The 2022 eruption of Hunga Tonga–
Hunga Ha‘apai, a volcano in the south
Pacific, had been a suspect in the global
temperature jump because of the vast
amounts of climate-warming water vapor
it injected into the stratosphere. But early
studies neglected the sulfate particles it also
sent into the upper atmosphere, which re-
flected light and canceled out the water va-
por’s warming effect, says Mark Schoeberl,
an atmospheric scientist at the Science and
Technology Corporation. “For 2022, it was
a nonevent. I have continued my computa-
tions into 2023—still a nonevent.”
Perhaps the best explanation for the
extra warming is the continued drop in
light-blocking pollution as society shifts to
cleaner sources of energy, says Tianle Yuan,
an atmospheric physicist at NASA’s God-
dard Space Flight Center. In 2022, satellites
began to detect this decrease from space
(Science, 22 July 2022, p. 353). In 2020, new
regulations from the International Mari-
A church in Sancourt,
France, is pictured at sunset
during a July 2023 heat wave.
p
g
time Organization added to the effect when
y
ships began to cut sulfur pollution—and
inadvertently curbed the light-reflecting
clouds that the sulfur particles help create
(Science, 4 August 2023, p. 467). A preprint
on Research Square suggests the loss of
these clouds alone can explain half of the
increase in the warming rate seen so far this
decade, says Yuan, who led the work. “[It]
would not account for all the warming we
see this year, but it would represent a sig-
y g
nificant additional warming.”
In a November 2023 paper, famed climate
scientist James Hansen suggested curbing
pollution has accelerated warming to 0.27°C
per decade, up from the 0.18°C per decade
y
rate experienced from 1970 to 2010. But the
acceleration has yet to show up in records
,
of heat in the ocean depths, which resist the
short-term fluctuations of the atmosphere
and offer a truer sense of long-term trends.
The mystery of the past year leaves pro-
jections for this year less certain than usual.
El Niño may inflate temperatures further,
pushing the world briefly past the arbitrary
1.5°C “limit” settled on by policymakers in
2015’s Paris agreement to protect small is-
land nations from extreme sea-level rise.
But extreme heat will again have to develop
over the northern oceans for the world to
breach the threshold—hardly a sure bet.
Regardless, the long-term warming pat-
tern is certain to continue, as it has for
decades—until fossil fuel burning ends. j
science.org SCIENCE

EARTH SCIENCE
U.S. undersea mapping is a boon for science
New maps of continental shelf could vastly expand U.S. territory and resource claims
By David Malakoff
T
he United States has unveiled the
results of a monumental undersea
mapping effort that could add 1 mil-
lion square kilometers of sea floor—
twice the area of California—to its
territory. In addition to enabling the
U.S. to claim valuable geological and bio-
logical resources, particularly in the Arctic,
the project has produced a wealth of sea-
floor data that are fueling a wide range of
scientific advances.
“It’s just a treasure trove of informa-
tion,” says marine mapping specialist
Derek Sowers of the Ocean Exploration
Trust, who has used the data to identify
deep-sea habitats that could be rich in
biodiversity. The maps have also shed
light on the geologic evolution of ocean
basins, identified areas at risk of pro-
ducing submarine mudslides that could
trigger tsunamis, and pointed to poten-
tial seafloor mineral deposits. “There
certainly are a lot of other potential
applications,” says geoscientist Larry
Mayer of the University of New Hamp-
shire, one leader of the effort.
The impetus to make the maps,
which were released last month by the
U.S. Department of State, came from
the United Nations Convention on the
Law of the Sea. One provision of that
1982 pact gives a coastal nation the
right to claim sea floor that sits out-
side its exclusive economic zone, which
stretches 200 nautical miles offshore, if
it can demonstrate that the territory is
a “natural prolongation” of its continen-
tal shelf. Defining this extended conti-
nental shelf (ECS) involves measuring
the thickness of marine sediments,
which are typically thicker on the continental
slope, where the shelf descends to the abyss,
and identifying the “foot” of the slope.
Two decades ago, many nations launched
seafloor mapping efforts in order to submit
an ECS claim to the U.N. (Science, 6 Decem-
ber 2002, p. 1877). In 2003, the U.S. joined the
GRAPHIC: M. HERSHER/SCIENCE
rush even though the Senate had not rati-
fied the pact, which Republicans opposed on
grounds of national security. Mayer and other
scientists argued that emerging technologies,
especially multibeam sonars that produce de-
tailed 3D images, offered an unprecedented
opportunity to document little known sea
SCIENCE science.org
floors off the U.S. “We have much better maps
of Mars than we do of most of our continen-
tal margin,” Mayer told Science at the time.
Federal agencies, including the National
Oceanic and Atmospheric Administration
and the U.S. Geological Survey (USGS), ul-
timately spent tens of millions of dollars on
nearly 50 mapping cruises, some held jointly
with Canada. The ships completed multi-
beam surveys of 3 million square kilometers
of sea floor and conducted acoustic surveys,
which use sound to map sediments, along
nearly 30,000 linear kilometers.
Pushing the boundaries
Newly charted extensions to the U.S. continental shelf hold
resources and scientific riches.
Current maritime boundaries Added shelf territory
Arctic
tic Ocean
Bering Sea
United States
Atlantic
Ocean
O Alaska is helping settle debates over
Pacific Ocean
0 2000
Gulf of Mexico
km
After analyzing terabytes of data, research-
ers identified ECS additions in seven areas
(see map, above). By far the largest is in the
Arctic, where the U.S. is claiming 520,400
square kilometers, about the area of Spain. In
the Atlantic, the claim totals 239,100 square
kilometers. The third largest chunk—176,300
square kilometers—is in the Bering Sea. In
places, the U.S. claims overlap those made by
Canada, Japan, and the Bahamas; diplomatic
negotiations will resolve those boundaries
(Science, 21 June 2019, p. 1120).
Though vast, the U.S. claims are fairly con-
servative from a technical and legal perspec-
N E WS
tive and aren’t likely to draw fierce objections,
says David Mosher, an emeritus geoscientist
at Natural Resources Canada. He helped with
the U.S. mapping effort and is also a member
of the U.N. commission that evaluates ECS
claims. For now, however, the U.S. isn’t sub-
mitting its maps to the commission, because
the Senate still hasn’t ratified the Law of the
Sea and some nations are likely to argue that
the U.S. is not entitled to have the commis-
sion vet its claims. Mosher adds that even
if the commission agreed to review the U.S.
p
maps, it could take decades to evaluate them.
Meanwhile researchers are feasting
on the data. For his 2020 doctoral dis-
sertation, for example, Sowers used an
algorithm to sort seafloor features along
the U.S. Atlantic coast into habitat types,
including mounds and seamounts that
g
may host deep water sponges and cor-
als. Such analyses, he notes, could help
U.S. officials identify rich seafloor eco-
systems that need protection.
y
Other Atlantic shelf research re-
vealed a surprising number of subma-
rine landslides. It even provided the
first complete picture of a massive
375-kilometer-long slide off North Car-
olina’s Cape Fear, which tumbled down
the continental slope 10,000 to 27,000
years ago.
Deborah Hutchinson, a marine geo-
physicist at USGS who helped lead
y g
several cruises, says the mapping off
how shifting tectonic plates formed
parts of the Arctic Ocean. “It was a huge
breakthrough to get ships into the Arc-
y
tic,” she says, noting that U.S. funding
for such cruises has been scarce. And in
,
the Bering Sea, seismic and other data
confirmed that parts of the sea floor
are composed of oceanic and not continental
crust. That “helped eliminate some hypothe-
ses about its formation,” marine geophysicist
Gail Christeson, who is a program director
for marine geology and geophysics at the Na-
tional Science Foundation.
Scientists from many fields also “piggy-
backed” on the mapping voyages, obtaining ev-
erything from seafloor rock and mud samples
to one of the first bits of icelike gas hydrate—
a potential energy source—recovered from
the Arctic. “The U.S. investment in ECS map-
ping,” Hutchinson says, “produced much
more than just maps.” j
12 JANUARY 2024 • VOL 383 ISSUE 6679 135
NE WS | I N D E P T H
By sampling helium gases in hundreds of springs across southern Tibet, researchers identified places where mantle rocks were closer to the surface.
GEOPHYSICS
Tectonic plate under Tibet may be splitting in two
Peeled-apart Indian Plate could be affecting earthquake hazards
By Maya Wei-Haas
T
he towering peaks of the Himalayas
are a geologic battleground—a slow-
motion collision of the Indian and
Eurasian tectonic plates. The crunch
began some 60 million years ago as
India, then an island, plowed into
Eurasia, buckling the surface and form-
ing the highest mountains on Earth. But
the peaks are just the noise and smoke of
the battle; tectonic maneuvers tens of kilo-
meters below them drive the clash—and
hold mysteries.
Continental tectonic plates, unlike their
dense oceanic cousins, are thick and buoy-
ant, so they don’t easily sink, or subduct,
into the mantle during collisions. Some sci-
entists believe the Indian Plate resists plung-
ing into the mantle and continues to slide
horizontally under Tibet. Others suggest the
most buoyant part of the Indian Plate rum-
ples like a rug along the front edge of the
collision, making it easier for the lower half
of the plate to sink into the depths.
136 12 JANUARY 2024 • VOL 383 ISSUE 6679
Now, a new analysis of earthquake waves
traveling beneath Tibet and telltale gases
rising to the surface points to yet another
possibility, one that in effect splits the dif-
ference between the two scenarios. Part of
the Indian Plate appears to be “delaminat-
ing” as it slides under the Eurasian Plate,
with the dense bottom part peeling away
from the top. The study also finds evidence
for a vertical fracture, or tear, at the bound-
ary between the peeled-apart section of the
slab and its intact neighbor.
“We didn’t know continents could be-
have this way and that is, for solid earth sci-
ence, pretty fundamental,” says Douwe van
Hinsbergen, a geodynamicist at Utrecht
University. The work, presented in De-
cember 2023 at the American Geophysical
Union conference and in a preprint posted
online, could help scientists better under-
stand the formation of the mighty Himala-
yas and perhaps even earthquake hazards
in the region.
Fabio Capitanio, a geodynamicist at
Monash University, cautions that plenty of
p
g
y
uncertainties remain about the process, and
the data are limited. “It’s just a snapshot,” he
y g
says. But the work is an important step to-
ward understanding how our modern land-
scape came to be, he says. “It’s definitely the
type of work that we need to move [forward].”
Scientists have long suggested tectonic
y
plates could unzip like this. The plates are
a layered combination of buoyant crust
,
and more dense upper mantle rock. When
squeezed and thickened, a plate might split
along the weak interface between the lay-
ers. But the process was mostly studied in
the interiors of thick continental plates and
simulated in computer models. “This is the
first time that … it’s been caught in the act
in a downgoing plate,” van Hinsbergen says.
The Himalayan collision is a promising
place to look for a plate being torn apart,
says Peter DeCelles, a geologist at the Uni-
versity of Arizona. Before the smashup be-
gan, the Indian Plate varied in thickness
and composition, which helps explain the
crescent shape of the 2500-kilometer-long
Himalayan front. He compares the ancient
science.org SCIENCE
 plate to a manta ray: a pair of thin wings
of oceanic crust flanking a thick middle of
continental crust. The thin oceanic slabs
readily plunged below the Eurasian Plate
while the thick continental crust plowed
headlong into Eurasia like a battering ram.
The difference in subduction speed
likely tugged and tore the Indian Plate in
multiple directions. Scientists have pro-
posed so many tears in recent years, “it’s
become a cottage industry,” quips study
author Simon Klemperer, a geophysicist at
Stanford University.
Klemperer was intrigued by a zone in
northeastern India, near Bhutan, where the
subduction zone curves, making it a prime
locale for potential tears. “That’s where
things get ugly,” he says. The suspicion
drove him on a multiyear quest for clues to
the subsurface violence.
One line of evidence came from isotope
measurements of helium that burbles up
in Tibetan springs. Klemperer and his col-
leagues, including collaborators in China,
traversed dirt roads and forded streams
to collect samples from some 200 natu-
ral springs across nearly 1000 kilometers
of southern Tibet. Gas samples rich in
helium-3, a light isotope trapped within
Earth at the time of its birth, signal mantle
rocks underfoot, whereas those depleted in
helium-3 likely rise from buried crust.
A stark pattern emerged when they
mapped the springs. To the south of a line
lay springs with crustal signatures, and to
the north lay springs with mantle finger-
prints. The team interpreted the line as the
farthest edge of the intact Indian Plate slid-
A geological battleground
The continental collision of the Indian and Eurasian tectonic plates
has created the Himalayas. New evidence suggests part of the Indian
Plate may be splitting away and plunging into the mantle.
Himalayas
Eurasian Plate
Indian Plate
Underplating
GRAPHIC: N. BURGESS/SCIENCE
Mantle
Delamination
Subduction
SCIENCE science.org
ing under Tibet before plunging into the
mantle. But in one area south of the line,
near the eastern border of Bhutan, a trio of
springs also contained mantle signatures—
a hint that one section of the Indian Plate
might be peeling apart, with hot mantle
rock flowing into the space in between.
Support for this picture came from an
analysis of earthquake waves rippling
across the boundary between crustal and
mantle rock. Waves recorded at hundreds
of seismic stations enabled the researchers
to construct images of structures under-
ground. Two blobs seen in one image seem
to hint at a lower slab of the Indian Plate
detaching from its top.
A more recent analysis based on a dif-
ferent set of earthquake waves points to a
tear on the western edge of the delaminated
slab. West of the proposed break, the bot-
tom of the Indian Plate appears to be some
200 kilometers deep, suggesting it is still
intact; to the east, where the slab splits in
two, mantle rock is flowing in around a
depth of 100 kilometers.
Almost every landmass on Earth was
built from a series of collisions like the Hi-
malayas, says Anne Meltzer, a seismologist
at Lehigh University. So understanding how
continents collide sheds light not only on
our modern landscape, but also the hazards
posed by earthquakes that can occur along
the ancient scars of continental crashes.
Klemperer points out the newly proposed
tear may also be influencing earthquake
hazards in Tibet today. A deep fracture in
the Tibetan Plateau known as the Cona-
Sangri rift overlies the tear—a tantalizing
hint that the tumult in the
Indian Plate’s underbelly
might somehow ripple to
the surface. Although the di-
rect link to quakes remains
uncertain, van Hinsbergen
notes that tears and peel-
ing of tectonic plates could
affect how and where stress
builds, and thus the poten-
tial for quakes.
Continents are complex
to study, Klemperer says, as
the ancient collisions that
built the modern landscape
left behind a network of
overlapping scars. But that’s
what makes the work excit-
ing, he says. “They have this
palimpsest of a billion years
Mantle of history.” He and other sci-
flow
entists are gradually learn-
ing to read it. j
Maya Wei-Haas is a science
journalist in Washington, D.C.
ANIMAL RESEARCH
Massive study
of dog aging
likely to lose
funding
Organizers hope to save
long-running project on
canine aging and longevity
By Mitch Leslie
p
S
cientists who study aging are howling
about the possible demise of one of
the field’s biggest studies, the Dog Ag-
ing Project. The effort has been prob-
ing cognitive and physical aspects of
aging in about 50,000 dogs and is
g
running a clinical trial to test a drug that
may boost their longevity. But organizers
say the project will probably lose funding
this year from the National Institute on Ag-
y
ing (NIA), which has furnished at least 90%
of its annual budget, now about $7 million.
“It is a big loss if this project in dogs
does not continue,” says gerontologist João
Pedro de Magalhães of the University of Bir-
mingham, who notes that large, long-lived
animal models promise valuable insights
into human aging. “It was going to be the
most informative study of aging that was
not done in humans,” says biogerontologist
y g
Steve Austad of the University of Alabama
at Birmingham. (Neither scientist has a
role in the research, but Austad’s 2-year-old
dachshund is a participant.)
Organizers are pessimistic about contin-
y
ued funding because the project received
marginal scores on its grant renewal ap-
,
plication late last year. They are striving to
raise money from other sources and have
launched a petition drive to convince the
director of the National Institutes of Health
to intervene to restore funding. “I’m doing
everything possible to keep [the project]
going in its current form,” says co-director
Daniel Promislow, an evolutionary geneti-
cist at the University of Washington (UW).
Although longevity labs teem with ro-
dents and other small animals, “Dogs are
probably the most powerful model for
studying the biology of aging,” says project
co-director Matt Kaeberlein, a former UW
geroscientist who is now CEO of the Seattle-
based biotech Optispan. Thanks to decades
12 JANUARY 2024 • VOL 383 ISSUE 6679 137
NE WS | I N D E P T H
of veterinary research, scientists know
more about how health changes over time
in dogs than they do in rodents. And unlike
lab-dwelling rodents, dogs are exposed to
the same environments as we are, and they
develop many of the same age-related ail-
ments, including heart disease and demen-
tia. “It’s remarkable how much dog aging is
teaching us about human aging,” says bio-
chemist and geneticist Laura Niedernhofer
of the University of Minnesota, who isn’t
connected to the project.
Other researchers and companies have
joined the pack. The Vaika Project tracked
the health of a group of 103 retired sled
dogs for 5 years until it shut down, and
the Golden Retriever Lifetime Study is
still following more than 3000 members
of that breed. Several companies are also
developing therapies to slow aging in dogs,
and one, San Francisco–based Loyal, has
received a preliminary Food and Drug Ad-
ministration endorsement for its drug—
although the company has not revealed
what’s in the treatment.
The Dog Aging Project began in 2014, but
research didn’t start in earnest until 2018,
when the effort received a 5-year grant for
nearly $29 million from NIA. Dog owners
fill out annual questionnaires on their pets’
health to chart the animals’ physical dete-
rioration. Some dogs get a closer look, fur-
nishing DNA and other samples and going
through tests of mobility and cognition. So
far, scientists have sequenced the genomes
of 1000 dogs and cached 14,000 tissue
samples. Published studies have tracked
dogs’ cognitive decline, gauged their sus-
ceptibility to tumors, and investigated
how their eating schedules affected their
health, among other topics. A clinical trial
is also underway to test whether the drug
rapamycin, which extends life in rodents,
The Dog Aging Project’s chief veterinary officer, Kate Creevy (foreground), examines a dog.
138 12 JANUARY 2024 • VOL 383 ISSUE 6679
also helps dogs live longer. “The project is
really just beginning to hit its stride,” says
its chief veterinary officer, Kate Creevy of
Texas A&M University.
Kaeberlein says that, despite the dis-
ruptions from the COVID-19 pandemic,
the project’s organizers thought they had
achieved enough for NIA to approve a re-
newal of their grant. But, “The reviewer
score wasn’t as positive as we had hoped
for,” Promislow says, almost certainly put-
ting it outside the cutoff for funding. Al-
though the agency hasn’t announced the
grant recipients for this cycle, “It’s very un-
likely that NIA will be able to fund us this
cycle,” he says. An NIA press representa-
tive said the agency does not comment on
grant deliberations.
Dog geneticist Heather Huson of Cornell
University, who isn’t connected to the study,
says such projects need long-term support.
“It takes you 5 years to start accumulating
data.” She knows the challenges firsthand
because she took part in the Vaika Project
before it shut down last year. The project
was a casualty of the war between Russia
and Ukraine, because philanthropies in Rus-
sia provided much of the money, Huson says.
To avoid that fate, the leaders of the Dog
Aging Project are trying to raise money
from other sources to tide the project over
for the next year, and they plan to reapply
for NIA funding in 2025. Promislow and the
other organizers are also creating a chari-
table foundation and hope to raise $40 mil-
lion to $50 million for an endowment that
would ensure continued funding.
The Vaika Project also tried to coax do-
nors to open their wallets for dog aging re-
search but failed, Huson cautions. Yet she
and others are hoping the larger project—
which has suddenly become an underdog—
can be rescued. j
PALEOGENOMICS
Ancient DNA
ties modern
diseases
to ancestry
Among Europeans, risk
of multiple sclerosis rises
with genes from Bronze Age
Yamnaya herders
p
By Andrew Curry
L
ike many diseases, multiple sclerosis
hits some populations harder than
others. For example, Scandinavians
are an estimated 17 times more likely
g
than people from sub-Saharan Africa
to develop the devastating chronic
disease, in which the body’s immune system
attacks nerves. “It’s very puzzling,” says Uni-
y
versity of Copenhagen paleogeneticist Eske
Willerslev. Some have speculated that “Vi-
king genes” or some aspect of Northern Eu-
rope’s diet or environment might boost risk.
New data from ancient skeletons show
part of the answer arose in the Bronze Age.
About 5000 years ago, people from the
steppes near the Black Sea, whom archaeo-
logists call the Yamnaya, moved west across
central and Northern Europe. In Scandi-
y g
navia and parts of Northern Europe, these
mobile cattle herders took just a few centu-
ries to largely replace the sedentary farmers
who had tilled the soil with stone tools for
millennia. The Yamnaya brought with them
y
not just a different way of life, but also genes
linked to a higher risk of multiple sclerosis
,
(MS), DNA from the bones shows.
“With MS, you can see it’s Yamnaya ances-
try, and the Yamnaya are basically Danes,”
says Willerslev, who led the research, pub-
lished this week in Nature. “Everything
fits beautifully.”
It’s not just MS. Willerslev, a pioneer in
the study of DNA from ancient remains, and
co-authors have completed a wide-ranging
genetic survey of early Europeans, from the
earliest modern humans on the continent
45,000 years ago to the later waves of farm-
ers and herders who swept in from the east.
In multiple papers, the team reports analyz-
ing the whole genomes of more than 1000
ancient individuals and comparing them
science.org SCIENCE
 with the UK Biobank, a massive data set of
health histories and DNA from more than
430,000 modern British people. The com-
parisons revealed how the proportion of an-
cestry from ancient populations predicted
differences in disease risk and physical at-
tributes such as height and body mass in
present-day Europeans. “There’s a striking
difference in a range of traits, whether skin
and eye color or dietary and mental health,”
says co-author Evan Irving-Pease, a compu-
tational geneticist at Copenhagen. “Subtle
differences in ancestry have a distinct im-
pact. … The legacy of these groups is still
very present in modern people.”
Because they analyzed ancient samples,
mostly from previously excavated skeletal
remains, the researchers could show when
and where key genetic vari-
ants first appeared and infer
how shifts in ancestry and
selective pressures such as
disease could explain them.
“Disentangling the main
ancestries helped us to de-
tect whether selection pro-
cesses happened, and when,”
says co-author Alba Refoyo
Martínez, a computational
geneticist at Copenhagen.
The studies are “an ex-
emplary tour de force,” says
population geneticist Lluís
Quintana-Murci of the Pas-
teur Institute. “These papers
illustrate how the intricate
interplay between ancient se-
lection and admixture events
has profoundly shaped …
present-day Eurasians.”
Willerslev and his col-
leagues started with a nar-
rower goal: to understand
what happened when hunter-
gatherer groups that had The DNA of Denmark’s Porsmose Man, pierced by arrows around 2600 B.C.E.,
survived the most recent ice came mostly from early farmers, not the herders arriving from the east.
age gave way to farmers, in
a 6000-year period known as the Neolithic
that began around 9000 B.C.E. in Europe.
That was when humans first began to settle
in densely populated villages and live close
to domestic animals. Researchers expected
those conditions would spread diseases, in-
cluding those originating in animals, and
therefore would have favored genetic adap-
tations against infection. “It’s a reasonable
PHOTO: DANISH NATIONAL MUSEUM
theory,” Irving-Pease says.
But, he says, “It’s not borne out by the
data.” The team did find what they suspect
are anti-infection adaptations, many in
a region of the genome called the human
leukocyte antigen complex, which regulates
immune response. But those variants pro-
liferated thousands of years after the first
SCIENCE science.org
farmers arrived, in people with lots of Yam-
naya ancestry. “That’s really surprising,”
says University of Kiel bioarchaeologist
Ben Krause-Kyora. “This classical arms race
between the human genome and bacteria
and viruses isn’t happening until the end of
the Neolithic.”
A recent preprint by members of the
same team suggests a reason: Many dis-
eases themselves were latecomers. Analysis
of pathogen DNA in the ancient bones sug-
gests the incidence of diseases passing from
animals to humans, such as the bubonic
plague and leptospirosis, did not rise until
the time of the first Yamnaya migrations,
even though animals had been domesti-
cated thousands of years earlier. The steppe
herders may have kept animals at higher
densities, or early Neolithic communities
may have been scattered widely, slowing the
spread of disease.
The timing could help explain the link
between Yamnaya ancestry and higher
rates of autoimmune diseases such as MS.
The same genes that bolster the immune
response to pathogens are thought to raise
the risk that a person’s immune system will
turn against their own cells. Those genes
were probably beneficial in the past, but
became problematic in the past century
when the advent of vaccines and sanita-
tion lowered exposure to dangerous patho-
gens. “It’s a compelling illustration of the
hypothesis that the contemporary high
prevalence of autoimmune diseases is, at
least in part, the result of past selection fa-
voring survival against infectious diseases,”
Quintana-Murci says.
Comparing ancient samples with the UK
Biobank’s database, the team was able to see
when other traits appeared in ancient popu-
lations and begin to think about why. “You
can see the impact of selection in real time,”
says Pontus Skoglund, a paleogeneticist at
the Francis Crick Institute. For example, a
gene variant that today brings a higher risk
of diabetes and high cholesterol appears
more frequently in samples from about
12,000 years ago and may have helped ice
age hunter-gatherers weather famine. Later,
ancient farmers contributed other variants,
linked to the ability to thrive on a more veg-
etarian diet, to the genetic makeup of mod-
ern Europeans.
Such knowledge has “impli-
p
cations for modern precision
medicine,” Krause-Kyora says.
“The legacy of ancient popula-
tions could explain why peo-
ple react differently to chronic
or infectious disease.”
The studies also bring an-
g
cient population shifts into
sharper focus. In two of the
Nature papers, researchers
use ancient DNA to track
y
the movement of the first
farmers from Anatolia north
into Europe, where hunter-
gatherers were already living.
Researchers had wondered
whether those encounters,
which happened over millen-
nia, were peaceful. But the
new data show that genes
from the farmers increase
y g
abruptly, in stuttering surges.
In many places, including
Denmark about 5900 years
ago, they completely replace
those of earlier hunter-
y
gatherer populations. The
suddenness and thoroughness of the change
,
“lets us seriously question whether the meet-
ing of populations was one of love and peace
and admixture,” Willerslev says. “At least
in Denmark, it’s a replacement, and a very
abrupt replacement—reminiscent of Europe-
ans entering the Americas, with disease and
violence and devastating effects for the pre-
vious population.”
The clutch of new papers is centered on
Europe. But growing numbers of ancient
DNA samples from Africa and Asia could
make it possible to apply similar approaches
elsewhere. “It would be good to expand to
other regions, and go further back in time,”
Refoyo Martínez says. “There’s still a lot to
look at and study.” j
12 JANUARY 2024 • VOL 383 ISSUE 6679 139
NE WS | I N D E P T H
SEISMOLOGY
Mysterious seismic swarm led up to Japan quake
Researchers study connection of rising fluids and quake flurry to magnitude 7.5 monster
By Dennis Normile
B
ig earthquakes sometimes bring
big surprises. The magnitude
7.5 earthquake that struck Japan on
New Year’s Day, killing more than
160 people, is no exception. Over the
past 3 years, tens of thousands of
small to moderate earthquakes, most barely
noticeable, had rattled the Noto Peninsula,
a finger of land that juts 100 kilometers
into the Sea of Japan from the west coast
of Honshu, Japan’s main island. Such earth-
quake swarms occur throughout the world,
typically in areas of volcanic or geothermal
A 1 January earthquake triggered the collapse of many traditional wooden buildings in Japan.
activity. But the swarms almost always ta-
per off and end with a whimper. The Noto
Peninsula swarm produced a bang.
“I can’t think of another earthquake
swarm globally that preceded such a large
event,” says Zachary Ross, a geophysicist at
the California Institute of Technology. Sci-
entists are now puzzling over the details of
that process and how the swarm may have
led to the 1 January shock. “There are many
questions to be resolved,” says Kyoto Uni-
versity geodesist Takuya Nishimura.
Earthquake swarms typically occur when
heat from deep underground produces high-
pressure fluid—water, gas, or magma—that
permeates fault systems, says University of
Tokyo seismologist Aitaro Kato. The fluid
lubricates faults, resulting in slow slips that
140 12 JANUARY 2024 • VOL 383 ISSUE 6679
produce small earthquakes. Unlike the brief
flurry of small quakes that often precedes a
major temblor, swarms continue rumbling
for days or years without generating an
identifiable main shock.
The ongoing collision of tectonic plates
off Japan’s east coast produces the coun-
try’s most damaging earthquakes, including
the 2011 magnitude 9 Tohoku earthquake
that also triggered a mammoth tsunami
(Science, 18 March 2011, p. 1375). But Japan’s
west coast is also seismically active, with an
intensely fractured and permeable fault
zone. Over the past 20 years, several magni-
tude 6 earthquakes occurred near the Noto
Peninsula, which lies about 320 kilometers
northwest of Tokyo. And Kato says a 2014
modeling study concluded a magnitude
7.6 event was possible.
But seismologists did not expect an
earthquake swarm there because the region
has no volcanic or geothermal activity to
generate high-pressure fluids. Nonetheless,
thousands of quakes each year have rattled
a small part of Noto’s northeastern tip since
November 2020, in an unusually long-lived
swarm. Studies suggest fluid is rising from
the upper mantle through faults. A 2023
study by Nishimura and colleagues found
the land above the swarm had risen by
70 millimeters, suggesting upwelling fluid
was swelling the crust at a depth of about
16 kilometers.
The overwhelming majority of the earth-
quakes in the swarm have been weaker than
magnitude 4. But magnitudes have gradu-
ally increased, Kato says. A magnitude 5.4
earthquake occurred in June 2022 and a
magnitude 6.5 event in May 2023 killed
one person and caused widespread damage.
Nishimura says the deep fault movements
associated with the earthquake swarm
probably added strain and stress to faults
higher up in the crust until they finally rup-
tured in the larger earthquakes.
p
Then came 1 January. It was the stron-
gest temblor to hit the Sea of Japan coast
since 1993. In addition to the deaths, scores
of people remain missing, and upward of
30,000 people are in emergency shelters.
Landslides blocked highways connecting
the peninsula to Japan’s main island, ham-
g
pering rescue and relief efforts. And ongoing
aftershocks threaten further destruction.
How the earthquake might be connected
to the swarm is unclear, Nishimura says.
y
Rising fluids might have greased the fault
that ruptured, for example, or slip from the
swarm could have loaded the fault with
stress. The earthquake epicenter is within
the swarm region, and researchers are now
plotting the distribution of the aftershocks
to determine whether the quake occurred
on faults involved in the swarm. Even if it
began there, the rupture extended far be-
yond the swarm region, both farther along
y g
the Noto Peninsula and undersea toward
Sado Island.
“It is important to learn how to distin-
guish an earthquake swarm that might lead
to a large earthquake from those that do not
y
pose such a threat,” Nishimura says. Some
scientists think that’s ambitious. “I think we
,
should take a wait-and-see attitude,” says
Robert Geller, a seismologist and professor
emeritus at the University of Tokyo.
The earthquake highlighted the vulner-
ability of Japan’s traditional wooden build-
ings. Engineers have long warned that
their heavy tile roofs, which stand up well
to Japan’s frequent typhoons, are prone to
collapse during earthquakes. (Authorities
say collapsing houses caused most of the
casualties.) And they are flammable. A fire
sparked by the quake in the city of Wajima
on Noto’s northern coast spread through a
historic market area in the center of town,
burning some 200 wooden buildings to
the ground. j
science.org SCIENCE

FEATURES
COVID’S COLD COUSINS
PHOTOS: (TOP TO BOTTOM) MATT ELLENBOGEN/LA JOLLA INSTITUTE FOR IMMUNOLOGY; SCOTT CAMAZINE/SCIENCE SOURCE
Four largely ignored coronaviruses circulate in humans without causing
great harm and may portend the future for SARS-CoV-2 By Jon Cohen
O
ver a few weeks in November
1889, a respiratory disease at-
tacked half the residents of St.
Petersburg, Russia, and it soon
began to race through Europe
and the rest of the world. Two
years later, in a spectacularly
detailed book, a British medi-
cal officer, H. Franklin Parsons,
described what was dubbed the “Russian
influenza” epidemic, which raged until
1894. People seemed to spread the disease
before developing symptoms, the young did
not suffer as much as the old, a dry cough
was common among the ill, some had a
“perversion of taste and smell,” and deaths
SCIENCE science.org
The human coronavirus dubbed OC43.
rose. Suspicions ran high that a pathogen
had jumped from an animal into humans.
Sound like COVID-19?
In 2005, scientists in Belgium proposed
that the earlier pandemic’s cause was not
N E WS
p
g
In the wake of COVID-19,
labs have ramped up
studies of four widespread
y
human coronaviruses
that cause common colds.
y g
y
,
an influenza virus, but rather a coronavi-
rus. Three years before their theory was
published, a coronavirus had passed from
an animal to humans, touching off a highly
lethal outbreak of what was called severe
acute respiratory syndrome (SARS). The dis-
ease spread from China and brought new
attention to these once-obscure viruses. The
Belgian team wondered whether something
similar happened in Russia more than a
century ago. Based on molecular clues, they
suggested that the once-deadly virus is still
circulating today, as a coronavirus known
as OC43 that in most people causes nothing
worse than a cold. So far there’s no direct
evidence to back the group’s theory, but two
12 JANUARY 2024 • VOL 383 ISSUE 6679 141
NE WS | F E AT U R E S
other teams soon hope to look at
tissue samples from the late 19th
century to see whether they can the natural reservoir for many of them—to other animals such as camels,
spot when the virus first became cows, or civets before jumping into humans. With SARS-CoV-2, raccoon
a human pathogen. dogs and other mammals sold at a Wuhan, China, market could have been
This upcoming search for
OC43’s roots is part of a flurry
of research, since COVID-19
erupted globally 4 years ago
Natural reservoir host
this month, on it and the three
other coronaviruses that cause
common colds. Long ignored
except by a tiny scientific com-
munity, these pathogens with
clunky, alphanumeric names—
NL63, 229E, and HKU1 are the
other three—are now getting
their due. Some groups are re-
examining how the viruses leapt
from animals to people, in part
to understand how SARS-CoV-2,
COVID-19’s cause, may have
emerged. Studying the four may
also illuminate whether other
coronaviruses discovered in wild
and domesticated animals pose
a threat to humanity. And some
scientists are exploring how im-
mune responses to these four
overlap and interact with the
response to SARS-CoV-2.
The four viruses currently
show up each fall and winter,
accounting for up to 30% of the
colds we endure. But all may
once have caused more seri-
ous disease, suggesting to some
virologists that they offer a hope-
ful glimpse of COVID-19’s future.
“Those four are the model system
of what’s ahead for us,” predicts
Lia van der Hoek, a virologist at
the Amsterdam University Medi-
cal Centers who in 2003 discov-
ered NL63. “SARS-CoV-2 is going
to become a common cold. At
least that’s what we want.”
THE FIRST HUMAN coronavirus was iso-
lated 6 decades ago from the runny
noses of English school boys. In the win-
ter of 1960–61, virologist David Tyrrell,
who ran the Common Cold Unit in the
United Kingdom, and co-workers looked
for viruses in the boys’ handkerchief goop.
When they couldn’t identify any known cold
virus, they inoculated adult volunteers with
extracts from the nasal washings to confirm
that something in the samples caused colds.
Yet nothing from those disease-bearing
samples would grow in standard culture
media. So they turned to an odd culture
system recently developed for certain influ-
enza and adenoviruses: cilia-bearing cells
142 12 JANUARY 2024 • VOL 383 ISSUE 6679
Family history
Coronaviruses have repeatedly spilled over from bats or rodents—
intermediate hosts, although that remains contentious.
COLD CORONAS
Intermediate host
Discovered | genus
? 2003 | alpha
?
KILLER COUSINS
? SARS-CoV-2
from the trachea—the natural habitat for
respiratory viruses—extracted from aborted
fetuses. One sample, dubbed B814, yielded a
new virus. “After considerable initial doubts
we now believe that the B814 strain is a vi-
rus virtually unrelated to any other known
virus of the human respiratory tract,”
Tyrrell and colleagues reported in 1965.
The next year Tyrrell sent samples of B814
to June Almeida, a talented electron micro-
scopist who did not have a university degree,
at St. Thomas’ Hospital in London. She re-
ported back that she had seen similar viral
particles in samples from chickens with in-
fectious bronchitis and mice with hepatitis,
although she had been unable to publish her
observations. “Referees said the
images she produced were just
bad pictures of influenza virus
particles,” wrote Tyrrell in a book
he co-authored, Cold Wars: The
Fight Against the Common Cold.
Her new images of B814
made a convincing case that
the various viruses were a re-
Human virus
lated, unrecognized group.
“So what should we call them?
‘Influenza-like’ seemed a bit
229E feeble, somewhat vague, and
1962 | alpha
probably misleading,” Tyrrell
recalled. But he and Almeida
noticed “a kind of halo sur-
rounding them … and so the
OC43
1966 | beta name coronavirus was born.”
Around the same time, infec-
tious disease specialists Dorothy
p
Hamre and John Procknow at
NL63 the University of Chicago were
conducting their own hunt for
new cold viruses in medical
students there. In 1966, they
reported having grown a virus,
HKU1 designated 229E, from a partici-
g
2004 | beta pant who had a “minor upper
respiratory illness.” They gave
samples to Tyrrell, whose team
intentionally infected people
y
with it and showed, again by
handkerchief counts, that 229E
SARS-CoV-1 caused a mild cold, à la B814.
2003 | beta The two viruses looked identical
under the microscope, but re-
searchers could adapt only 229E
to a cell line—and B814 was lost
to history before any genetic
MERS-CoV
2012 | beta
comparison could take place.
The researchers behind a
y g
long-running cold study at the
U.S. National Institutes of Health
(NIH) in 1967 reported what
would prove to be a clearly dis-
2020 | beta
tinct second coronavirus, OC43.
y
“We advertised to the employees
at the NIH to come by Building
,
7, third floor if you had a cold, and we would
be very delighted to wash out your nasal pas-
sages and collect the fluids,” remembers Ken
McIntosh, then a young medical doctor who
ran the project in the lab of Robert Chanock.
Once again, electron microscopy showed
a virus similar in shape to the one that
causes avian infectious bronchitis. (Initially,
McIntosh could only grow it in the organ cul-
ture medium Tyrrell had used—hence the OC
in the isolate’s name—but it, too, was eventu-
ally adapted to a cell line.)
Yet research on the new viruses languished.
“Working with them was so awkward and dif-
ficult that nobody wanted to do it,” McIntosh
says. By January 2003, only a few hundred
science.org SCIENCE
 studies had appeared about human corona-
viruses, and most who did coronavirus re-
search were interested in those that sickened
animals. “[Coronavirus] people who studied
human medicine were rare,” says Leiden
University virologist Eric Snijder, who recalls
struggling that January to draw scientists to
a meeting he co-organized on nidoviruses,
the order that includes coronaviruses.
Then, in April 2003, researchers reported
that the deadly, atypical pneumonia spread-
ing through China, soon to be called SARS,
was caused by a coronavirus. As the dis-
ease began to sicken people elsewhere and
triggered international alarm, last-minute
registration for the May meeting jumped
David Tyrrell in
1966 searched
for common cold
viruses by squirting
nasal washings
into the noses of
volunteers.
from 130 to 170, and SARS was added to the
schedule. A human coronavirus had finally
caught the wider scientific community’s at-
tention, and two more were soon uncovered.
Van der Hoek found the one she called
NL63 in a nasal sample from a 7-month-
old girl in the Netherlands who recently
had fever, pink eye, and a runny nose. Ron
Fouchier’s lab at nearby Erasmus Medi-
cal Center simultaneously uncovered what
PHOTO: PA IMAGES/ALAMY
seemed to be the same virus, and both team’s
findings appeared online within a few weeks
of each other in the early spring of 2004. Be-
fore the year was out, a team led by clinical
microbiologist Patrick Woo at the Univer-
sity of Hong Kong discovered another hu-
SCIENCE science.org
man coronavirus, HKU1, in a 71-year-old
man who had an unexplained pneumonia.
Both van der Hoek and Woo, now at Na-
tional Chung Hsing University, doubt that
there are more human coronaviruses cir-
culating widely that researchers have yet
to detect. “For years and years and years,
people have screened respiratory samples
… and no other [common cold] coronavirus
has been identified,” van der Hoek says. “I’m
convinced that there are only those four.”
But some veteran coronavirologists are
more circumspect. “How could there only
be four?” asks Susan Weiss at the University
of Pennsylvania, who has studied corona-
viruses for 40 years. “It doesn’t make sense
to me.” The University of Iowa’s Stanley
Perlman, another coronavirus old-timer,
says it’s important to keep looking for new
human ones. “In 2002, we felt we were
finished when we had 229E and OC43,”
Perlman says. “We always get deceived
when we think we’re finished.”
WITHIN A FEW YEARS of the discovery of the
coronavirus that caused SARS, scientists
had mapped out a convincing origin sce-
nario. A virus in civets and raccoon dogs
sold in marketplaces in southern China
matched the one that sickened humans,
and a virus later found in bats looked like
its ancestor. This triggered an international
push to sample bats and other animals for
coronaviruses that might pose threats to
humans, leading to thousands of viral se-
quences being cataloged. Although most
of these coronaviruses have been identi-
fied only by sequencing fragments of their
genome—getting intact viruses that grow
in culture is often difficult—the viral fam-
ily is clearly abundant in many species. And
other mammals appear to be the source of
all known cold-causing coronaviruses.
The Belgian researchers studying the
1890s pandemic, for example, sequenced
the genome of OC43 and found “remark-
able” genetic similarities to a coronavirus
found in cows. Using the estimated muta-
tion rates of the bovine virus and OC43,
they created a molecular clock and calcu-
lated that the two viruses shared a common
ancestor somewhere around 1890. (The
p
range went from 1815 to 1918.) The timing
led the scientists to wonder whether the bo-
vine cousin hopped into humans as a much
more lethal pathogen and over time became
the relatively mild OC43 seen today.
“It seemed like an interesting coincidence
that when we estimated the divergence time
g
of the bovine virus and human OC43 it was
basically spot on the date you would expect
with the Russian flu epidemic,” says Philippe
Lemey of KU Leuven, a co-author of the
y
study, published in the Journal of Virology.
He and his colleagues pointed out that be-
tween 1870 and 1890 an epidemic of pneu-
monia in cows led to “massive culling” of
the animals in industrialized countries. This
provided “ample opportunity for the culling
personnel to come into contact with bovine
respiratory secretions” that could have con-
tained OC43’s precursor, they wrote.
In 2022, a French team published a study
y g
in Microbial Biotechnology reporting “very
preliminary” biological evidence that sup-
ports the OC43 hypothesis: They found
antibodies to the virus in the dental pulp of
World War I soldiers who were alive at the
y
time of the Russian flu and died in battle
in 1914.
,
None of the analyses that link OC43 to
the Russian flu persuade Michael Worobey,
a University of Arizona evolutionary
biologist who has collaborated with Lemey
on high-profile studies about the origin of
SARS-CoV-2. “I see it as extremely unlikely,”
Worobey says. As he argued in a 2014 paper
in the Proceedings of the National Academy
of Sciences, “compelling evidence” ties the
global outbreak to a specific influenza viral
variant—including a study of stored sam-
ples from people born as far back as 1876
that found antibodies to a novel flu virus
from the time of the pandemic.
Worobey is now hoping to resolve the
debate by obtaining archived tissue from
12 JANUARY 2024 • VOL 383 ISSUE 6679 143
NE WS | F E AT U R E S
people seen at a London hospital
around 1890 and looking for linger-
ing genetic sequences of influenza or
coronaviruses. A research group from
Spain has identified “suitable sam-
ples” from that time period as well,
at the Basque Museum of the History
of Medicine and Science. It plans to
probe them soon.
The other common cold corona-
viruses are also thought to have leaped
from animals. Hipposideros bats in
Ghana harbor a relative of 229E, a team
led by virologist Christian Drosten,
now at the Charité University Hospi-
tal of Berlin, reported in 2009. The
researchers estimated that the bat
virus and 229E have a common ances-
tor that dates back to between 1659
and 1803, suggesting that’s the period
when it found a way to humans.
Like the SARS virus, it may have
come via an intermediate species.
Studies, some done by Drosten’s
team, found 229E relatives in healthy
dromedary camels in the Arabian
Peninsula and Africa, firming up the
theory. Drosten’s team also charted a
bat-to-camel-to-human pathway for
the highly lethal coronavirus that
causes Middle East respiratory syn-
drome, which was first recognized
in 2012. A similar scenario also looks
likely for SARS-CoV-2, which some
evidence suggests may have passed
from bats to people via an animal
host such as raccoon dogs or other
susceptible species known to have
been sold in a Wuhan, China, food
market that had the earliest cluster
of COVID-19 cases.
The other two cold coronaviruses
have less certain origins. NL63 has an
ancestor found in tricolored bats in
Maryland. A genetic comparison with
the bat virus suggests NL63 crossed
into humans 563 to 822 years ago,
according to a 2012 estimate in the
Journal of Virology. HKU1 has the murki-
est evolutionary history, but its genetic
sequence clusters close to the murine hepa-
titis virus, suggesting it has a rodent origin.
In a chapter about human coronaviruses
that Drosten and co-authors wrote for Ad-
vances in Virus Research in 2018, they
noted it was “peculiar” that no great apes
other than humans have their own corona-
viruses. “This absence provides further
support to the suspicion that contact with
domestic animals may have been essential
in human acquisition of most or all endemic
CoV,” they concluded. Anastasia Vlasova, a
virologist at Ohio State University, may soon
have further evidence for that theory as
144 12 JANUARY 2024 • VOL 383 ISSUE 6679
Electron microscopist June Almeida (top, in 1963) discovered that
the human virus dubbed B814 (bottom) resembled certain chicken
and mouse viruses in its shape and surrounding “halo.”
she’s leading an effort to find novel corona-
viruses in farmers who handle livestock.
“There probably is fairly frequent trans-
fer of zoonotic coronaviruses into the hu-
man population,” says J. Glenn Morris, an
epidemiologist who heads the Emerging
Pathogens Institute at the University of
Florida. But many then fail to spread fur-
ther, he suspects. Indeed, over the years
Morris, Vlasova, and others have identified
coronaviruses from cows, dogs, cats, and
pigs that appear to have infected people,
then petered out.
To Gregory Gray, an epidemiologist at
the University of Texas Medical Branch who
helped Vlasova uncover a canine corona-
virus that infected a few Malaysians,
humanity is under constant, low-level
siege from the viruses. “I think there
are certainly other animal corona-
viruses circulating that are challeng-
ing human immune systems.”
WHEN SARS-COV-2 BEGAN to gallop
around the world, researchers won-
dered whether our immune memories
of its four milder relatives could blunt
the impact of the ferocious new virus.
All coronaviruses share the same ba-
sic repertoire of proteins, suggesting
immune responses built up over re-
peated exposure to colds might ease
COVID-19. The evidence is mixed.
For one thing, the surface pro-
tein of SARS-CoV-2, called spike,
differs markedly from the ones
p
that stud its cold-causing cousins.
As a result, antibodies to the cold
coronaviruses don’t prevent infec-
tions with SARS-CoV-2 or blunt the
symptoms it causes. A report in the
6 September 2023 issue of Science
Translational Medicine even sug-
g
gests that previous exposure to OC43
might leave people with antibodies to
its spike that can interfere with the
immune system’s attempt to make
y
antibodies against the SARS-CoV-2
surface protein, increasing the risk
of developing the lasting, debilitat-
ing symptoms known as Long Covid.
Yet a bevy of studies early in the
pandemic showed that other immune
memories of the common cold corona-
viruses did help. “It has been well
established that prepandemic, some
people had preexisting immune re-
y g
activity to SARS-CoV-2, and it was
consequential,” says immunologist
Alessandro Sette. His group at the
La Jolla Institute for Immunology is
among several to have reported that,
y
in test tube experiments, T cells from
people who had never been infected
,
by SARS-CoV-2 could sometimes recognize
and destroy other cells infected by the vi-
rus. “We and others have shown that, at
least in some cases, this could be mapped to
similarities between common cold [corona-
virus] sequences and SARS-CoV-2 se-
quences,” Sette says.
Another study found that health care
workers who had T cell responses to certain
coronavirus proteins, besides spike, that are
similar in the cold viruses and SARS-CoV-2
appeared to abort infections with the latter.
Other research documented that household
contacts of people with SARS-CoV-2 had a
lower risk of becoming infected themselves
if they had T cells that reacted to proteins in
science.org SCIENCE

Was the “Russian flu” pandemic from 1889 to 1894, depicted here in the United Kingdom’s Illustrated Police News, actually caused by a coronavirus now linked to mild colds?
the viral capsule of OC43 and HKU1. Immu-
nity to the common cold cousins also seems
to lead to less severe COVID-19, and Sette’s
group showed that it improved responses to
COVID-19 vaccines.
Early in the pandemic, this preexisting
coronavirus immunity might have signifi-
cantly reduced the toll of SARS-CoV-2. But
it may have little importance today, Sette
says, because “the vast majority of the
planet has been exposed to SARS-CoV-2 and
vaccinated against SARS-CoV-2.”
Infectious disease specialist Manish
Sagar of Boston University and co-workers
have flipped this issue on its head, asking
whether immunity to SARS-CoV-2 protects
against the common cold. They looked for
the cold-causing coronaviruses in nasal
swabs of nearly 5000 people who came to
the Boston Medical Center between No-
vember 2020 and October 2021. People
who had prior SARS-CoV-2 infections were
50% less likely to have symptomatic dis-
ease from one of the four, they reported in
a bioRxiv preprint on 24 October 2023. T
ILLUSTRATION: SCIENCE HISTORY IMAGES/ALAMY
cells that targeted two of the internal pro-
teins of OC43, the most frequently found
cold coronavirus in their study, likely ex-
plained the benefit.
But van der Hoek has also examined
the cross-immunity question and come to
a different conclusion. In the fall of 2021
her team began to test respiratory samples
to see whether SARS-CoV-2 affected the
presence of common cold coronaviruses.
After the Netherlands ended its COVID-19
SCIENCE science.org
lockdowns, all four of the common cold
coronaviruses returned, according to her
unpublished analysis. “I don’t think that
SARS-CoV-2 has any effect on their circula-
tion,” she says.
TO VAN DER HOEK, the significance of the
“other” coronaviruses is different: She
thinks they foreshadow the likely future of
SARS-CoV-2. She is struck by how sharply
severe disease and death from SARS-CoV-2
has dropped over the past 4 years, shifting
its status from a widely feared killer to yet
another human coronavirus that, at least in
people under age 65 who have no comor-
bidities, causes little acute harm. Indeed
for many, Long Covid has become more of a
worry than immediate hospitalization.
The early ferocity of the virus has much
to do with the fact that aside from some
possible modest protection from previous
colds, the world population in January 2020
was immunologically blindsided by the new
infection. But van der Hoek suspects that an
evolutionary “trade-off” has also defanged
SARS-CoV-2: As the virus has spread to bil-
lions of people, it may have become less vir-
ulent so it can spread more readily. “When
viruses jump species, they are not adapted
to their hosts at all, and they don’t take into
account that the host must survive for them
to survive,” she says.
Each of the four common cold corona-
viruses, she contends, probably came in
lethally hot and then cooled down. “This
must have happened with all four of them,
p
g
and this is just number five,” she says.
“Coughing in your own bed is bad for trans-
mission of an acute respiratory virus. As
y
soon as they start to adapt to their hosts,
they let those infected people walk on the
streets and go shopping.”
But evolutionary biologist Jemma
Geoghegan at the University of Otago is
skeptical. Geoghegan co-authored a Decem-
ber 2018 article in Nature Reviews Genetics
that calls into question the entrenched idea
that emerging viruses become less viru-
lent to persist. “I think the classic view is
y g
wrong,” says Geoghegan, whose article of-
fers several examples of viruses—including
HIV—that did not weaken over time.
She notes that SARS-CoV-2 begins to
spread before people develop symptoms
y
and often doesn’t even sicken the immuno-
logically naïve—which means there’s little
,
evolutionary pressure for it to become less
virulent. “There’s no selection for this re-
duced virulence/transmission trade-off.”
The procession of SARS-CoV-2 variants
contributes to Geoghegan’s skepticism.
Delta was more virulent than the original
virus that emerged in Wuhan. Omicron, the
next to emerge, took over because it spreads
more quickly, not because it’s milder. There’s
no sign of the supposed trade-off, she says.
So put an asterisk on the notion that
SARS-CoV-2 is heading down an evolution-
ary path to becoming as docile as OC43
and the other cold coronaviruses. “Omicron
still hospitalizes and kills lots of people,”
Geoghegan says. “It’s not there yet.” j
12 JANUARY 2024 • VOL 383 ISSUE 6679 145
 INSIGHTS
PERSPECTIVES
IMMUNOLOGY
Decoding the
language of immunity
Optimized transfer RNA (tRNA) codon
use can speed up antibody generation
By Raymond A. Alvarez1,2 and
Louisa K. James3
A
ntibodies are critical for human
health, providing long-lived and ex-
quisitely specific immunity to patho-
gens. Plasma cells secrete tens of thou-
sands of antibody molecules every
second and sustain this output con-
tinuously for several decades. Plasmablasts
and plasma cells [collectively, antibody-
secreting cells (ASCs)] are generated follow-
ing the activation and differentiation of B
cells, which involves complete reprogram-
ming of the transcriptional machinery and
remodeling of the endoplasmic reticulum
and Golgi apparatus. ASCs have co-opted
the unfolded protein response (UPR), a
146 12 JANUARY 2024 • VOL 383 ISSUE 6679
stress response, to accommodate the excep-
tional rate of protein synthesis and secretion
required. On page 205 of this issue, Giguère
et al. (1) reveal that mRNA encoding anti-
body genes is enriched with codons recog-
nized by modified transfer RNA (tRNA).
This codon optimization is accompanied by
an increase in tRNAs with complementary
modifications, which serves to enhance an-
tibody biosynthesis by ASCs. This has wider
implications for therapeutic protein produc-
tion, vaccine design, and beyond.
Eighteen of the 20 amino acids are en-
coded by two or more synonymous codons.
Organisms display preferential biases in the
selection of synonymous codons, which is
thought to play a role in efficient transla-
tion. Although dependent on several differ-
p
g
y
y g
y
ent factors, codon usage bias is particularly
prominent in highly expressed genes and
,
appears to correlate with the copy number
or abundance of cognate tRNAs. The num-
ber of tRNAs not only differs from species
to species but also can vary even across
tissues and developmental stages in mul-
ticellular organisms. In humans, there are
several hundred tRNA genes with varying
copy numbers spread across the genome.
Furthermore, posttranslational modifica-
tions of tRNAs, including so-called “wobble
modifications,” increase the range of syn-
1Department of Medicine, Icahn School of Medicine
at Mount Sinai, New York, NY, USA. 2Ichor Biologics, LLC, New
York, NY, USA. 3Centre for Immunobiology, Blizard Institute,
Queen Mary University of London, London, UK.
Email: louisa.james@qmul.ac.uk
science.org SCIENCE
 Specific codon usage in mRNAs (blue) and
modification of transfer RNAs (tRNAs; center)
enhance antibody production.
onymous codons that can be translated, for
example, inosine-34 (see the figure). Giguère
et al. found that, relative to other genes, an-
tibody mRNA sequences from both human
and mouse ASCs display a conserved bias
toward codons recognized by inosine-34–
modified tRNA and that ASCs are enriched
with tRNA with inosine modifications. In
a mouse model, reducing the frequency of
inosine-34–dependent codons resulted in
decreased antibody production, confirming
that coordination of inosine-34–dependent
codon bias and inosine tRNA abundance
is a mechanism to enhance antibody bio-
synthesis. Understanding the regulatory
mechanisms governing tRNA modification
requires further research.
This coordination may extend beyond
antibody production, potentially offering
insight into fundamental regulatory mecha-
nisms governing cellular fitness and func-
tion. Adenosine deaminase acting on RNA
(ADAR) enzymes, which are responsible
for catalyzing RNA adenosine-to-inosine
posttranscriptional modifications, have
been associated with a wide range of hu-
man diseases, including cancer and neu-
rological, metabolic, and autoimmune dis-
eases (2, 3). Humans express three ADAR
proteins that are highly conserved across
vertebrates. ADAR1 is essential for mam-
malian development (4). Ablation of Adar1
in mice leads to lethal defects during em-
bryonic development, notably with severe
defects in hematopoiesis (5). In human
embryonic stem cells (hESCs), inhibiting
ADAR1 expression reduces their ability to
differentiate into neurons (6). Analogous
to ASCs, during hematopoiesis, stem cells
rapidly synthesize high levels of proteins
to undergo cellular differentiation and
(re)generate new or damaged tissues and
maintain tissue homeostasis. Recent studies
have suggested that RNA editing is critical
for regulating stem cell fate and function
(7), but how dynamic regulation of tRNA
profiles dictates cell fate decisions, influenc-
ing differentiation versus proliferation, is
unknown. Indeed, the correlations drawn
from antibody production may be relevant
to understanding how tRNA modifications
may orchestrate the delicate balance be-
tween cell differentiation and maintenance
GRAPHIC: N. BURGESS/SCIENCE
of pluripotency, offering new perspectives
on the regulatory networks governing cel-
lular fate determination.
In the case of cancer, the metabolic de-
mands are somewhat analogous to those
observed in ASCs. Increased ADAR1 expres-
sion and A-to-I editing activity have been as-
SCIENCE science.org
sociated with progression in several cancers,
including leukemias and solid tumors (8, 9).
Although increases in ADAR-mediated ino-
sine modifications have been identified, al-
terations in tRNA pool availability and their
impact on cellular translational dynam-
ics remain undefined. The parallels drawn
from antibody production may reveal the
potential relevance of tRNA adaptations in
elevated metabolism and proliferation ob-
Boosting antibody generation
In transfer RNAs (tRNAs), deamination of
adenosine-34 (A34) to generate inosine-34 (I34)
by the enzyme adenosine deaminase acting
on RNA 1 (ADAR1) extends the range of pairing
from U-ended codons to include A- and
C-ended codons on eight different tRNAs.
Anti-codon
tRNA tRNA
C G A A34
G C U
ADAR1 mRNA
Codon
tRNA
C G I I34
G C U G C C G C A
Optimizing translation
I34-dependent codons and inosine-
containing tRNAs are enriched in
antibody-secreting cells (ASCs),
resulting in enhanced
antibody secretion.
l34 dependence Antibodies
Naïve Memory
ASC
B cell B cell
Differentiation
served in aggressive cancers and may also
provide insights into their initiation and
evolution, offering new avenues for under-
standing and therapeutic development tar-
geting aberrant translational processes.
Giguère et al. revealed that codon usage
plays a role in the selection of B cells into
the memory compartment. Following acti-
vation, B cells undergo affinity-based selec-
tion to expand rare clones with the stron-
gest antigen-binding variable regions. By
examining the human postvaccination an-
tibody repertoire, Giguère et al. identified
an enrichment in I34-dependent codons
in the mRNA encoding the antigen binding
domains (variable regions) of plasma cells
and memory B cells relative to the naïve B
cell repertoire. Notably, these insights hold
promise for optimizing protein expression in
the mammalian cell lines used for manufac-
turing recombinant proteins, such as mono-
clonal antibodies for clinical use. Strategies
such as media formulation, selective pres-
sure, and cell and vector engineering have
all been used to increase expression and im-
prove the stability of these cell factories (10,
11). Although these strategies have enhanced
stability and yields, a greater understanding
of the natural process that governs antibody
secretion from ASCs may yield considerable
advances. For instance, designing constructs
that tailor codon usage to cell-specific tRNA
expression may revolutionize recombinant
protein production, especially the synthe-
sis of antibodies. Alternatively, modifica-
tions to the expression profiles of tRNA in
producer cell lines could also be engineered
to complement the requirements of particu-
p
lar monoclonal antibody sequences, enhanc-
ing the efficacy and yields of therapeutic
protein production.
Similarly, understanding how specific
codons interact with the tRNA pool opens
avenues for more efficient design of mRNA-
based therapeutics with enhanced protein
g
expression and potentially vaccine efficacy
and duration of immunity. Indeed, tailoring
mRNA sequences to align with tRNA avail-
ability may enable optimal vaccine designs
y
and fine-tune protein expression systems.
Deciphering the intricate network of signal-
ing pathways and regulatory factors driving
tRNA remodeling in ASCs is crucial; under-
standing how tRNA modifications affect
translational fidelity, efficiency, and protein
folding during antibody synthesis warrants
further investigation.
Whether altered tRNA pools influence
broader aspects of cellular fitness beyond
y g
antibody production remains unknown. The
study by Giguère et al. provides a founda-
tional understanding of tRNA remodeling
and codon usage in humoral immunity,
representing a paradigm shift in our under-
y
standing of translation dynamics. j
,
RE FE REN C ES AN D N OT ES
1. S. Giguère et al., Science 383, 205 (2024).
2. W. Slotkin, K. Nishikura, Genome Med. 5, 105 (2013).
3. B. Song, Y. Shiromoto, M. Minakuchi, K. Nishikura, Wiley
Interdiscip. Rev. RNA 13, 1665 (2022).
4. J. C. Hartner et al., J. Biol. Chem. 279, 4894 (2004).
5. Q. Wang, J. Khillan, P. Gadue, K. Nishikura, Science 290,
1765 (2000).
6. T. Chen et al., Cell Res. 25, 459 (2015).
7. D. Lu, J. Lu, Q. Liu, Q. Zhang, Biomark. Res. 11, 61 (2023).
8. K. Fritzell, L. D. Xu, J. Lagergren, M. Öhman, Semin. Cell
Dev. Biol. 79, 123 (2018).
9. X. Xu, Y. Wang, H. Liang, Curr. Opin. Genet. Dev. 48, 51
(2018).
10. B. Bachhav, J. de Rossi, C. D. Llanos, L. Segatori,
Biotechnol. Bioeng. 120, 2441 (2023).
11. S. Xiao, J. Shiloach, M. J. Betenbaugh, Curr. Opin. Struct.
Biol. 26, 32 (2014).
10.1126/science.adn1067
12 JANUARY 2024 • VOL 383 ISSUE 6679 147
I NS I GHTS | P E R S P E C T I V E S
PHYSICS
Electrons catch light pulses on the fly
Energy exchange between electrons and photons enables ultrafast probing of materials
By Albert Polman1 and
F. Javier García de Abajo2,3
A
lthough free electrons are widely used
as probes to analyze the morphology,
atomic structure, and optical prop-
erties of nanoscale materials using
transmission or scanning electron
microscopy, they can also interact
strongly with light and, thus, sample the
distribution of optical fields in and around
those materials with high spatial resolution
(1–3). Specifically, in photon-induced near-
field electron microscopy (PINEM), intense
light fields bring an electron
into multiple energy states si-
multaneously—a quantum-me-
chanical superposition state—
which are then measured with
an electron spectrometer. On
page 168 of this issue, Yang et al.
(4) report the use of PINEM to
examine the creation of special
light pulses known as solitons
in an optical integrated circuit.
The electron–soliton interac-
tion shapes the electron’s prob-
ability distribution in space
and time, thereby enabling new
ways to probe ultrafast dynam-
ics in matter with an electron
microscope.
In quantum mechanics, a
free-space electron is described
as a wave packet characterized by a spati-
otemporal distribution of probability den-
sity—the likelihood of an electron being pres-
ent at a given position in space as a function
of time. A distinct feature of PINEM is that,
by tailoring the light field, electron–light
interaction enables control over this distri-
bution of the electron probability density.
Starting with an incident electron of energy
E0 and a smooth spatial density distribution
curve, its interaction with light expands the
energy spectrum, creating energy states
(sidebands) that are separated from E0 by
positive and negative multiples of the pho-
ton energy "v. After the initial observation
of PINEM and associated electron sideband
1Center for Nanophotonics, NWO Institute AMOLF,
Amsterdam, Netherlands. 2ICFO-Institut de Ciencies
Fotoniques, The Barcelona Institute of Science and
Technology, Castelldefels, Spain. 3ICREA-Institució Catalana
de Recerca i Estudis Avançats, Barcelona, Spain.
Email: a.polman@amolf.nl
148 12 JANUARY 2024 • VOL 383 ISSUE 6679
spectra (1), a well-defined coherence relation
was established between the electron waves
at different sideband energies for every elec-
tron (2, 3). This coherence has important im-
plications for propagation of the electron be-
cause sidebands with different energies have
different electron velocities, and therefore,
the electron wave packet that is formed by
the superposition of all sidebands undergoes
a spatiotemporal reshaping as the electron
propagates (3). Specifically, the probability
density distribution of each electron takes
the shape of a train of pulses of short dura-
tion compared with the period of the light,
Probing electron–light interaction
A high-energy beam aligned with a microring cavity can probe the ultrafast evolution
of optical soliton pulses. Strong electron–photon interaction tailors the production
of quantum-mechanical electron wave packets in space and time.
Laser light
Continuous-wave
pump direction
Circulating light
e–
Electron beam Microring
which is a few femtoseconds in the visible
regime. PINEM thus pushes the combined
spatiotemporal resolution of electron micro-
scopes to the subfemtosecond and nanom-
eter (sub-fs/nm) regime. Besides reshaping
the electron probability density, the side-
band amplitudes provide a direct measure of
the intensity associated with the optical near
field traversed by the electron. Yang et al.
leveraged this effect to measure the intensity
associated with the generation of solitons.
The creation of strong electron–light in-
teractions in PINEM requires intense opti-
cal fields, which are commonly achieved by
using ultrashort intense laser pulses under
far-field illumination conditions (5). Intense
fields are also achieved by using optical mi-
croring waveguides made of highly transpar-
ent silicon nitride and fabricated on a silicon
chip through lithographic techniques
When light is injected into a microring, it cir-
culates hundreds of times inside the ring cav-
ity, thereby increasing the light intensity to
a high value and enabling PINEM to be per-
formed with a low-power continuous-wave
laser light source (7). This capitalizes on the
ability of integrated optical circuits to com-
bine waveguides, multiplexers, and other op-
tical components to process information en-
tirely using light, rather than using electrons
as in common electronic integrated circuits.
Yang et al. exploited a fascinating property
of some optical materials such as silicon ni-
tride—a relatively strong nonlinear optical
p
response. The speed of light in silicon nitride
depends on light intensity even for moderate
laser powers. Above a certain
power threshold, such nonline-
arity causes light to propagate
in complex ways. This includes
the formation of Kerr solitons
g
(8), light pulses generated from
a continuous laser that circulate
Soliton pulse in the microring cavity.
Yang et al. brought together
y
Photonic chip free electrons and nonlinear
optics by using PINEM with
200-keV electrons to probe Kerr
solitons in a silicon nitride mi-
e– croring of hundreds of microm-
eters in diameter. This estab-
Electron wave lished a disruptive approach to
packet explore electron–light interac-
tions and to shape the free-elec-
tron probability density. The
y g
structure is pumped with a
continuous-wave laser at a wavelength com-
monly used for on-chip telecommunication
applications (1.55 µm). The authors aligned
the electron beam (which travels in vacuum
y
along a straight path) and the ring such that
the electron interacts twice along the cavity
,
circumference, thereby probing the optical
field at two different moments in time (see
the figure). The delay between the two inter-
actions is controlled by the electron beam
distance from the cavity center (i.e., zero de-
lay when the beam is tangent to the ring, and
a maximum delay determined by the ring
diameter when the beam intersects the ring
center). This provides a means to examine
the temporal propagation and phase of light
as it circulates the ring cavity. In particular,
the phase is proportional to the delay, thus
allowing Yang et al. to observe characteristic
(6). Ramsey interference patterns of light as the
distance between the two points of light–
electron interaction is gradually changed.
science.org SCIENCE
 These patterns evolve between constructive
and destructive interference as the delay-in-
duced phase varies from 0 to p. The soliton
signature emerges as a broad background
signal superimposed on the Ramsey pattern
and involves higher light intensities (opti-
cal fields accumulated in the pulse-shaped
soliton) that, consequently, produce high-
er-order sidebands.
Nonlinear optical behavior can also be
observed through the formation of chaotic
light fields, trains of solitons, or individual
solitons, depending on how the input light
wavelength and its power are tuned. These
phenomena are driven by the silicon nitride
nonlinear optical response (5, 8). As the laser
wavelength and power were varied, Yang et
al. could probe the transitions between non-
linear optical regimes because of the strong
sensitivity of PINEM to the light field inten-
sity and its distribution around the ring. The
data of Yang et al. could be explained through
a combination of well-established theoretical
models for both the nonlinear optical re-
sponse of microcavities (9) and the electron–
photon interaction (2, 3). Simulations based
on these models were in good quantitative
agreement with the authors’ experiments.
The interaction of electrons with en-
hanced light fields enabled by the microring
geometry creates opportunities to obtain
previously inaccessible information on light
propagation inside integrated optical cir-
cuits. This includes the degree and spatial
distribution of coherence associated with
nonlinearly generated light pulses such as
solitons. Furthermore, electron–soliton in-
teraction enables a disruptive approach to
shaping the electron probability density in
space and time. Also, solitons circulating
the microring cavity can interact with a
continuous electron beam at a high repeti-
tion rate approaching the terahertz, which
is inaccessible with current ultrafast optics
technology. The resulting electron modula-
tions could be synchronized with the optical
excitation, presenting new ways to perform
electron microscopy and probe ultrafast dy-
namics in material systems with sub-fs/nm
spatiotemporal resolution. j
GRAPHIC: F. PETZSCHNER, ADAPTED BY K. HOLOSKI/SCIENCE
RE F ER E NC ES AND NOTES
1. B. Barwick et al., Nature 462, 902 (2009).
2. F. J. García de Abajo et al., Nano Lett. 10, 1859 (2010).
3. A. Feist et al., Nature 521, 200 (2015).
4. Y. Yang et al., Science 383, 168 (2024).
5. M. Liebtrau et al., Light Sci. Appl. 10, 82 (2021).
6. Y. Liu et al., Science 376, 1309 (2022).
7. J.-W. Henke et al., Nature 600, 653 (2021).
8. T. Herr et al., Nat. Photonics 8, 145 (2014).
9. L. A. Lugiato, R. Lefever, Phys. Rev. Lett. 58, 2209 (1987).
AC KN OWL EDG M ENTS
The authors receive support from European Research Council
(ERC) 789104/eNANO; ERC 01019932/QEWS; and European
Commission FET-Proactive 101017720/EBEAM.
10.1126/science.adn1876
SCIENCE science.org
MEDICINE
Practical challenges for
precision medicine
The prediction of individual treatment responses
with machine learning faces hurdles
By Frederike H. Petzschner
P
recision medicine promises treat-
ments tailored to individual patient
profiles. Machine learning models
have been heralded as the tools to ac-
celerate precision medicine by sifting
through large amounts of complex
data to pinpoint the genetic, sociodemo-
graphic, or biological markers that predict
the right treatment for the right person at
the right time. However, the initial enthu-
siasm for these advanced predictive tools is
now facing a sobering reality check. On page
164 of this issue, Chekroud et al. (1) show that
machine learning models that predict treat-
ment response to antipsychotic medication
among individuals with schizophrenia in one
clinical trial failed to generalize to data from
new, unseen clinical trials. The findings not
only highlight the necessity for more strin-
gent methodological standards for machine
learning approaches but also require reexam-
ination of the practical challenges that preci-
sion medicine is facing.
What predicts whether a patient will ben-
efit from a particular treatment? The answer
may lie in their genetics, biology, sociode-
mographic background, social environment,
past experiences, or a myriad of other po-
tential factors. Machine learning techniques
Individual treatment prediction using machine learning
Supervised machine learning for individual treatment prediction is based on the development of classifiers.
To prevent overfitting, model validation is crucial, typically achieved through cross-validation or out-of-sample
validation. Out-of-sample validation requires a completely independent dataset and is more resource-intensive,
but this approach is less susceptible to overfitting and can provide more generalizable results.
Dataset Training Test
Cross-validation
Independent
dataset
Out-of-sample
validation
have the capacity to analyze large datasets
and identify the most effective combination
of features that accurately predicts a vari-
able of interest. They thus offer a promising
avenue for discovering relevant features or
biomarkers that predict individual treatment
responses. Typically, this involves training the
model on a dataset for which the outcome,
p
such as the response to a given treatment, is
already known. This is known as supervised
learning. One common pitfall of this method
is overfitting. Overfitting occurs when a
model is too flexible relative to the data it is
trained on, which limits its generalizability.
A sign of overfitting is when the model ac-
g
curately predicts outcomes on the data it was
trained on but performs poorly on new, un-
seen data. To address the issue of overfitting,
it is essential to validate models on unseen
y
data. Cross-validation is a widely used tech-
nique for this purpose. It involves repeatedly
dividing the data into subsets, training the
model on one subset, and then evaluating its
prediction accuracy on the remaining “held-
out” data (see the figure).
However, cross-validation is not infallible.
Chekroud et al. revealed that models trained
to predict responses to antipsychotic medica-
tion in schizophrenia within a specific clinical
y g
trial using cross-validation failed to predict
treatment responses in other independent
y
,
Training
Training data from single patients Classification
“Responder”
“Nonresponder”
Validation
“Responder?”
12 JANUARY 2024 • VOL 383 ISSUE 6679 149
I NS I GHTS | P E R S P E C T I V E S
clinical trials. One reason that cross-valida-
tion can inadvertently result in overfitting
the held-out data is that the modeler, through
iterative model adjustments, may eventually
use all the available data. The issue is likely
more widespread than typically acknowl-
edged. For example, a comprehensive review
of 116 studies across various psychiatric diag-
noses found signs of overfitting specifically
in studies with small sample sizes (<50 par-
ticipants) (2). Small sample sizes also cause
large variance in cross-validation results, and
although these issues are well known in sta-
tistics and machine learning, many studies
still do not follow best practices to improve
the outcomes of cross-validation (3).
A reliable way to assure the generalizability
of machine learning models lies in validating
their predictive accuracy on a truly indepen-
dent, untouched validation sample, known
as out-of-sample validation. Often, this ap-
proach is not used in clinical studies owing
to the challenges associated with acquiring
larger datasets and the need for stringent
rules governing data acquisition and usage.
However, the study by Chekroud et al. adds to
a growing body of evidence that underscores
the necessity of these more robust validation
standards to avoid overly optimistic results
from machine learning models that fail to
generalize to wider clinical contexts.
Even with models that are properly vali-
dated and supported by large sample sizes,
attempts to predict the clinical outcome or
treatment response for individual patients
can be unreliable. In the study by Chekroud
et al., even when data from multiple clini-
cal trials were pooled to train the model, its
predictions still failed to generalize to a new
independent trial. The reasons for this are
complex and multifaceted. A primary factor
is the inherent heterogeneity in data from
clinical populations. This issue is particularly
prominent in psychiatric disorders, which
are typically defined by sets of symptoms
(syndromes). Patients with the same diagnos-
tic label may exhibit vastly different symptom
profiles that warrant different treatments.
Moreover, identical symptoms in different
individuals might have distinct biological
underpinnings and thus require different
therapeutic strategies (4). Basing machine
learning models purely on diagnostic labels
without taking this type of heterogeneity into
account can lead to inaccuracies when pre-
dicting effective treatment strategies.
A promising approach to address this chal-
lenge is to stratify patients into more pre-
cisely defined categories, for example, based
on underlying symptom causes. This can be
achieved, in part, through the use of theory-
Robert J. and Nancy D. Carney Institute for Brain
Science, Brown University, Providence, RI, USA.
Email: frederike_petzschner@brown.edu
150 12 JANUARY 2024 • VOL 383 ISSUE 6679
driven computational models that aim to
describe underlying disease mechanisms, a
method gaining traction in the field of com-
putational psychiatry. These models are in-
creasingly being used alongside data-driven
machine learning techniques, forming pow-
erful tools to tackle the issue of heterogeneity
in patient populations (5, 6).
Another form of heterogeneity may stem
from systematic differences across studies,
locations, or time points. As a result, predic-
tions of machine learning models trained on
data from a specific context—a population,
country, setting, or time period—might rely
on features that are associated but not caus-
ally related with a clinical outcome in a given
study but are not predictive in other contexts.
One way to address this heterogeneity is to
pool data across multiple studies and sites.
Unreliable predictions may also be the
result of outdated outcome measures. Many
existing symptom scores are based on ques-
tionnaires that may no longer align with
understanding of the disease and potentially
lead to inaccurate assessments of treatment
response. For example, the positive and nega-
tive syndrome scale (PANSS) used in the
clinical trials from Chekroud et al. is gradu-
ally being supplanted by more contemporary
assessment tools, specifically in the context
of negative symptoms in schizophrenia (7).
If a questionnaire fails to fully capture the
true disease burden, it might not accurately
detect genuine improvements resulting from
treatment. This discrepancy can lead to mis-
classification of who has or has not benefited
from the treatment, which hinders the accu-
rate training of the machine learning model.
Similar to the heterogeneity within diagnos-
tic categories, outcome measures will become
more accurate with increasing insight into
the underlying disease mechanism.
The challenges of using machine learning
to predict individual treatment response in
medicine, specifically in the context of psy-
chiatry, stem from a complex interplay of is-
sues related to model validation standards,
diagnostic heterogeneity, and the relevance
of outcome measures used. Addressing these
challenges is essential for impactful clinical
research and to enable progression toward
effective precision medicine. j
R EFER ENC ES A ND N OT ES
1. A. M. Chekroud et al., Science 383, 164 (2024).
2. T. Wolfers et al., Neurosci. Biobehav. Rev. 57, 328 (2015).
3. G. Varoquaux et al., Neuroimage 145 (Pt B), 166 (2017).
4. K. E. Stephan et al., Lancet Psychiatry 3, 77 (2016).
5. M. P. Paulus, W. K. Thompson, Psychopharmacology 238,
1231 (2021).
6. Q. J. M. Huys et al., Nat. Neurosci. 19, 404 (2016).
7. B. Kirkpatrick et al., Schizophr. Bull. 32, 214 (2006).
ACK NOWL EDG M E N TS
F.H.P. was supported by the Brainstorm Program at the
Robert J. and Nancy D. Carney Institute for Brain Science.
10.1126/science.adm9218
MATERIALS SCIENCE
The bumpy
road to friction
control
The frictional properties
of material interfaces can
be rationally designed
By Viacheslav Slesarenko1,2 and
Lars Pastewka1,2
F
riction controls daily life, often with-
p
out being noticed. It allows walking
without slipping, holds sandcastles
together, and determines the per-
ceived cleanliness of hair. Little resis-
tance is desired when pedaling bikes,
yet the expectation of pulling the brakes is
to stop moving. Overall, machines use 20%
g
of the world’s energy production to over-
come frictional resistance (1). Present-day
strategies to tune friction, derived from
more than a century of engineering in-
y
sights, often involve the lubrication of in-
terfaces with oils or greases. On page 200
of this issue, Aymard et al. (2) report an
alternative strategy of rationally design-
ing the frictional properties of interfaces.
Their approach to friction control may
lead to the development of surfaces that
adapt to the environment in real time.
Aymard et al. show that small bumps of
identical radii (3) constitute simple build-
y g
ing blocks that can be combined into a
frictional metainterface. By using many
such bumps on a surface and adjust-
ing their height distribution, the authors
could prescribe a desired, even nonlinear,
y
dependence of the frictional force that re-
sists sliding motion on the external load
,
that pushes the sliding interfaces together.
The effect of surface topography on
friction has long been known. Charles-
Augustin Coulomb, one of the founders of
tribology (the science of friction), wrote in
1779 about the interlocking of asperities
(4), the name given to “bumps” on rough
surfaces. Surface topography determines
the amount of actual contact that two
bodies make. Thus, two bodies typically
1Department of Microsystems Engineering, University of
Freiburg, 79110 Freiburg, Germany. 2Cluster of Excellence
livMatS, Freiburg Center for Interactive Materials and
Bioinspired Technologies, University of Freiburg, 79110
Freiburg, Germany. Email: viacheslav.slesarenko@livmats.
uni-freiburg.de; lars.pastewka@imtek.uni-freiburg.de
science.org SCIENCE
 Metamaterials hint at the future of frictional metainterfaces exploited for designing their metainter-
Metamaterials progressed from simple repetition of identical unit cells to the assembly of various cells faces. Classical metamaterials can be con-
with individually reprogrammable geometries and properties (top). Frictional metainterfaces evolved from sidered frozen in time and space, unable
simple textures to surfaces with combinatorial topography (bottom). By making each individual bump to change behavior after fabrication. By
reprogrammable, adaptive frictional metainterfaces are within reach. contrast, building unit cells with geom-
etries that can be altered makes it possible
Mechanical metamaterials
to encode not one but multiple structure-
property relationships in the same materi-
als and switch between these relationships
on demand (9). Such metamaterials might
rely on buckling, stimuli-responsive mate-
rials or on electromechanical actuators for
Classical Combinatorial Reprogrammable triggering reprogramming to adjust me-
chanical behavior (10, 11).
Embedding such active elements into
Classical Combinatorial Reprogrammable metainterfaces would enable control of
friction and facilitate frictional adaptivity.
Indeed, simple forms of such reconfigu-
rable interfaces have already been real-
Frictional ized to control wettability (12). Frictional
metainterfaces

touch only on the highest peaks of their
topographies. The contact area of random
interfaces increases proportional to the ap-
plied load because the surfaces deform to
conform (5). Friction force is typically pro-
area versus normal load, tuning the fric-
tional response. By this controlled sepa-
ration of the small-scale and bump-scale
response, Aymard et al. could “program”
the frictional response of a metainterface.
p
adaptivity would have numerous applica-
tions, such as touch displays with haptic
feedback. Current attempts to control fric-
tion center around electrochemistry (13)
or electroadhesion (14). A metainterface
approach would supersede the chemical
route because control of not just simply

portional to contact area, which is why the
friction coefficient—the ratio of friction
force to normal force—is constant.
Random roughness is present on all
The idea of programming material be-
havior through geometry and properties of
discrete building blocks is ingrained in the
field of metamaterials. The first implemen-
g
friction but the whole nonlinear depen-
dency of friction force on normal load
appears within reach. Key engineering
challenges will revolve around reliability

y
natural and man-made interfaces (6), but tation of this concept in mechanics dates because friction is typically accompanied
understanding the influence of roughness back to the 1987 work of the engineering by wear. Miniaturization of active ele-
on friction is complicated scientist Roderic Lakes, ments will also be challenging, but either
because roughness is scale who proposed a material microsystems or stimuli-responsive mate-
free; there is no single “…control of not just composed of repeating rials may offer solutions to this. The road
length scale that stands
out. To exert control, en-
simply friction but identical unit cells and ca-
pable of imitating unusual
toward friction control holds a bumpy, but
bright, future. j
gineers have long explic- the whole nonlinear auxetic behavior that is
RE FE REN CES A ND N OT ES
itly introduced geometric observed in some foams
structures with well-de- dependency of (7). The geometry of the
1. K. Holmberg, A. Erdemir, Friction 5, 263 (2017).
2. A. Aymard, E. Delplanque, D. Dalmas, J. Scheibert,

y g
fined length scales into
interfaces. For example,
friction force on unit cell defines the macro-
scopic behavior of the en-
Science 383, 200 (2024).
3. R. Sahli et al., Proc. Natl. Acad. Sci. U.S.A. 115, 471 (2018).
honing produces trenches,
and laser patterning pro-
normal load appears tire metamaterial, enabling
control over relations be-
4. C. A. Coulomb, Théorie des Machines Simples
(Imprimerie de Huzard-Courzier, 1821).
5. B. N. J. Persson, Phys. Rev. Lett. 87, 116101 (2001).
duces dimples. Although within reach.” tween structure and prop- 6. T. D. B. Jacobs, L. Pastewka, MRS Bull. 47, 1205 (2022).

the effect of these struc-
tures on lubricated interfaces can be mea-
sured, their functional role is debated,
impeding the design of surfaces with de-
sired properties.
The bumpy metainterfaces of Aymard
et al. also explicitly introduce a length
scale, but in a very controlled manner and
for unlubricated contacts. The effects of
roughness, mechanics, and chemistry on
the frictional properties of the individual
GRAPHIC: N. BURGESS/SCIENCE
erties (see the figure).
Over time, mechanical metamaterials
constructed from a single unit cell gradu-
ally morphed into structures that com-
bined unit cells with different geometries
and properties within the single metama-
terial. Just like Aymard et al.’s metasur-
faces that have bumps instead of unit cells,
this enabled the programming of specific
complex mechanical responses through
internal organization. The price for versa-
y
7. R. Lakes, Science 235, 1038 (1987).
8. S. Bonfanti, R. Guerra, F. Font-Clos, D. Rayneau-
,
Kirkhope, S. Zapperi, Nat. Commun. 11, 4162 (2020).
9. X. Xia, C. M. Spadaccini, J. R. Greer, Nat. Rev. Mater. 7,
683 (2022).
10. H. Yasuda et al., Nature 598, 39 (2021).
11. R. H. Lee, E. A. B. Mulder, J. B. Hopkins, Sci. Robot. 7,
eabq7278 (2022).
12. M. Specht, M. Berwind, C. Eberl, Adv. Eng. Mater. 23,
2001037 (2021).
13. F. Bresme, A. A. Kornyshev, S. Perkin, M. Urbakh, Nat.
Mater. 21, 848 (2022).
14. M. Ayyildiz, M. Scaraggi, O. Sirin, C. Basdogan, B. N. J.
Persson, Proc. Natl. Acad. Sci. U.S.A. 115, 12668 (2018).

bump were experimentally characterized tility is a necessity for advanced modeling

(3) and fed into a machine-learned model
that was used to design the aggregate re-
sponse of the metainterface, which consists
of many bumps. The height distribution of
the bumps was then optimized to yield a
desired functional relationship of contact
methods to search for the structure that fa-
cilitates the required behavior. Luckily, this
transition nicely coincided with progress
in machine learning that has become vital
for the efficient inverse design of metama-
terials (8)—a progress that Aymard et al.
AC KN OW LE DG M E N TS
The authors receive funding from the Deutsche
Forschungsgemeinschaft under Germany’s Excellence
Strategy (grant EXC-2193/1–390951807).
10.1126/science.adn1075

SCIENCE science.org 12 JANUARY 2024 • VOL 383 ISSUE 6679 151

I NS I GHTS
P OLICY FORUM
CLIMATE POLICY
Climate risk assessments must
engage with the law
Legal actions determine the allocation and magnitude
of climate-related financial risk exposures
By Thom Wetzer1,2, Rupert Stuart-Smith1,
Arjuna Dibley1,3,4
C
limate-related financial risk is the
dominant frame through which many
companies, investors, and regulators
engage with climate change. We ar-
gue that developments in legal ac-
tion mean that the basis for these
assessments, which focus on physical and
transition risks (1), is no longer accurate.
Accounting for the legal system substan-
tially alters the distribution of climate-re-
lated risk between firms, governments, and
the public. Drawing on analysis of climate
litigation, regulatory enforcement, and
other legal action, we propose a framework
that accounts for how legal action shifts
or amplifies physical and transition risk ex-
posures and creates additional climate risk
exposures. We then preview five qualitative
and quantitative approaches that can be ap-
plied to assess the implications of legal ac-
tion for firms’ climate-related risk exposure.
The financial implications of climate
change are widely recognized. Policy re-
sponses include mandating that firms
evaluate and disclose climate risk and cen-
tral bank stress tests on asset values under
climate change scenarios. Climate-risk as-
sessments underpinning such policies focus
on physical risks—expected financial losses
from climate change impacts on assets—
and transition risks—legislative, regulatory,
technological, market, and reputation-
related drivers of emission reductions and
adaptation that introduce financial risk for
nonaligned firms.
Rising physical and transition risks,
coupled with a perceived lack of urgency
in addressing their drivers, has led climate-
related litigation and regulatory enforce-
ment action to proliferate: More than 2485
climate lawsuits have been filed worldwide
(see the figure) (2). The objectives of the
claims, brought in more than 52 national
jurisdictions, encompass enforcement or
strengthening of emission-reduction com-
mitments and policies; challenging con-
152 12 JANUARY 2024 • VOL 383 ISSUE 6679
struction or operation of high-emissions
assets; and penalizing inadequate due dili-
gence, management, or disclosure of finan-
cial risks, which may alter transition risk.
Other legal actions seek to hold firms finan-
cially responsible for adaptation costs or
losses caused by their emissions. Litigation
can also stymie climate action by challeng-
ing governments’ climate policies or in-
creasing their costs through compensation
claims. An illustrative sample of cases is
summarized in table S1.
Legal developments other than litigation
may also alter firms’ climate risk exposure;
for example, an agreement similar to the
master settlement agreed with the tobacco
industry [US$159 billion paid to US states
through April 2023 (3)] could involve large
financial transfers from emitters to com-
pensate for physical risks and impacts.
Litigation against corporations and fi-
nancial institutions is in its early stages. For
the cases that pose the largest financial risk,
successes have been limited to date, and for
damages claims, most are still pending. This
may change as greater risks from climate
change materialize, case numbers grow, and
evidence of climate change evolves. How-
ever, existing climate risk assessments do
not account for the effect of litigation and
regulatory enforcement action in full.
LEGAL RISK IN CLIMATE RISK
ASSESSMENTS
The relevance of legal action to climate
risk is widely accepted. In 2015, Mark Car-
ney, then governor of the Bank of England,
described “liability risks” as a distinct cli-
mate-risk category, noting that such risks
will “increase as the science and evidence
of climate change hardens” [(4), p. 10].
That abstract acknowledgement has only
barely begun to enter technical analyses,
let alone policy. Regulators are starting to
develop methods that account for climate-
related legal risk, but such approaches re-
main in their infancy (5).
The International Sustainability Standards
Board, established in 2021 to develop inter-
nationally accepted standards on climate
risk, considers legal risk as a part of tran-
sition risk. The Network for Greening the
Financial System (NGFS), a network of
central banks working on climate risk sce-
nario analysis, considers legal risks to be
subsets of physical and transition risk (5).
In contrast to their extensive treatment of
physical and transition risks, these bodies
provide little or no detail on how to evalu-
ate climate-related legal risk, suggesting
that its operationalization is at best pe-
ripheral in practice.
Policy frameworks that recognize cli-
mate-related legal risk as a separate cat-
egory also offer limited methodological
clarity. Mandatory climate risk disclosure
policies for firms in Australia, Canada, the
European Union (EU), and New Zealand
recognize “liability risks” as a distinct
p
category, as do prudential supervision
regimes applying to EU banks and insur-
ers, but they do not explicate how firms
should evaluate such risks. Some jurisdic-
tions’ guidance merely notes that firms
subjected to ongoing climate litigation
should disclose the associated risks when
g
they are likely to affect a company’s finan-
cial performance. In the rare instances for
which detailed guidance is provided, it
takes a narrow view of legal impact. For
y
example, the Bank of England’s Climate
Biennial Exploratory Scenario (CBES), the
first climate stress test to systematically
incorporate legal risk, focused on litiga-
tion impacts on selected insurance claim
payouts (6).
The state of the academic literature is
similarly embryonic. Studies consider cat-
egories of (possible) cases against finan-
cial institutions and associated risks (7),
y g
the role of legal action as a transmission
channel for physical and transition risk
(6), or how litigation and liability concerns
might incentivize adaptation investment
(8). However, existing literature does not
y
assess the implications of legal action for
climate-risk exposures comprehensively.
,
Consequently, existing assessments inaccu-
rately represent the effect of legal action on
the distribution and scale of firms’ climate
risk exposure.
MAGNITUDE AND ALLOCATION
A more systematic account of climate-
related legal risk challenges the domi-
nant logic of climate risk assessment that
equates ownership of assets exposed to
physical and transition risks with financial
exposure to those risks (9). Instead, legal
action may shift risk to other entities and
alter the magnitude of these risks.
Successful litigation would redirect costs
incurred because of climate change (in-
science.org SCIENCE
 Climate-related legal action is on the rise globally
More than 100 climate lawsuits have been filed per year globally since 2015. These lawsuits have varied
objectives, including accelerated state and corporate climate action and accountability for past emissions.
Data include cases filed to November 2023. Rest of world
Climate lawsuits filed by year (globally)
300
250
200
150
100
50
0 manage risks or emissions, or from regula-
2002 2009 2016 2023
“Other” includes all other regions, notably New Zealand, France, Mexico, and Indonesia.
cluding of adaptation measures) to green-
house gas emitters (Lliuya v. RWE and
People of the State of California v. Big Oil).
Such claims would (partially) internalize
firms’ externalities, transferring physical
risk exposure to large emitters, potentially
supplemented by punitive damages. Re-
cent cases sought to pass costs of defend-
ing such lawsuits onto insurers (Aloha
Petroleum v. AIG). Other cases aim to hold
fossil-fuel companies liable for climate
change impacts resulting from misleading
CREDITS: (GRAPHIC) D. AN-PHAM/SCIENCE; (DATA) SABIN CENTER FOR CLIMATE CHANGE LAW’S GLOBAL CLIMATE CHANGE LITIGATION DATABASE
or deceptive actions that stymied climate
policy—for example, by creating uncer-
tainty about climate science (Connecticut
v. Exxon Mobil Corp.). These cases would
internalize costs of deception or political
lobbying. Lawsuits have not yet held cor-
porate emitters liable for impacts of their
activities, but developments in attribution
(10) and social science research improve
their likelihood of success.
Legal action may also insulate firms
from transition risk exposure. Transition
risk assessments typically presume that
climate change mitigation policies intro-
duce costs to asset owners—for example,
by causing asset stranding. However,
firms’ transition risks may shift to govern-
ments if legal entitlements to compensa-
tion exist for policy-induced asset strand-
ing. For example, the Energy Charter
Treaty (ECT), an international agreement
to facilitate energy investment and trade,
includes protections for incumbent en-
ergy investments (11). In 2021, RWE initi-
1Oxford Sustainable Law Programme, Smith School of
Enterprise and the Environment, University of Oxford,
Oxford, UK. 2Faculty of Law, University of Oxford, Oxford,
UK. 3Sustainable Finance Hub, Melbourne Climate Futures,
University of Melbourne, Melbourne, Australia. 4Melbourne
Law School, University of Melbourne, Melbourne, Australia.
Email: thom.wetzer@law.ox.ac.uk
SCIENCE science.org
United States
Climate litigation by jurisdiction as of November 2023
807 1678
Australia (134)
United Kingdom (100)
Brazil (77)
European Union (69)
Germany (54)
Canada (35)
Other (338)
ated arbitration proceedings against the
Dutch government under the ECT (RWE
AG and RWE Eemshaven Holding II BV
v. Kingdom of the Netherlands), seeking
€1.4 billion in compensation for a ban on
operating coal-fired power stations from
2030. In 2022, Rockhopper Exploration
was granted €190 million in compensation
from the Italian government under ECT ar-
bitration proceedings. It is estimated that
oil and gas investors may seek more than
US$340 billion in legal claims because of
governments’ climate policies (11). Existing
contracts (such as power purchase agree-
ments) and creative use of corporate law
[such as strategic insolvency to avoid tort
liability (12)] may also protect parties from
transition risk or liability.
Legal action can raise transition risk by
accelerating mitigation and asset strand-
ing through strengthened firm or state
climate policies, enforcing existing policy,
and translating domestic or international
climate policy into duties for companies
(Milieudefensie v. Shell) (table S1). These
effects will be most pronounced for the
firms least aligned with domestic or inter-
national climate policy. Legal action thus
amplifies transition risks and accelerates
their manifestation—for example, by as-
sets becoming stranded. This increases the
present value of these risks, which are oth-
erwise contingent on uncertain and often
protracted policy implementation, by re-
ducing the effects of discounting.
Individual firms may be exposed to am-
plified risks directly, but the possibility of
legal action can also raise risk perception
and borrowing costs. Successful litigation
against one firm can set precedents or
demonstrate the success of a legal strat-
egy, raising (perceived) risks for similarly
situated firms in the same jurisdiction.
Legal action can also amplify transition
risk by rendering voluntary commitments
legally binding. Signatories of the Glasgow
Financial Alliance for Net Zero cited this as
a reason to loosen their commitments (13).
Legal challenges to government policy or
regulation may affect firms’ transition risk
exposure indirectly by triggering policy
changes (Neubauer et al. v. Germany) such
as reduced fossil fuel subsidies, enhanced
emission standards, or more stringent dis-
closure requirements.
Climate-related legal actions also intro-
duce financial risks that are not directly
related to firms’ underlying transition and
physical risk exposure. These additional
risks derive primarily from obligations to
tory requirements to disclose climate risk
and not to make inadequate or misleading
p
statements about risk management, invest-
ment policies (such as “greenwashing”),
and Paris-alignment of business plans.
Firms could face such liability regardless
of the size of their emissions footprint or
physical or transition risk exposures. US
banks BNY Mellon and Goldman Sachs
g
agreed with the Securities and Exchange
Commission (SEC) to pay US$1.5 million
and US$4 million fines, respectively, after
the SEC concluded that the banks’ commu-
y
nications about their environmental, so-
cial, and governance (ESG) investment pol-
icies were misleading. Similar enforcement
action hit German asset manager DWS, in
which an ongoing investigation resulted in
the resignation of its CEO. Although these
risks are growing in importance, they have
had modest financial consequences for the
firms affected to date.
y g
DEVELOPING ASSESSMENTS
Not accounting for the law in climate risk
evaluations is increasingly untenable.
Nevertheless, a recent survey of central
banks indicated that most struggle with
y
assessments of climate-related legal risk:
93% of respondents did not yet quantify its
,
impact (5).
Firms’ climate-related legal exposures
are contingent on legal action and out-
comes that in turn depend on available
legal causes of action, the likelihoods that
a case is brought and is successful, the rem-
edy sought, and the likelihood of effective
enforcement. Moreover, exposures vary be-
tween actors; high-emitting firms may be
targeted by liability claims and indirectly
experience consequences of challenges to
government policy, whereas banks may be
scrutinized for investment and risk-man-
agement decisions. Uncertainty in each
of these areas may appear to create an in-
tractable constraint on assessments and
12 JANUARY 2024 • VOL 383 ISSUE 6679 153
I NS I GHTS | P O L I C Y F O RU M
quantification of contingent liabilities of
climate-related legal risk. Indirect risk ex-
posures—for example, by holding securi-
ties in a firm targeted by climate litigation
or affected by new legal precedents—fur-
ther complicates assessments.
No single approach can assess legal risk
comprehensively, and the suitability of dif-
ferent methods depends on the source and
nature of risk in question, the particular
legal arrangements of each jurisdiction,
and the nature of the claim and the rem-
edy provided. To illustrate how legal action
can be accounted for in climate risk assess-
ment, we propose five complementary ap-
proaches for evaluating firms’ exposures,
each with its strengths and shortcomings.
Market impacts of legal judgments can
be retrospectively analyzed—for exam-
ple, by assessing stock price movements
around climate litigation events. Such
market movements represent the market’s
best estimate of the financial consequences
of legal action. Analyses that focus on mar-
ket movements sidestep the need to under-
stand how risk materializes and provide
a baseline indication of the impact of cli-
mate litigation on firm value. However, the
approach should be adopted cautiously.
Event studies are context specific, so their
findings may extrapolate poorly to other
firms, jurisdictions, and causes of action,
especially given rapidly evolving climate
litigation strategies, legal precedents, and
societal expectations.
Estimates of climate impacts, based on
the social cost of carbon or on the attri-
bution of specific impacts, can be used to
quantitatively estimate legal risk associ-
ated with efforts to hold defendants liable
for climate change. A back-of-the-envelope
analysis using the social cost of carbon in-
dicates that Chevron’s emissions to 2010
produce a potential liability of more than
US$8.6 trillion (see supplementary mate-
rials). Alternatively, attribution analyses
indicate firms’ contributions to specific
climate-related disasters. Under this ap-
proach, Shell would have an annual liabil-
ity of US$0.55 billion if held liable for their
contribution to two events with attribut-
able losses equal to those of Hurricane
Harvey (see supplementary materials).
Both approaches are rooted in scientific
analysis and legal judgment that relate, re-
spectively, to estimates of the proportion of
a firm’s externality that legal action might
internalize and the number of disasters for
which legal action will hold firms liable.
Punitive damages are also possible; these
estimates may therefore be a lower bound.
The financial impacts of legal action that
mandates accelerated mitigation can be as-
sessed by using sector-specific or technol-
154 12 JANUARY 2024 • VOL 383 ISSUE 6679
ogy-specific marginal abatement cost esti-
mates. This approach estimates the costs
of firms’ transitions with and without spe-
cific legal decisions. The likelihood of suc-
cess of these claims, on which this quan-
titative analysis is contingent, depends on
the existence of obligations to follow a spe-
cific sectoral or firm-level transition path-
way. These have not been explicitly articu-
lated in many countries, but Article 15 of
the EU’s proposed Corporate Sustainability
Due Diligence Directive could require firm-
level transition plans and would harden
legal expectations around the speed of
firms’ transitions. Without explicit obliga-
tions, judicial evaluation of the alignment
of firms’ transition plans with legal duties
depends on judges’ interpretations of na-
tional and international law (14). The judg-
ment in Milieudefensie v. Shell hinged on
new interpretations of firm-level transition
requirements implied by international law,
illustrating that judicial interpretation can
evolve quickly and unpredictably, compli-
cating ex ante risk assessments.
Qualitative assessments of climate-re-
lated legal risk could evaluate implications
of developing judicial doctrine and prec-
edent, legislation, and scientific advances
as well as trends in litigation outcomes to
characterize risks faced by specific firms.
Such analyses elucidate the distribution
and magnitude of climate-related legal
risks across firms and scenarios, facilitat-
ing risk management (5)—for example,
within an investment portfolio—and in-
form and contextualize quantitative as-
sessment approaches.
MOVING FORWARD
The emergence of unexpected new interpre-
tations of open legal norms and the rapid
evolution of legislative practice that alters
rights and obligations means that varia-
tions in climate-related legal risk assess-
ments are to be expected. These differences
can form the backbone of legal transition
scenarios, much in the same way that cli-
mate-risk analysts already use physical and
transition risk scenarios.
Designing legal transition scenarios,
an approach that is new to the legal and
climate risk professions, raises complex
questions regarding individual doctrinal
and legislative developments, how rep-
resentative scenarios can be developed
at the jurisdictional level, and how to in-
tegrate legal and climate transition sce-
narios. Such scenarios would also need to
capture, as with physical and transition
risks, nonlinear changes, including dra-
matic shifts in legal paradigms that could
substantially change risk exposures as
future legal outcomes diverge from those
seen to date. Legal outcomes may also vary
between jurisdictions. Nonetheless, such
an approach would embed legal risk
analysis at the heart of climate transition
scenario analysis. In its CBES exercise,
the Bank of England was the first to ask
insurers to quantify financial impacts of
hypothetical legal cases, an early use of
scenario-based approaches.
Policy-makers, investors, and firm man-
agers have accepted the need to under-
stand climate risk exposures. Doing so dili-
gently will require more engagement with
the law through interdisciplinary research
that couples legal reasoning with financial
analysis and climate science. Else, we will
continue to fly blind in our treatment of cli-
mate risk. j
R E F E R E N C ES A N D N OT ES
1. Task Force on Climate-related Financial Disclosures
p
(TCFD), Recommendations of the Task Force on Climate-
related Financial Disclosures (TCFD, 2017).
2. J. Setzer, C. Higham, “Global trends in climate change lit-
igation: 2023 snapshot” (Grantham Research Institute
on Climate Change and the Environment and Centre for
Climate Change Economics and Policy, London School
of Economics and Political Science, 2023).
3. National Association of Attorneys General (NAAG),
g
Payments to States since Inception through April 20
2023 (NAAG, 2023).
4. M. Carney, “Breaking the tragedy of the horizon—
Climate change and financial stability” (Bank of
England, 2015); https://www.bankofengland.co.uk/-/
y
media/boe/files/speech/2015/breaking-the-tragedy-
of-the-horizon-climate-change-and-financial-stability.
pdf.
5. NGFS, “Report on micro-prudential supervision of
climate-related litigation risks” (NGFS, 2023); https://
www.ngfs.net/sites/default/files/medias/documents/
ngfs_report-on-microprudential-supervision-of-
climate-related-litigation-risks.pdf.
6. Bank of England, “Results of the 2021 Climate
Biennial Exploratory Scenario (CBES)” (Bank of
England, 2022); https://www.bankofengland.co.uk/
stress-testing/2022/results-of-the-2021-climate-
y g
biennial-exploratory-scenario.
7. J. Solana, Green Finance 2, 344 (2020).
8. S. Barker, J. Dellios, E. Mulholland, “Liability risk sand
adaptation finance” (MinterEllison, 2021); https://www.
unepfi.org/wordpress/wp-content/uploads/2021/04/
UNEPFI-Climate-Change-Litigation-Report-Lowres.pdf.
9. G. Semieniuk et al., Nat. Clim. Chang. 12, 532 (2022).
y
10. R. F. Stuart-Smith et al., Nat. Clim. Chang. 11, 651 (2021).
11. K. Tienhaara, R. Thrasher, B. A. Simmons, K. P. Gallagher,
,
Science 376, 701 (2022).
12. J. Macey, J. Salovaara, Stanford Law Rev. 71, 879 (2019).
13. S. Morris, K. Bryan, O. Walker, Financial Times 21
September 2022.
14. B. J. Preston, J. Environ. Law 33, 1 (2021).
AC K N OW L E D G M E N TS
The authors thank O. Bisel and S. Leonard for excellent
research assistance and J. Armour, S. Fankhauser, N. Ranger,
J. Rogelj, and T. Schuermann for valuable feedback on the
manuscript. T.W. and R.S.-S. acknowledge support from
the Oxford Martin School and the Oxford Sustainable Law
Programme, and A.D. acknowledges support from the
Australian Research Council and Melbourne Climate Futures.
SU P P L E M E N TA RY M AT E R I A LS
science.org/doi/10.1126/science.adg0598
10.1126/science.adj0598
science.org SCIENCE

B O OKS et al .
SCIENCE LIVES
A family history of the Universe
Musings on natural phenomena merge with tales from
a science journalist’s personal life
By Mary Ellen Hannibal
T
he voice of Nell Greenfieldboyce may
well already be in your head. The long-
time National Public Radio (NPR) sci-
ence reporter steadily brings listeners
news from deep inside the cell, the
ocean, space, and more. In Transient
and Strange: Notes on the Science of Life, her
essays elucidate tornadoes, meteors, black
holes, and fleas, among other phenomena of
our shared world. She also probes
deeply into her own past and
psyche, interweaving facts with
personal history. In print as in ra-
dio, her voice is wry, charming, in-
formative, and, indeed, sometimes
strange. (The title references Walt
Whitman’s poem “Year of Meteors.”)
The prompting to bring often
lofty science findings down to the Transient and Strange: head in the sand, as we often do
private realm is instigated when
Greenfieldboyce becomes a parent.
PHOTO: PIERRE-PAUL FEYTE/BIOSPHOTO/MINDEN PICTURES
“I suddenly had a new audience. A
much smaller audience—only two,
instead of millions,” she writes, “but with
higher emotional stakes.”
In the book’s opening essay, “The Symbol
of a Tornado,” Greenfieldboyce recounts her
children’s ratcheting fears of being swept
away by a twister, anxieties she does not
exactly quell. These little geniuses (as most
children are) counter her half-hearted reas-
The reviewer is the author of Citizen Scientist: Searching for
Heroes and Hope in an Age of Extinction (The Experiment,
2016). Email: maryellenhannibal@gmail.com
SCIENCE science.org
surances with uncomfortable counterargu-
ments. She suggests that they are relatively
safe in Washington, DC, but her kids find
out that tornadoes have been recorded in all
50 states and the nation’s capital. The whole
family plunges into the details of tornadoes—
what they are, what they do, and where they
strike. Fun facts emerge.
They learn, for example, that Army Signal
Service meteorologist John Finley pioneered
tornado studies, opining that “the tornadic
disturbance is as old as the world
itself, if we are to believe that the
appearance of the atmosphere
was coincident with the creation
of the earth.” Trying to stave
off widespread panic, in 1905
the Weather Bureau Stations
Regulations prohibited forecast-
ing tornadoes at all. Putting one’s
Notes on the
with climate change, for exam-
Science of Life
Nell Greenfieldboyce ple, is nothing new.
Norton, 2024. 224 pp. In “What Else is There?,”
Greenfieldboyce’s hospitalized
father struggles with words. He wants to
know what she is wearing around her neck:
a meteorite. She uses the occasion to delve
into the long history of human interactions
with rocks from space. The Black Stone,
venerated by billions of Muslims and em-
bedded in a wall at Mecca, is considered by
many to be a meteorite. Sacred rocks must
provide some clear indication that they
are special, which is not always evident.
Folklorist Oliver Farrington asserts that
The author’s children’s fascination with tornadoes
prompted a family deep dive.
meteorites are revered in direct relation to
whether they were observed falling or not.
A found meteorite is likely to be treated ca-
sually, as an ordinary rock. The mysteries
and mundanities of physical life continually
shift in our minds.
Greenfieldboyce’s 3-year-old daughter re-
jects her mother’s assertion that she is a part
of the Universe. “No,” her daughter insists,
“I’m not, I am in the universe.” Whitman
himself asks, “What am I myself but one of
your meteors?” Yet of course, like meteors,
not only are we in the Universe but we help
compose it. In her essays, Greenfieldboyce
surfs the dichotomy astutely. We are both
integral and apart.
The question remains: What do our per-
p
sonal foibles, travails, and general experi-
ences have to do with any of this quantifi-
able realm? We are made of the same stuff
and yet so often feel very separate from it
indeed. The essay “A Very Charming Young
Black Hole” attempts to integrate a trou-
bling adolescent experience with, yes, notes
g
on gravitationally collapsed objects.
In her recounting, Greenfieldboyce’s
parents snore obliviously above in the Old
Faithful Inn at Yellowstone while their
y
12-year-old daughter carries on a provoca-
tive flirtation with a mildly drunk college-
age man. She tells us that she lied about her
age to the man, ratcheting up the sexual
banter as if she were three times older and
highly experienced. The reader worries how
the scenario will play out. Invoking black
holes in this narrative seems to suggest that
Greenfieldboyce is still something of a mys-
tery to herself; it is a relief to learn she is
y g
unharmed as it concludes.
“My Eugenics Project” tells the inter-
twining stories of Greenfieldboyce’s at-
tempts to get pregnant and her husband’s
need for a kidney transplant. Here the sci-
y
ence is focused on biological circumstances
that, for all our contemporary mastery of
,
this Universe, remain mysterious and con-
tingent. Her concern that she and her hus-
band will pass his genetic condition on to
the next generation is soon eclipsed when
multiple attempts at in vitro fertilization
fail to produce a viable embryo. Her hus-
band’s seemingly sudden need for dialysis
adds to the pressure: “Our marriage now
had become two simultaneous medical
campaigns, one to produce a kid and one to
procure a kidney.”
Our ordinary lives are often heartrending,
complex, and mostly unknowable. This is the
mystery that Greenfieldboyce celebrates. j
10.1126/science.adm9749
12 JANUARY 2024 • VOL 383 ISSUE 6679 155
I N SI G HTS | B O O K S
ROBOTICS
Envisioning artificial intelligence
Scholars consider a new translation of the prescient play
that coined the term “robot”
By Ed Finn
S
cience and science fiction offer a
shared gift to humanity: the words
to name new things and ideas and,
by naming them, making them real.
The word “robot,” for example, is so
pervasive today that it might be easy
to forget that it was invented by the Czech
playwright Karel Čapek, who coined the
term in his 1920 play R.U.R.: Rossum’s
Universal Robots. Within 2 years of its
first performance in Prague, R.U.R. had
been translated into 30 languages, and
the word “robot,” which was preserved
across various translations, quickly be-
came ubiquitous.
R.U.R. and the Vision of Artificial
Life, a new volume edited by Czech
chemical engineer Jitka Čejková, is wel-
come because it reminds us that words
have histories too, and we ignore them
at our peril. Čapek’s neologism is an
adaptation of a Czech word meaning
“serf” or “servant” that is linked both
to the service obligation that peasants
would perform for feudal landowners
and etymologically to Czech words for
“woman,” “child,” and “orphan.” When
the robot arrived in our shared imagina-
tion of the future, it was already a form
of indentured labor, a figure whose per-
sonhood is marginal, if it exists at all.
The most important contribution of
the volume is a new translation of the
play by Štěpán S. Šimek, a theater direc-
tor and scholar who manages to capture
the surreal weirdness of Čapek’s dark
comedy of errors while making the text
accessible to contemporary audiences.
Šimek presents a faithful rendition of
the entire play, leaving the question of
emendations and adaptations to fu-
ture dramaturges and giving his readers the
chance to understand R.U.R. more or less as
Čapek intended.
R.U.R. is fascinating and bizarre, particu-
larly for readers who might think of robots as
mechanical beings plated in chrome, as they
are in Fritz Lang’s Metropolis, or as cyber-
tronic machines, as they are in Isaac Asimov’s
stories. As many of the essays in this volume
The reviewer is at the Center for Science and the
Imagination, Arizona State University, Tempe, AZ 85287, USA.
Email: edfinn@asu.edu
156 12 JANUARY 2024 • VOL 383 ISSUE 6679
R.U.R. and the Vision
of Artificial Life
Karel Čapek,
translated by Štěpán Šimek,
edited by Jitka Čejková
MIT Press, 2024. 312 pp.
point out, Čapek envisioned the rise of arti-
ficial life not out of silicon but synthetic bi-
ology, and the robots in his play are visually
indistinguishable from humans.
Rossum’s Universal Robots is a family
business built on a secret formula for cre-
ating synthetic workers. The orders keep
flooding in: robots to replace human work-
ers, robots to fight wars, robots to replace
the humans killed in the wars. Meanwhile,
Robot Asimo offers flowers to a bust of playwright Karel Čapek.
human fertility plunges to zero. Predictably,
the robots throw off their chains and erase
humanity from the Earth. The play ends as
strangely as it begins, with two robots depart-
ing like Adam and Eve, offering up the prom-
ise of new life from the ashes of humanity’s
destructive ambition.
Like Mary Shelley’s Frankenstein, R.U.R.
is breathtakingly prescient in anticipating
the central questions of artificial intelligence
and synthetic biology decades before the dis-
covery of DNA or the transistor. The blithe
executives of R.U.R. sound exactly like tech
evangelists today, predicting the end of work
and an age of abundance. The humans acquit
themselves poorly, of course, blinded by greed
and badly constructed moral frameworks.
Čapek also disentangles intelligence from
personhood. One of the play’s best lines is
spoken when Rossum’s CEO points out that
the robots have a phenomenal memory
but never come up with anything original:
“They’d make great college professors,
p
now that I think about it.” The anxieties
of human identity, intelligence, and rel-
evance feel fresh and familiar in the era
of ChatGPT.
Anticipating countless future plots,
Čapek’s characters repeatedly mistake
robots for humans and vice versa. As a
g
comedy, R.U.R. mercilessly undercuts
the moral distinctions its human char-
acters painstakingly assemble to justify
their greed and ambition. The audience
y
gets read in on that irony as the world
outside the factory falls apart, pushing
us to address our profound uncertainty
about the stakes of scientific creativity.
What is life, and what do we owe to our
creations?
The rest of the volume digs into
these questions through an editorial
introduction and more than 150 pages
of supplemental essays. These range in
y g
quality, but all strive to connect R.U.R.
to contemporary science in society. In
an incisive contribution, Jana Horáková,
a professor of new media, reads the
frequent mixing up of humans with ro-
y
bots as a prescient warning of the “ro-
botization of humanity.” In our decades
,
of thickening the layers of computation
and automation girdling the world, we
bend ourselves to the wheel of mecha-
nization as often as we let the machines
liberate us from drudgery.
We desperately want to know ourselves,
and so we create new tools in order to find
the boundaries of agency, selfhood, and mo-
rality. We even imagine personhood in our
tools. Čapek’s masterpiece reminds us first
that just because we can does not mean we
should. But he also tells us what we must do:
care for our creations and for one another
and take responsibility for the full conse-
quences of our technological dreams. j
10.1126/science.adl5502
science.org SCIENCE

LET TERS
Edited by Jennifer Sills
Portugal’s farmland bird
crisis requires action
The outlook for farmland bird conserva-
tion in Portugal is dire. The great bustard
(Otis tarda), little bustard (Tetrax tetrax),
and Montagu’s harrier (Circus pygargus)—
three priority species under the European
Union’s Birds Directive (1)—have declined
by 50 to 80% over the past 10 years (2–4).
One-third of common agricultural birds
are also decreasing (5). Portugal must take
action to reverse these trends.
Reductions in farmland bird populations
are mostly attributable to a nationwide
shift from arable land to specialized live-
stock production (mainly beef ) and perma-
nent crops (e.g., olive groves). The transi-
tion began in 2003 when Portugal stopped
subsidizing cereal production but main-
tained subsidies for livestock within the
context of the European Union’s Common
Agricultural Policy reform (6). In addi-
PHOTO: FREDERIC DESMETTE/BIOSPHOTO/MINDEN PICTURES
tion to the direct loss of open habitat, this
change has resulted in the abandonment
of extensive rotational crop farming—a
matrix of cereals, legumes, and fallow land.
The reduction of landscape diversity and
vegetation complexity depleted the birds’
food supplies (mainly vegetation and inver-
tebrates) and breeding cover (which was
previously more plentiful given the taller
crops and lack of overgrazing). To fulfill
cattle dietary needs, livestock production
also involves growing fodder crops, which
are cut earlier in the season than cereal
SCIENCE science.org
crops, destroying nests and harming both
chicks and adult birds (2, 7).
Between 1999 and 2008, Portugal inte-
grated key bird conservation sites located
in farmland areas into the Natura 2000
protected network. The country also intro-
duced specific voluntary agri-environmental
measures designed to maintain or enhance
key species’ conservation status, along with
financial incentives to encourage imple-
mentation. However, measures intended to
enhance habitats containing cereal crops
(8, 9) are less effective as grazing systems
expand and intensify (6). Meanwhile, rec-
ommended measures to protect nests can-
not compete financially with current agri-
cultural practices (6), resulting in minimal
adherence among farmers.
By failing to reverse the decline of priority
species and manage Natura 2000, Portugal
is in clear noncompliance with the European
Union’s Birds Directive (10). The country
must incentivize farmers to reinstate appro-
priate habitats for farmland birds by widely
implementing agri-environmental measures.
This action can only be achieved through
close collaboration between the Ministries of
Environment and Agriculture and effective
engagement with stakeholders. To judge the
program’s success, monitoring frameworks
should be put in place to oversee agri-
environmental measures and conservation
initiatives, determine the degree to which
farmers are implementing them, gauge their
efficacy, and facilitate prompt adjustments.
Only a joint effort by those responsible for
agriculture and those responsible for conser-
vation can reverse the precipitous decline of
farmland birds in Portugal.
Land use changes in Portugal threaten farmland
birds such as the little bustard (Tetrax tetrax).
João P. Silva1,2,3*, João Gameiro1,2,4, Francesco
Valerio1,2,5, Ana T. Marques1,2,4
1Research Centre in Biodiversity and Genetic
Resources (CIBIO)/Research Network in
Biodiversity and Evolutionary Biology (InBIO),
Universidade do Porto, 4485-661 Vairão, Portugal.
2Program in Genomic, Biodiversity and Land
Planning (BIOPOLIS)/CIBIO, 4485-661 Vairão,
Portugal. 3Estação Biológica de Mértola/CIBIO,
7750-329 Mértola, Portugal. 4CIBIO/InBIO,
Instituto Superior de Agronomia, Universidade
de Lisboa, 1349-017 Lisbon, Portugal. 5EaRSLab,
University of Évora, 7000-671 Évora, Portugal.
*Corresponding author. Email: jpsilva@cibio.up.pt
RE FE REN C ES AN D N OT ES
1. EU, “Directive 2009/147/EC of the European
Parliament and of the Council of 30 November 2009
on the conservation of wild birds (Codified version)”
(2009); https://eur-lex.europa.eu/legal-content/EN/
TXT/?uri=CELEX:32009L0147.
2. J. P. Silva et al., Sci. Rep. 13, 10005 (2023).
p
3. J. C. Alonso, C. Palacín, Bird Conserv. Int. 32, 523 (2022).
4. J. Gameiro et al., “1o Censo Nacional da Águia-caçadeira
Circus pygargus: Resultados Finais” (2023); https://
steppebirdsmove.com/wp-content/uploads/2023/11/
Gameiro-et-al_2023_1o-Censo-Nacional-Aguia-
cacadeira-resultados-finais.pdf [in Portuguese].
5. H. Alonso, J. Andrade, J. Teodósio, A. Lopes, “O estado
das aves em Portugal 2022. 2a Edição” (Lisboa, 2022)
g
[in Portuguese].
6. P. F. Ribeiro, Agric. Ecosyst. Environ. 183, 138 (2014).
7. N. Faria, M. B. Morales, J. E. Rabaça, Eur. J. Wildl. Res. 62,
663 (2016).
8. T. Pinto-Correia, Landsc. Urban Plan. 50, 95 (2000).
9. M. Pinto, P. Rocha, F. Moreira, Biol. Conserv. 124, 415
y
(2005).
10. “Zero apresenta queixa à Comissão Europeia por
causa do declínio das aves estepárias,” Público (2022);
https://www.publico.pt/2022/01/28/local/noticia/
zero-apresenta-queixa-comissao-europeia-causa-
declinio-especies-esteparias-1993501 [in Portuguese].
10.1126/science.adn1390
Good governance can save
y g
China’s mine ecosystems
In China, there are an estimated 99,000
abandoned mines, about 11,700 of which
are metal mines (1). The abandoned mines
y
threaten ecosystems and pose health haz-
ards to people and animals (2, 3). Unclear
,
goals and expensive pollution mitigation
challenges impede ecological restoration
efforts. Stakeholders must ensure that all
environmental activities are coordinated to
maximize restoration efficiency.
China has taken steps to restore eco-
systems around abandoned mines. In
2018, the Chinese government created the
Department of Territorial Space Ecological
Restoration, which includes areas with
abandoned mines in its purview. The
Overall Plan for Major Projects for the
Protection and Restoration of National
Important Ecosystems (2021–2035) (4)
designates mine ecological restoration as a
key task (5). However, these measures are
12 JANUARY 2024 • VOL 383 ISSUE 6679 157
I N SI G H TS | L E T T E R S
inadequate because they fail to account for
the pollution that mining leaves in its wake
or the costs required to remediate it (6, 7).
Mining activities leave behind complex
soil and water environments. Organic and
inorganic pollutants are found in solid
waste, abandoned land, groundwater,
surface water, and other interconnected
systems (7, 8). Although some encouraging
advances have been made in cultivating
plants that sequester metals (9), much work
remains to be done to improve the effective-
ness of contaminant removal in some areas.
Removing the pollutants triggered by
metal mines is expensive. In many cases,
the companies that built the mines have
left, leaving the burden of environmental
remediation to the local government. To
minimize the investment required, stake-
holders tend to look for rapid and efficient
strategies, such as repurposing abandoned
land for production instead of restoring
ecological spaces (10). These solutions allow
investors to quickly recover the upfront
financial deficit and generate profits, but
the land might remain contaminated (11).
Effective ecological restoration would
simultaneously address environmental
pollution, but such efforts are hindered
by fragmented governance. Government
departments and private enterprises are
working toward disparate goals, such as
geological environmental protection, pol-
lution control, soil erosion prevention,
and biodiversity conservation. Because
there is no mechanism to coordinate
systematic restoration goals, these
varied efforts can undermine efficient
collaboration.
At the 2023 National Conference on
Ecological and Environmental Protection in
Beijing, attendees discussed the importance
of synergistic governance and high-level
protection (5). China should take immedi-
ate action to streamline ecological restora-
tion and pollution mitigation efforts in
areas with abandoned mines. Stakeholders
must ensure that all environmental activi-
ties are coordinated to maximize restora-
tion efficiency.
Wei-Bo Ma1, Hai-Dong Li1*, Shao-Gang Lei2, Xiang-
Hua Xu3, Sheng-Guo Xue4
1Nanjing
Institute of Environmental Sciences,
Ministry of Ecology and Environment, Nanjing
210042, China. 2Engineering Research Center
of Ministry of Education for Mine Ecological
Restoration, China University of Mining and
Technology, Xuzhou 221116, China. 3Nanjing
University of Information Science & Technology,
Nanjing 210044, China. 4School of Metallurgy and
Environment, Central South University, Changsha
410083, China.
*Corresponding author. Email: lihd2020@163.com
R E F E R E N C ES A N D N OT ES
1. J. D. Zhang, F. R. Hao, Acta Ecol. Sin. 40, 7921 (2020)
[in Chinese].
158 12 JANUARY 2024 • VOL 383 ISSUE 6679
2. K. Bennett, in Mine Closure 2016: Proceedings of the 11th
International Conference on Mine Closure, A. B. Fourie,
M. Tibbett, Eds. (Australian Centre for Geomechanics,
Perth, 2016), pp. 241–252.
3. Z. Li et al., Sci. Total Environ. 468, 843 (2014).
4. National Development and Reform Commission and
Ministry of Natural Resources, “Major Projects for
the Protection and Restoration of National Important
Ecosystems Master Plan (2021–2035)”; https://www.
gov.cn/zhengce/zhengceku/2020-06/12/
content_5518982.htm [in Chinese].
5. China Daily, “Xi stresses building beautiful China,
advancing modernization featuring harmony between
humanity and nature”; https://www.chinadaily.com.
cn/a/202307/19/WS64b7d62ca31035260b817596.
html.
6. Ministry of Ecology and Environment of the People’s
Republic of China, “Minister of Ecology and Environment
Huang Runqiu goes to Hunan, Guizhou, and Chongqing
to research manganese pollution control work”;
https://www.mee.gov.cn/ywdt/hjywnews/202310/
t20231010_1042783.shtml [in Chinese].
7. H. Liu et al., Sci. Tot. Environ. 825, 154004 (2022).
8. Z. Bian et al., Science 337, 702 (2012).
9. H. Ali et al., Chemosphere 91, 869 (2013).
10. H. Li, W. Ma, G. Hu, Theory and Practice of Mine
Restoration Ecology (China Environment Publishing
Group, Beijing, 2022) [in Chinese].
11. H. Xu et al., Ecol. Indic. 154, 110630 (2023).
10.1126/science.adn2422
Nature protection must
precede restoration
Globally, billions of dollars are spent to
restore terrestrial, freshwater, and marine
systems (1). In addition to governmental
financing and nongovernmental organi-
zations, transnational companies invest
substantial funds in nature restoration to
position themselves as leaders in offsetting
their environmental impact (2). However,
the overall success of restoration often
falls below expectations (3) because the
human pressures that led to ecosystem
degradation have not been addressed.
Legal protection of ecosystems can reduce
harmful human activities (4), but restora-
tion plans often move forward without
corresponding protections in place. Strict
legal protection should always accompany
restoration efforts.
The European Union provides one
example of restoration plans that lack
associated protective measures. In
November 2023, Europe provisionally
agreed to a new nature restoration law,
pledging to restore habitats and species
comprising at least 20% of land and sea
by 2030 and all ecosystems requiring res-
toration by 2050 (5). Although multiple
strategies and directives highlight the
importance of strict protection of eco-
systems, implementation by EU member
states remains limited (6). In Denmark, for
instance, nutrient runoff drives eutrophi-
cation and hypoxia in marine systems (7),
and the country is developing widespread
marine restoration programs. However,
reduction of nutrient quantities remains
limited and does not meet EU directives
(7). As a result, the legislation will likely
struggle to protect both natural and
restored marine ecosystems.
Australia, where intensified land use fre-
quently conflicts with ecosystem protection
(8), suffers from similarly short-sighted
decisions. The country has implemented
innovative restoration methods, but in
2020, the government agreed to not inter-
vene with the development of 69 coal
mining projects or 45 gas mining projects
(9). Together with the ongoing expansion
of land clearing (10), these actions put
immense pressure on terrestrial, freshwa-
ter, and marine habitats, undermining any
restoration plans.
Ecosystem restoration is undeniably
p
beneficial—restoring 15% of lands glob-
ally could prevent 60% of expected species
extinctions (11). However, strict protection
of natural and restored ecosystems should
be prioritized. Recovery after restoration
is lengthy and complex (3), and rates of
global habitat destruction, fragmentation,
g
and exploitation largely surpass habitat
restoration and protection (12). The devel-
opment of an environmentally sustainable
economy will likely take decades. Unless
y
strict legal protection is enforced now,
more ecosystems will likely be lost than
restoration can revitalize.
Deevesh A. Hemraj1*, Melanie Bishop2, Jacob
Carstensen1, Dorte Krause-Jensen3,4, Peter A. U.
Stæhr1,4, Bayden D. Russell5,6,7
1Department of Ecoscience, Aarhus University,
DK-4000 Roskilde, Denmark. 2School of Natural
Sciences, Macquarie University, Sydney, NSW,
Australia. 3Department of Ecoscience, Aarhus
University, DK-8000 Aarhus C, Denmark. 4Center
y g
for Marine Nature Restoration, Roskilde, Denmark.
5The Swire Institute of Marine Science and Area
of Ecology and Biodiversity, School of Biological
Sciences, The University of Hong Kong, Hong Kong
SAR, China. 6Institute for Climate and Carbon
Neutrality, The University of Hong Kong, Hong
Kong SAR, China. 7The Joint Laboratory for Marine
y
Ecology and Environmental Sciences, The Swire
Institute of Marine Science, The University of Hong
,
Kong, Hong Kong SAR, China.
*Corresponding author.
Email: adhemraj@ecos.au.dk
R E F E R E N C ES A N D N OT ES
1. B. Bodin et al., Restor. Ecol. 30, e13515 (2022).
2. T. A. C. Lamont et al., Science 381, 1053 (2023).
3. D. A. Hemraj et al., Sci. Adv. 8, eabp8747 (2022).
4. H. O. Pörtner et al., Science 380, eabl4881 (2023).
5. Council of the European Union, “Press release 852/23”
(2023); https://europa.eu/!PxTr6k.
6. V. Hermoso et al., Environ. Sci. Pol. 127, 263 (2022).
7. C. C. Hoffmann et al., Ecol. Eng. 155, 105863 (2020).
8. B. K. Sovacool et al., PLOS Clim. 2, e0000221 (2023).
9. R. MacNeil, G. A. Edwards, Aust. J. Int. Aff. 77, 19 (2023).
10. E. Brunton et al., Science 382, 275 (2023).
11. B. B. N. Strassburg et al., Nature 586, 724 (2020).
12. P. Jaureguiberry et al., Sci. Adv. 8, eabm9982 (2022).
10.1126/science.adn0543
science.org SCIENCE
 RESEARCH
IN S CIENCE JOU R NA L S Edited by Michael Funk

p
g
CONSERVATION

y
Assessing the forest for the trees

E
fforts to set conservation priorities and evaluate protection
activities often depend on assessments of species’ con-
servation statuses, such as the International Union for
Conservation’s Red List of Threatened Species. Assessments
require detailed data, considerable time, and expertise.
de Lima et al. used an automated, quantitative method to assess
species based on the Red List criteria and applied it to nearly
5000 tree species from the Atlantic Forest, a relatively data-rich
biodiversity hotspot in South America. They classified over 80% of
endemic species as threatened and 13 species as possibly extinct.
Data to estimate population reductions, which are not available in
many tropical areas, were key to assessing threatened status for
many species. —BEL
Science p. 219, 10.1126/science.abq5099

y g
Assessment reveals that a majority of tree species in the Atlantic Forest in eastern Brazil are threatened by extinction.

CANCER IMMUNOLOGY Tumor growth results in the in both biological and synthetic MEMBRANES

Saposins and antitumor
immunity
Cancers develop mechanisms to
evade detection and subsequent
killing by the immune system.
One such strategy involves
suppressing the presentation
PHOTO: LEONARDO MERCON/SHUTTERSTOCK
addition of complex sugar chains
to prosaposin molecules in DCs,
leading to prosaposin secretion
and the depletion of lysosomal
saposins. Targeting tumor DCs
with an engineered form of pro-
saposin enhances the antitumor
immune response and improves
chemistry, it is comparatively
rare for both O atoms from O2 to
end up in molecules of the target
product. Instead, one of them is
typically diverted to a co-product
to facilitate the primary reaction.
Hoque et al. now show that an
electrochemical approach can
y
Fast diffusion
,
in a molecular sieve
High-volume olefin separations
are a core chemical process,
and efficient production in high
purity is challenging. Cui et al.
developed a crystalline porous

of tumor-associated antigens.
Sharma et al. report that a protein
called prosaposin is key for trig-
gering immunity against tumors.
Immune sentinels called dendritic
cells (DCs) engulf dead cancer
cells and use prosaposin to pro-
cess tumor-derived antigens for
subsequent display and activa-
tion of protective T lymphocytes.
tumor control. —PNK improve atom economy in this
Science p. 190, 10.1126/science.adg1955 context. Specifically, they paired
oxygen reduction with water
oxidation to produce manganese
ELECTROCHEMISTRY oxo species at each electrode,
which in turn oxidized thioethers
Charging ahead to sulfoxides. The net reaction
with both oxygens transferred both O atoms without
Although elemental oxygen is electron consumption. —JSY
an extremely common reactant Science p. 173, 10.1126/science.adk5097
coordination polymer, ZU-609,
as a molecular sieve for propane
and propylene separation.
ZU-609 was prepared by the
reaction of 1,2-ethanedisulfonate
(EDS) and 4,4’-dipyridylsulfide
(DPS) with copper(II) ions, and
EDS anions acted as pillars to
connect a two-dimensional
network of DPS and copper

SCIENCE science.org 12 JANUARY 2024 • VOL 383 ISSUE 6679 159

R ES E ARCH
ALSO IN SCIENCE JOURNALS
CANCER IMMUNOLOGY
Neutrophil
reprogramming in tumors
Neutrophils are the most abun-
dant white blood cell population in
the body. Although different types
of neutrophils frequently gather in
the tumor microenvironment, it is
currently unclear how they might
coordinate to support tumor
growth. Using an experimental
model of pancreatic cancer,
Ng et al. report populations of
tumor-infiltrating neutrophils
that converged to develop into a
single long-lived “T3” cell subset.
T3 neutrophils prompted the
growth of new blood vessels,
which enhanced tumor survival
in areas with low oxygen and
limited nutrients. Depleting T3
neutrophils or inhibiting their
angiogenic function reduced
tumor growth. The researchers
suggest that neutrophils could be
reprogrammed within the tumor
microenvironment into a single
functional state, and modula-
tion of protumoral neutrophil
responses may have therapeutic
potential. —PNK
Science p. 163, 10.1126/science.adf6493
SOLAR CELLS
Toward viable perovskite-
silicon tandems
Tandem perovskite-silicon solar
cells, in which the perovskite
layer is tuned to absorb the
higher-frequency end of the
solar spectrum to complement
absorption of the silicon cell, can
surpass the power-conversion
efficiency of the best single-
junction silicon cells. However,
for widespread adoption, the
levelized cost of electricity from
tandems must be lower than
that of mainstream silicon tech-
nology. Aydin et al. reviewed the
efforts needed to commercialize
these devices, including issues
with stability, process scale-up
and throughput, cell-to-module
integration, and field perfor-
mance evaluation. —PDS
Science p. 162, 10.1126/science.adh384
159-B 12 JANUARY 2024 • VOL 383 ISSUE 6679
Edited by Michael Funk
NEUROSCIENCE
Limitations of AI-based
predictive models
A central promise of artificial
intelligence (AI) in health care
is that large datasets can be
mined to predict and identify the
best course of care for future
patients. Unfortunately, we do
not know how these models
would perform on new patients
because they are rarely tested
prospectively on truly indepen-
dent patient samples. Chekroud
et al. showed that machine learn-
ing models routinely achieve
perfect performance in one
dataset even when that dataset
is a large international multisite
clinical trial (see the Perspective
by Petzschner). However, when
that exact model was tested in
truly independent clinical trials,
performance fell to chance
levels. Even when building what
should be a more robust model
by aggregating across a group of
similar multisite trials, subse-
quent predictive performance
remained poor. —PRS
Science p. 164, 10.1126/science.adg8538;
see also p. 149, 10.1126/science.adm9218
OPTICS
Controlling electron
beams
Electron microscopes provide
imaging capability on the tiniest
of scales. The electron beams
that scatter off the samples are
generally energetically stable
and spatially uniform. Being
able to modulate the beam to
access spatiotemporal informa-
tion about the sample would be
extremely useful but is techni-
cally challenging. Yang et al.
demonstrate that the nonlinear
optical states induced in a
microresonator can interact with
the electron beam and imprint
the nonlinear optical states onto
the beam (see the Perspective
by Polman and García de Abajo).
This interaction provides access
to ultrafast modulation of the
electron beam and broadens
the application of electron
microscopes for spatiotempo-
ral imaging and spectroscopy.
—ISO
Science p. 168, 10.1126/science.adk2489;
see also p. 148, 10.1126/science.adn1876
TRIBOLOGY
Picking the friction
Friction is ubiquitous but
surprisingly difficult to deter-
mine outside of trial-and-error
methods for many applications.
Aymard et al. developed a general
strategy that should allow the
frictional relationship to be
designed in advance (see the
Perspective by Slesarenko and
Pastewka). The strategy requires
understanding the behavior of
individual bumps called asperi-
ties and then modeling how
groups of those bumps affect
friction. The authors provide
several examples of tuning the
frictional response in a model
system. The strategy should be
broadly applicable across differ-
ent material combinations and
scales. —BG
Science p. 200, 10.1126/science.adk4234;
see also p. 150, 10.1126/science. adn1075
IMMUNOLOGY
Plasma cells hobbled
without their wobble
Plasma cells generate antibod-
ies at considerable rates, often
over the course of many years.
Although antibody production
is a well-regulated process, little
is known about how the plasma
cell’s translational machin-
ery adapts to this formidable
task. Giguère et al. report that
transfer RNA (tRNA) and codon
usage are congruently optimized
as B cells differentiate from
a naïve phenotype to plasma
cells (see the Perspective by
Alvarez and James). The tRNA
pool in plasma cells is enriched
in inosine (I34)–modifiable
anticodons, which allows
one tRNA to decode mul-
tiple synonymous codons by
“wobble” (non–Watson-Crick)
base-pairing. These findings
suggest that codon usage and
cell-specific tRNA pools may be
an attractive future target for
both the exogenous production
of pharmaceutical antibodies
and the rational design of vac-
cines. —STS
Science p. 205, 10.1126/science.adi1763;
see also p. 146, 10.1126/science.adn1067
CONSERVATION
Still in danger
Over the past decade, the
plight of the world’s sharks
p
has received much attention,
resulting in increased regula-
tion and finning bans. However,
whether this increased attention
has translated into improved
outcomes for sharks is unclear.
Worm et al. estimated fishing-
g
induced mortality globally and
found that, overall, it has con-
tinued to increase over the past
10 years. Finning bans had little
y
impact, but fishing regulations
did reduce mortality. — SNV
Science p. 225, 10.1126/science.adf8984
HEPATITIS
Activating antiviral T cells
Promoting CD8+ T cell responses
to hepatitis B virus (HBV) with
interleukin-2 (IL-2) may offer
y g
an approach to clear the virus;
however, traditional IL-2–based
therapies are plagued by toxici-
ties and unintended activation
of other immune cell popula-
y
tions. Andreata et al. developed
and tested a CD8 cis-targeted
,
IL-2 (CD8-IL2) that selectively
activates CD8+ T cells without
affecting regulatory T cells or
natural killer cells. The authors
found that CD8-IL2 treatment
exhibited antiviral activity in a
preclinical model of chronic HBV,
reducing viremia. Further data
in human cells and cynomolgus
monkeys supported the specific-
ity and safety of this molecule.
These data support the further
development of CD8-IL2 for
chronic HBV. —CM
Sci. Transl. Med. (2024)
10.1126/scitranslmed.adi1572
science.org SCIENCE
R ES EA RCH | I N S C I E N C E J O U R NA L S

nodes. The angular geometry IMMUNOLOGY

of the dipyridyl ligands created
one-dimensional pores of about
0.5 nanometers in diameter that
enabled high sorption and diffu-
sion for propylene while propane
was excluded. —MSL
Science p. 179, 10.1126/science.abn8418
NANOMATERIALS
Two sides to epitaxy
Epitaxial growth of a crystalline
film normally proceeds from
one substrate. Cui et al. report
that two molybdenum disulfide
(MoS2) layers can both impose
orientation effects on gold. The
authors grew gold nanoparticle
layers on one MoS2 substrate
Alcohol metabolizer and IN OTHER JOURNALS
antiviral
RNA-based viruses activate the Edited by Caroline Ash
receptor RIG-I, which stimulates and Jesse Smith
antiviral signaling through the
aggregation of the mitochondrial
protein MAVS. Sun et al. found
that this antiviral response was
enhanced by ALDH1B1, a mito-
chondrially localized aldehyde
dehydrogenase involved in etha-
nol metabolism. ALDH1B1 levels
were increased by infection with
various RNA-based viruses,
which limited viral replication.
ALDH1B1 promoted the binding
of MAVS to RIG-I and MAVS
aggregation. ALDH1B1-deficient

and then covered the nanopar-
ticles with a second MoS2 film.
Upon heating, the nanopar-
ticles flattened to nanodisks.
For small twist angles between
the two substrates (about 7
degrees), the nanometer-thick
p
mice developed more severe
lung damage and were less likely
to survive infection with influ-
enza A. —WW
Sci. Signal. (2024)
10.1126/scisignal.adf8016

g
gold nanodisks adopted an
orientation intermediate to both IMMUNOLOGY
substrates. This alignment was
driven in part by the chemical
Tracking the fate of gut

interaction of gold with sulfur.
—PDS
Science p. 212, 10.1126/science.adk5947
ANTHROPOLOGY
A lost “city”
When intact, the Amazonian
forest is dense and difficult to
penetrate, both on foot and with
y
immune cells
There is growing evidence that
the interplay among enteric
microbiota, dietary antigens,
and immune cells located in
the gut has systemic impacts
in both health and disease. To
better understand how the gut
mediates interorgan commu-
nication, a clearer delineation CHAOS THEORY rapidly changing environments.

scanning technologies. Over
the past several years, however,
improved light detection and
ranging scans have begun to pen-
etrate the forest canopy, revealing
of gut immune cell migration
to the periphery is needed.
Galván-Peña et al. used Kaede
photoconvertible mice and
single-cell RNA sequencing to
Interactions
causing chaos
Chaos theory has been applied to
good effect in mathematics and
y g
These insights could be useful in
investigating antibiotic resistance,
among other phenomena. —CA
Curr. Biol. (2023)
10.1016/j.cub.2023.11.002

previously unknown evidence of
past Amazonian cultures. Rostain
et al. describe evidence of such an
agrarian Amazonian culture that
began more than 2000 years ago.
They describe more than 6000
earthen platforms distributed in
a geometic pattern connected
by roads and intertwined with
agricultural landscapes and river
drainages in the Upano Valley.
Previous efforts have described
mounds and large monuments in
Amazonia, but the complexity and
extent of this development far
surpasses these previous sites.
—SNV
Science p. 183, 10.1126/science.adi6317
map the migration of immune
cells from the colon at homeo-
stasis, after gut injury, and in
the context of extraintestinal
inflammation. This approach
revealed much greater dynamic
immunocyte turnover than
was previously thought and
uncovered distinctive patterns
of migration that depended on
the nature and location of the
inflammatory lesions. These
findings may help to explain how
perturbations in the gut micro-
biota can engender such distant
and potent effects. —STS
Sci. Immunol. (2024)
10.1126/sciimmunol.adi0672
physics but less so in cell biology.
Environmental change can trigger
seemingly random responses in
cells, but whether this effect could
in fact be explained by chaos
theory is unclear. Choudhary
et al. developed a deterministic
model to predict gene expression
dynamics and tested it using a
microfluidic device containing
growing Escherichia coli cells
experimentally exposed to oxida-
tive stress. Molecular feedbacks
from cell growth dynamics and
cell-cell interactions gener-
ate chaotic, not stochastic,
gene regulation that allowed
cell populations to respond to
y
,
NEUROIMMUNOLOGY
Immunity, blood pressure,
and cognition
Hypertension is a leading cause
of cognitive impairment. However,
the mechanisms underlying this
link between the blood vessels
and the brain remain unclear.
Santisteban et al. found that
cognitive dysfunction in a mouse
model of hypertension depended
on the production of the cytokine
interleukin-17 (IL-17) from T cells
in the dura mater (one of the
membranes surrounding the

160 12 JANUARY 2024 • VOL 383 ISSUE 6679 science.org SCIENCE


brain). This IL-17 was released into
the cerebrospinal fluid and acti-
vated receptors on macrophages.
The authors found that depleting
these macrophages or delet-
ing the relevant IL-17 receptor
prevented cognitive impairment
without altering blood pres-
sure. The findings highlight new
potential therapeutic targets for
impaired cognition in individuals
with hypertension, although it is
not yet clear whether the same
mechanism seen in mice exists in
humans. —SAL
Nat. Neurosci. (2023)
10.1038/s41593-023-01497-z
PLANT SCIENCES
Auxin at the cell surface
Auxin is a plant hormone that
regulates plant growth and
development. Decades of work
SCIENCE science.org
CELESTIAL MECHANICS
A possible gap in the
outer Solar System
T
he Kuiper Belt extends to
about 50 astronomical
units (au) from the Sun.
The few bodies known
beyond 50 au have highly
uncertain orbits because
observations cover only a
small fraction of their orbital
periods (several centuries).
de la Fuente Marcos and de la
Fuente Marcos show that the
current distance and radial
velocity of those bodies have
smaller uncertainties than the
absolute sizes of their orbits.
They used this approach to bet-
ter constrain the distribution
of known objects, identifying a
gap at about 70 au. The statisti-
cal significance of the gap was
low, but if it is real, it could be
caused by the gravitational
influence of a planet. —KTS
Mon. Not. R. Astron. Soc. (2024)
10.1093/mnrasl/slad132
A gap in the distribution of minor
objects orbiting the Sun beyond the
Kuiper Belt might indicate the presence
of a large body beyond the heliopause.
has proposed auxin binding signaling pathway are initiated at
protein 1 (ABP1) as a receptor the cell surface and integrate with
for extracellular auxin perception the canonical intracellular auxin
in the apoplast, which mostly signaling pathway. —AWa
comprises the fluid-filled spaces Cell (2023)
of cell walls. However, this role has 10.1016/j.cell.2023.10.017
been debated because mutants
lack obvious phenotypes. Yu et
EPILEPSY
al. have identified two auxin-
binding proteins, ABL1 and ABL2, Stiffening seizures
that localize to the apoplast and Mechanoreceptors, including
interact with transmembrane PIEZO1, have been shown to
kinases (TMKs), which are known play a crucial role in the cellular
players in mediating extracellular response to mechanical forces.
auxin perception and signaling. However, their roles in the brain
ABL1/2-TMKs form an auxin- in physiological and pathologi-
sensing module in the apoplast cal conditions remain elusive.
that has functions both overlap- Garcia et al. used brain samples
ping and distinct from the ABP1 from patients with temporal
protein. Experiments showed that lobe epilepsy to investigate the
ABL1/ABP1 and TMK1 are co- relationship among PIEZO1
receptors for extracellular auxin in expression, brain inflammation,
the apoplast. These findings hint and epilepsy. In epileptic tissue,
at how auxin perception and the the increased expression of the
12 JANUARY 2024 • VOL 383 ISSUE 6679
proinflammatory cytokine tumor
necrosis factor-a could induce
PIEZO1 expression in glial cells,
indicating that mechanosensa-
tion in glial cells may play a role
in the detrimental effects of
brain inflammation on seizure
activity. —MMa
Neurobiol. Dis. (2023)
10.1016/j.nbd.2023.106297
LIGAND DESIGN
Deep learning
for full pockets
A common way of targeting pro-
tein pockets for drug discovery is
by fragment-based screening, in
which small molecules with low
p
affinity can be identified and then
joined to form higher-affinity
binders. Powers et al. leveraged
advances in machine learning to
perform in silico fragment expan-
sion. A starting fragment served
as a seed, and new fragments
g
were attached to optimally fill
the binding pocket. The authors
used a neural network scoring
model based on protein-ligand
y
structures and generated pre-
liminary predictions based on
known ligand-protein complexes
that are promising early starting
points for validation and optimi-
zation. —MAF
ACS Cent. Sci. (2023)
10.1021/acscentsci.3c00572
y g
ELECTROCHEMISTRY
Zeroing in on nickel
Nickel has long been an important
catalyst component for a wide
variety of chemical reactions.
y
In recent years, there has been
even greater emphasis on its
,
prospective use as a replacement
for rarer, more expensive metals
such as palladium and platinum.
However, often overlooked in
these ever-widening applica-
tions is the need for hazardous
reagents such as metal hydrides
to reduce salts to Ni(0) before
catalysis. Rubel et al. introduce a
systematically optimized, conve-
nient electrochemical protocol to
produce common Ni(0) com-
plexes. A flow process confers
robust scalability. —JSY
Angew. Chem. Int. Ed. (2023)
10.1002/anie.202311557
161
RES EARCH

◥ trial manufacturing without invoking device

REVIEW SUMMARY
SOLAR CELLS
Pathways toward commercial perovskite/silicon
tandem photovoltaics
Erkan Aydin*, Thomas G. Allen, Michele De Bastiani, Arsalan Razzaq, Lujia Xu, Esma Ugur,
Jiang Liu, Stefaan De Wolf*
BACKGROUND: Photovoltaics is projected to
play a key role in averting the anticipated cat-
astrophic effects of climate change thanks to
its cost competitiveness, continued technolog-
ical advancements, and the abundance of solar
energy. Practical deployment has relied on
crystalline silicon solar cells for many years;
yet, this technology is now approaching prac-
ufacturing practice and offer independent ope-
ration of the submodules.
For monolithic tandems, a range of vacuum-
and solution-based processing technologies
have been developed for the perovskite photo-
absorber and its contacts. Often, the choice
for a given deposition technology is dictated
by the surface morphology of the substrates,
performance penalties. Sustainability and the
cost of feedstock materials are also crucial con-
siderations for large-scale production.
However, the most critical effort needed
to move toward commercialization lies in
improving device stability and performance
evaluations in realistic deployment environ-
ments. Specifically, cell-to-module integra-
tion requires attention, including ensuring
the chemical compatibility of encapsulants
and the mechanical strength of perovskite-
contacting materials to achieve durable mod-
ules. Early outdoor tests of tandems have
already provided valuable insights into degra-
dation mechanisms caused by factors such
as ion migration at grain boundaries and
phase decomposition under the influence of
heat, light, temperature cycles, and voltage bias-
ing. Reliability tests specific to perovskite/
silicon tandem modules are required because

tical efficiency limits. To overcome these lim-
itations and harness the full potential of
photovoltaics, perovskite/silicon dual-junction
tandem solar cells have emerged as a prom-
ising solution toward the pairing of a high
power conversion efficiency with cost-effective
which may depend on the chosen silicon
bottom-cell technology. Although spin coat-
ing is a common choice to achieve record
cells, its throughput is problematic, which
highlights the need to develop deposition
technologies that are compatible with indus-
p
existing standard tests for silicon modules may
not adequately address their distinctive char-
acteristics. Recently, new test protocols, such
as maximum power point tracking at ele-
vated temperatures, have been developed to
assess the stability of the perovskite subcell.

g
manufacturing. Introduced in 2013, ini-
tial perovskite/silicon tandem cells ex- OUTLOOK: Perovskite/silicon tandems have
hibited a modest performance. However, Single Historical data progressed beyond the laboratory stage,
through relentless research and develop- junction Future projections and recently, several companies have

y
ment efforts, perovskite/silicon tandem 100 announced plans for pilot processing
Module cost (US $ per W)

cells have now achieved certified power
conversion efficiencies exceeding 33%
for laboratory-scale devices, which is
now higher than the theoretical limit
of any single–junction cell technology.
With a demonstrated high efficiency
potential and prospects for further
enhancements, perovskite/silicon tan-
dems have now entered the path toward
lines. It is expected that the challenges
related to scaling the process will be over-
10 come within a few years. However, for
High-cost
and short
perovskite/silicon tandems to become
warranties a commercially viable option, it is cru-
1 cial to bring down the levelized cost of
electricity (LCOE) to a competitive level
Tandem with market-established silicon photo-
0.1 voltaics. Despite the advantages of higher-

y g
commercialization. performance modules toward a lower
cost per generated watt, uncertainties
ADVANCES: We review notable reported over module durability, energy yield, and
30
Module warranty (years)

advances toward translating laboratory- reliability of emerging photovoltaics tech-

scale tandem performance to industry- nologies (like perovskites) may cancel out
grade modules. In this context, various the translation toward a lower LCOE.

y
20 Low-cost
device configurations have been ana- and long These uncertainties will also compromise

lyzed, including two-terminal, three-
terminal, and four-terminal tandems.
Beyond this, bifacial tandems and triple-
junction cells have also been explored.
So far, most efforts have focused on
monolithic two-terminal tandems, where
the perovskite subcell is directly fab-
ricated onto a silicon solar cell, with
both connected in series using an in-
ternal junction. Tandems can also be
made by mechanically stacking a semi-
transparent perovskite cell onto a sili-
con cell. Monolithic tandems require
fewer materials-processing steps, where-
as mechanically stacked tandems more
easily integrate with existing silicon man-
10
0 ing with accurate energy-yield forecast-
1970 1980 1990 2000 2010
Evolution and projection of photovoltaic module cost and
warranty duration over time. (Bottom) Transition from
high-cost and short warranties in the 1980s to the current
scenario of low-cost, long warranties. The previous years’ data
were retrieved from Jordan et al., Prog. Energy 4, 022002
(2022), and future predictions are adapted from the International
Technology Roadmap for Photovoltaic 2022. (Top) Perovskite/
silicon tandems should expand on these trends to become a
mainstream photovoltaic technology.
,
warranties the bankability of perovskite/silicon tan-
dem technologies. Hence, accelerated deg-
radation tests—targeted specifically to
perovskite technology—and outdoor test-
2020 2030 ing will be instrumental to advance the
market entry of this technology.
▪
The list of author affiliations is available in the full article online.
*Corresponding author. Email: erkan.aydin@kaust.
edu.sa (E.A.); stefaan.dewolf@kaust.edu.sa (S.D.W.)
Cite this article as E. Aydin et al., Science 383,
eadh3849 (2024). DOI: 10.1126/science.adh3849
READ THE FULL ARTICLE AT
https://doi.org/10.1126/science.adh3849

Aydin et al., Science 383, 162 (2024) 12 January 2024 1 of 1

RES EARCH

◥ door testing coupled with accurate energy

REVIEW yield forecasting are critically needed to re-
duce such uncertainties to translate perform-
SOLAR CELLS ance gains at the module level to a lower LCOE
and to thereby advance the market entry of
Pathways toward commercial perovskite/silicon this technology.

tandem photovoltaics Tandem module configurations

Most experimental efforts have focused on
Erkan Aydin*, Thomas G. Allen, Michele De Bastiani†, Arsalan Razzaq, Lujia Xu, Esma Ugur, monolithic tandems in the two-terminal (2T)
Jiang Liu‡, Stefaan De Wolf* configuration with a perovskite cell fabricated
on top of a c-Si solar cell, series-connected by
Perovskite/silicon tandem solar cells offer a promising route to increase the power conversion an internal junction (Fig. 1A) (7). Although
efficiency of crystalline silicon (c-Si) solar cells beyond the theoretical single-junction limitations at monolithic modules require fewer transparent
an affordable cost. In the past decade, progress has been made toward the fabrication of highly electrodes and BOM and BOS components
efficient laboratory-scale tandems through a range of vacuum- and solution-based perovskite compared with other tandem configurations,
processing technologies onto various types of c-Si bottom cells. However, to become a commercial their optimal performance requires current
reality, the transition from laboratory to industrial fabrication will require appropriate, scalable input matching, which demands the tuning of the
materials and manufacturing processes. In addition, perovskite/silicon tandem research needs to top cell bandgap through compositional engi-
increasingly focus on stability, reliability, throughput of cell production and characterization, cell-to- neering of the perovskite. Eliminating the
module integration, and accurate field-performance prediction and evaluation. This Review discusses need for bandgap tuning is possible when in-

p
these aspects in view of contemporary solar cell manufacturing, offers insights into the possible corporating a third terminal, for instance by
pathways toward commercial perovskite/silicon tandem photovoltaics, and highlights research using c-Si bottom cells with interdigitated back
opportunities to realize this goal. contacts. However, this strategy introduces in-
creased complexity in both processing and in-
terconnections. (8, 9).

T
he large-scale deployment of photovol- A high PCE is paramount because it is a Tandems can also be constructed by the

taics (PV) witnessed in recent years has
been driven by a sharp decline in PV
module prices (dropping >85% since
2010), increased power conversion effi-
strong lever toward lowering all relative balance-
of-module (BOM) and balance-of-systems (BOS)
costs, such as the module glass, lamination
materials, frame, wiring, inverters, land, and
g
mechanical stacking of the subcells. In this
case, the subcells are manufactured indepen-
dently, which offers greater freedom in process-
ing but adds costs. More transparent electrodes

ciencies (PCEs) and system lifetimes, as well
as inherently low variable running costs (1).
All of these factors help to lower the levelized
cost of electricity (LCOE) from PV, which is a
cost metric that accounts for the overall sys-
tem costs and energy generated over a PV
system’s lifetime. Yet, PV technologies based
on crystalline silicon (c-Si), which currently
dominate the market, are approaching their
theoretical and technical PCE limits. Today,
installation costs. A straightforward way to
further increase the PCE of c-Si PV is by stack-
ing one or more cost-effective wider-bandgap
solar cells on top of a c-Si device to use the
solar spectrum more effectively. For instance,
dual-junction tandems that stack two solar
cells can theoretically yield PCEs of >40% (3, 4).
Perovskite solar cells (PSCs) are promising for
such tandem integration owing to their tuna-
ble bandgap (which is needed to maximize the
y
are required [usually in the form of transpar-
ent conductive oxides (TCOs)], which may also
increase parasitic absorption losses (10). More-
over, the large perovskite submodule requires
several film patterning steps (Fig. 1B), usually
by laser scribing. Finally, the perovskite and
c-Si subcells need to be electrically isolated
through appropriately chosen module lami-
nation materials (Fig. 1C).
Mechanically stacked tandems come in prin-

the record PCE for c-Si PV has reached 26.8%
for cells (the theoretical maximum is ~29.4%)
and >24% for modules, and mass-produced
commercial modules are now sold with PCEs
in the range of 22 to 24% (2). Advanced cell
interconnection and module integration schemes
spectral efficiency) (5) combined with their
potential for high performance (small-area,
single-junction devices have reached PCEs of
>26%) and their potential for low-cost man-
ufacturing (2).
Silicon-based tandems are predicted by the
y g
ciple in a four-terminal (4T) configuration. This
allows the submodules to operate indepen-
dently but requires two dc-ac inverters, which,
along with extra cabling, adds further to the
total system cost. Yet, a 2T module configu-
ration is also possible for mechanically stacked

y
can help to further reduce cell-to-module losses, International Technology Roadmap for Photo- tandems by connecting the two submodules

but ultimately c-Si PV technologies are nearing
their practical PCE limits. One way to further de-
crease the LCOE of PV is through improve-
ments in PCEs beyond what can be achieved
with c-Si technology alone.
KAUST Solar Center (KSC), Physical Sciences and
Engineering Division (PSE), King Abdullah University of
Science and Technology (KAUST), Thuwal 23955-6900,
Kingdom of Saudi Arabia.
*Corresponding author. Email: erkan.aydin@kaust.edu.sa (E.A.);
stefaan.dewolf@kaust.edu.sa (S.D.W.)
†Present address: Centre Suisse d’Électronique et de Microtechnique
(CSEM), Neuchâtel, Switzerland.
‡Present address: College of Energy, Soochow Institute for Energy
and Materials Innovations, Soochow University, Suzhou, 215006,
Jiangsu, China.
voltaic (ITRPV) to enter the market in the next
few years (6). However, mainstream c-Si mod-
ule technology already delivers reliable prod-
ucts at a cost of less than $0.25 per watt with a
scaled and rapidly expanding manufacturing
capacity that exceeded 600 GW per year at the
end of 2022 (6). To be market competitive,
perovskite/silicon tandems must demonstrate
clear benefits in terms of cost and LCOE. De-
spite the advantages of higher-performance
modules toward a lowered cost per nameplate
watt of PV systems, uncertainties over module
durability and energy yield may compromise
the bankability of perovskite/silicon tandems.
Hence, accelerated degradation tests, targeted
specifically to perovskite technology, and out-
,
either in series or in parallel. This requires
current matching or voltage matching of the
constituent submodules, respectively; in either
case, a single inverter is sufficient. One advan-
tage of voltage matching is the less stringent
demand on bandgap engineering and better
performance stability against spectral and
operating temperature changes (11). Overall,
for mechanically stacked tandems, perovskite
submodules may be integrated with minimal
adjustments to commercial c-Si module man-
ufacturing, which could enable faster entry into
the mainstream PV market. Nevertheless, their
monolithic counterparts may offer decisive
benefits in terms of device performance, re-
liability, and cost.

Aydin et al., Science 383, eadh3849 (2024) 12 January 2024 1 of 13

RES EARCH | R E V I E W

p
g
y
Fig. 1. Perovskite/silicon tandem module connection schemes. (A to C) Schematic illustrations of monolithic (A) and mechanically stacked (B) tandem
solar modules, where the perovskite submodule is fabricated onto the front glass, and possible perovskite/silicon tandem module types (C). The modules

y g
are shown with their Sun-facing side upward. Wirings have been shown between the positive (+) and negative (−) terminals of the modules or submodules. All
of the module connections in the voltage-matched and current-matched mechanically stacked tandems are located in a separate junction box (not shown
here). 3T, three-terminal.

Monolithic 2T tandems are frequently re- and enable a fair comparison between the dif- cell can be used as a recombination junction

y
ported with independently certified PCEs, with ferent tandem concepts. (14, 15) layer with only minor modifications (e.g., by

33.9% being the current record (2). For 4T
tandem cells, no standard measurement pro-
tocol exists yet; their PCEs are usually reported
as summed PCEs of separately measured sub-
cells, where the bottom cells are illuminated
by light filtered through a dummy perovskite
top cell stack (12, 13). Moreover, 4T tandem
PCEs are often reported for much larger bot-
tom cells (1 to 4 cm2) relative to the top cells
(typically ~0.1 to 1 cm2) to maximize the sub-
cell PCEs (12). These anomalous procedures
lead to substantial variations in the reported
results. Hence, establishing a consensus around
measurement practices and standardized in-
dependent certification is needed to boost con-
fidence in reported 4T tandem device results
Device scaling
Silicon bottom-cell choices
Silicon heterojunction (SHJ) solar cells in a
both-sides contacted layout (where electrons
and holes are collected at opposite sides of
the cell) have been the preferred choice to con-
struct monolithic perovskite/silicon tandems,
including the recent record tandems to date.
The attractiveness of SHJ technology lies in
its high operating voltages that result from
passivating contacts, efficient light-trapping
features (10, 16), inherently bifacial nature,
and fabrication process that can easily enable
p-i-n or n-i-p tandem configurations (17). More-
over, in a 2T tandem, the top TCO of the SHJ
,
keeping its thickness <20 nm) (18). Figure 2A
plots the rapid PCE progress of monolithic
perovskite/silicon tandems based on SHJ bot-
tom cells.
Despite the proven status of SHJ technology
as the most efficient c-Si cell and its potential
for high throughput production rates of ~10,000
wafers per hour, the ITRPV predicts that its
market share may only reach 15% by 2030.
This falls well below the predicted production
capacity of contemporary mainstream c-Si
solar cells which are based on high-temperature
diffused junctions, such as passivated emitter
and rear contact (PERC) and tunneling oxide–
based passivating-contact (TOPCon) technolo-
gies (6, 19). Hence, building tandems from such

Aydin et al., Science 383, eadh3849 (2024) 12 January 2024 2 of 13

RES EARCH | R E V I E W

A B
36 2.1
Ideal perovskite Eg of 1.66 eV
Longi for 75 C operating temperature
34 and band edge absorption

Ugur et al.
KAUST

Mariotti et al.
KAUST

Al-Ashouri et al.
32 EPFL/CSEM HZB

Liu et al. Aydin et al.

HZB 2.0
30
PCE (%)

HZB
Oxford PV
28 Oxford PV
Oxford PV

Xu et al.
26 EPFL/CSEM

Chin et al.
Oxford PV/Oxford/HZB 1.9

Chen et al.
24 Chen et al.
Stanford/ASU
Mazzarella et al.
22

Kim et al.
2016 2017 2018 2019 2020 2021 2022 2023 2024

VOC (V)
Sahli et al.
Year 1.8
Jost et al.

Aydin et al.
C 36
Albrecht et al.
De Bastiani et al.

Hou et al.
Perovskite-SHJ Tandem
Longi Perovskite-PERC Tandem Werner et al. Sahli et al.
34 KAUST
Longi Perovskite-TOPCon Tandem 1.7 Ugur et al. Nogay et al.
SHJ Single-Junction Cell Werner et al.
32

p
HZB
EPFL/CSEM
30 Bush et al.
Werner et al.
PCE (%)

EPFL/CSEM

28 Oxford PV
CSEM
KAUST 1.6
26 Longi Single-Junction SHJ
Ideal perovskite Eg of 1.69 eV
Mailoa et al.
for 25 C operating temperature
24 CSEM
and band edge absorption

g
Nankai
HZB
CSEM CSEM Ideal perovskite
22 UNSW Eg of 1.73 eV at 25 C
HZB 1.5
UNSW (no band edge absorption)
20
0.1 1 10 100 1.50 1.55 1.60 1.65 1.70 1.75

y
Device area (cm2) Eg (eV)

Fig. 2. Trends in perovskite/silicon tandem solar cell performance.
(A) Historical evolution of monolithic 2T perovskite/silicon tandem solar cell
record PCEs (all data are retrieved from the National Renewable Energy Laboratory’s
best cell efficiencies chart). (B) Perovskite bandgaps of previously reported monolithic
perovskite/silicon tandem solar cells. The dashed lines indicate ideal tandem bandgaps
without and with corrections for the temperature dependence of the band
edge absorption in the top and bottom cell, as detailed in (11). The references
labeled with red-colored text are for bifacial tandem cells. Eg, bandgap.
(C) Published PCEs of monolithic perovskite/silicon tandem cells as a function
of their size.

diffused-junction technologies may be appealing PERC or TOPCon technologies often are found TOPCon cells, which highlights that perovskite/
for today’s c-Si manufacturers, where the re- to aid in deactivating bulk defects of boron- silicon tandems are technology-agnostic in terms

quired recombination junction could be accom-
plished by replacing the antireflection/surface
passivation stack on top of diffused junctions
with either a sputtered TCO or a doped poly-
silicon stack (7). Nevertheless, Auger recombi-
nation and free carrier absorption losses—both
doped wafers. However, in the foreseeable
future, advancements in research and devel-
opment (R&D) are anticipated to lead to a cost
reduction of n-type wafers, ultimately match-
ing the price of p-type wafers. On the other
hand, the bulk lifetime of p-type wafers is now
y g
of appropriate bottom cells.
Perovskite deposition technology
As already alluded to, the translation of research-
scale perovskite/silicon tandems to industrial
prototypes differs for monolithic and mechan-

y
caused by the high carrier concentrations in also approaching that of n-type material by ically stacked configurations. In mechanically

the highly doped, diffused regions—pose a per-
formance penalty when comparing PERC and
TOPCon technologies with SHJ solar cells.
Currently, SHJ solar cells are almost exclu-
sively fabricated using n-type c-Si wafers ow-
ing to the high bulk carrier lifetime needed to
achieve high open-circuit voltage (VOC) and fill
factor (FF) values. However, n-type wafers
are still more expensive compared with their
p-type counterparts because of the smaller
segregation coefficient of phosphorus (the do-
pant in n-type wafers) compared with boron
(the dopant in p-type wafers), which limits the
usable portion of the ingot and the reuse of
the Czochralski (Cz) crucibles. Furthermore, the
high-temperature processing steps inherent to
replacing boron with gallium as the dopant,
so it is expected that the electronic quality
and cost of both types of wafers will become
comparable.
Notable examples of monolithic perovskite/
silicon tandem devices with larger cell areas are
summarized in Fig. 2C together with laboratory-
scale examples. In 2023, Oxford PV achieved a
cell PCE of 28.6% for a commercial-sized tan-
dem (258.15 cm2), likely using a SHJ bottom
cell, surpassing the record for single-junction
c-Si cells (26.81% by Longi) (15). On the other
hand, Hanwha Q-Cells announced a non–SHJ-
based bottom-cell technology for their planned
perovskite/silicon tandem pilot lines, and Jinko
Solar announced 32.33% tandem cells on n-type
,
stacked tandems, the perovskite submodules
are fabricated onto the planar module glass,
with similar choices for scalable deposition
techniques as for regular thin-film perovskite
modules (20, 21). Thin-film PV modules are
often in a superstrate configuration, which im-
plies use of the module glass on the light-facing
side. To pass the International Electrotechnical
Commission IEC 6170 reliability standards,
the module glass should remain intact in case
of mechanical failure, which calls for the use of
tempered glass. However, such glass usually
features surface imperfections, which may not
be an issue for vacuum-based film deposition
methods but can affect film uniformity in the
case of solution processing—for instance when

Aydin et al., Science 383, eadh3849 (2024) 12 January 2024 3 of 13

RES EARCH | R E V I E W

using slot-die coating, a scalable linear printing
technique with continuous ink supply, as cur-
rently pursued by several companies.
For monolithic tandems, to be aligned with
current industrial PV processing practices, re-
cent efforts have focused on depositing perov-
skite top cells onto textured c-Si bottom cells.
This approach poses additional challenges for
perovskite processing, especially when using
solution-based techniques (22–25). Almost all
industrial single-junction c-Si cells are fabri-
cated from monocrystalline Cz silicon wafers,
which are wire-cut from a silicon ingot, re-
sulting in a grooved, defective surface (Fig. 3A).
With alkaline wet-chemical etching, the saw
damage is removed followed by texturing of
the surface to decrease reflection losses and
promote the internal trapping of weakly ab-
sorbed photons to improve the solar cell’s in-
frared response (26). Overall, this results in a
wavy surface morphology onto which random
pyramid texture is superimposed (27), as shown
in the scanning electron microscopy (SEM) im-
ages in Fig. 3, B to D. Although laboratory-scale
tandem cells often use front side–polished c-Si
wafers, because this enables easier top-cell pro-
cessing, industrial tandems will likely have to
use double side–textured wafers given their
optical benefits and the prohibitive cost of
wafer polishing. We note that the texturing
size for tandems with solution-processed pe-
rovskites represents a compromise between
light-trapping capabilities and perovskite top-
cell performance. Too-thin perovskites may
leave shunt paths through exposed pyramids
(Fig. 3E); too-thick perovskites yield poor car-
rier collection (12). By contrast, conformal coat-
ing techniques are not subject to limitations
imposed by the texturing size (Fig. 3G).

p
g
y
B C D

y g
Shunt paths G
E

y
,

F
Full coverage

Fig. 3. Surface topography of industrial silicon wafers. (A) Sketch of the
commonly observed surface features on commercial Cz wafers and resulting
perovskite film morphologies. (B to D) Cross-sectional SEM images of the c-Si
wafers for various surface topographies, such as as-cut (B), saw damage removal
applied (C), and random pyramidal textured Cz wafers in alkaline solution (D).
(E and F) SEM cross-sectional images of the micrometer-scale pyramids with
solution-processed perovskite on top and pyramids coming out from the
perovskite layer, which causes a shunt on the devices (E), and well-optimized
pyramids with end-to-end coverage (F). The images were reproduced from (106)
and (107), respectively. The scale bars for both images [(E) and (F)] are 2 mm.
(G) Conformal coated perovskites on textured interfaces by hybrid technique.
The images are used with permission from Chin et al. (108).

Aydin et al., Science 383, eadh3849 (2024) 12 January 2024 4 of 13

RES EARCH | R E V I E W
Spin coating of perovskite-precursor inks,
combined with antisolvent treatments for film
crystallization and additive engineering for de-
fect passivation, has so far led to the highest-
performing perovskite single-junction and
perovskite/silicon tandem solar cells (28, 29).
However, materials waste and limited scal-
ability and throughput undermine the com-
mercial prospects of this fabrication route.
These limitations may be overcome by slot-
die coating, for which the suitability for top-
cell fabrication, even on textured surfaces, has
been reported for laboratory-scale tandem de-
vices. Yet, several challenges arise related to
slot-die coating, which likely explains the lack
of large-scale demonstrations to date (24, 30).
First, deposition of the perovskite through
slot-die coating is incompatible with antisol-
vent treatments, so film crystallization has
been accomplished by temperature control
(31) or gas quenching (32). Second, dedicated
ink compositions are needed to obtain uni-
form coatings and morphologies over larger
areas, where surface terminations of the sub-
strate must strike a balance between wettabil-
ity to form closed films and hydrophobicity
to enhance crystal growth (33, 34). Third, the
leading and trailing edges of linear-printed
coatings often have uncontrolled thicknesses,
morphologies, and electrical properties, for
which mitigation strategies still need to be
formulated (35). This is of particular concern
when coating on device-active silicon bottom
cells and may compromise tandem performance.
Finally, given the modulated textured surface
of silicon wafers or surface imperfections of
tempered glass (e.g., flatness), ensuring that
the perovskite consistently covers all of the sur-
face features may require thicknesses of multi-
ple micrometers (Fig. 3F), which could limit
effective charge-carrier extraction (12, 36, 37).
Overcoming this challenge requires effective
defect-passivation strategies combined with
control of the substrate morphology.
Physical vapor deposition techniques, such
as thermal evaporation, can yield directional
material transfer, which is attractive for scalable,
conformal coatings on textured surfaces with ac-
curate film-thickness control (38, 39). Yet, fully
evaporated perovskites have rarely been re-
ported even for single-junction PSCs (40–42),
which is likely in part because of the high capi-
tal expenditure (CapEx) and low throughput
of laboratory-scale tools. For halide perov-
skites, one-step vacuum deposition requires
coevaporation of multiple precursors (such as
PbI2, MAI, FAI, and MABr). Greater insight
into the compositional control of deposited films
would be a first critical step toward closing the
large performance gap between evaporated
and solution-processed PSCs (43). Sequen-
tial deposition of the perovskite constituents
in a linear evaporator, followed by thermal
annealing, could simplify the compositional
Aydin et al., Science 383, eadh3849 (2024) 12 January 2024
control (39, 44). However, the integration of
additive engineering strategies, often used in
solution-processed perovskites to passivate de-
fects and fully unlock the optoelectronic prop-
erties of the bulk material, remains difficult in
this approach.
Hybrid sequential deposition, such as through
vacuum deposition of a porous inorganic tem-
plate, followed by solution- (e.g., slot-die or spray
coating) or gas-based (chemical vapor transport)
conversion into the desired perovskite phase
combines the advantages of vacuum- and
solution-based techniques in providing con-
formal coverage on textured surfaces (Fig. 3G),
thickness control, and facile compositional
tuning (45–47). Moreover, additive engineer-
ing approaches can be introduced during the
solution processing step, which facilitates
defect passivation and improved perovskite
crystallization (47–50). Guaranteeing full in-
filtration of the solution into the mesoscopic
template to yield full conversion into the pe-
rovskite phase remains the biggest challenge
for this approach.
Overall, linear evaporation, slot-die coating,
or their combination in a hybrid approach may
be the most promising candidates for scaled
perovskite deposition for monolithic tandems.
Yet, in the absence of a well-established scaling
platform, alternative techniques merit deeper
exploration too, such as pulsed laser deposi-
tion (PLD) or even magnetron sputtering—
both of which enable perovskite deposition
with stoichiometric material transfer from
target to substrate (51).
Finally, we also note the importance of pro-
cess robustness when scaling technologies,
with consistent batch-to-batch reproducibility
for depositing the perovskite absorber layer,
passivation methods, and depositing contact-
ing materials. Ultimately, the most adequate
processing technique should yield a combina-
tion of high device performance, throughput,
and reproducibility.
Reducing electrical shunts in perovskite subcells
During the processing of perovskites, recombination-
active defects may form that decrease the device
voltage. Pinholes and cracks, as well as compo-
sitional inhomogeneities, can result in unde-
sired local shunts within the perovskite cells.
Such defects can result from the substrate mor-
phology, surface energy, and coating-related
inhomogeneities. The probability of forming
a defective region scales with the device area
and will translate into PCE losses for indus-
trially sized perovskite PV modules compared
with laboratory-scale record cells (52) if not
properly addressed. This effect is shown in
Fig. 2C for monolithic tandems, which need
to scale to areas of industrial c-Si wafer sizes
(>100 cm2). The impact of localized shunts may
be minimized by engineering the contact resist-
ance at the recombination junction/perovskite
interface, as well as the use of recombination
junctions with high lateral resistance, such as
those based on doped nanocrystalline silicon
(nc-Si) films (Fig. 4A) (46) or thinner and there-
fore more resistive TCOs (53).
In the case of mechanically stacked tandems,
the perovskite top cell will cover dimensions as
large as the module glass onto which it is coated
(>1 m2). To pattern the perovskite top cell and
reduce resistive losses at the module level, sev-
eral laser scribing steps (usually three, called
P1, P2, and P3 in specific processing orders)
are necessary. These steps divide the cell into
smaller units and interconnect the adjacent
cells electrically in series, as shown in Fig. 4B.
Such compartmentalizing also reduces the im-
pact of local shunts on the overall module per-
formance. An additional fourth scribe (P4) to
subdivide the cells can isolate underperform-
ing areas. The benefit of this approach is seen
through an overall improved electrolumines-
p
cence (EL) response from a module, evidenced
by a decrease in dark zones (implying low EL
emission and thus low PV performance) and
an increase in pink zones (implying high EL
emission and high PV performance) (Fig. 4B).
An alternative strategy to minimize the impact
g
of imperfections may lie in the fabrication of
smaller-sized perovskite subcell coupons, in-
tegrated into tandems at the module lamina-
tion level. Despite these mitigation strategies,
y
developing large-scale, current leakage–free
perovskite films is essential for reproducible
large-scale perovskite manufacturing.
Energy yield and module reliability
Perovskite bandgap engineering
Initially, monolithic tandems used CH3NH3PbI3,
an early-generation halide perovskite with a
bandgap of ~1.55 eV (54, 55). Mixed iodide-
bromide perovskites can yield bandgaps of
y g
~1.73 eV, which is the ideal value for monofa-
cial monolithic tandems based on theoretical
PCE-limit calculations (11, 56). However, pe-
rovskites with bandgaps wider than 1.68 eV
usually suffer from light-induced phase segre-
gation, which remains an unresolved challenge
y
(57). Moreover, deviations from standard test
,
conditions (STC) (25°C, 1-sun intensity, AM1.5G
spectrum) as occurring in the actual outdoor
environment, such as temperature and spec-
tral variations, affect the current generation in
each of the subcells in tandems; spectrally de-
pendent optical losses at the device level (such
as through reflection and parasitic absorption)
can have a similar effect.
The ideal perovskite bandgap to maximize
the power output will depend on these devia-
tions (Fig. 2B). For example, the actual operat-
ing temperatures of modules can be as high
as 60°C in sunny climates (58). The effect of
temperature-induced broadening and narrow-
ing of the bandgaps of perovskite and silicon,
respectively, causes the ideal perovskite bandgap
5 of 13
RES EARCH | R E V I E W

A B No P4 With P4

P4 scribing lines

Transparent electrode
Electron/hole selective contact

p
Perovskite absorber

Electron/hole selective contact

Transparent electrode
Glass or flexible transparent
substrate

g
Fig. 4. Electrical shunt in tandem solar cells. (A) Schematic illustrations of scribing (horizontal lines) process aims to isolate the underperforming
the shunting paths in TCO-based recombination junctions, and the tunneling areas into smaller cells (image is used with the permission of CubicPV). The
junctions show localizing the shunts as a result of the low lateral conductivity. magnified area (left) shows the sketch of the P1, P2, and P3 scribing steps

y
(B) Photographs (top) and photoluminescence images (bottom) of perovskite to build a series-connected string of thin-film cells [used with permission
minimodules aimed at mechanically stacked tandem utilization. The P4 from (109)].

to be smaller than 1.68 eV for monofacial mono-
lithic tandems under such conditions. This
could offer increased resilience against light-
induced phase segregation because less bromide
is needed to obtain such smaller bandgaps (11).
Therefore, the intended geographical location
of operation may become an additional param-
5% relative, which means less PCE retention
after 25 years) in combination with different
variable module costs (0, 10, and 30%) to com-
pare the LCOE of tandems with c-Si modules
(Fig. 5B) (60). If a similar annual degradation
rate can be guaranteed (0.4% relative), and as-
suming only a 30% additional cost resulting
decomposition of the perovskite as the primary
issues that require attention. Substantial pro-
gress has been made in addressing critical sta-
bility issues [e.g., damp-heat and maximum
power point tracking (MPPT) at elevated tem-
perature] for low-bandgap (~1.55 eV) cells
based on FAPbI3 (63, 64). However, the use of

eter to consider when designing commercial
tandems. Whether such market-specific bandgap
tailoring will truly be needed must be determined
through accurate energy yield and cost calcu-
lations, based on realistic climate and device
data, and validated by extensive field testing.
from the perovskite subcell processing, then the
tandem module PCE should be >24% to become
competitive with mainstream c-Si PV. Under
similar cost assumptions, if the annual degra-
dation rate is 2% relative, a >32% tandem mod-
ule PCE will be needed to be competitive. Given
y g
wide-bandgap perovskites (~1.67 to 1.70 eV) in
perovskite/silicon tandems with mixed halide
(iodide, bromide, and chloride) compositions
remains a more complex challenge.
Module reliability and lifetime predictions

y
unavoidable cell-to-module losses, the required To determine the durability of perovskite/silicon

LCOE versus stability
To be commercially competitive, the warranty
of perovskite/silicon tandem modules should
be on par with that of mainstream c-Si–based
modules, which at present is typically about
25 years, although some manufacturers already
offer up to 40 years (Fig. 5A) (59). A simple LCOE
calculation considering 25 years of service gives
insight into acceptable annual degradation rates
for perovskite/silicon tandems. As a reference,
we used a 22%-PCE c-Si module with a 0.4%
relative annual degradation rate for our LCOE
calculations, which is representative of the com-
mercial state of the art (representing >90%
relative PCE retention over 25 years). We then con-
sidered higher annual degradation rates (2 and
tandem cell PCEs will need to be even higher
(61). To date, the longest reported annual deg-
radation rate of small-area (1 cm2, 21.4% initial
PCE for encapsulated cell) perovskite/silicon
tandems based on outdoor data is >17% rela-
tive. This large value underlines the urgency of
improving the stability of perovskite/silicon
tandem solar cells rather than merely enhanc-
ing their PCEs (62).
Outdoor testing of perovskite/silicon tandem
cells provides valuable insights into actual sta-
bility concerns that may arise during field op-
erations, covering synchronously all aspects
of degradation such as heat, light, humidity,
temperature cycles, and voltage biasing. Initial
tests have identified ion migration and phase
,
tandem modules intended to last for more than
25 years, accelerated tests that can give predic-
tions within a substantially shorter time frame
are required. c-Si PV manufacturers have estab-
lished protocols that effectively anticipate po-
tential early failures that may arise during field
operations. For instance, the IEC 61215 protocol
consists of a range of controlled-laboratory ac-
celerated degradation tests, such as damp-heat,
thermal cycling, humidity-freeze, ultraviolet
(UV) light exposure, and potential induced
degradation (PID) testing, among others. Confi-
dence in this protocol is based on statistical
feedback from reported long-term outdoor
test data, which remain scarce for perovskite/
silicon tandems (Fig. 5C).

Aydin et al., Science 383, eadh3849 (2024) 12 January 2024 6 of 13

RES EARCH | R E V I E W

A B C
Today's champion tandem cell PCE
100

Module outdoor test periods (Years)

1.5 0% Extra Cost Goal: 30 years stability with >80-92% PCE retention
10% Extra Cost 30
Assuming 30% module PCE
30
Module warranty (Years)

30% Extra Cost Aydin et al. 2020

Module cost (US$ per W)

(79% retention, first outdoor data)

LCOEtandem / LCOEc-Si
28
10 De Bastiani et al. 2021
(51% retention)
20
5 % / year 26
Liu et al. 2021
1 (95% retention)
10 1 2
2 % / year Babics et al. 2023
(83.4% retention)
Today's c-Si 0
0.4 % / year
0 0.1 0.8 PV panels

1970 1980 1990 2000 2010 2020 2030 2020 2022 2024 2026 2028 2030 2032
22 30 33.9
Years PCE (%) Years

D E F

p
g
Fig. 5. Reliability and LCOE of perovskite/silicon tandem modules. (A) Market perovskite/silicon tandem modules outdoor data. The PCE retention goal is based
status of the c-Si PV industry in terms of cost and module warranties. The previous on the longest product warranties in the market. (D) Sketch of state-of-the-art
years’ data are retrieved from Jordan et al. (110), and future predictions are encapsulation for perovskite/silicon tandem solar cell modules. (E) Photograph of

y
adapted from ITRPV 2022. The PV prices are retrieved from (111). From the 1970s, the failed encapsulation for 156 mm by 156 mm perovskite/silicon tandem cells because
market status changed from high-cost and short warranties to low-cost, long warranties. of contact delamination. (F) Detailed view of contact delamination at the C60/SnO2
(B) Relative LCOEs of perovskite/silicon tandems versus silicon cells, based on module interface, part of the perovskite top contact, resulting from module lamination at 120°C
degradation rates and additional perovskite submodule costs. (C) Reported [adapted with permission from (71), copyright 2022 American Chemical Society].

Additionally, some perovskite-specific degra- can stay operational under field conditions. pathway may be through additive engineer-
dation modes, although known, have yet to be Accurate degradation rates are crucial to de- ing during perovskite processing that targets
included in mainstream industrial testing pro- termine warranty costs for the modules. To the removal of initial degradation products,
tocols, and others doubtlessly remain to be dis- achieve this, longer outdoor tests are necessary, which, however, will require new materials-
covered. For example, simultaneous exposure covering realistic device areas (larger than science insights.

y g
to light and elevated temperature is a known 1 cm2 and with maximized active area/wafer Furthermore, typical mainstream PV encap-
critical cause of failure of PSCs, but this mech- size ratios) at both the module and system sulants, such as ethylene vinyl acetate (EVA),
anism is not thoroughly probed by the IEC levels (consisting of strings of modules). polyolefin elastomer (POE), coextruded EVA/
61215 protocol. To gauge its impact, the US ini- POE/EVA, and thermoplastic polyurethane (TPU),
tiative Perovskite PV Accelerator for Commer- Module packaging require lamination at 120° to 140°C for 15 min
cializing Technologies (PACT) recommends Mainstream c-Si module encapsulation con- under a mild vacuum (68). However, a temper-

y
four sequences of light exposure at elevated sists of vacuum lamination of strings of cells ature of >110°C is usually sufficient to induce

,
temperatures (1-sun, 75°C for 1000 hours) with between a glass sheet at the front and a poly- degradation of the perovskite, the organic in-
module loss analysis after every 250 hours (65). meric back sheet at the rear, with two layers of terlayers, and their interfaces (70), which has
This is similar to the so-called light and ele- encapsulant. The porous nature of typical poly- spurred the search for alternative encapsulants
vated temperature–induced degradation (LeTID) meric back sheets is not suitable for perovskite- or more thermally stable perovskite cells.
test applied to c-Si PV modules defined by IEC based PV because of moisture ingress—water Moreover, the forces that occur during lami-
TS63342. Determining the activation energy rapidly degrades perovskite devices. Instead, nation and temperature cycles (Fig. 5E), which
of degradation of the perovskite cells through the polymeric back sheet needs to be replaced can be driven by a mismatch in coefficients of
MPPT tests at elevated temperatures can also with a second glass sheet, and module edges thermal expansion, present another challenge.
help to predict the operational lifetimes of the are to be sealed with rubber (Fig. 5D) (68). How- Specifically, the ubiquitous fullerene (C60) used
tandem cells (66, 67). ever, in this case, initial degradation products as the electron transport layer (ETL) in the pe-
Table 1 provides an overview of the aging of perovskite absorbers, such as iodine gas, iodo- rovskite top contact is only weakly bonded to
protocols that have been proposed to date for methane, and hydroiodic acid, may remain the adjacent layers and can easily be fractured
perovskite modules. Correlation analysis be- trapped inside the module. These degradation and peeled off. Figure 5F shows an example of
tween these laboratory tests and field per- products can then further accelerate module delamination of perovskite top contact in tan-
formance can be used to assess how long these degradation, particularly in the interconnec- dems at their C60/SnO2 interface, where SnO2
devices—specifically the perovskite subcells— tions between different cells (69). One mitigation is widely used as a buffer layer against the

Aydin et al., Science 383, eadh3849 (2024) 12 January 2024 7 of 13

RES EARCH | R E V I E W

Table 1. Summary of critical aging tests considered for the long-term stability and reliability predictions of perovskite/silicon tandem solar cells.

Reports for mechanically

Reports for stacked perovskite/
Critical tests Conditions monolithic perovskite/ silicon tandems or Research pathway
silicon tandems single-junction
perovskite modules
Damp heat 95% PCE retention after Passed one cycle (93) Passed one cycle (112) Passing damp heat test
testing at 85°C and 85% and three cycles (113) for one time might not
relative humidity for 1000 hours perovskite single-junction be sufficient to ensure
module, and one cycle for >30 years of field stability. More
small perovskite pixel (64). correlation analysis is needed.
............................................................................................................................................................................................................................................................................................................................................
Thermal cycling 95% PCE retention after 98.8% of its initial Passed for perovskite Interfacial toughening
−40° to +85°C, 200 cycles PCE after 200 thermal single-junction methods or new materials
cycles (tabbing was module (112). should be developed that can
copper tape) (114) resist expansion and shrinking
during thermal cycles (115).
............................................................................................................................................................................................................................................................................................................................................
Humidity freeze 95% PCE retention after No test data reported No test data reported Moisture-resistant and
−40° to +85°C and 85% for tandem modules. Passed perovskite-compatible

p
relative humidity, 50 cycles for small single-junction encapsulation materials
perovskite pixels (116). should be developed for
freezing-heating cycles.
............................................................................................................................................................................................................................................................................................................................................
Outdoor stability 95% PCE retention 40 days, 90% No test data reported for Longer (>>1 year) outdoor
after 60 kWh/m2 total PCE retention (117) tandem modules. For tests at different climates,
solar irradiance single-junction realistic module sizes, and the

g
pixel, 1-year data (118). multiple number of module
tests are required to
correlate them with
laboratory tests.

y
............................................................................................................................................................................................................................................................................................................................................
MPPT at elevated 95% PCE retention after No test data reported No test data reported for This is not a well-defined standard
temperature 65° to 85°C, MPPT, tandem modules. protocol but is critical for
>1000 hours 97%-PCE retention at 65°C understanding the halide
after 1200 hours for segregation issue. The
small perovskite pixel (119). ideal testing temperature
should be chosen according
to the defect activation energies
of the perovskite layers.
............................................................................................................................................................................................................................................................................................................................................
PID 60°C, 85% humidity, ~50% loss after No test data reported for Perovskite film quality
1000-V load for a 24 hours (75) modules. For 1-cm2 cell, should be improved, and new

y g
period of 96 hours no degradation observed encapsulants and ionic
at ±500 V (120). barrier layers should
be developed (120–122).
............................................................................................................................................................................................................................................................................................................................................
Reverse bias degradation Hotspot test, partial Negligible degradation No loss for perovskite For monolithic tandems,
or full shading after biasing at minimodule with carbon- contact quality should be
−2.0 V for 12 hours (123)

y
of subcells based electrodes improved. For thin film modules,
(area of 56.8 cm2) (124). bypass diodes should be

,
integrated to the film stacks.
............................................................................................................................................................................................................................................................................................................................................

sputtering damage during the top TCO depo- to further enhance the long-term stability of deployment. Bifacial c-Si modules are rapidly
sition (71). This problem might be mitigated by perovskite/silicon tandem modules. Finally, becoming the dominant utility-scale technology
materials that create stronger interfaces, such adequate encapsulants may enhance the robust- (76). Calculations show that adopting the bifa-
as by functionalizing C60 and using low–elastic ness of perovskite modules against PID (75). cial concept also enhances the energy yield of
modulus encapsulants (72). perovskite/silicon tandems (77). Such calcula-
Care needs to be taken with regards to the Emerging concepts tions require accurate knowledge of the front
by-products released from encapsulants dur- Bifacial tandems and rear irradiance, as well as potential (non-
ing the curing process and device operation Bifacial solar cells have transparent rear con- uniform and temporal) shading from the sur-
that may also degrade the perovskite absorber tacts to exploit the diffuse and reflected light roundings (78). Bifacial tandems may also offer
(73, 74). Developing new encapsulants with from the environment and the ground (referred superior stability because they require the use
lower water vapor transmission rates and to as albedo) to enhance their current output of lower-bandgap (<1.60 eV) perovskites, usually
higher volume resistivity might be necessary by 10 to 30% relative, depending on the specific with bromide-lean compositions (79, 80).

Aydin et al., Science 383, eadh3849 (2024) 12 January 2024 8 of 13

RES EARCH | R E V I E W
For deployment, specific ground covering
with increased albedo may be considered,
provided that the costs are offset by gains in
LCOE (81). The LCOE of bifacial tandems may
further benefit from solar-tracking technolo-
gies (82). Single-axis tracking can increase the
annual energy yield of bifacial solar panels
by an additional 30 to 40% relative with only
~10% extra cost for the tracker system (76, 83).
Despite growing interest in bifacial tandems,
no consensus regarding reporting cell or module
performance is yet in place.
Triple-junction and all-perovskite tandems
With the existing materials library, PCEs ex-
ceeding 37% should be achievable for dual-
junction perovskite/silicon tandem solar cells
(the theoretical limit is ~44%) (84). Higher
practical values are doubtlessly possible with
triple-junction perovskite/perovskite/silicon
tandems (85), which have a theoretical limit of
49.4% (86). In monolithic triple-junction ar-
chitectures, current matching is again needed,
which requires an ~1.45-eV bandgap for the
middle cell, which is attainable through mixed-
metal lead-tin perovskite compositions with
a relatively high tin content. However, tin-based
perovskites are susceptible to oxidation-induced
degradation. For the top cells, an ~1.95-eV band-
gap is needed, for which mixed-halide perov-
skite compositions with a relatively high Br
content can be used. Yet, these perovskites
undergo light-induced phase segregation and
lead to higher VOC deficits at the device level
(87). Inorganic perovskites may avoid phase
segregation but will need to be made proces-
sable at relatively low temperatures (~100°C)
for monolithic integration (88).
An additional hurdle for triple-junction tan-
dems arises from the need to fabricate perovskite
subcells on top of other perovskite subcells, which
can lead to processing compatibility challenges
similar to those encountered when scaling dual-
junction perovskite/perovskite tandem cells. For-
tunately, the advances achieved in perovskite/
perovskite tandem cells can be leveraged toward
triple-junction tandems. If the present limi-
tations of the mid- and wide-bandgap perovskite
can be solved, it may even become appealing to
develop efficient and stable perovskite formula-
tions with bandgaps near that of c-Si (1.12 eV)
to construct all-perovskite tandems. Such ma-
terials would open a route to the elimination
of c-Si from the device (86). Yet, for this, the
realization of efficient and stable Sn-based pe-
rovskites becomes essential (89), which is likely a
long-term challenge.
Manufacturing sustainability
Major challenges toward the industrializa-
tion of perovskite/silicon tandems relate to
performance at scale, reliability, and CapEx.
However, the cost and sustainability of the raw
materials involved and operational expendi-
Aydin et al., Science 383, eadh3849 (2024) 12 January 2024
tures (OpEx) are also important for scaled
manufacturing. For example, all module con-
figurations require at least one TCO, where
indium-based TCOs have ideal properties such
as high transparency, conductivity, and dura-
bility (90). Although recent studies have claimed
that indium production may meet terawatt PV
goals in certain scenarios, it is crucial to remain
vigilant regarding potential supply risk issues
from indium scarcity (91, 92). Minimizing the
thickness and number of required TCOs in the
device, where 2T tandems have an advantage,
can lower such risks (53).
Tandem solar cells contain several charge-
selective contacts, and the choice of the contact
materials is largely determined by the device po-
larity. In p-i-n tandems, molecular self-assembled
monolayers (e.g., 2PACz, Me-4PACz, or deriv-
atives thereof), show promise as low-loss hole
transport layers (HTLs). However, their large-
area performance and long-term stability have
not yet been investigated thoroughly. In the
front contacts of monolithic tandems, C60 is
frequently used as an ETL. Despite this, C60 is
performance limiting owing to high parasitic
absorption losses in the UV, recombination
when interfaced directly with perovskite, and—
as already stated—mechanical reliability issues
(71, 93–95). Combined, this argues for its replace-
ment, but alternatives remain largely unidentified.
Vacuum deposition of metal oxides as the ETL
(such as SnO2 or other n-type metal oxides) di-
rectly onto the perovskite layers (without C60)
may be the most attractive industrial approach;
however, this poses challenges related to the
chemical stability in their interface with perov-
skite (96). It can thus be anticipated that initial
commercial perovskite/silicon tandems will rely
on C60 as an ETL, but alternatives are needed.
Perovskite/silicon tandems can be fabricated
in the n-i-p configuration as well, where the
perovskite ETL is deposited on the silicon bot-
tom cell, and the HTL side faces the Sun. In this
case, initial tandems relied on organic mate-
rials, in particular spiro-OMeTAD; however,
their production is complex and costly. More-
over, these organic molecules often need to be
doped to activate their charge-selective prop-
erties, but this results in device stability issues.
Inorganic HTLs, such as CuSCN or NiOx, may be
promising alternatives if they can avoid dam-
aging the perovskite layers underneath (97).
The scarcity of silver as the primary material
for grid-electrode metallization is an increasing
concern for c-Si PV (98), and silver consumption
in wafer-based tandem solar cells could be re-
laxed compared with SHJ cells because the
currents are halved. However, cost-efficient,
low-temperature silver pastes, or alternatives
using copper, are required.
Lead leakage from the module and the use of
hazardous solvents in manufacturing are the
two major concerns for perovskite-based tan-
dem PVs (99), but they might be mitigated by
module inspection and replacement before cat-
astrophic failure. Moreover, recent studies have
argued that in the case of catastrophic module
failure, the environmental impact will only be
moderate (100, 101). For solution-processed
perovskites, toxicity concerns arising from the
solvents [such as dimethylformamide (DMF)
and dimethyl sulphoxide (DMSO)] should be
mitigated through waste management and pu-
rification and recycling processes (102). Al-
ternatively, green solvents can be considered,
together with vacuum deposition approaches,
provided that the performances of the perov-
skite layer are preserved.
Market status: Challenges and opportunities
Tandem PV technologies have been projected
to take a more than $10 billion market share by
2032 (6). However, predicting a realistic date for
successful market entrance of perovskite/silicon
tandems remains challenging (103). Currently,
p
both start-ups and established PV companies
are exploring this technology, pursuing a var-
iety of different fabrication routes and module
configurations (104). Table 2 summarizes the
highest perovskite/silicon tandem PCEs at dif-
ferent scales, either as a published study or a
g
company announcement.
In any scenario, the technology must success-
fully pass the various technology readiness lev-
els (TRLs) before commercialization is possible,
y
including laboratory research (TRLs 1 to 5),
pilot line (TRLs 5 to 8), and mass production
(TRLs 8 and 9), which are shown in Fig. 6A.
The PV industry has accumulated substantial
knowledge and expertise in both thin-film and
silicon technologies over the years, which ne-
cessitated extensive and costly R&D efforts
spanning several years, as shown in Fig. 6B.
Much of this can be leveraged in terms of time
and budget requirements for perovskite/silicon
y g
tandem technology. Nevertheless, substantial
and sustained funding is essential for its suc-
cessful commercialization. Currently, an R&D
laboratory focused on perovskite subcell (or
submodule) development can be established
with a budget of less than ~€5 million. Mean-
y
while, several companies have already announced
,
investments exceeding $100 million for estab-
lishing pilot lines with limited information
disclosure (Table 3). For reference, the ap-
proximate CapEx budget for 1 GW of SHJ pro-
duction in China is ~€50 million, based on the
most up-to-date information available. The given
expenses could vary considerably depending
on the location. It is anticipated that the ex-
penses linked to gigawatt-scale manufactur-
ing of perovskite subcells will soon decrease to
below €100 million, with further reductions
anticipated in the future.
One of the key enablers in cost reduction in
c-Si PV production has been the notable in-
creases in production throughput, with c-Si
processing steps reaching production rates of
9 of 13
RES EARCH | R E V I E W

Table 2. Summary of the state-of-the-art perovskite/silicon tandem cells for various device architectures. Here, PGDBIFI,20 denotes 0.2 sun rear
illumination. A dash indicates data unavailable.

Conditions 2T 3T 4T
PCEs for ~1 cm2 33.9% (Longi) (2) 23.5% (SNU) (125) 29.2% (Solliance) (126)
............................................................................................................................................................................................................................................................................................................................................
PCEs at larger scale 28.6% (OxfordPV, 258 cm2) (15) 29.56% (CSEM-Meyer Burger, 24.5 cm2) (127) 28.4% (Kaneka, 64 cm2) (15)
............................................................................................................................................................................................................................................................................................................................................
PGDBIFI,20 for bifacial tandems 27.1 mW cm−2 (80) – 30.5 mW cm−2 (14)
............................................................................................................................................................................................................................................................................................................................................

A Advanced
Early proof- Early Early Advanced Industrializa-
Creation proof-of- Feasibility Pilot Maturity
of-concept scaling feasibility scaling tion
concept

Laboratory stage Pilot line stage Mass production

B 1012 D
Cumulative cost per TRL (Euro)

1010
Billion €

108

p
106
Million €
4
10

102

100
0 1 2 3 4 5 6 7 8 9

g
TRL of c-Si PV technologies
C
Required Capex for 1 GW capacity (%)

6 x 3600 wph
100

80 4 x 5500 wph

y
3 x 8000 wph
60

2 x 10000 wph
40

20

0
Number of equipment x Equipment througput

Fig. 6. Industrialization aspects of perovskite/silicon tandems. (A) TRLs of CapEx for ~1-GW physical vapor deposition line for various numbers of equipment and

y g
various stages of PV development. (B) Cost of the development of silicon PV capacity, demonstrating that increasing the equipment throughput reduces the CapEx
technologies at different TRL levels, each usually spanning several years (with (the data are used with permission from Von Ardenne GmbH, https://vonardenne.de/).
permission from Fraunhofer, https://www.fraunhofer.org/). Perovskite/silicon (D) Illustration of representative monolithic and mechanically stacked tandem
tandem technologies are expected to benefit from these developments, lowering solar cell processing lines. The inset shows the most promising scalable techniques
the required cost to become technologically mature. (C) The required for perovskite layer deposition in high-throughput processing lines.

y
,
more than 10,000 wafers per hour. Similar have packaging procedures similar to those of far typically require several seconds and power
processing throughputs will be required for c-Si single-junction modules and only neces- stabilization.
perovskite/silicon tandems to reach com- sitate wafer-scale processing steps for the pe-
petitive LCOE values. Figure 6C illustrates an rovskite subcells. A recent techno-economic REFERENCES AND NOTES
example that increasing the throughput of a analysis corroborates that the production cost 1. N. M. Haegel et al., Terawatt-scale photovoltaics: Transform
physical vapor deposition line by approx- of mechanically stacked tandems might be global energy. Science 364, 836–838 (2019). doi: 10.1126/
imately threefold can reduce the CapEx by higher compared with that of their mono- science.aaw1845; pmid: 31147512
2. National Renewable Energy Laboratory, “Best Research-Cell
>50% and thus decrease the LCOE. In Fig. lithic counterparts (105). Efficiency Chart”; https://www.nrel.gov/pv/cell-efficiency.html.
6D, conceptual processing lines for mechan- Additional challenges arise in characteriz- 3. B. A. Veith-Wolf, S. Schäfer, R. Brendel, J. Schmidt,
ically stacked and monolithic tandem cells ing industrially produced perovskite/silicon tan- Reassessment of intrinsic lifetime limit in n-type crystalline
silicon and implication on maximum solar cell efficiency.
are depicted. Mechanically stacked tandems dem technology. Current-voltage (I–V) analysis Sol. Energy Mater. Sol. Cells 186, 194–199 (2018).
require extra processing steps for the perov- within light exposure at the millisecond time- doi: 10.1016/j.solmat.2018.06.029
skite submodules, including multiple laser scale is commonly used for industrial char- 4. S. Schäfer, R. Brendel, Accurate Calculation of the
Absorptance Enhances Efficiency Limit of Crystalline
scribing, and they demand intricate module acterization of c-Si cells. New methods for
Silicon Solar Cells With Lambertian Light Trapping.
assembly lines, which leads to increased over- fast and reliable measurements of perovskite/ IEEE J. Photovolt. 8, 1156–1158 (2018). doi: 10.1109/
all CapEx. On the other hand, monolithic cells silicon tandems must be developed, which so JPHOTOV.2018.2824024

Aydin et al., Science 383, eadh3849 (2024) 12 January 2024 10 of 13

RES EARCH | R E V I E W

Table 3. Overview of consortia and companies developing monolithic perovskite/silicon tandem solar cells. Note, although the information provided in
the tables is provided in good faith, the authors, editors, and publishers cannot accept direct responsibility for any errors or omissions.

Country Company or consortia Budget Duration Remarks

Europe PEPPERONI Project €18.8 million 2022–2026 17 partners from 12 countries aim to bring the TRL level of 2T tandems
from 4 to 5 up to TRL 7, with a long-term goal of reaching gigawatt-scale
manufacturing in Europe.
............................................................................................................................................................................................................................................................................................................................................
Europe NEXUS Project €3.6 million 2022–2025 12 partners from nine countries aim to reach >30% tandem module PCE in
2T configuration by using dry perovskite processing.
............................................................................................................................................................................................................................................................................................................................................
Europe PrEsto Project €4.25 million 2021–2024 Nine partners from three countries aim to identify suitable
production processes for scaled 2T tandems and perform
techno-economic evaluations.
............................................................................................................................................................................................................................................................................................................................................
UK Oxford PV $142 million Since 2010 Oxford PV is the first perovskite-based tandem PV company and
received investment from multiple sources. The company currently holds
the highest certified PCE (28.6%) for industry-size 2T tandems (274 cm2).
The budget data are acquired from www.crunchbase.com.
............................................................................................................................................................................................................................................................................................................................................
USA TEAMUP Project $12.2 million Since 2023 12 partners from academic, industrial, and federal laboratories in
the US aim to enhance the stability of both 2T and 4T tandems
in real-world conditions.

p
............................................................................................................................................................................................................................................................................................................................................
USA ADDEPT Project $11.25 million Since 2023 Eight partners from academic institutes and companies from the US aim
to enhance the performance and durability of both 2T and 4T tandem
cells through machine learning and robot screening.
............................................................................................................................................................................................................................................................................................................................................
USA PV PACT Project $9 million Since 2021 12 partners including national laboratories, universities, and companies
from the US aim to develop performance and reliability test
protocols for industrial applications.

g
............................................................................................................................................................................................................................................................................................................................................
China Auner >$14 million Since 2017 Auner is a start-up company and claims a 32.44% PCE for 1 cm2 and
30.83% certified PCE on 25 cm2 aperture area. The company reports
that an industrial-size pilot line is under construction.
............................................................................................................................................................................................................................................................................................................................................
China Jinko Solar >$6.6 million Since 2021 Jinko ranks first in cumulative PV module shipment

y
in the world. The company reported 32.33% PCE for
perovskite/TOPCon 2T tandems, with limited information about
other tandem research activities.
............................................................................................................................................................................................................................................................................................................................................
China Trina Solar >$17 million Since 2021 Trina claims >31% laboratory-scale PCE, and the company is one of the
leading enterprises (top three) in silicon solar cell and PV modules.
The given budget is for the initial investment.
............................................................................................................................................................................................................................................................................................................................................
China Longi No information No information Longi is the highest–market value PV company in China, holding the
record for the PCE achieved in SHJ cells (usually used as bottom
cells). The company reported 33.9% PCE for ~1-cm2 2T tandems,
and there is no other public release.
............................................................................................................................................................................................................................................................................................................................................

y g
China Tongwei No information Since 2022 Tongwei stands as the foremost company in the shipment of polysilicon
and c-Si cells, asserting an impressive 31.1% PCE for 2T tandems,
albeit with only limited public disclosure.
............................................................................................................................................................................................................................................................................................................................................
Japan Kaneka No information No information Kaneka is an expert in amorphous silicon and has
held the world record for interdigitated back contact (IBC) cells for

y
several years. Also, the company has background with a-Si thin-film
tandem cells. Kaneka released 29.2% PCE (~1 cm2) for 2T

,
tandems with no other public release about R&D activities.
............................................................................................................................................................................................................................................................................................................................................
Korea Hanwha Q Cells $100 million 2023–2024 This Korea-based company, with its research hub in Germany,
has disclosed 29.9% PCE achievement for 2T tandem solar
cells, covering an area of 1 cm2. They have also earmarked
$100 million for a pilot production line in Jincheon, Korea,
and they are one of the partners of the PEPPERONI project.
............................................................................................................................................................................................................................................................................................................................................
Singapore SERIS and REC Solar $57 million 2023–2025 SERIS and Singapore-based panel manufacturer REC aim to develop a
>30%-PCE 2T tandem module in the REC@NUS corporate R&D laboratory.
............................................................................................................................................................................................................................................................................................................................................

5. Z. Yu, M. Leilaeioun, Z. Holman, Selecting tandem partners for 7. M. De Bastiani et al., Recombination junctions for efficient 9. P. Wagner, P. Tockhorn, S. Hall, S. Albrecht, L. Korte, Performance of
silicon solar cells. Nat. Energy 1, 16137 (2016). doi: 10.1038/ monolithic perovskite-based tandem solar cells: Physical Monolithic Two- and Three-Terminal Perovskite/Silicon Tandem
nenergy.2016.137 principles, properties, processing and prospects. Mater. Horiz. 7, Solar Cells Under Varying Illumination Conditions. Sol. RRL 7,
6. VDMA Photovoltaic Equipment, “International 2791–2809 (2020). doi: 10.1039/D0MH00990C 2200954 (2023). doi: 10.1002/solr.202200954
Technology Roadmap for Photovoltaic (ITRPV)” (2023); 8. E. L. Warren et al., A Taxonomy for Three-Terminal Tandem 10. Z. C. Holman et al., Infrared light management in high-efficiency
https://www.vdma.org/international-technology-roadmap- Solar Cells. ACS Energy Lett. 5, 1233–1242 (2020). silicon heterojunction and rear-passivated solar cells. J. Appl.
photovoltaic. doi: 10.1021/acsenergylett.0c00068 Phys. 113, 013107 (2013). doi: 10.1063/1.4772975

Aydin et al., Science 383, eadh3849 (2024) 12 January 2024 11 of 13

RES EARCH | R E V I E W
11. E. Aydin et al., Interplay between temperature and bandgap 35.
energies on the outdoor performance of perovskite/silicon
tandem solar cells. Nat. Energy 5, 851–859 (2020).
doi: 10.1038/s41560-020-00687-4 36.
12. B. Chen et al., Enhanced optical path and electron diffusion length
enable high-efficiency perovskite tandems. Nat. Commun. 11,
1257 (2020). doi: 10.1038/s41467-020-15077-3; pmid: 32152324
13. S. Kim et al., Over 30% efficiency bifacial 4-terminal 37.
perovskite-heterojunction silicon tandem solar cells with
spectral albedo. Sci. Rep. 11, 15524 (2021). doi: 10.1038/
s41598-021-94848-4; pmid: 34330966
14. G. Coletti et al., Bifacial Four-Terminal Perovskite/Silicon 38.
Tandem Solar Cells and Modules. ACS Energy Lett. 5,
1676–1680 (2020). doi: 10.1021/acsenergylett.0c00682
15. M. A. Green et al., Solar cell efficiency tables (version 62).
Prog. Photovolt. Res. Appl. 31, 651–663 (2023). doi: 10.1002/ 39.
pip.3726
16. T. G. Allen, J. Bullock, X. Yang, A. Javey, S. De Wolf, Passivating
contacts for crystalline silicon solar cells. Nat. Energy 4, 40.
914–928 (2019). doi: 10.1038/s41560-019-0463-6
17. S. De Wolf, A. Descoeudres, Z. C. Holman, C. Ballif, High-
efficiency Silicon Heterojunction Solar Cells: A Review. Green 2, 41.
7–24 (2012). doi: 10.1515/green-2011-0018
18. J. Zheng et al., Efficient monolithic perovskite–Si tandem
solar cells enabled by an ultra-thin indium tin oxide interlayer.
Energy Environ. Sci. 16, 1223–1233 (2023). doi: 10.1039/ 42.
D2EE04007G
19. A. Razzaq, T. G. Allen, W. Liu, Z. Liu, S. De Wolf, Silicon
heterojunction solar cells: Techno-economic assessment and
opportunities. Joule 6, 514–542 (2022). doi: 10.1016/ 43.
j.joule.2022.02.009
20. N.-G. Park, K. Zhu, Scalable fabrication and coating methods
for perovskite solar cells and solar modules. Nat. Rev. Mater.
5, 333–350 (2020). doi: 10.1038/s41578-019-0176-2 44.
21. Z. Li et al., Scalable fabrication of perovskite solar cells. Nat.
Rev. Mater. 3, 18017 (2018). doi: 10.1038/natrevmats.2018.17
22. Y. Hou et al., Efficient tandem solar cells with solution-
processed perovskite on textured crystalline silicon. Science 45.
367, 1135–1140 (2020). doi: 10.1126/science.aaz3691;
pmid: 32139544
23. B. Chen et al., Blade-Coated Perovskites on Textured
Silicon for 26%-Efficient Monolithic Perovskite/Silicon Tandem 46.
Solar Cells. Joule 4, 850–864 (2020). doi: 10.1016/
j.joule.2020.01.008
24. A. S. Subbiah et al., High-Performance Perovskite Single-
Junction and Textured Perovskite/Silicon Tandem Solar Cells 47.
via Slot-Die-Coating. ACS Energy Lett. 5, 3034–3040 (2020).
doi: 10.1021/acsenergylett.0c01297
25. F. Fu et al., Monolithic Perovskite-Silicon Tandem Solar 48.
Cells: From the Lab to Fab? Adv. Mater. 34, e2106540 (2022).
doi: 10.1002/adma.202106540; pmid: 35060205
26. K. Jäger, J. Sutter, M. Hammerschmidt, P.-I. Schneider, 49.
C. Becker, Prospects of light management in perovskite/silicon
tandem solar cells. Nanophotonics 10, 1991–2000 (2020).
doi: 10.1515/nanoph-2020-0674 50.
27. E. Köhnen et al., 27.9% Efficient Monolithic Perovskite/
Silicon Tandem Solar Cells on Industry Compatible Bottom
Cells. Sol. RRL 5, 2100244 (2021). doi: 10.1002/
solr.202100244 51.
28. J. Park et al., Controlled growth of perovskite layers with
volatile alkylammonium chlorides. Nature 616, 724–730
(2023). doi: 10.1038/s41586-023-05825-y; pmid: 36796426
29. P. Tockhorn et al., Nano-optical designs for high-efficiency 52.
monolithic perovskite-silicon tandem solar cells. Nat.
Nanotechnol. 17, 1214–1221 (2022). doi: 10.1038/
s41565-022-01228-8; pmid: 36280763 53.
30. K. Xu et al., Slot-Die Coated Triple-Halide Perovskites for
Efficient and Scalable Perovskite/Silicon Tandem Solar Cells.
ACS Energy Lett. 7, 3600–3611 (2022). doi: 10.1021/ 54.
acsenergylett.2c01506; pmid: 36277135
31. Y. Deng et al., Surfactant-controlled ink drying enables
high-speed deposition of perovskite films for efficient 55.
photovoltaic modules. Nat. Energy 3, 560–566 (2018).
doi: 10.1038/s41560-018-0153-9
32. Y. Yu et al., A Review on Gas-Quenching Technique for 56.
Efficient Perovskite Solar Cells. Sol. RRL 5, 2100386 (2021).
doi: 10.1002/solr.202100386
33. W.-Q. Wu et al., Bilateral alkylamine for suppressing charge 57.
recombination and improving stability in blade-coated
perovskite solar cells. Sci. Adv. 5, eaav8925 (2019).
doi: 10.1126/sciadv.aav8925; pmid: 30873433
34. K. Liu et al., Zwitterionic-Surfactant-Assisted Room-Temperature 58.
Coating of Efficient Perovskite Solar Cells. Joule 4, 2404–2425
(2020). doi: 10.1016/j.joule.2020.09.011
Aydin et al., Science 383, eadh3849 (2024) 12 January 2024
A. Verma et al., Towards industrialization of perovskite solar
cells using slot die coating. J. Mater. Chem. C 8, 6124–6135
(2020). doi: 10.1039/D0TC00327A
M. Rai, L. H. Wong, L. Etgar, Effect of Perovskite Thickness on
Electroluminescence and Solar Cell Conversion Efficiency.
J. Phys. Chem. Lett. 11, 8189–8194 (2020). doi: 10.1021/
acs.jpclett.0c02363; pmid: 32893636
A. Farag et al., Efficient Light Harvesting in Thick Perovskite
Solar Cells Processed on Industry-Applicable Random
Pyramidal Textures. ACS Appl. Energy Mater. 5, 6700–6708
(2022). doi: 10.1021/acsaem.1c04028
M. Roß et al., Co-Evaporated Formamidinium Lead Iodide Based
Perovskites with 1000 h Constant Stability for Fully Textured
Monolithic Perovskite/Silicon Tandem Solar Cells. Adv. Energy
Mater. 11, 2101460 (2021). doi: 10.1002/aenm.202101460
H. Li et al., Sequential vacuum-evaporated perovskite solar
cells with more than 24% efficiency. Sci. Adv. 8, eabo7422
(2022). doi: 10.1126/sciadv.abo7422; pmid: 35857518
J. Li et al., Highly Efficient Thermally Co-evaporated
Perovskite Solar Cells and Mini-modules. Joule 4, 1035–1053
(2020). doi: 10.1016/j.joule.2020.03.005
C. Momblona et al., Efficient vacuum deposited p-i-n and
n-i-p perovskite solar cells employing doped charge transport
layers. Energy Environ. Sci. 9, 3456–3463 (2016).
doi: 10.1039/C6EE02100J
Y.-H. Chiang et al., Vacuum-Deposited Wide-Bandgap Perovskite
for All-Perovskite Tandem Solar Cells. ACS Energy Lett. 8,
2728–2737 (2023). doi: 10.1021/acsenergylett.3c00564;
pmid: 37324541
M. Kroll et al., Insights into the evaporation behaviour
of FAI: Material degradation and consequences for perovskite
solar cells. Sustain. Energy Fuels 6, 3230–3239 (2022).
doi: 10.1039/D2SE00373B
J. Feng et al., High-throughput large-area vacuum deposition for
high-performance formamidine-based perovskite solar cells.
Energy Environ. Sci. 14, 3035–3043 (2021). doi: 10.1039/
D1EE00634G
S. Siegrist et al., Triple-cation perovskite solar cells
fabricated by a hybrid PVD/blade coating process using
green solvents. J. Mater. Chem. A 9, 26680–26687 (2021).
doi: 10.1039/D1TA07579A; pmid: 34976398
F. Sahli et al., Fully textured monolithic perovskite/silicon
tandem solar cells with 25.2% power conversion efficiency.
Nat. Mater. 17, 820–826 (2018). doi: 10.1038/s41563-018-0115-4;
pmid: 29891887
X. Luo et al., Efficient Perovskite/Silicon Tandem Solar Cells on
Industrially Compatible Textured Silicon. Adv. Mater. 35, e2207883
(2023). doi: 10.1002/adma.202207883; pmid: 36599055
E. Aydin, M. De Bastiani, S. De Wolf, Defect and Contact
Passivation for Perovskite Solar Cells. Adv. Mater. 31, e1900428
(2019). doi: 10.1002/adma.201900428; pmid: 31062907
F. Sahli et al., Vapor Transport Deposition of Methylammonium
Iodide for Perovskite Solar Cells. ACS Appl. Energy Mater. 4,
4333–4343 (2021). doi: 10.1021/acsaem.0c02999
L. Mao et al., Fully Textured, Production-Line Compatible
Monolithic Perovskite/Silicon Tandem Solar Cells
Approaching 29% Efficiency. Adv. Mater. 34, e2206193
(2022). doi: 10.1002/adma.202206193; pmid: 35985840
N. Rodkey et al., Pulsed Laser Deposition of Cs2AgBiBr6:
From Mechanochemically Synthesized Powders to Dry,
Single-Step Deposition. Chem. Mater. 33, 7417–7422 (2021).
doi: 10.1021/acs.chemmater.1c02054; pmid: 34602745
Y. Rong et al., Challenges for commercializing perovskite
solar cells. Science 361, eaat8235 (2018). doi: 10.1126/
science.aat8235; pmid: 30237326
E. Aydin et al., Enhanced optoelectronic coupling for
perovskite/silicon tandem solar cells. Nature 623, 732–738
(2023). doi: 10.1038/s41586-023-06667-4; pmid: 37769785
J. P. Mailoa et al., A 2-terminal perovskite/silicon multijunction
solar cell enabled by a silicon tunnel junction. Appl. Phys. Lett.
106, 121105 (2015). doi: 10.1063/1.4914179
S. Albrecht et al., Monolithic perovskite/silicon-heterojunction
tandem solar cells processed at low temperature. Energy
Environ. Sci. 9, 81–88 (2016). doi: 10.1039/C5EE02965A
G. E. Eperon, M. T. Hörantner, H. J. Snaith, Metal halide
perovskite tandem and multiple-junction photovoltaics. Nat. Rev.
Chem. 1, 0095 (2017). doi: 10.1038/s41570-017-0095
J. Xu et al., Triple-halide wide-band gap perovskites with
suppressed phase segregation for efficient tandems.
Science 367, 1097–1104 (2020). doi: 10.1126/science.
aaz5074; pmid: 32139537
L. Xu et al., Heat generation and mitigation in silicon solar
cells and modules. Joule 5, 631–645 (2021). doi: 10.1016/
j.joule.2021.01.012
59. J. F. Weaver, “How the numbers may work for 50-year solar”
(PVMagazine, 2021); https://www.pv-magazine.com/2021/
11/05/how-the-numbers-may-work-for-50-year-solar/.
60. S. J. Andrews, B. Smith, M. G. Deceglie, K. A. Horowitz,
T. J. Silverman, “NREL Comparative PV LCOE Calculator”
(2021); https://www.nrel.gov/pv/lcoe-calculator/.
61. J. Qian, M. Ernst, N. Wu, A. Blakers, Impact of perovskite
solar cell degradation on the lifetime energy yield and
economic viability of perovskite/silicon tandem modules.
Sustain. Energy Fuels 3, 1439–1447 (2019). doi: 10.1039/
C9SE00143C
62. M. Babics et al., One year outdoor operation of monolithic
perovskite/silicon tandem solar cell. Cell Rep. Phys. Sci. 4,
101280 (2023). doi: 10.1016/j.xcrp.2023.101280
63. S. M. Park et al., Engineering ligand reactivity enables
high-temperature operation of stable perovskite solar cells.
Science 381, 209–215 (2023). doi: 10.1126/science.adi4107;
pmid: 37440655
64. R. Azmi et al., Damp heat-stable perovskite solar cells with
tailored-dimensionality 2D/3D heterojunctions. Science 376,
73–77 (2022). doi: 10.1126/science.abm5784; pmid: 35175829
65. L. Schellhaas, J. Stein, PACT Perovskite PV Module Stress
Testing Protocol Version 0.0. (Perovskite PV Accelerator for
Commercializing Technologies, 2022); https://pvpact.sandia.
gov/publications-and-protocols/.
66. X. Zhao et al., Accelerated aging of all-inorganic, interface-
stabilized perovskite solar cells. Science 377, 307–310
p
(2022). doi: 10.1126/science.abn5679; pmid: 35709247
67. Q. Jiang et al., Towards linking lab and field lifetimes of
perovskite solar cells. Nature 623, 313–318 (2023).
doi: 10.1038/s41586-023-06610-7; pmid: 37696288
68. M. De Bastiani et al., All Set for Efficient and Reliable
Perovskite/Silicon Tandem Photovoltaic Modules? Sol. RRL 6,
2100493 (2022). doi: 10.1002/solr.202100493
69. Y.-H. Lin et al., A piperidinium salt stabilizes efficient metal-
g
halide perovskite solar cells. Science 369, 96–102 (2020).
doi: 10.1126/science.aba1628; pmid: 32631893
70. G. Divitini et al., In situ observation of heat-induced
degradation of perovskite solar cells. Nat. Energy 1, 15012
(2016). doi: 10.1038/nenergy.2015.12
y
71. M. De Bastiani et al., Mechanical Reliability of Fullerene/Tin
Oxide Interfaces in Monolithic Perovskite/Silicon Tandem
Cells. ACS Energy Lett. 7, 827–833 (2022). doi: 10.1021/
acsenergylett.1c02148
72. R. Cheacharoen et al., Design and understanding of
encapsulated perovskite solar cells to withstand temperature
cycling. Energy Environ. Sci. 11, 144–150 (2018).
doi: 10.1039/C7EE02564E
73. M. D. Kempe et al., Acetic acid production and glass
transition concerns with ethylene-vinyl acetate used in
photovoltaic devices. Sol. Energy Mater. Sol. Cells 91,
315–329 (2007). doi: 10.1016/j.solmat.2006.10.009
74. R. Cheacharoen et al., Encapsulating perovskite solar cells to
withstand damp heat and thermal cycling. Sustain. Energy
y g
Fuels 2, 2398–2406 (2018). doi: 10.1039/C8SE00250A
75. L. Xu et al., Potential-induced degradation in perovskite/
silicon tandem photovoltaic modules. Cell Rep. Phys. Sci. 3,
101026 (2022). doi: 10.1016/j.xcrp.2022.101026
76. R. Kopecek, J. Libal, Bifacial Photovoltaics 2021: Status,
Opportunities and Challenges. Energies 14, 2076 (2021).
doi: 10.3390/en14082076
y
77. J. Lehr et al., Energy yield of bifacial textured perovskite/
silicon tandem photovoltaic modules. Sol. Energy Mater. Sol.
,
Cells 208, 110367 (2020). doi: 10.1016/j.solmat.2019.110367
78. A. Onno et al., Predicted Power Output of Silicon-Based
Bifacial Tandem Photovoltaic Systems. Joule 4, 580–596
(2020). doi: 10.1016/j.joule.2019.12.017
79. M. De Bastiani et al., Efficient bifacial monolithic
perovskite/silicon tandem solar cells via bandgap
engineering. Nat. Energy 6, 167–175 (2021). doi: 10.1038/
s41560-020-00756-8
80. E. Ugur et al., Front-contact passivation through 2D/3D
perovskite heterojunctions enables efficient bifacial
perovskite/silicon tandem solar cells. Matter 6, 2919–2934
(2023). doi: 10.1016/j.matt.2023.05.028
81. M. T. Patel, R. Asadpour, J. Bin Jahangir, M. Ryyan Khan,
M. A. Alam, Current-matching erases the anticipated
performance gain of next-generation two-terminal
Perovskite-Si tandem solar farms. Appl. Energy 329, 120175
(2023). doi: 10.1016/j.apenergy.2022.120175
82. M. Babics et al., Unleashing the Full Power of Perovskite/
Silicon Tandem Modules with Solar Trackers. ACS Energy
Lett. 7, 1604–1610 (2022). doi: 10.1021/
acsenergylett.2c00442
12 of 13
RES EARCH | R E V I E W
83. C. D. Rodríguez-Gallegos et al., Global Techno-Economic
Performance of Bifacial and Tracking Photovoltaic Systems.
Joule 4, 1514–1541 (2020). doi: 10.1016/j.joule.2020.05.005 99.
84. T. Leijtens, K. A. Bush, R. Prasanna, M. D. McGehee,
Opportunities and challenges for tandem solar cells using
metal halide perovskite semiconductors. Nat. Energy 3, 100.
828–838 (2018). doi: 10.1038/s41560-018-0190-4
85. J. Werner et al., Perovskite/Perovskite/Silicon Monolithic
Triple-Junction Solar Cells with a Fully Textured Design.
ACS Energy Lett. 3, 2052–2058 (2018). doi: 10.1021/ 101.
acsenergylett.8b01165
86. I. Almansouri, A. Ho-Baillie, S. P. Bremner, M. A. Green,
Supercharging Silicon Solar Cell Performance by Means of
Multijunction Concept. IEEE J. Photovolt. 5, 968–976 (2015). 102.
doi: 10.1109/JPHOTOV.2015.2395140
87. E. T. Hoke et al., Reversible photo-induced trap formation in
mixed-halide hybrid perovskites for photovoltaics. Chem. Sci. 6, 103.
613–617 (2015). doi: 10.1039/C4SC03141E; pmid: 28706629
88. Z. Wang et al., Suppressed phase segregation for triple-
junction perovskite solar cells. Nature 618, 74–79 (2023).
doi: 10.1038/s41586-023-06006-7; pmid: 36977463 104.
89. M. T. Hörantner et al., The Potential of Multijunction
Perovskite Solar Cells. ACS Energy Lett. 2, 2506–2513
(2017). doi: 10.1021/acsenergylett.7b00647 105.
90. M. Morales-Masis, S. De Wolf, R. Woods-Robinson, J. W. Ager,
C. Ballif, Transparent Electrodes for Efficient Optoelectronics.
Adv. Electron. Mater. 3, 1600529 (2017). doi: 10.1002/
aelm.201600529 106.
91. B. A. McNulty, S. M. Jowitt, Barriers to and uncertainties in
understanding and quantifying global critical mineral and
element supply. iScience 24, 102809 (2021). doi: 10.1016/
j.isci.2021.102809; pmid: 34308296 107.
92. Y. Zhang, M. Kim, L. Wang, P. Verlinden, B. Hallam, Design
considerations for multi-terawatt scale manufacturing of
existing and future photovoltaic technologies: Challenges and
opportunities related to silver, indium and bismuth 108.
consumption. Energy Environ. Sci. 14, 5587–5610 (2021).
doi: 10.1039/D1EE01814K
93. J. Liu et al., Efficient and stable perovskite-silicon tandem solar 109.
cells through contact displacement by MgFx. Science 377,
302–306 (2022). doi: 10.1126/science.abn8910; pmid: 35737811
94. J. Warby et al., Understanding Performance Limiting Interfacial
Recombination in pin Perovskite Solar Cells. Adv. Energy Mater. 110.
12, 2103567 (2022). doi: 10.1002/aenm.202103567
95. E. Aydin et al., Sputtered transparent electrodes for optoelectronic
devices: Induced damage and mitigation strategies. Matter 4, 111.
3549–3584 (2021). doi: 10.1016/j.matt.2021.09.021
96. A. F. Palmstrom et al., Interfacial Effects of Tin Oxide Atomic 112.
Layer Deposition in Metal Halide Perovskite Photovoltaics.
Adv. Energy Mater. 8, 1800591 (2018). doi: 10.1002/
aenm.201800591 113.
97. N. Arora et al., Perovskite solar cells with CuSCN hole
extraction layers yield stabilized efficiencies greater
than 20. Science 358, 768–771 (2017). doi: 10.1126/science.
114.
aam5655; pmid: 28971968
98. A. u. Rehman et al., Electrode metallization for scaled
perovskite/silicon tandem solar cells: Challenges
Aydin et al., Science 383, eadh3849 (2024) 12 January 2024
and opportunities. Prog. Photovolt. Res. Appl. 31, 429–442
(2023). doi: 10.1002/pip.3499
G. S. Han et al., Cutting-Edge Studies Toward Commercialization
of Large Area Solution-Processed Perovskite Solar Cells. Adv.
Mater. Technol. 8, 2201387 (2023). doi: 10.1002/admt.202201387
F. Schmidt, L. Ledermann, A. Schäffer, H. J. Snaith, M. Lenz,
Rapid sequestration of perovskite solar cell-derived lead in
soil. J. Hazard. Mater. 436, 128995 (2022). doi: 10.1016/
j.jhazmat.2022.128995; pmid: 35525217
S. A. Khalifa et al., Fate and Exposure Assessment of
Pb Leachate from Hypothetical Breakage Events of Perovskite
Photovoltaic Modules. Environ. Sci. Technol. 57, 5190–5202
(2023). doi: 10.1021/acs.est.2c05815; pmid: 36966415
R. Vidal et al., Assessing health and environmental impacts of
solvents for producing perovskite solar cells. Nat. Sustain. 4,
277–285 (2021). doi: 10.1038/s41893-020-00645-8
T. D. Siegler et al., The Path to Perovskite Commercialization:
A Perspective from the United States Solar Energy
Technologies Office. ACS Energy Lett. 7, 1728–1734 (2022).
doi: 10.1021/acsenergylett.2c00698
L. Duan et al., Stability challenges for the commercialization
of perovskite–silicon tandem solar cells. Nat. Rev. Mater. 8,
261–281 (2023). doi: 10.1038/s41578-022-00521-1
M. Hull, J. Rousset, V. S. Nguyen, P.-P. Grand, L. Oberbeck,
Prospective Techno-Economic Analysis of 4T and 2T Perovskite
on Silicon Tandem Photovoltaic Modules at GW-Scale Production.
Sol. RRL 7, 2300503 (2023). doi: 10.1002/solr.202300503
M. De Bastiani et al., Monolithic Perovskite/Silicon Tandems
with >28% Efficiency: Role of Silicon-Surface Texture on
Perovskite Properties. Adv. Funct. Mater. 33, 2205557
(2023). doi: 10.1002/adfm.202205557
E. Aydin et al., Ligand-bridged charge extraction and enhanced
quantum efficiency enable efficient n–i–p perovskite/silicon
tandem solar cells. Energy Environ. Sci. 14, 4377–4390 (2021).
doi: 10.1039/D1EE01206A
X. Y. Chin et al., Interface passivation for 31.25%-efficient
perovskite/silicon tandem solar cells. Science 381, 59–63
(2023). doi: 10.1126/science.adg0091; pmid: 37410835
C. Altinkaya et al., Tin Oxide Electron-Selective Layers for
Efficient, Stable, and Scalable Perovskite Solar Cells.
Adv. Mater. 33, 2005504 (2021). doi: 10.1002/
adma.202005504; pmid: 33660306
D. C. Jordan, N. Haegel, T. M. Barnes, Photovoltaics module
reliability for the terawatt age. Prog. Energy 4, 022002
(2022). doi: 10.1088/2516-1083/ac6111
Our World in Data, “Solar (photovoltaic) panel prices”;
https://ourworldindata.org/grapher/solar-pv-prices.
Solliance, “Scaled perovskite solar modules pass three critical
stability tests” (2020); https://www.solliance.eu/2020/scaled-
perovskite-solar-modules-pass-three-critical-stability-tests/.
Microquanta, “Microquanta demonstrates high stability
beyond IEC standards” (2020); https://www.microquanta.
com/en/newsinfo/112E5B4ABC180F9D/.
G. Wang et al., Molecular engineering of hole-selective layer
for high band gap perovskites for highly efficient and stable
perovskite-silicon tandem solar cells. Joule 7, 2583–2594
(2023). doi: 10.1016/j.joule.2023.09.007
115. Z. Dai et al., Interfacial toughening with self-assembled monolayers
enhances perovskite solar cell reliability. Science 372, 618–622
(2021). doi: 10.1126/science.abf5602; pmid: 33958474
116. L. Shi et al., Gas chromatography-mass spectrometry
analyses of encapsulated stable perovskite solar cells.
Science 368, eaba2412 (2020). doi: 10.1126/science.
aba2412; pmid: 32439657
117. J. Liu et al., 28.2%-efficient, outdoor-stable perovskite/silicon
tandem solar cell. Joule 5, 3169–3186 (2021). doi: 10.1016/
j.joule.2021.11.003
118. Q. Emery et al., Encapsulation and Outdoor Testing of
Perovskite Solar Cells: Comparing Industrially Relevant
Process with a Simplified Lab Procedure. ACS Appl. Mater.
Interfaces 14, 5159–5167 (2022). doi: 10.1021/
acsami.1c14720; pmid: 35108814
119. S. Yang et al., Stabilizing halide perovskite surfaces for solar cell
operation with wide-bandgap lead oxysalts. Science 365,
473–478 (2019). doi: 10.1126/science.aax3294; pmid: 31371610
120. K. Brecl et al., Are Perovskite Solar Cell Potential-Induced
Degradation Proof? Sol. RRL 6, 2100815 (2022).
doi: 10.1002/solr.202100815
121. Z. Purohit et al., Impact of Potential-Induced Degradation on
Different Architecture-Based Perovskite Solar Cells. Sol. RRL
5, 2100349 (2021). doi: 10.1002/solr.202100349
122. J. Carolus et al., Potential-Induced Degradation and Recovery
of Perovskite Solar Cells. Sol. RRL 3, 1900226 (2019).
doi: 10.1002/solr.201900226
p
123. Z. Xu et al., Reverse-bias resilience of monolithic perovskite/
silicon tandem solar cells. Joule 7, 1992–2002 (2023).
doi: 10.1016/j.joule.2023.07.017
124. D. Bogachuk et al., Perovskite Photovoltaic Devices with
Carbon-Based Electrodes Withstanding Reverse-Bias
Voltages up to −9 V and Surpassing IEC 61215:2016
International Standard. Sol. RRL 6, 2100527 (2022).
doi: 10.1002/solr.202100527
g
125. I. J. Park et al., A Three-Terminal Monolithic Perovskite/Si
Tandem Solar Cell Characterization Platform. Joule 3,
807–818 (2019). doi: 10.1016/j.joule.2018.11.017
126. E. Bellini, “Solliance hits 29.2% efficiency on perovskite/
silicon tandem solar cell” (PV Magazine, 2021); https://www.
y
pv-magazine.com/2021/11/01/solliance-hits-29-2-efficiency-
on-perovskite-silicon-tandem-solar-cell/.
127. L.-L. Senaud et al., “High Efficiency 3-Terminal Perovskite/
Silicon Tandems Based on Tunnel Junction IBC” paper
presented at The 8th World Conference on Photovoltaic
Energy Conversion, Milan, Italy, 26 to 30 September 2022.
AC KNOWLED GME NTS
Competing interests: The authors declare no competing
interests. License information: Copyright © 2024 the authors,
some rights reserved; exclusive licensee American Association for
the Advancement of Science. No claim to original US government
works. https://www.science.org/about/science-licenses-journal-
article-reuse
y g
Submitted 28 February 2023; accepted 1 December 2023
10.1126/science.adh3849
y
,
13 of 13
RES EARCH

◥ reprogramming to converge into the T3 neu-

RESEARCH ARTICLE SUMMARY trophil state, which was terminally differentiated
and expressed the surface marker dcTRAIL-R1.
CANCER IMMUNOLOGY dcTRAIL-R1 up-regulation in tumor-naïve neu-

Deterministic reprogramming of neutrophils

trophils could be induced by exposure to tumor-
conditioned medium in vitro or entry into the

within tumors
tumor in vivo, and was accompanied by the
expression of T3-specific genes. More impor-
tantly, this phenomenon was independent of
Melissa S. F. Ng†*, Immanuel Kwok†, Leonard Tan†, Changming Shi†, Daniela Cerezo-Wallis, their initial maturation phenotype. These find-
Yingrou Tan, Keith Leong, Gabriel F. Calvo, Katharine Yang, Yuning Zhang, Jingsi Jin, Ka Hang Liong, ings thus underscore the capability of neutro-
Dandan Wu, Rui He, Dehua Liu, Ye Chean Teh, Camille Bleriot, Nicoletta Caronni, Zhaoyuan Liu, phils to adopt a new functional phenotype,
Kaibo Duan, Vipin Narang, Iván Ballesteros, Federica Moalli, Mengwei Li, Jinmiao Chen, Yao Liu, overlaying it onto their existing differentiation
Lianxin Liu, Jingjing Qi, Yingbin Liu, Lingxi Jiang, Baiyong Shen, Hui Cheng, Tao Cheng, Veronique Angeli, stage. The T3 phenotype was strongly correlated
Ankur Sharma, Yuin-han Loh, Hong Liang Tey, Shu Zhen Chong, Matteo Iannacone, Renato Ostuni, with a prolonged lifespan, with dcTRAIL-R1+
Andrés Hidalgo, Florent Ginhoux, Lai Guan Ng* neutrophils persisting for more than 5 days
within the tumor. Furthermore, T3 neutrophils
were mainly localized to a unique hypoxic-
INTRODUCTION: Neutrophils are the first re- mors in a murine orthotopic model of pancre- glycolytic niche within the tumor, where they
sponders to infection and injury and are rapidly atic cancer. Tumor neutrophil states identified optimally exerted their pro-angiogenic function.
recruited to affected tissues in large numbers to from these analyses were validated by multi- This indicated that neutrophil reprogramming

enact their protective function. As such, neu-
trophils were historically perceived as a homo-
geneous and transient population. Recently,
however, a diverse array of neutrophil states
has been reported in cancer, varying in their
maturation, surface marker expression, and
parametric flow cytometry, and spatial mapping
at the RNA and protein levels were performed
to reveal their localization within the pancre-
atic tumors. In vitro and in vivo approaches
were then used to examine how the tumor en-
vironment shapes neutrophil phenotype, life-
p
plays a critical role in enabling their survival
under hypoxia, oxidative stress, and metabolic
perturbations within the tumor microenviron-
ment. Specifically, T3 neutrophils expressed
high levels of vascular endothelial growth fac-
tor alpha (VEGFa) and substantially enhanced

transcript profiles. The relationship between
these neutrophil states and their organiza-
tion into a unified protumoral response have
yet to be elucidated, limiting the therapeutic
span, and protumoral functions.
RESULTS: We identified three distinct neutro-
phil states within the tumor microenvironment,
g
blood vessel formation within the tumor core, and
only coinjection of T3 neutrophils with tumor
cells accelerated tumor growth. Therefore, the ab-
lation of either T3 neutrophils or VEGFa inhibits

targeting of neutrophils in cancer.
RATIONALE: To identify the mechanisms by
which disparate neutrophil states are coordi-
nated into a concerted protumoral response, we
used single-cell RNA sequencing and ATACseq
(assay for transposable chromatin sequencing)
on neutrophils from various organs and tu-
Developmental trajectory
T1, T2, and T3, which were epigenetically and
transcriptionally distinct from neutrophils in
the bone marrow, spleen, and blood. By assess-
ing nuclear morphology and maturation status,
we determined that immature and mature neu-
trophils infiltrating the tumor differentiated into
transitional T1 and T2 populations, respective-
ly. T1 and T2 neutrophils underwent further
Tumor reprogramming
y
this growth enhancement. Finally, all three tu-
mor neutrophil states were observed across mouse
models and in multiple human cancers, with the
T3 signature predicting poorer patient outcomes
in two independent human pancreatic cancer
cohorts and other solid tumor types.
CONCLUSION: By examining neutrophils in the
context of their ontogeny, we uncovered their
intrinsic flexibility in adapting to environmen-

y g
tal signals regardless of their initial maturation
Mature stage. This implies that neutrophils infiltrating
a tissue niche follow a common path, merging
Immature Growth
Angiogenesis their different functional states into a single
terminal phenotype as guided by the tissue.
Transitional VEGFα Within the tumor, this deterministic program

y
states
likely ensures a continual supply of pro-angiogenic

,
Reprogrammed T3 neutrophils that fuel tumor growth. Our find-
neutrophils ings thus demonstrate how short-lived effector
cells such as neutrophils effectively tailor their
functions to accommodate tissue require-
ments, highlighting the untapped possibilities
of targeting the local neutrophil response as
Stromal Hypoxic-glycolytic niche Tumor immunotherapy.
▪
Tumor-infiltrating neutrophils undergo convergent reprogramming into pro-angiogenic neutrophils
The list of author affiliations is available in the full article online.
that support tumor growth. In cancer, both immature and mature neutrophils infiltrate the tumor. After *Corresponding author. Email: nglaiguan@renji.com (L.G.N.);
entering the tumor microenvironment, these neutrophils undergo differentiation, leading to the formation of melissa_ng@immunol.a-star.edu.sg (M.S.F.N.)
†These authors contributed equally to this work.
transitional populations. Through reprogramming, these populations ultimately converge into a terminal neutrophil
Cite this article as M. S. F. Ng et al., Science 383, eadf6493
state. Reprogrammed neutrophils strongly express VEGFa and localize to a unique hypoxic-glycolytic niche near (2024). DOI: 10.1126/science.adf6493
the tumor core. This places them in an optimal position to exert their pro-angiogenic function within hypoxic and nutrient-
poor tumor regions, thereby promoting tumor growth. The emergence of tumor reprogramming reflects the adaptability READ THE FULL ARTICLE AT
of neutrophils to environmental cues, allowing them to consolidate their protumoral responses. https://doi.org/10.1126/science.adf6493

Ng et al., Science 383, 163 (2024) 12 January 2024 1 of 1

RES EARCH

◥ (12, 21, 22). Collectively, this spectrum of neu-

RESEARCH ARTICLE trophil states, encompassing phenotypic and
maturation differences, is proposed to make
CANCER IMMUNOLOGY up the functional diversity of neutrophils in

Deterministic reprogramming of neutrophils

cancer (4–15).
Here, we used multi-omics approaches in

within tumors
pancreatic tumors at single-cell transcriptional
and spatial resolution to examine neutrophil
functional diversity in the context of their on-
Melissa S. F. Ng1*†, Immanuel Kwok1†, Leonard Tan1†, Changming Shi2†, Daniela Cerezo-Wallis3,4, togenetic order, origin, and influence from
Yingrou Tan1,5, Keith Leong1, Gabriel F. Calvo6, Katharine Yang1, Yuning Zhang7,8, Jingsi Jin2, tissue signals. We unexpectedly found that
Ka Hang Liong1, Dandan Wu9, Rui He2, Dehua Liu1, Ye Chean Teh1, Camille Bleriot10,11, although various populations of neutrophils
Nicoletta Caronni12, Zhaoyuan Liu9, Kaibo Duan1, Vipin Narang1, Iván Ballesteros3, Federica Moalli13,14, entered the tumor, it was only inside the tu-
Mengwei Li1, Jinmiao Chen1, Yao Liu15, Lianxin Liu15, Jingjing Qi16,17, Yingbin Liu16,17, Lingxi Jiang18,19,20, mor that neutrophils entered a convergent tra-
Baiyong Shen18,19,20, Hui Cheng21, Tao Cheng21, Veronique Angeli7,8, Ankur Sharma22,23,24, Yuin-han Loh25, jectory that directed them toward a specific
Hong Liang Tey5,26,27, Shu Zhen Chong1,28, Matteo Iannacone13,14,29, Renato Ostuni12,29, Andrés Hidalgo3,4, protumoral state. Our findings challenge cur-
Florent Ginhoux1,9,10,30, Lai Guan Ng2,1,28* rent models and suggest the potential of tar-
geting neutrophils for cancer immunotherapy.
Neutrophils are increasingly recognized as key players in the tumor immune response and are associated
with poor clinical outcomes. Despite recent advances characterizing the diversity of neutrophil states Intratumoral neutrophils converge into
in cancer, common trajectories and mechanisms governing the ontogeny and relationship between a distinct transcriptional state

p
these neutrophil states remain undefined. Here, we demonstrate that immature and mature neutrophils To investigate neutrophil heterogeneity in can-
that enter tumors undergo irreversible epigenetic, transcriptional, and proteomic modifications cer, we used an orthotopic mouse model of pan-
to converge into a distinct, terminally differentiated dcTRAIL-R1+ state. Reprogrammed dcTRAIL-R1+ creatic ductal adenocarcinoma (PDAC). Using
neutrophils predominantly localize to a glycolytic and hypoxic niche at the tumor core and exert a pancreatic cancer cell line previously estab-
pro-angiogenic function that favors tumor growth. We found similar trajectories in neutrophils across lished from a tumor in the Pdx1Cre; KrasG12D/+;
multiple tumor types and in humans, suggesting that targeting this program may provide a means Trp53R172H/+ (KPC) genetically modified mouse

g
of enhancing certain cancer immunotherapies. model (23), cultured cells were orthotopically
transplanted into the pancreas, which grew in

N
situ to form a tumor characterized by copious
eutrophils play a substantial role in the trum of neutrophil states in cancer, with differ- neutrophil infiltration (14). Single-cell RNA

immune response to infection and in-
jury. As one of the first cells to enter a
damaged site from the circulation, the
rapid recruitment of large numbers of
neutrophils into the tissue is key for their pro-
tective function (1). This process is co-opted in
pathological settings such as cancer, in which
persistent neutrophil infiltration into the tu-
mor has been consistently associated with
poorer patient outcomes (2). Many studies have
ences in density (4, 5), surface markers (6–10),
and transcript expression (11–13).
Neutrophil heterogeneity also exists in the
bone marrow (BM) in the form of various mat-
uration stages. Numerous studies have shown
that granulocytic progenitor cells differenti-
ate sequentially into precursor, immature, and
finally mature neutrophils, and each subset of
cells has distinct functional capabilities (14–19).
Tumor-induced chronic inflammation triggers
y
sequencing (scRNAseq) was performed on
total CD11b+CD115–Ly6G+ cells from the BM,
spleen, blood, and pancreatic tumors to char-
acterize neutrophil heterogeneity in the con-
text of their ontogenetic order and origin (n = 2;
Fig. 1A and fig. S1, A and B). Uniform manifold
approximation and projection (UMAP) analysis
revealed that neutrophils from the BM, spleen,
and blood clustered according to their develop-
mental states, from pre-neutrophils (preNeus)

associated protumoral functions to neutro-
phils, which in this context have been referred
to as granulocytic myeloid-derived suppressor
cells (3). This uniform view has since been
disrupted by the identification of a wide spec-
the premature egress of these precursors into
the circulation and subsequently into the tu-
mor (4, 14, 17, 19, 20), and extramedullary sites
such as the spleen (13, 14), have been proposed
to be priming sites of protumoral neutrophils
y g
to immature neutrophils (IMM 1 and 2) and
mature neutrophils (MAT 1 to 5) (Fig. 1B and
fig. S1C). Integration with a previous dataset
containing healthy mouse neutrophils from
matching tissues (19) showed that despite

y
,
1
Singapore Immunology Network (SIgN), Agency for Science, Technology and Research (A*STAR), Singapore. 2Shanghai Immune Therapy Institute, Renji Hospital, Shanghai Jiao Tong University
School of Medicine, Shanghai, China. 3Area of Cell & Developmental Biology, Centro Nacional de Investigaciones Cardiovasculares Carlos III, Madrid, Spain. 4Vascular Biology and Therapeutics
Program and Department of Immunobiology, Yale University School of Medicine, New Haven, CT, USA. 5National Skin Centre, National Healthcare Group, Singapore. 6Department of Mathematics
& MOLAB-Mathematical Oncology Laboratory, University of Castilla-La Mancha, Ciudad Real, Spain. 7Immunology Translational Research Program, Department of Microbiology and Immunology,
Yong Loo Lin School of Medicine, National University of Singapore, Singapore. 8Immunology Program, Life Science Institute, National University of Singapore, Singapore. 9Shanghai Institute of
Immunology, Shanghai Jiao Tong University School of Medicine, Shanghai, China. 10INSERM U1015, Institut Gustave Roussy, Villejuif, France. 11CNRS UMR8253, Institut Necker des Enfants
Malades, Paris, France. 12Genomics of the Innate Immune System Unit, San Raffaele-Telethon Institute for Gene Therapy (SR-Tiget), IRCCS San Raffaele Scientific Institute, Milan, Italy. 13Division
of Immunology, Transplantation, and Infectious Diseases, IRCCS San Raffaele Scientific Institute, Milan, Italy. 14Experimental Imaging Centre, IRCCS San Raffaele Scientific Institute, Milan, Italy.
15
Department of Hepatobiliary Surgery, The First Affiliated Hospital of USTC, Division of Life Sciences and Medicine, University of Science and Technology of China, Anhui, China. 16Department of
Biliary and Pancreatic Surgery, Renji Hospital, Shanghai Jiao Tong University School of Medicine, Shanghai, China. 17Shanghai Institute of Cancer Biology, Renji Hospital, Shanghai Jiao Tong
University School of Medicine, Shanghai, China. 18Department of General Surgery, Pancreatic Disease Center, Ruijin Hospital, Shanghai Jiao Tong University School of Medicine, Shanghai, China.
19
Research Institute of Pancreatic Diseases, Shanghai Key Laboratory of Translational Research for Pancreatic Neoplasms, Shanghai Jiao Tong University School of Medicine, Shanghai, China.
20
State Key Laboratory of Oncogenes and Related Genes, Institute of Translational Medicine, Shanghai Jiao Tong University, Shanghai, China. 21State Key Laboratory of Experimental Hematology,
National Clinical Research Center for Blood Diseases, Haihe Laboratory of Cell Ecosystem, Institute of Hematology & Blood Diseases Hospital, Chinese Academy of Medical Sciences & Peking
Union Medical College, Tianjin, China. 22Harry Perkins Institute of Medical Research, QEII Medical Centre, Nedlands, Western Australia, Australia. 23Curtin Medical School, Curtin University,
Bentley, Western Australia, Australia. 24Curtin Health Innovation Research Institute, Curtin University, Bentley, Western Australia, Australia. 25Institute of Molecular and Cell Biology (IMCB),
A*STAR (Agency for Science, Technology and Research), Singapore. 26Lee Kong Chian School of Medicine, Nanyang Technological University, Singapore. 27Yong Loo Lin School of Medicine,
National University of Singapore, Singapore. 28Department of Microbiology and Immunology, National University of Singapore, Singapore. 29Vita-Salute San Raffaele University, Milan, Italy.
30
Translational Immunology Institute, SingHealth Duke-NUS Academic Medical Centre, Singapore.
*Corresponding author. Email: nglaiguan@renji.com (L.G.N.); melissa_ng@immunol.a-star.edu.sg (M.S.F.N.)
†These authors contributed equally to this work.

Ng et al., Science 383, eadf6493 (2024) 12 January 2024 1 of 16

RES EARCH | R E S E A R C H A R T I C L E

Fig. 1. Neutrophils infiltrate A scRNAseq workflow

pancreatic tumors and undergo Generate tumor Harvest selected tissues Sort CD11b+Ly6G+ neutrophils 10X V3 3' scRNAseq
further differentiation and Orthotopic PDAC Bone marrow Hashtag Demultiplex,
tumor cell injection label downstream
converge upon a transcriptionally Blood
individual analysis
Spleen tissues
distinct T3 neutrophil state. Tumor
(A) Schematic showing scRNAseq
workflow. CD11b+Ly6G+ neutrophils B Bone marrow (BM) Spleen Blood Tumor
T3
were sorted from the BM, spleen, preNeu T1
blood, and pancreatic tumor of
MAT 5 T2
tumor-bearing mice 6 weeks after MAT 1 IMM 1
MAT 3
MAT 4
orthotopic injection (n = 2 biological IMM 2
MAT 2
UMAP2

replicates). Each sample was

individually tagged with cell-hashing
UMAP1
antibodies before being pooled for
analysis with 10X V3 3′ scRNAseq
(see also fig. S1A and the materials C Maturation signature D Diffusion map E RNA velocity

and methods). (B) UMAP projection preNeu preNeu

of total neutrophils in the BM,
spleen, blood, and tumor show strong or
urs

p
rec
enrichment of three clusters (T1, P preNeu
Immature
T2, and T3) in the tumor. Louvain re IMM 1
m atu IMM 2
Immature
Im

Diffusion component 2
Diffusion component 2
clustering was performed, and colors MAT 1(BM)
e MAT 2 T1 T1
correspond to the clusters identified. a tur
M MAT 3 Mature
Mature
(C) The neutrophil maturation MAT 4
MAT 5
score can be used to identify Louvain T1 T2

g
clusters along their differentiation or T2
Tu
m T3 T2
trajectory. Histograms show the
T3 T3
0 1
Maturation score Diffusion component 1 Diffusion component 1
module score of the maturation gene
signature for each cluster identified

y
in (B), with scores closest to 1 being F ATACseq Immature
BM,BM,Spleen,
Blood
G Open chromatin regions (OCRs) at T3 genes
Immature
the most mature. Dotted line in Tumor WT Tumor
Mature
gray defines the cutoff for mature BM,BM, BM BM Spleen Blood Tumor
100 Spleen,
Blood IMM MAT IMM MAT IMM MAT IMM MAT IMM MAT
neutrophils and is set at the lower
PC3

0
: 7.9

20 Tumor
bound of the Mature 2 cluster. -10
0
10 0 mouse:
0 Mature 0 0
3%

BM
-20
(D) Low dimensional embedding of Tumor -1 Spleen
7%

0 00
20 0
.6

-2 Blood
46

00 7
all neutrophils using a diffusion map 1 0 - 30 Tumor
1:

0
PC

0
PC

-4 Hypoxia:
2:

00 WT
approach reveals a branch point -1 00
19

mouse:
00 Alkbh5, Bnip3,
.5

-2 BM
8%

between mature and tumor neutrophils. Bnip3l,Cited2, 6

Ddit4, Egln3,
Scatterplot shows diffusion components
H T3 enriched TF regulons Hilpda, Hmox1,

y g
1 and 2. Light gray arrows are from scRNAseq Itpr2, Npepps,
T3 T2 T1 MAT IMM pre P4ha1, Rbpj, 5
used to indicate the trajectory from Scaled Rgcc, Vegfa
Sap30 (514)
AUC
precursor, immature, and mature Ddit3 (31)
Atf4 (867) score
Mafk (86) 3 Glycolysis:
neutrophil states. Black arrows denote Bhlhe41 (195) 2
Ddit4,
4
Usf2 (1632) 1
the branching into T3 neutrophils Atf5 (436) 0 Eno1, Hk2,
Hif1a (248) Ier3, Pfkp
from T1 and T2 states. (E) RNA -1

y
Zmiz1 (861)
Atf3 (1326) -2 3
velocity suggests the convergent Jun (1664) Angiogenesis:

,
Tgif1 (476)
differentiation of T1 and T2 into T3 Maff (463) Hk2, Hmox1,
Bhlhe40 (1971) Itgb1, Lgals3,
neutrophils. RNA velocity vectors Rest (138) 2
Egr1 (2600) Rhob, Thbs1,
Nfil3 (1096) Vegfa
were projected on the diffusion map Atf1 (1412)
Creb5 (143)
embedding with velocity vectors Fos (800) 1
Junb (1510)
terminating in the tumor and mature Cebpb (1454)
Nfe2l2 (49)
neutrophils. (F) Principal component Fosb (238)
Runx1 (66) 0
analysis of ATACseq indicated that -2kb 0 +2kb genomic dist.(bp)

changes in chromatin accessibility

established in tumor neutrophils clustered them away from other neutrophil
subsets. ATACseq was performed for immature (circles) and mature (squares)
neutrophils sorted from the BM (gray and dark blue, n = 3 each), spleen
(light blue, n = 3 each), blood (red, n = 3 mature, n = 2 immature), and tumor
(orange, n = 3 each) in wild-type (WT) or tumor-bearing (tumor) mice. Numbers
denote the number of biological replicates used. (G) OCRs matched to differentially
expressed T3 genes have increased accessibility only in immature and mature
tumor neutrophils. Heatmap shows intensity of fragment mapping across
indicated neutrophil subsets; histograms indicate average mapping intensity. T3
genes linked to hypoxia, glycolysis, and angiogenesis are annotated (see also
table S2). (H) Heatmap showing scaled AUC scores of the top 25 transcription
factor regulons enriched in T3 neutrophils computed by PySCENIC (see also
table S2). Numbers in brackets denote the number of genes assigned to the
transcription factor regulon. Transcription factors that also had increased motif
enrichment in tumor-associated immature and mature bulk neutrophil
populations are shown in bold.

Ng et al., Science 383, eadf6493 (2024) 12 January 2024 2 of 16

RES EARCH | R E S E A R C H A R T I C L E
potential perturbations from the tumor, no new
developmental clusters or trajectories emerged
in neutrophils from tissues of the tumor-bearing
mice (fig. S1, D and E). This suggests a robust
adhesion of neutrophils to a single develop-
mental trajectory within the tissues where they
develop and circulate. By contrast, neutrophils
acquired specific transcriptional profiles in the
tumor (Fig. 1B and fig. S1C) and formed three
distinct clusters (T1, T2, and T3) separate from
the other compartments.
Because immature and mature neutrophils
can infiltrate tumoral tissue (14, 15), we inves-
tigated the relationship between the maturation
and functional states of the T1 to T3 neutrophil
populations using a neutrophil maturation
gene signature (table S1) curated by Xie et al.
(19); each cluster in our dataset was graded
by a maturation score. T1 neutrophils had a
maturation score comparable to immature
neutrophils (i.e., the IMM 2 cluster), whereas
T2 neutrophils had a maturation score sim-
ilar to mature neutrophils (i.e., the MAT 1 to
5 clusters) (Fig. 1C). Next, we sought to un-
cover more potential links between neutrophil
clusters in the tumor and the other tissue com-
partments by using a diffusion map approach
that orders cells on the basis of transitional
probabilities and better preserves differentia-
tion trajectories (24, 25). We found that tumor
neutrophils deviated from the steady-state
neutrophil developmental trajectory (Fig. 1D).
Specifically, T1 and T2 neutrophils progressed
along the tumor-specific branch and converged
at the T3 state, suggesting that T1 and T2 may
both give rise to the T3 population, which is
predicted to be the most terminally differen-
tiated subset. RNA velocity (26) indicated a
pseudotime progression from immature neu-
trophils to T1 and mature neutrophils to T2,
suggesting that T1 and T2 neutrophils are
transitory states from immature and mature
neutrophils that have migrated into the tumor
(Fig. 1E). By contrast, T3 neutrophils exhibited
a maturation score that was in between that of
the T1 and T2 neutrophils (Fig. 1C), represent-
ing a potential admixture of T1 and T2 cells.
Accordingly, we observed that velocity vectors
originating from both T1 and T2 populations
terminated in the T3 cluster (Fig. 1E). These
findings thus corroborate earlier studies dem-
onstrating that both mature and immature
neutrophils infiltrate tumor tissue (4) and be-
come T1 and T2 neutrophils, respectively. Fur-
ther, our data predict that these stages are
transitional, because both T1 and T2 neutro-
phils remained amendable for further differ-
entiation into T3 neutrophils.
Epigenetic reprogramming
of tumor-infiltrating neutrophils
To confirm that T3 neutrophils can be repro-
grammed from both immature (T1) and ma-
ture (T2) tumor neutrophils, we performed
Ng et al., Science 383, eadf6493 (2024) 12 January 2024
ATACseq (assay for transposase-accessible chro-
matin followed by sequencing) on sorted CD101–
(immature) and CD101+ (mature) neutrophils
(14) from various tissue compartments of con-
trol and tumor-bearing mice. Principal com-
ponent analysis revealed that both immature
and mature tumor neutrophils clustered sep-
arately from corresponding subsets of neutro-
phils from other tissues (Fig. 1F), indicating
that signals from the tumor microenviroment
imprint specific chromatin accessibility changes
within infiltrating neutrophils. We then iden-
tified all open chromatin regions (OCRs) that
were differentially accessible in immature and
mature tumor neutrophils relative to the non-
tumor neutrophil subsets (table S2). Increased
chromatin accessibility for genes up-regulated
in T3 neutrophils (T3 genes) in both immature
and mature tumor neutrophils (Fig. 1G) in-
cluded genes involved with hypoxia, glycolysis,
and angiogenesis, such as Vegfa and Hk2 (fig.
S2A). These changes in chromatin were un-
detectable in neutrophils from other tissues,
which still shared similar accessibility for ca-
nonical neutrophil genes such as Cebpa and
Gfi1 (fig. S2B).
To elucidate putative upstream regulators of
the T3 program, we used pySCENIC (27) to iden-
tify transcription factor regulons predicted in
T3 neutrophils that had minimal presence in
the other neutrophil subsets (Fig. 1H). This
was then correlated to motifs found in the bulk
ATACseq analysis (fig. S2C), which revealed T3-
specific transcription factors with enriched mo-
tifs in immature and mature tumor neutrophil
differentially accessible OCRs (fig. S2, D and
E). Among the T3-specific transcription fac-
tors, Mafk, Nfe2l2, and Atf3 were implicated in
regulating metabolic and oxidative stress typ-
ical of the tumor microenvironment (fig. S2, D
and E). Correlation of the underlying chroma-
tin accessibility with transcription factors pre-
dicted to govern the T3 state suggested that
these epigenetic changes initiated within T1
and T2 neutrophils are reinforced by transcrip-
tion factor activity as they transition toward a
T3 state. These findings suggest that tumor-
infiltrating neutrophils can be reprogrammed
by the tumor microenvironment regardless of
their stage of maturity by converging transcrip-
tional and chromatin trajectories toward a dis-
tinct T3 state.
CD101 and dcTRAIL-R1 discriminate tumor
neutrophil states
Having shown that immature T1 and mature T2
neutrophils in the tumor undergo transcrip-
tional and epigenetic reprogramming toward
the T3 neutrophils, we next assessed whether
we could distinguish these subsets by protein
expression. This was particularly important be-
cause increases in RNA expression in neu-
trophils typically precedes protein expression,
especially for genes tightly regulated in neu-
trophil development, such as those for granule
proteins (14). T1, T2, and T3 neutrophils had
differential expression of surface marker genes,
suggesting that a combination of surface marker
protein staining could be used to identify them
(fig. S3A). Using a multiparametric flow cy-
tometry approach, live CD45+ cells from the PDAC
tumor were screened for 249 cell surface mark-
ers. Then, all captured cells were clustered with
InfinityFlow (28, 29), and the coexpression pat-
terns of all tested markers were evaluated (fig.
S3B), which revealed two distinct neutrophil clus-
ters (Fig. 2A). High levels of immuno-modulatory
or suppressive markers such as dcTRAIL-R1,
PD-L1 (CD274) (6), CD14 (13), CD371 (30), VISTA
(31), and CD39 (32), distinguished Cluster 2 neu-
trophils from those of Cluster 1 (Fig. 2B). These
markers correlated with surface marker genes
enriched in T3 neutrophils (fig. S3C), suggesting
that they may serve to identify the T3 state. In-
deed, we found that dcTRAIL-R1 was expressed
p
mainly by a population of tumor neutrophils
(Fig. 2C) while having minimal expression in
tumor-infiltrating monocytes and macrophages
and in neutrophils in the BM, spleen, and blood
(Fig. 2C and fig. S3D). By contrast, staining
for other markers such as VISTA and CD14
g
showed expression across multiple neutrophil
subsets and was less restricted to tumor neu-
trophils (Fig. 2D and fig. S3D). dcTRAIL-R1
expression in tumor-infiltrating neutrophils
y
was also conserved across several other tu-
mor types, including orthotopic breast can-
cer [median: 20.4%; interquartile range (IQR):
18.0 to 23.1%] and orthotopic lung cancer
(median: 12.1%; IQR: 4.54 to 19.5%) mouse
models (fig. S3E). Across all experimental tu-
mor models examined, dcTRAIL-R1 expres-
sion was similarly limited to a subset of tumor
neutrophils, suggesting a possible candidate
marker for T3 neutrophils.
y g
Consistent with its protein expression pat-
tern, gene expression and RNA velocity (indic-
ative of active gene transcription) for the gene
encoding dcTRAIL-R1 (Tnfrsf23) were highest
in the T3 neutrophil cluster (fig. S4A). Therefore,
we attributed the T3 population to the dcTRAIL-
y
R1+ cluster (cluster 2), whereas the T1 and T2
,
populations were likely contained within the
dcTRAIL-R1– cluster (cluster 1) identified in our
InfinityFlow analysis (Fig. 2A). Because T1 and
T2 neutrophils were transcriptionally identified
as immature and mature, respectively, in our
scRNAseq data (Fig. 1D), we next evaluated
CD101 expression, which separates immature
neutrophils from mature neutrophils (14). Clus-
ter 1 exhibited a continuum of CD101 expression
(fig. S4B), suggesting that CD101 can be used
to distinguish T1 and T2 neutrophils within
the dcTRAIL-R1– population. Using CD101 and
dcTRAIL-R1 to identify the different neutro-
phil types by flow cytometry (Fig. 2E), we iso-
lated putative T1, T2, and T3 neutrophils and
discovered that putative T1 neutrophils had
3 of 16
RES EARCH | R E S E A R C H A R T I C L E

Fig. 2. Combinatorial A Gated on live CD45+ cells in the tumor,

249 predicted markers by InfinityFlow:
B dcTRAIL-R1 PD-L1 CD14
CD101 and dcTRAIL-R1 Z score
0 24
Z score
0 4
Z score
0 8
-2.1 1.4
expression identifies Eosinophils
1000 Ly6G intensity (log2MFI) Cluster 2
T1, T2, and T3 neutro-
phils. (A) UMAP Macrophages/ NK
projection of live CD45+ 800 DCs
Neutrophil
immune cells within the Monocytes Cluster 2 Cluster 1
PDAC tumor reveals 600
UMAP 2

CD8 T CD371 (MICL) VISTA CD39

two clusters of neutro- Z score Z score Z score
1000 0 5 0 8 0 4
phils. High-parameter 400
flow cytometry was 750

UMAP 2
performed on single-cell 200 B 500
CD4 T cells
suspensions of pooled Neutrophil 250
tumors (n = 10 biological 0 Cluster 1
0
replicates) using the 0 200 400 600 800 1000 0 250 500 750 1000
LEGENDScreen panel UMAP 1 UMAP 1

(BioLegend). Live CD45+

well-compensated FCS C Peripheral tissue neutrophils F Gene signature expression

p
Blood Spleen BM
were first analyzed and 10 5
0.44% 10 0.96% 5
10 0.98% 5

then exported out for

T3 T2 T1
4 4 4
10 10 10
dcTRAIL-R1+ dcTRAIL-R1- dcTRAIL-R1-
analysis using the 3 3 3Immature/mature Mature Immature
10 10 10
InfinityFlow package in 0 0 0

R. UMAP shows Ly6G -10 3

-10 0 10 3
10 10
-10
3 4
-10 0 10 5
10
3

10
-10
3 3
-10 0 10 10 4
10 5
3

Atf3 3 3 4 5

expression intensity Ccl3

g
Tumor Ccl4
imputed by InfinityFlow Neutrophils Monocytes Macrophages Cd274
from high (red) to low 10 5
37.7% 10 8.94% 5 10 9.85% Cstb
5

Cxcl3
(blue). Clusters are
T3 genes

10 4 10 4 10 Hcar24
Ly6G

annotated by canonical Hilpda

3
3 10
10 3 10
Hk2

y
surface marker expres- 0 0 0
Hmox1
-10 3 -10 3 -10 Ier3
3

sion (see also fig. S3B). -10 0 10 3

10 3
-10 0 4
10 10 3
-10 0 3
10 10 4
Jun 3 3 4

dcTRAIL-R1 Plin2
(B) Neutrophil cluster 2
Spp1
has increased expression Tgif1
of immunosuppressive
D Validated markers Tnfrsf23
Vegfa
and/or immunomodula- dcTRAIL−R1 Zeb2
Scaled
MFI CD371 Ldha
tory surface markers 1.00 Mif
CD14
compared with cluster 1. 0.75
0.50 CD39 Cd300ld
Surface marker expres- Cxcr2
0.25 VISTA Dusp1
sion intensities are Gbp2
0.00 PD-L1

y g
shown for curated Ifitm1
T2 genes
Macs.
Mono.
Neu.

BM Spleen Blood Il1b

Tum.

surface markers (clock-

Tum.

Tum.
WT

WT

WT

Isg15
wise): dcTRAIL-R1, Jaml
Peripheral tissues Tumor Junb
PD-L1, CD14, CD371, Msrb1
VISTA, and CD39. Data Osm
S100a6
are represented as a E Sorting for transcriptomic validation

y
Selplg
Slpi
Z score based on Total live,CD45+

,
predicted log2 mean neutrophils in tumor Ltc4s
T1 genes

Mmp8
fluorescence intensity dcTRAIL-R1+ dcTRAIL-R1- Mmp9
Ppia
(MFI) from high (red) to Prr13
low (blue) (see also Mature Immature Ptma
Retnlg
fig. S3C). (C) dcTRAIL-R1
Scaled log counts
expression marks and Putative T3 Putative T2 Putative T1
+/- + -
is restricted to a sep- CD101 CD101 CD101
dcTRAIL-R1+ dcTRAIL-R1- dcTRAIL-R1- -3 -2 -1 0 1 2
arate population of
tumor-infiltrating neu-
trophils. Representative contour flow cytometry plots (top) show dcTRAIL-R1 (E) Proposed gating strategy to isolate T1, T2, and T3 neutrophils by dcTRAIL-R1 and
expression against Ly6G expression in the tumor for indicated cell populations. CD101 expression from the tumor. (F) Sorted dcTRAIL-R1+ tumor neutrophils have the
(D) Heatmap shows scaled MFI for markers in (B), scaled between 0 and highest expression of the T3 transcriptional signature. Heatmap shows scaled Nanostring
1 across all populations, with 0 being the lowest MFI. Populations were analyzed gene counts (normalized against internal positive controls and housekeeping genes)
by flow cytometry, in which total neutrophils were gated as CD11b+CD115–Ly6G+ for T1 (n = 4), T2 (n = 4). and T3 (n = 4) neutrophils sorted according to (D). Numbers
in the BM, spleen, blood, and tumor. Tumor macrophages (CD11b+Gr-1–F4/80hiMHCIIhi) represent the biological replicates across two independent experiments. Genes belonging
and monocytes (CD11b+Ly6G–Ly6Chi) were gated accordingly (see also fig. S3D). to either the T1, T2, or T3 transcriptional signature are indicated (see also fig. S4D).

Ng et al., Science 383, eadf6493 (2024) 12 January 2024 4 of 16

RES EARCH | R E S E A R C H A R T I C L E
a toroidal nuclear morphology resembling
immature neutrophils from the BM (fig. S4C).
Putative T2 neutrophils had hypersegmented
nuclei similar to those of mature BM neutro-
phils (fig. S4C), whereas putative T3 neutro-
phils contained cells with hypersegmented or
toroidal nuclear morphology (fig. S4C), con-
gruent with our findings that T3 neutrophils are
reprogrammed from both immature and mature
neutrophils. Next, we found that dcTRAIL-R1+
neutrophils showed the highest expression of T3
genes (table S3), whereas the dcTRAIL-R1–CD101–
and dcTRAIL-R1–CD101+ populations were strong-
ly enriched for genes associated with T1 and
T2 neutrophils, respectively (Fig. 2F and fig.
S4D). Finally, all three neutrophil populations
were identified in all experimental tumor mod-
els examined (fig. S4E), and the frequencies of
the T1 to T3 populations in PDAC as determined
by scRNAseq were comparable to the median
frequencies of putative T1 to T3 populations
isolated by flow cytometry (fig. S4, E and F).
Thus, dcTRAIL-R1 and CD101 expression pheno-
typically divides tumor neutrophils into three
distinct populations and successfully recapit-
ulates the T1 to T3 populations defined in our
transcriptomic and epigenetic approaches.
Spatial compartmentalization
of neutrophils in tumors
Spatial mapping of immune microenviron-
ments in tumors has been highly informa-
tive by providing insights into their diversity
and functionality on the basis of their local-
ization within the tumor (33). To gain a better
understanding of how tumor neutrophils are
affected by their localization in the tumor, we
first investigated possible functional differ-
ences between the three tumor neutrophil sub-
sets by performing gene ontology (GO) analysis
on differentially expressed T1, T2, and T3 genes
(table S4). T1 neutrophils were enriched for
pathways relating to transcription and trans-
lation (including genes governing ribosomal
biogenesis such as Npm1) and oxidative phospho-
rylation (proton membrane transport) (Fig. 3A),
which is consistent with the immature pheno-
type of T1 neutrophils (14). T2 neutrophils were
enriched in pathways related to transcriptional
regulation, amide and reactive oxygen species
metabolism, and immune responses, as well
as type I interferon genes, including Ifit1, Ifit2,
Ifit3, and Isg15 (Fig. 3B), which likely reflects
neutrophil activation upon tumor infiltration.
Finally, T3 neutrophils featured pathways of
cell stress and survival, including response to
hypoxia, oxidative stress, and glycolysis, (Fig. 3C),
suggesting adaptation to the tumor environ-
ment. T3 neutrophils were also enriched for
angiogenic genes, including Vegfa, Thbs1 (34),
and Lgals3 (35), which, when combined, is sug-
gestive of a strong protumoral role.
We next characterized how neutrophils with
distinct transcriptional states would interact
Ng et al., Science 383, eadf6493 (2024) 12 January 2024
with the neighboring tumor microenvironment.
Using 4′,6-diamidino-2-phenylindole (DAPI) and
pan-cytokeratin (panCK) costaining, we noted
that our orthotopic tumors were mainly composed
of tumor cell regions (PanCKhighDAPIhigh), fibrotic
and/or necrotic regions (panCKhighDAPI–/low),
and sparse stromal regions along the edges
(panCK–/lowDAPIhigh) (fig. S5, A and B). We per-
formed spatial transcriptomics on four cryosec-
tions of three different PDAC tumor samples
(Fig. 3D). As expected, necrotic regions with low
numbers of DAPI+ cells contained low-quality
counts, which were manually annotated and
filtered to mitigate RNA contamination from
nearby spots and to improve downstream analy-
ses (fig. S5, C and D). We next assessed the tu-
mor architecture by using BayesSpace (36) to
identify transcriptionally similar neighborhoods
within the tumor (fig. S5E). Clustering and
UMAP embedding revealed that 10 clusters
were shared across all four sections (fig. S5E).
Indicative of the complexity of the tumor en-
vironment, the distributions of spatial clusters
identified were different in the two cryosections
[regions of interest 1 and 2 (ROI1 and ROI2)]
taken from different regions of the same tumor
(fig. S5E). Each spatial cluster had specific func-
tional pathways associated with their distribution
(fig. S5F). For example, regions enriched for
pathways related to epithelial-to-mesenchymal
transition (EMT, clusters 1 and 10), typically
associated with the invasive tumor front (37), were
present closer to the tumor periphery, where-
as regions associated with hypoxia (cluster 6)
and cell cycle progression (G2-M checkpoint,
cluster 2) were located close to the core of the
tumor (fig. S5, E and F).
Using Cell2location (38) with a pre-annotated
scRNAseq dataset (fig. S6A) allowed us to map
the full transcriptional signature of each cell
cluster within the tumor and estimate cell type
abundances within each spot. To ensure the
accuracy of neutrophil assigned spots, we as-
certained the presence of Ly6G immunoflu-
orescence staining falling within the spot as a
selection cutoff (fig. S6B). Cell2location re-
vealed that T1, T2, and T3 neutrophils mapped
to different regions within the tumor (Fig. 3E),
with visualization of T3 enriched spots on the
UMAP embedding indicating that T3 neutrophils
had the highest enrichment in spatial cluster 6,
which was associated with hypoxia (Fig. 3F). T3
neutrophils were also observed to be mostly
distributed adjacent to PanCKhighDAPI–/low
necrotic zones (Fig. 3G), consistent with their
localization to cluster 6 (fig. S5E). By con-
trast, T1 neutrophils were highly enriched in
the cluster 1 (EMT) and clusters 7 and 8 (IL-2
and STAT5 signaling, respectively) regions,
whereas T2 neutrophils were mostly found in
cluster 3 associated with p53 signaling (Fig. 3F).
Both T1 and T2 annotated spots were generally
localized closer to the tumor periphery com-
pared with T3 neutrophils (Fig. 3G). Our data
suggest that neutrophils are spatially organized
within the tumor, and T3 neutrophils occupy a
tumor niche that is distinct from that of T1 and
T2 neutrophils.
Reprogrammed neutrophils occupy a hypoxic
and glycolytic tumor niche
Because T3, but not T1 or T2, neutrophils had
transcriptional and epigenetic profiles that
were enriched for the hypoxia, glycolysis, and
angiogenesis pathways, we evaluated whether
this was reflected in their localization to hy-
poxic, glycolytic, or angiogenic regions within
the tumor, respectively. All spots within the
tumor region were first scored for GO path-
ways corresponding to glycolysis (fig. S6C), hy-
poxia (fig. S6D), and angiogenesis (fig. S6E), with
a 50th percentile cutoff used to determine
low and high regions. We then determined
the frequency of neutrophil-containing spots
falling into low and high regions for each
p
tumor neutrophil subset. T3 neutrophils were
found at greater frequencies in high scoring
regions for glycolysis (68.8 ± 19.7%), hypoxia
(53.3 ± 4.65%), and angiogenesis (62.2 ± 14.2%)
(Fig. 3H). By contrast, most T2 (fig. S6F) and T1
(fig. S6G) neutrophil-containing spots fell into
g
regions with low glycolysis, hypoxia, and angio-
genesis scores, indicating that partitioning of
the neutrophil subsets potentially results from
T3 neutrophils occupying a specific regional
y
hypoxic and glycolytic tumor niche.
We next set out to validate these observa-
tions at single-cell resolution because the cur-
rent Visium spatial transcriptomic resolution
is limited to 55-mm spots. We used MACSima
imaging cyclic staining (MICS) (Fig. 4A), in
which iterative immunofluorescence staining
and photobleaching allows for the evaluation
of large numbers of antibody targets within
the same tissue slice (39). Optimization was
y g
first performed to determine working markers
within orthotopic pancreatic tumor tissues (fig.
S7, A and B), given that MICS staining times
were relatively short at 10 min. Testing revealed
that CD29, Galectin-3, and CD44, which typ-
ically mark fibroblasts (40–42), displayed the
y
greatest staining at peripheral stromal and
,
desmoplastic (panCK–/lowDAPIhi) regions (fig.
S7C), whereas CD31 and CD105 marked vessels
within the stromal and tumor region (panCK+).
CD105 was able to better resolve intratumoral
vessels (fig. S7C), consistent with their role in
marking tumor-associated endothelium (43).
To identify hypoxic regions within the tumor,
Hif1a, CD39, and CD73 staining was evaluated.
CD73, an adenosine monophosphate ectoen-
zyme, has been found to be directly induced by
hypoxia and Hif1a in cancer (44–47). Immuno-
fluorescence analysis revealed diverse staining
patterns of CD73 in both stromal and desmo-
plastic regions and tumor (panCK+) regions
(fig. S7C), resembling the spatial distribution
of hypoxia gene signature scores observed in
5 of 16
RES EARCH | R E S E A R C H A R T I C L E

A T1 enriched GO pathways B T2 enriched GO pathways C T3 enriched GO pathways

Immune Regulation of cell Response to

Regulation of response Response to hypoxia
rRNA Proton transmembrane ROS metabolic proliferation oxidative stress
processing transport process Protein folding
Neg. regulation
Regulation of of translation by Translation Positive regulation
Nucleoside cellular amide RNA pol. II of angiogenesis
disphophate metabolic
phosphory- process Glycolytic
3b, process
lation Rtp4, Ifits, Crsnp1,
Atf3,Bhlhe40, Ier2, Fo Hk2,Pfk
Btgs2,Egr1
Tgm2 Pfkp,Aldo l, Rps8,Rps5,Rps3a1,
a,
Ddit4,Eno1
Nr4a1 Rps3,Rps20,Rps19
Hk2 Trmt112 Id1 Marcksl1
Il1b Rplp1, Rpl35
Ier3 Rps7 Jun
Junb, Cox17, Gapdh,Ier3, Rpl32, Rpl14
Mif Rps24 Pim1, Cish, Socs3, Mif,Pgam1,
Nfil3,Plk3, Il1r2,Igtp, Ifi47 Rpl12, Rpl10, Fau
Nme1 Rps17
Ppp1r15a,il Pgk1,Tpi1
Rpl35 Rbpj,Sk Eif1
Nme2 Rpl26 a
Egln
1 Zfp 3 l1a, Eef1a
Pgam i1 Rpl1 Tnf Rh 36l1 Hmox1,Ials3, 1
Tp p 1 oh Itgb1,Lg,Rhob
Pn
No 4 Xbp 6 U
Sn p10 3 N sp1 Ptgs2 Rps28, Rps26, Rplp1,
5e Rp u13 Zfp gcc t1 Tribfkbia8, P Rpl8,Rpl38, Rpl36,
Atp g1 R ta 1 Ptg 1, , Ly lac8 egfa,
5 2 R l35 S bs Rpl35a,Rpl30,Rpl26,
Avg. Atp p5g 5j N pl1 Avg s2 Tes 6a, , Avg. Thbs1,V aja1,
t p R pm 0a Th rn, d , Canx,Dn bp1a Rpl22,Rpl19,Rpl18
log2FC A At tp5j2 ps 1
log2FC Il1 1
1, l2a log2FC Fk

Fthdkn ,Zbp oc
8
Rp l10 l

R p s1
Rp tp5

C tg1 1,S
Clelec4 14, ,
am c

C d l6
A Etf1, Eif4a1,Eif3a,
Rp psa

B rem
Ic 1, B

1, 1a,C 1, s1
s2 6

Hsp90aa1,
2.10 3.40 2.90

4, ,C Cc
Rp 1
A

T
R s5

c4 e,
l1

Rps2,Eif5, Denr,
1
2

Rp s28

Du d
od

Samb20,Pm, id
n
Hsp90b1,

27 crl2 l4,
Rpl58

Rp 27l
l1

Ra p,Os fkb
Ctss,
Rps

Cxcl30,Cxcl2,

Pn biz,N
Zfp36l2,Npc1
l3

Cd1,C , Cc
1.88 2.92

2,

sp 9,
Rpl6

Rps

Hspa1a,

Nfk
A
2.52

Npepps
Itpr2
Gbp7,
Rps25
Rplp0

Isg20,Isg15,
Irgm1,Il1a,
Rps2
Rps10

Ifit3,Ifit2,Ifit1
Ifi204
Ddit3,Bax,
Rps15
Rps19
Rp

Rps14

,Gbp

1,
sn1 tafr,
Hspa1b,

Cc Ccl3
Atf4,Atf3,
27

Csf1,
1.66 2.44 2.14 Hspa5,P4hb, Acod1,Rpl21,

,Rs
Rgcc,Rbpj

Gbp5,
Ppia,St13 Mdm2,Ppp1r15a,

Cxcl1
Alkbh5,Bnip3,
1.44 1.96

ad2
Dnajb9
1.76 Bnip3l,Cited2,
Egln3, Egr1

,
1.22 1.48
1.38 Fam162a, Higd1a
1.00 1.00
1.00

D Harvest orthotopic
PDAC tumors
E Cell type deconvolution G Mapped neutrophils T1 T2 T3 H T3 localization
Mapped
6 weeks post injection neutrophils ROI 1 (Tumor 1) ROI 2 (Tumor 1) * Glycolysishigh
80 Glycolysislow
Periphery
T1

Glycolysis
T2
T3 re 60
UMAP2

re Co
Necrotic Co

Pe
core 40

p
rip
Pe
UMAP1

he
rip
Freeze and 20

ry
he
cryosection
F Spatial clustering

ry
ry
Co

Co
1: EMT
rip
he * Hypoxiahigh
re

10x Visium

re
2: G2-M checkpoint ry

Frequency (%)
Spatial Gene Expression 3: p53 signaling ph
e Pe Hypoxialow
4: Angiogenesis
P eri 55

Hypoxia
5: Myc targets
6: Hypoxia
UMAP2

7: IL-2/STAT5 50
Data 8: signaling
analysis 9: IFN response
10: EMT 2 ROI 3 Periphery ROI 4 Periphery
UMAP1 45
Cell type

g
deconvolution
(Cell2location)
T1 mapped
* Angiogenesishigh
80 Angiogenesislow
Spatial clustering

Angiogenesis
and enhancement
(BayesSpace) T2 mapped 60

y
Gene 40
Core

signature
Core

mapping T3 mapped
20
(UCell) Core
1 2 3 4 5 6 7 8 9 10 Core High Low
Cluster UCell score

Fig. 3. Spatial compartmentalization of T1, T2, and T3 neutrophils in the neutrophils localize to different spatial clusters. Projection of T1, T2, and T3
pancreatic tumor. (A) Chord diagram showing differentially expressed genes enriched spots identified by Cell2location on merged UMAP derived from
(DEGs) in T1 neutrophils that are enriched for GO pathways linked to transcription BayeSpace enhanced clustering analysis of tumor sections (n = 4 biological
and oxidative phosphorylation. (B) Chord diagram showing DEGs in T2 neutrophils replicates). (F) Merged UMAP representation of spots of tumor sections were
that are enriched for GO pathways linked to metabolism and immune response. analyzed and color-coded according to BayesSpace-identified clusters (top).
(C) Chord diagram showing DEGs in T3 neutrophils that are enriched for GO Violin plots show frequency of T1, T2, and T3 neutrophils enriched spots that

y g
pathways linked to survival and angiogenesis. In (A) to (C), bars associated with map to each cluster (bottom). (G) Spatial mapping of T1, T2, and T3 neutrophils
each gene are colored by strength of fold change of differential expression and are across tumor sections (n = 4) by Cell2location. Black lines denote the outline
sized based on the number of pathways with which it interacts (see also table S4 of the section; gray colored areas indicated the excluded DAPI–panCKhi regions
for DEG lists). (D) Spatial transcriptomic analysis workflow. PDAC tumors were annotated to be fibrotic and or necrotic. Spots are filtered based on Ly6G+
isolated and cut into quarters, where the sharp edges denote the core facing regions, staining (see also fig. S6B). (H) Quantification of percentages of deconvoluted T3

y
before flash-freezing. Fresh frozen PDAC tumors were sectioned and placed neutrophil-enriched spots falling into high- or low-scoring spots for GO ontology

,
on 10X Visium slides containing spatially barcoded capture spots. After processing pathways: glycolysis (GO: 0061621), hypoxia (GO: 001666), and angiogenesis
and sequencing, the data were clustered spatially (BayesSpace) and cell type (GO: 0045766). Center line of boxplots show median, box hinges represent 25th
deconvolution was performed (Cell2location). Gene signatures of various biological and 75th percentiles, and whiskers extend to minimum and maximum values.
processes were then probed and mapped with the UCell package. (E) Tumor *P < 0.05 by one-tailed t test.

our 10X Visium analysis (fig. S6B). By contrast, flow cytometric strategy (Fig. 2D), and were ically evaluated staining for the tumor vascula-
Hif1a did not stain within the MICS time frame, successfully annotated in both confocal (fig. ture (CD105), hypoxia (CD73), and glycolysis
whereas CD39 mostly identified vessels (fig. S7D) and MICS (fig. S7E). (GLUT1) across all five ROIs (Fig. 4C and fig.
S7C). Similarly, GLUT1 (glucose transporter 1) We performed MICS for five different ROIs S8, A and B). CD105+ blood vessels were present
staining showed regional restriction (fig. S7C) across two different orthotopic pancreatic tu- as a diffuse network scattered throughout the
similar to glycolysis signature scores (fig. S6B). mors at 4 to 6 weeks after injection. ROIs were tumor (Fig. 4C and fig. S8, A and B). A gradient
Thus, CD73 and GLUT1 representative stain- selected to unequivocally cover an entire region of CD73 staining was apparent throughout
ing was used to define the hypoxic and gly- from the left to right margin encompassing the tumor, whereas GLUT1 staining was more
colytic tumor niche. T1, T2, and T3 neutrophils the core of the tumor (fig. S8, A and B), and tightly localized; however, GLUT1hi areas strong-
were identified by staining for Ly6G, CD101, we observed staining across the full panel for ly colocalized with CD73hi regions marking
and dcTRAIL-R1, respectively, based on our all markers surveyed (fig. S9). We next specif- a distinct hypoxic and glycolytic niche (Fig.

Ng et al., Science 383, eadf6493 (2024) 12 January 2024 6 of 16

RES EARCH | R E S E A R C H A R T I C L E

Fig. 4. T3 neutrophils occupy a hypoxic- A B ROI 1

Tumor E
glycolytic niche in pancreatic tumors. dg
e

(A) Workflow diagram of multiplexed

imaging using MICS technology. Cryo-
sections of halved PDAC tumors were or

C
e

placed on slides, fixed, and permeabilized dge

rE

o
before ROI selection with DAPI staining.

m
Tu
Sections were stained for 10 min per C D CD73 & GLUT1 F
cycle containing antibodies in FITC, PE,
and APC. After scanning, sections were
photobleached and scanned to subtract
background signals. After imaging the
desired cycles, images were registered,
segmented, and exported for conventional
flow cytometric annotation of cell types,
which were further used for spatial statistical
analysis with SPIAT. (B) Tumor pictograph
showing ROI1 selection area. (C) Immuno-
fluorescent image of ROI1 with indicated
stain markers of respective tumor regions.
Scale bar, 100 mm. (D) Left, Expression

p
marker intensity map of CD73 and 2.08e−17

GLUT1. Right, Coexpression plot of CD73 1.56e−17

1.04e−17

5.20e−18

and GLUT1 marker intensities. (E) Spatial 5.93e−24

mapping of each annotated tumor E

neutrophil subsets in ROI1. (F) Comapping
of neutrophil subsets in ROI1. (G) Left,

g
Gating strategy of CD45– CD105– tumor
regions demarcated by CD73 and GLUT1.
Right, Mapped gated regions on
segmented data of ROI1. (H) Spatial

y
statistical analysis of segmented tumor
neutrophils (n = 5 biological replicates).
Left, Average minimum distance of
each segmented neutrophil subset to
hypoxichighglycolytichigh regions. Center
line of boxplots show median, box hinges
represent 25th and 75th percentiles,
and whiskers extend to minimum and 100um

maximum values. *P < 0.05, **P < 0.01

by Mann-Whitney U test. Right, Colocal-
G H Nearest Distance to Co-localisation with

y g
Hypoxichigh Glycolytichigh Hypoxichigh Glycolytichigh
Hypoxichigh Glycolytichigh

ization of tumor neutrophils with Gated on CD45negCD105neg

segmented cells Region Region
hypoxichighglycolytichigh regions measured
1.5 T1 T2 T3
using a normalized mixing score for each Hypoxichigh
Glycolytichigh
Normalised Mixing Score

radius of interaction. Dots represent 3000

Mean Min Distance (Pixels)

mean across five replicates. Error bars 1.0

indicate SEM. ***P < 0.001 by Mann-

y
2000
Whitney test, corrected for multiple
Tumor

,
comparisons with a false discovery rate 0.5
1000
(FDR) of 1% using the two-stage step up
Stromal

(Benjamini, Krieger, and Yekutieli)

0 0.0
method. T1 T2 T3 50 100 150 200 250 300 350 400 450 500
Radius of interaction

4D and fig. S8, A and B). Annotation of T1
(Ly6G+CD101–dcTRAIL-R1–) and T2 (Ly6G+
CD101+dcTRAIL-R1–) neutrophils (fig. S8C)
revealed that they were distributed throughout
the tumor parenchyma in two different patterns
reflective of their likely routes into the tumor:
bordering the tumor edges in CD73– stromal
regions or diffused throughout GLUT1–CD73+
regions, both of which contained a substantial
network of CD105+ vessels (Fig. 4E and fig. S8, Next, we agnostically quantified the three
A and B). By contrast, T3 (Ly6G+CD101–/+dcTRAIL- neutrophil subgroups’ relative locations with-
R1+) neutrophils across all ROIs were more in the tumor. Using marker intensities on seg-
likely to cluster together at the CD73hiGLUT1hi mented cells, we excluded CD45+ immune
hypoxic and glycolytic niche (Fig. 4E), thus cells and CD105+ endothelial cells and sub-
giving rise to a clear spatial segregation of divided the remaining cells into three distinct
the three neutrophil populations, as suggested regions: the stromal niche (CD73–GLUT1–), the
by our 10X Visium analysis (Fig. 4F and fig. tumor parenchyma niche (CD73+GLUT1–),
S8, A and B). and the hypoxichighglycolytichigh tumor niche

Ng et al., Science 383, eadf6493 (2024) 12 January 2024 7 of 16

RES EARCH | R E S E A R C H A R T I C L E
(CD73hiGLUT1hi) (Fig. 4G and fig. S8C). In all five
ROIs, T3 neutrophils had the shortest average
minimum distance to the hypoxichighglycolytichigh
niche compared with T1 and T2 neutrophils
(Fig. 4H). We then assessed normalized mix-
ing scores, which quantify the proportion of
direct touches between the target regions and
reference neutrophil cells (target-reference)
against reference-reference interactions (48).
T3 neutrophils had the greatest mixing scores
within hypoxichighglycolytichigh tumor regions,
which was conserved with each radius of area
investigated (Fig. 4H). Scores for the stromal
region showed a trend for highest scores in T1,
followed by T2, with T3 being minimally de-
tected within the stroma (fig. S8D). Finally, T2
neutrophils trended toward higher propor-
tions of interspersion within the tumor paren-
chyma compared with T1 and T3 neutrophils
(fig. S8D). These observations suggest that
T1 and T2 neutrophils are mostly located
away from the hypoxichighglycolytichigh regions,
in contrast to T3 neutrophils. Therefore, our
spatial data imply that T3 neutrophils likely
migrate into and occupy the specialized hypoxic-
glycolytic tumor niche. This is consistent with
our finding that transcription factor regulons
enriched in T3 neutrophils (Fig. 1H) include
those up-regulated in response to hypoxia (e.g.,
Hif1a and Bhlhe40) and to metabolic and ER
stress (e.g., Atf4, Ddit3, Atf3, and Nfe2l2). Thus,
epigenetic and transcriptional up-regulation
of these pathways in T3 neutrophils could
confer an added survival advantage within
these tumor microenvironments. Collective-
ly, our data provide a spatial representation
by which tumor neutrophils converge upon
the occupancy of a specialized tumor niche
upon reprogramming.
Deterministic reprogramming of neutrophils
in the tumor
Because both mature and immature neutro-
phils were predicted to give rise to T3 neutro-
phils inside the tumor, we examined the impact
of the maturation state on the acquisition
of the T3 signature. We first isolated immature
and mature neutrophils from WT mice and
cultured them with tumor conditioned me-
dium (TCM), which allowed us to mimic the
infiltration of immature and mature neutro-
phils into the tumor microenvironment (Fig. 5A).
We then measured up-regulation of dcTRAIL-R1
on cultured neutrophils across multiple time
points as a proxy to estimate the acquisition of
the T3 profile (Fig. 5A). After 3 days, a substan-
tial proportion of BM immature and mature
neutrophils cultured in TCM were dcTRAIL-R1+,
but not those cultured in control media [i.e.,
complete Dulbecco’s modified Eagle’s medium
(cDMEM)] (Fig. 5A). Specifically, exposure to
TCM could trigger dcTRAIL-R1 expression 24
hours into culture and prolonged survival up
to 3 days compared with culture in cDMEM
Ng et al., Science 383, eadf6493 (2024) 12 January 2024
(fig. S10, A and B). We obtained similar results
regardless of tissue origin or maturation state,
such that neutrophils isolated from the BM,
spleen, and circulation cultured in TCM all
exhibited dcTRAIL-R1 up-regulation and con-
current increased survival (Fig. 5B and fig. S10, A
and B). Furthermore, exposure to TCM induced
the up-regulation of the T3 gene signature in
immature and mature neutrophils (Fig. 5C).
We then assessed whether culture in hypoxic
conditions would enhance neutrophil sur-
vival or dcTRAIL-R1 up-regulation of WT neu-
trophils. Culture in hypoxic conditions slightly
augmented neutrophil survival in vitro, espe-
cially for mature neutrophils (fig. S10C), but
did not induce dcTRAIL-R1 expression (fig.
S10D), indicating that hypoxic conditions are
not the main driver of T3 reprogramming.
Adoptive transfer of CD45.1+ immature and
mature neutrophils into tumor-bearing mice
(Fig. 5D) further confirmed that exposure to
the tumor microenvironment was necessary
for the up-regulation of dcTRAIL-R1 (Fig. 5E),
and dcTRAIL-R1 up-regulation in vivo followed
the same kinetics as in vitro (Fig. 5F and fig.
S11A). Therefore, exposure to tumor-derived
stimuli supports extended survival and
the acquisition of the T3 phenotype in WT
neutrophils.
Both TCM-cultured immature and mature
neutrophils up-regulated dcTRAIL-R1 expres-
sion and the T3 gene signature to an equal
degree, indicating that immature neutrophils
could be reprogrammed directly into T3 neu-
trophils (Fig. 5, A and C). RNA velocity analysis
suggested that this occurs in a stepwise fashion
as immature neutrophils transit into T1 neu-
trophils and subsequently differentiate into T3
neutrophils (Fig. 1E). As expected, transfer of
CD45.1+ immature neutrophils resulted in the
appearance of dcTRAIL-R1–CD101– T1 neutro-
phils within the tumor 1 day after transfer (fig.
S11B). By contrast, only dcTRAIL-R1–CD101+ T2
neutrophils, not T1 neutrophils, were observed
in the tumor with CD45.1+ mature neutrophil
adoptive transfer (fig. S11B). Both T1 and T2
populations derived from transferred cells
were undetected after 3 days within the tumor
because most of these transferred cells had
completed their reprogramming into dcTRAIL-
R1+ T3 neutrophils (fig. S11C). We further cor-
roborated the transitory nature of T1 and T2
subsets by in vitro culture (fig. S11, D and E).
Isolated T1 and T2 neutrophils up-regulated
dcTRAIL-R1 expression after 24 hours in cul-
ture. Unlike tumor-naïve neutrophils, dcTRAIL-
R1 up-regulation was independent of the culture
medium used for both T1 (fig. S11F) and T2 (fig.
S11G) neutrophils. Therefore, T1 and T2 neutro-
phils reflect immature and mature neutrophil
populations captured in the process of T3
differentiation and thus do not require any
further input from tumor soluble factors. Our
findings suggest a deterministic program
within neutrophils that enables them to
acquire the T3 phenotype independent of
their maturation stage.
T3 neutrophils are long-lived, terminal
effectors within the tumor microenvironment
To better understand the temporal regulation
of neutrophil differentiation into the T3 state
within the tumor microenvironment in vivo,
we administered 5-bromo-2′-deoxyuridine (BrdU)
intravenously at selected time points before
harvest, pulse-labeling all proliferating neu-
trophil precursors (Fig. 5G). This approach de-
fines a time stamp at the point of labeling,
allowing us to evaluate all time points simulta-
neously at the time of harvest, thereby mini-
mizing batch effects. Modeling the disappearance
rate from the peak of BrdU+ signal (fig. S12, A
and B) and incorporating the entire temporal
interval measured allowed the prediction of
neutrophil half-life and lifespans (fig. S12C),
p
reflecting their dwell time within each tissue
compartment (Fig. 5G).
The earliest peak of BrdU+ neutrophils was
observed in the BM and spleen, because of
active granulopoiesis within these organs in
the tumor-bearing state, whereas recruitment
g
of BrdU-labeled neutrophils into the blood
and subsequently the tumor placed their peak
of recruitment at ~4 to 5 days after labeling
(Fig. 5G). Because the dwell times of neutro-
y
phils in the BM and spleen are complicated by
continual neutrophil production, we focused
our comparison between the blood and tumor
neutrophils. Blood neutrophils had a half-life
of 31.4 hours, reflecting an increased transit
time in circulation compared with previously
predicted times for WT blood neutrophils
(49). Even so, tumor neutrophils had an even
longer half-life [41.8 hours; 95% confidence
interval (CI) = 39.6 to 46.6] and a predicted
y g
life span of 135 hours (up to 5.625 days; 95%
CI = 126.5 to 142.4) (Fig. 5G and fig. S12, C
and D), representing a marked extension of
neutrophil life span within the tumor micro-
environment. Compared with the unmarked
BrdU– fraction, the proportion of BrdU+ dcTRAIL-
y
R1+ neutrophils increased over time, surpassing
,
baseline levels at day 6 after labeling (fig. S12, E
and F). Newly recruited BrdU+ tumor neutro-
phils already expressed dcTRAIL-R1 at 1 day
after labeling (fig. S12E), suggesting that T3
reprogramming is initiated upon tumor entry.
dcTRAIL-R1 expression steadily increased and
was the highest in neutrophils at 15 days after
labeling (Fig. 5, H and I), confirming that
acquisition of the T3 phenotype was associ-
ated with their extended lifespans in vivo. Our
results show that a notable increase in the du-
ration of neutrophil half-life and residence in
the tumor compared with those in nontumor
tissues (fig. S12D), indicating that neutrophils can
survive long enough within the tumor to undergo
reprogramming and sustained persistence
8 of 16
RES EARCH | R E S E A R C H A R T I C L E

Fig. 5. Reprogramming within the A immature/

Sorted
Culture in cDMEM Flow B dcTRAIL-R1 expression C T3 gene signature
tumor environment results in mature vs TCM (tumor cytometry BM Immature BM Mature
neutrophils conditioned media) at D1, D3 cDMEM *** ***
long-lived, terminally differentiated 80 TCM *** ***
cDMEM TCM

% dcTRAIL-R1+ of live cells

T3 neutrophils. (A) Experimental 60

Log normalized Counts

105 D3 105 7.5
setup of in vitro culture of sorted 40 *** ***
104
Immature 104
(CD101-) 20
neutrophils from WT mice in cDMEM 103 103
0
versus TCM. Representative flow 6.45%
0
77.0%
0
5.0
-10
3 Blood Mature -10
3 Spleen Mature
cytometry plots show dcTRAIL-R1 0 104 105 0 104 105

D3 80
expression increases on sorted
105
*105
***
104 Mature 60 104
2.5
immature and mature WT neutrophils (CD101 ) +

Ly6G
103 40 103

from the BM after 3 days of culture in 0

20 0 *
75.1%
-10 -10

TCM but not in cDMEM. (B) Neutro- 3

0 0104 105
3
0 104 105
D0 D3 D0 D3
dcTRAIL-R1 D1 D3 D1 D3 WT BM IMM WT BM MAT
phils cultured in TCM up-regulate
dcTRAIL-R1 over time. Line plots show
the percentage of dcTRAIL-R1+ cells
D Sort CD45.1+ Immature CD45.1+ CD45.1+
Mature E Blood Tumor
D3 Immature transferred F Immature
transferred
Mature
transferred
gated as in (B), dots indicate the BM neutrophils transferred transferred 105 105
Gated on Tumor
median, and error bars indicate Q1 and live, Ly6G Spleen + 104 104

% dcTRAIL−R1+
CD45.1 6.06% 79.6% 100 Blood 100 +
Q3 intervals for neutrophil subsets neutrophils
103
0
103
0
80 80
cultured in cDMEM (dotted line) and Inject i.v. into
60 60
-103
3
-10 0 103 104 105
-103
-103 0 103 104 105
PDAC tumor D3 Mature transferred
TCM (solid line) over 1 and 3 days.

*
*

*
bearing mice 40 40 10 5
10 5

Each group contains the following

p
104 104

Ly6G
80.4% 20 20
number of samples: day 1 cDMEM: BM Flow cytometry
0 0
103
0
103
0
Blood Spleen Tumor D1 D3 D1 D3 -103 -103
immature (n = 8), BM mature (n = 8), at D1, D3
dcTRAIL-R1 Days post transfer
-10 3 0 10 3
104
10 5
-103 0 103 104 105

spleen mature (n = 8), blood mature

(n = 3); day 1 TCM: BM immature
(n = 8), BM mature (n = 8), spleen G Orthotopic
PDAC injection
BrdU label
for timepoints
Harvest all
timepoints H Gated on live, Ly6G+
tumor neutrophils I dcTRAIL-R1 gMFI
+
Fold vs = BrdU gMFI
mature (n = 8), blood mature (n = 5); 0 6 weeks BrdU +
( BrdU- gMFI (

g
D15 baseline
BrdU-
day 3 cDMEM: BM immature (n = 10), D15 D12 D8 D6 D4 D3 D2 D1
BM mature (n = 10), spleen mature BM Spleen D12 2.5 *
(n = 10), and blood mature (n = 4); 100 100
*
%BrdU+ (normalized to max)

day 3 TCM: BM immature (n = 10), BM D8

(dcTRAIL-R1 gMFI)
Fold vs baseline

y
50 50 2.0
mature (n = 10), spleen mature (n = 10), **
D6
blood mature (n = 5) performed 0 0
0 5 10 15 0 5 10 15
across seven independent experiments. Blood Tumor D4 1.5
*
*P < 0.05, **P < 0.01, ***P < 0.001 by
100 t½= 31.4h 100 t½= 41.8h
Mann-Whitney U test. (C) Neutrophils t5%= 91.5h t5%= 135.0h
D3
1.0
cultured in TCM up-regulate the 50 50
D2
T3 gene signature. Scatter dot plots 0 0
for T3 gene signature expression in 0 5 10 15 0 5 10 15 -10 0 10 10 10 2 3 4 6 8 12 15 3 3 4 5

Days post BrdU labelling dcTRAIL-R1 Days post BrdU labelling

BM immature (WT BM IMM) and
mature (WT BM MAT) neutrophils that

y g
were freshly sorted (D0) or cultured J Sorted T3 neutrophils K dcTRAIL-R1 expression L T3 gene signature
for 3 days (D3) (n = 3 for all samples). D0 D1, D3 n.s. n.s. n.s. n.s. 10 n.s.
n.s.
culture 100
Each dot denotes a single gene, lines
Log normalized Counts
% of total neutrophils

8
denote the mean, and error bars Sorted cDMEM TCM
75
indicate the SEM. ***P < 0.001 by 6
Loss of T3
Wilcoxon signed-rank test with

y
phenotype? 50
4
Bonferonni’s correction, with compar-

,
106 106
isons indicated on the graph. 105
25
105
2
(D) Experimental setup for transfer of 104 104

CD45.1+ neutrophils into pancreatic 0

Ly6G

103 0 103
91.7% 94.3% 0 d EM TCM MEM TCM
0 d M M
rte rte ME TC
tumor-bearing mice. Sorted WT 10-3
10 -3
0
S o
10
cD
3
M
4
10
cD 10 5
10-3
10-3 0
So
103 104
cD
105

BM immature and mature neutrophils dcTRAIL-R1 D0 D1 D3 D0 D1

were intravenously injected into WT
PDAC tumor-bearing mice. At 1 and 3 days after transfer, CD45.1+ neutrophils
were evaluated within the blood, spleen, and tumor for dcTRAIL-R1 expression by
flow cytometry. (E) Up-regulation of dcTRAIL-R1 expression was restricted
to the tumor. Representative flow plots show dcTRAIL-R1 expression present on
transferred CD45.1+ immature and mature neutrophils present in the blood or
tumor at 3 days after transfer. (F) Line plots show proportion of WT BM CD45.1+
immature (n = 4 at day 1 and n = 3 at day 3) or mature neutrophils (n = 3
at both time points) expressing dcTRAIL-R1 performed across two independent
experiments. Each dot denotes a single gene, lines denote the mean, and error
bars indicate SEM. *P < 0.05 by Kruskal-Wallis test with Dunn’s post test.
(G) Experimental setup for BrdU pulse labeling in tumor-bearing mice. WT mice
received orthotopic injection of the PDAC cells, and the tumor was allowed
to grow. At days 15 (n = 4), 12 (n = 4), 8 (n = 5), 6 (n = 3), 4 (n = 4), 3 (n = 4),
2 (n = 3), and 1 (n = 4) before the harvest, mice were injected with BrdU, thus
labeling proliferating neutrophil precursors within the BM and spleen. Data were
collected across three independent experiments. At day 42 (6 weeks) after
injection, mice were sacrificed and BrdU+ neutrophils within the BM, spleen, blood,
and tumor were quantified. BrdU percentages at all time points were then normalized
to the maximal BrdU+ percentage value for each tissue, which was set to
100%. Dots represent mean expression, with error bars denoting 95% CIs.

Ng et al., Science 383, eadf6493 (2024) 12 January 2024 9 of 16

RES EARCH | R E S E A R C H A R T I C L E

A mathematical model capturing the full temporal window was fitted to estimate
half-life (t1/2) and life span (t5%) for each organ, denoted on each plot in
hours (see also fig. S12C and the materials and methods). (H) BrdU labeled
neutrophils up-regulate dcTRAIL-R1 over time. Histograms show geometric MFI
(gMFI) of dcTRAIL-R1 within BrdU+ (orange) and BrdU– (gray) neutrophils at
days 2, 3, 4, 6, 8, 12, and 15 after BrdU labeling. (I) Quantification of gMFI in (H).
Line plots show fold change of dcTRAIL-R1 gMFI of BrdU+ against BrdU–
neutrophils. BrdU– neutrophils served as a measure of baseline dcTRAIL-R1 gMFI
within the tumor. Each dot denotes the mean, with error bars indicating SEM.
*P < 0.05, **P < 0.01 by Mann-Whitney U test, one-tailed, alternative = “greater.”
(J) Experimental setup of in vitro culture of sorted T3 neutrophils from PDAC
mice in cDMEM or TCM at 1 or 3 days. Representative flow cytometry plots show
that dcTRAIL-R1 expression is retained on T3 neutrophils after 1 day of culture in
both cDMEM and TCM. (K) Boxplots show quantification of frequency of
dcTRAIL-R1+ neutrophils (n = 5 each performed across three independent
experiments) in (I). Each dot represents one biological replicate, center line of
boxplots show median, box hinges represent 25th and 75th percentiles, and
whiskers extend to minimum and maximum values. P = n.s. (not significant) by
Kruskal-Wallis followed by many-to-one U test comparing against the D0
time point. (L) T3 neutrophils cultured overnight do not down-regulate the T3
gene signature. Scatter dot plots for T3 gene signatures in sorted, cDMEM-,
or TCM-cultured neutrophils (n = 3 each). Each dot denotes a single gene, lines
denote the mean, and error bars indicate SEM. P = n.s. by Kruskal-Wallis
followed by Dunn’s post test.

within the tumor as long-lived dcTRAIL-R1+ T3
neutrophils.
To determine whether the T3 state is tran-
sitory or is stably maintained once it has been
acquired, we isolated and cultured T3 neutro-
phils overnight and up to 3 days in the pres-
ence or absence of TCM (Fig. 5J). Survival of
T3 neutrophils in culture was not affected by
neutrophil plugs within the same mouse (fig.
S13B), and this increase in vascularization, al-
though modest, was consistent across all mice
observed in the assay (fig. S13C). By contrast,
co-injection of T1 and T2 neutrophils did not
enhance vascular flow (fig. S13, B and C), in-
dicating that the T3 population contains the
most potent angiogenic ability (Fig. 3C).
trols showed greater distribution along the
tumor edges (Fig. 6H and movie S2). Normal-
ization for tumor size revealed that both T3
and WTBM MAT–co-injected tumors had sim-
ilar densities of CD31+ blood vessels (fig. S13F)
but substantially differed in their distribution
(Fig. 6I). Although T3 neutrophils were the
highest expressors of Vegfa in the pancreatic

the medium type (fig. S12G), and dcTRAIL-R1
expression was maintained despite the ab-
sence of tumor-derived factors (Fig. 5K), with
expression levels comparable to freshly iso-
lated T3 neutrophils at both time points (fig.
S12H). Additionally, T3 neutrophils did not
The tumor triggers angiogenesis to main-
tain the supply of oxygen and other nutrients
to the core of the tumor to sustain continual
growth (54). To test whether T3 neutrophils
residing in their hypoxic-glycolytic niche at
the tumor core support this angiogenic switch,
p
tumor, Vegfa expression was also detected in
macrophage populations (fig. S13G), indicat-
ing that neutrophils may not be the sole pro-
angiogenic contributor for tumor growth support.
Nonetheless, blockade of T3 neutrophils with
an anti–dcTRAIL-R1 antibody decreased tumor

down-regulate their gene signature in culture
(Fig. 5L). This in vitro evidence indicates that
once neutrophils have been reprogrammed into
T3 neutrophils, they do not revert their pheno-
we modified our PDAC model by implanting
PDAC tumors subcutaneously to enable longi-
tudinal tumor size measurements. We then
assessed whether injection of neutrophils within
g
growth in T3-co-injected tumors compared with
isotype controls (Fig. 6J), indicating that T3 neu-
trophils are the predominant cells responsible
for increasing the tumor growth rate, most likely

type in the absence of supportive tumor factors
and they represent the terminally differentiated
neutrophil population within the tumor.
T3 neutrophils are pro-angiogenic and promote
tumor growth
Prior studies have identified key roles for neu-
trophils in tumor progression, especially through
the promotion of tumor angiogenesis (50–52).
We therefore sought to determine the func-
the tumor affected their subsequent growth (Fig.
6D). PDAC cells co-injected with T3 neutrophils
formed tumors that grew rapidly (Fig. 6D),
whereas co-injection with other neutrophil
types had little influence on tumor growth. T3-
co-injected tumors had a 100% tumor growth
rate (Fig. 6E, as evaluated in fig. S13D) and had
the greatest mass at the end point (Fig. 6F). T2-
co-injected tumors showed the lowest tumor
incidence rates (5 of 12) (Fig. 6F), suggestive of
y
through vascular remodeling.
Evidence for conserved neutrophil
reprogramming across tumor types
and human PDAC
Phenotypic differences in tumor-infiltrating
neutrophils have been observed across mouse
and human cancers, and multiple subsets have
been characterized by differential surface marker
or transcriptome expression (6–13). Whether

tional specialization of T1 to T3 neutrophils in
promoting tumor growth. Given their pheno-
typic stability, distribution near angiogenic re-
gions, and the expression of a pro-angiogenic
transcriptional signature (Fig. 3C), we specu-
lated that T3 neutrophils promoted angio-
increased tumor rejection and consistent with
the expression of proinflammatory genes in T2
neutrophils detected in our scRNAseq dataset
(Fig. 3B). T3-co-injected tumors continued to
grow rapidly for up to 3 weeks after the last
injection of neutrophils, suggesting that even
y g
such neutrophil profiles are conserved across
tumor types and species remains unclear. We
therefore assessed whether T1 to T3 neutrophil
states can be detected in previously published
mouse cancer scRNAseq datasets containing
annotated neutrophils (11, 13). We projected

y
genesis from their hypoxic-glycolytic niche to after their disappearance, transferred T3 neu- these datasets onto a reference UMAP embed-

support continual tumor growth. By contrast,
we anticipated that the transient T1 and T2
neutrophil states would not yet feature pro-
angiogenic ability. T3 neutrophils had the high-
est transcript (Fig. 6A) and protein expression
(Fig. 6, B and C) of Vegfa (vascular endothelial
growth factor a) compared with the other neu-
trophil subsets in the tumor and periphery. We
then evaluated whether T3 neutrophils had a
greater capacity to induce blood vessel forma-
tion in vivo using a modified Matrigel plug
assay and measured angiogenesis by the rate
of vascular flow measured by Doppler imag-
ing (53) (fig. S13A). Matrigel plugs co-injected
with T3 neutrophils showed the greatest
flux intensity compared with control WT BM
trophils generated long-lasting effects that
sustained this accelerated growth rate. Neu-
tralization of VEGFa in our model resulted in
the reduction of growth in T3-co-injected tumors,
but had no impact on the growth of WT BM
mature (WTBM MAT)–co-injected tumors
(Fig. 6G). To determine whether this growth
enhancement was due to increased angiogenesis
driven by T3 neutrophils, we optically cleared
T3- and WTBM MAT–co-injected tumors (fig.
S13E) and visualized the intratumoral three-
dimensional (3D) vessel network through CD31
staining (Fig. 6H). T3-co-injected tumors had
greater CD31 staining density toward the tu-
mor core (Fig. 6H and movie S1), whereas the
vessel network in WTBM MAT–co-injected con-
,
ding, mapping each cell on the same UMAP
space as our dataset (fig. S14A). As in the PDAC
model, neutrophils within Lewis lung carcinoma
tumors (13) could be assigned to the T1 to T3
states, whereas spleens from WT or tumor-
bearing mice mostly contained nontumor neu-
trophil clusters (fig. S14B). An intermediate
neutrophil subset identified in cancer-bearing
individuals [PMN2 in (13)] was enriched in the
preNeu and IMM 1 and 2 clusters (fig. S14C),
whereas T1 to T3 clusters scored highly for the
tumor-specific neutrophils (PMN3; fig. S14D).
Similarly, reanalysis of a different scRNAseq
dataset showed that neutrophils from their
KP1.9 tumor-bearing lung model (11) mapped
as T1 to T3 clusters, whereas the normal lung

Ng et al., Science 383, eadf6493 (2024) 12 January 2024 10 of 16

RES EARCH | R E S E A R C H A R T I C L E

Fig. 6. Protumoral T3 neutrophils A Vegfa Exp. density C 100 * ** D

promote tumor growth and associate *
Co-inject tumor

0. 25
0. 50
0. 00

5
80

+
07
Tumor volume

0
0
0
± neus.(s.c.)

0.
with poorer patient outcomes. (A) T3

% VEGF
60 D0 Co-injected:
neutrophils have the highest expression 40
D3
PBS control
WTBM IMM
T1
T2
*
Boost
of Vegfa transcripts. UMAP projection 20 neus. 1000 WTBM MAT T3

UMAP2

Volume (mm3)
0
of total neutrophils in BM, spleen, blood,
UMAP1
Spleen T1 T2 T3
Tumor Measure
D7 750 **
and tumor show expression density of tumor
B
+
Gated on live, Ly6G neutrophils 500
Vegfa. (B) T3 neutrophils have the volume D14
Spleen T1 T2 T3
*
highest expression of VEGFa. Repre- 106

105
106

105 D21
106

105
250 106
*
105
Endpoint
sentative flow plots show intracellular 104 104
volume/
104 104

103 103 weight 103 14 21 28

103
D28
VEGFa protein staining in neutrophils 10.2% 8.12% 66.4% 88.5% Days post tumor injection

Ly6G
0 0 0 0
-103 -103 -103 -103

from the spleen (n = 3) and T1, T2, and VEGF

-103 0 103 104 105 -103 0 103 104 105 -103 0 103 104 105 -103 0 103 104 105

T3 neutrophils (n = 5 biological repli-

cates) in the pancreatic tumor across
E Tumor incidence G VEGF inhibition H 3D vessel quantification
two independent experiments. (C) Box- Tumor No tumor Co-inject Boost Measure
PBS WTBM WTBM tumor + neus. tumor vol. T3 co-injected tumor
plots quantify VEGFa expression as 87.5% IMM
70%
MAT
83.3%
T3/WTBM + ab. to endpoint 0.63X 2X
MAT neus.
shown in (B). Each dot represents one T1
85.7%
T2
58.3%
T3
100%
± ab. CD31 Left margin Core

biological replicate, center line of boxplots D0 D3 D7 D14 D21

Left margin
show median, box hinges represent F Endpoint weight Co-injected antibody:
Isotype
T3: WTBM Isotype
25th and 75th percentiles, and whiskers 2.5 * α -VEGFα MAT: α -VEGFα
1500 µM
extend to minimum and maximum * ** 1500 µM

p
1000
2.0
values. *P < 0.05, **P < 0.01 by Kruskal-
Weight (g)

WTBM MAT co-injected tumor

Volume (mm3)
750
1.5 Left margin Core
Wallis followed by many-to-one U test n.s. CD31

1.0 n.s. n.s. n.s. n.s. 500 n.s.

Right margin
comparing against T3 neutrophils. (D) T3

Left margin
neutrophils promote rapid tumor growth. 0.5 250

Schematic of experimental setup to 0 0

PBS IMM MAT T1 T2 T3 7 14 7 14
determine the ability of neutrophils to Days post last ab. injection
1500 µM 1500 µM

g
WT BM Tumor
promote tumor growth in vivo. Equal
numbers of neutrophils and PDAC cells
were mixed in Matrigel before sub-
I Left margin Core Right margin
J dcTRAIL-R1 inhibition
Co-inject Co-injected antibody:
tumor + T3 neus. Isotype α -dcTRAIL-R1
cutaneous injection into the right flank. Centre Co-injected with + isotype/
(normalized to quantile vol.)

y
Neutrophils in the tumor was boosted * T3 α-dcTRAIL-R1
800
0.20 WTBM MAT D0
CD31+ volume

on day 3 (D3) and D7 by direct

Volume (mm3)
D3 600
subcutaneous injection into the visible
0.15
* Boost
T3 neus.
*
matrigel plug–tumor. Tumor growth 0.10 + antibody 400
D7
was measured by calipers weekly until 0.05 200
Measure D14
the day 28 end point. Line plots show 0.00 tumor 0
volume of measured subcutaneous 0 0 0 0 0 0 0 0 0 0 10 20 30 40 50 60 70 80 90 00 volume to 7 14
10 9 8 7 6 5 4 3 2 1 1 endpoint D21
Quantiles from centre Days post last ab. injection
tumors co-injected with PBS (n = 8), WT
BM immature (WTBM IMM, n = 10), K Overall survival
TCGA (PAAD) dataset Australia (PACA) dataset
L TCGA Pan-cancer dataset
(Liu et al., 2018)
WT BM mature (WTBM MAT, n = 12), and 1.0 T3 sig. T3 sig. high high

y g
(652 days) (427 days)
T1 (n = 7), T2 (n = 12), and T3 (n = 7) 0.8 T3 sig. T3 sig.
Scored
low
for Scored for Scored for low

(1332 days) (709 days) T3 signature T2 signature T1 signature

neutrophils. Data were collected across 0.6
Cumulative survival

five independent experiments. Dots show 0.4 BRCA

0.2 SKCM
means, with error bars indicating SEM p.val = 0.045* p.val = 0.026* COAD
0.0
over all three measurement time points. KIRC
Disease-free survival ACC
*P < 0.05, **P < 0.01 by Kruskal-Wallis

y
1.0
T3 sig. T3 sig. SARC high high

test followed by many-to-one U test 0.8 T3 sig. (293 days)

LUAD low

,
low
T3 sig.
(476 days) HNSC
comparing all other conditions to PBS 0.6
PAAD
0.4
co-injection. (E) Quantification of PDAC LGG
0.2 ESCA
subcutaneous tumor incidence after p.val = 0.055 p.val = 0.022* CESC
0.0
neutrophil co-injection (see fig. S13D for 0 00 00 00 00 500 00
5 10 15 20 2 30
0
50
0
10
00 500 000
1 2
1 2 3 1 2 3 1 2 3
quantification method). Piecharts show Days Days Hazard ratio

frequency of mice with tumors (light

gray) or no tumor growth (dark gray) at day 28 end point in (D). (F) Tumors
co-injected with T3 neutrophils had the biggest weights compared other
conditions. Each dot represents one biological replicate, with boxplots showing
median with IQR and whiskers indicating lowest and highest measurement.
P < 0.05* by Kruskal-Wallis test followed by many-to-one U test comparing all
other conditions with PBS co-injection. (G) Antibody neutralization of VEGFa
reduces T3-mediated acceleration of tumor growth. Schematic of experimental
setup with sorted T3 neutrophils are co-injected with PDAC tumor with anti-
VEGFa targeting antibody (n = 10) or isotype control (n = 7). WT BM mature
neutrophils were used as controls (anti-VEGFa targeting antibody, n = 4, isotype,
n = 5). Data were collected across three independent experiments. Neutrophils
were further injected on D3 and D7 with the corresponding antibody added,
and tumor growth was measured at D14 and D21, 7 and 14 days after tumor
injection. Boxplots show median tumor volume, where each dot represents one
biological replicate, box hinges represent the interquartile range, and whiskers
extend to the minimum and maximum values. P < 0.05* by by Mann-Whitney
U test. (H) Visualization of CD31 vessels within T3 and WT BM MAT co-injected
subcutaneous tumors. Subcutaneous tumors from (A) for T3 (n = 3) and
WTBM MAT (n = 3 biological replicates) were dissected, optically cleared, and
permeabilized, stained with anti-CD31 antibody, and imaged in 3D at 0.65×

Ng et al., Science 383, eadf6493 (2024) 12 January 2024 11 of 16

RES EARCH | R E S E A R C H A R T I C L E

(100% of total volume) and 2× (35% of total volume starting from midpoint)
resolution. Data were collected across three independent experiments.
Representative 3D immunofluorescence images show T3 (top, quarter) and
WTBM MAT (bottom, whole) co-injected tumors, with tumor margins marked out
in white dotted lines and indicated. (I) Subcutaneous PDAC tumors co-injected
with T3 neutrophils have greater CD31 vessel density within the tumor core.
Quantification of CD31 staining intensity at 2× resolution as in (E). CD31 staining
was surfaced with a seedpoint of 16.2, and binned in 10% quantiles according
to the distance from left or right margins toward the tumor core. To account
for differences in tumor sizes, CD31 staining intensity was further normalized
over total slice volume for each quantile. Line plots represent volume-normalized
staining intensity, with dots representing the mean and errors bars indicating
SEM for T3 (orange)– or WTBM MAT (gray)–co-injected tumors. *P < 0.05
by Mann-Whitney U test, one-tailed, alternative = “greater.” (J) Antibody-
mediated blockade of T3 neutrophils reduces their ability to promote rapid
tumor growth. Schematic of experimental setup. Sorted T3 neutrophils were
co-injected with PDAC tumor with anti–dcTRAIL-R1 targeting antibody (n = 8) or
isotype control (n = 8). T3 neutrophils were further injected on D3 and D7
with the corresponding antibody added, and tumor growth was measured at D14
and D21, 7 and 14 days after tumor injection. Data were collected across four
independent experiments. Boxplots show median tumor volume, with each point
representing one biological replicate, box hinges represent the interquartile
range with whiskers extending to minimum and maximum values. *P < 0.05 by
Mann-Whitney U test. (K) The T3 neutrophil gene signature is associated
with poorer patient OS and DFS in pancreatic cancer. Kaplan-Meier plots show
OS (top) and DFS (bottom) for patients in the TCGA-PAAD and PACA-AU
datasets. Patients were split into high and low expression of the T3 curated
signature. Median OS and DFS survival are represented on the graph when
available. Events are represented by vertical lines and were defined from days to
death (from initial pathological diagnosis) or days to first event (from initial
treatment, for TCGA, and from clinical disease-free diagnosis for PACA-AU).
*P < 0.05 as calculated by log-rank test. (L) The T3 neutrophil gene signature
is associated with poorer patient OS in a subset of solid human cancers.
Each data set within the curated TCGA Pan Cancer dataset was scored for T1,
T2, and T3 signatures (see also table S4). Forest plots show hazard ratio (HR)
scores associated with patient OS that were significant (P < 0.05) by Cox
proportional hazards test across all signatures. Dots indicate the calculated HR
and whiskers indicate 95% CIs.

tissue was mostly enriched for mature neu- tissue (fig. S15B). Our findings suggest that cancers, including pancreatic cancer [from the

trophils (fig. S14E). Likewise, neutrophil types
identified as tumor-specific neutrophil clusters
in this model [mN3-6 in (11)] corresponded to
T1 to T3 clusters (fig. S14F), confirming that the
tumor microenvironment induces a prototypi-
cal transcriptional trajectory that is determin-
tumor-induced reprogramming of neutrophils
is conserved in humans.
Because T3 neutrophils promote the growth
of pancreatic tumors, we hypothesized that
the genetic signature associated with the T3
state might be predictive of pancreatic can-
p
The Cancer Genomics Atlas Pancreatic Adeno-
carcinoma (TCGA-PAAD); see Fig. 6L and table
S7 for P values and confidence intervals]. By con-
trast, T1 and T2 signatures were both protective
against (lower hazard ratios, HRs) or associated
with (higher HRs) patient death depending on

istic in nature. By contrast, the signature of
neutrophils from healthy lungs (49) was large-
ly independent of tumor-induced transcrip-
tional changes (fig. S14G). These data suggest
cer outcomes in human patients. To test this,
we performed survival analysis on two inde-
pendent pancreatic cancer cohorts from The
Cancer Genome Atlas (TCGA) and Interna-
g
the type of solid tumor (Fig. 6L and table S7),
which is consistent with their nature as transi-
tional subsets in the process of reprogramming.
These data support a model in which tumor-

that reprogramming of tumor neutrophils is
conserved across different tumor types, and
that T3 neutrophils represent terminal differ-
entiated tumor neutrophils.
To examine cross-species conservation of the
neutrophil tumoral trajectory found in mice,
we investigated whether our neutrophil clas-
sification could account for the heterogeneity
present in human tumors. We examined in-
dependent scRNAseq datasets (55, 56) from
tional Cancer Genome Consortium Pancreatic
Cancer-Australia (PACA-AU) by scoring pa-
tients on the basis of high and low expression
of each signature (see the materials and meth-
ods for scoring criteria). Patients with high
expression of the T3 neutrophil signature had
poorer overall survival (OS) across both data-
sets, with a median of 652 (TCGA) and 427
(PACA-AU) days, respectively (Fig. 6K), and
this was independent of potential confound-
y
educated T3 neutrophils drive tumor progres-
sion in both mouse and human cancers.
Discussion
Environmental cues can fine-tune immune re-
sponses by inducing cell recruitment and ex-
pansion of immune cells, generating productive
responses with adequate numbers and spe-
cialized phenotypes. Unable to further prolifer-
ate, neutrophils rely on their ability to swiftly

two human PDAC cohorts (fig. S14, H and I)
and mapped them to a simplified reference
UMAP embedding through label transfer (fig.
S14J). T2 and T3 neutrophils were amply labeled
in both datasets, and a smaller cluster of T1
neutrophils was also identified (fig. S14, K and
ers such as patient gender, age, and tumor
stage (table S5). Similarly, when disease-free
survival (DFS, assessed as time to an adverse
event from the initial treatment) was evaluated,
patients with high T3 neutrophil signature ex-
pression had reduced DFS (Fig. 6M), with equal
y g
mobilize into tissues to perform their func-
tions effectively, and in diseases such as can-
cer, neutrophils at various maturation stages,
tissue origins, and phenotypes are recruited into
the tumor in large numbers (12–14, 17, 19). Given
the likely coexistence of multiple tumor neutro-

y
L). These tumor neutrophil subsets were pre- distribution of potential confounders across phil states with different functional pheno-

dominantly enriched in the pancreatic tumor
and not in adjacent normal pancreatic tissue
(fig. S14K) or in the peripheral blood (fig. S14L).
A pro-angiogenic neutrophil type expressing
genes linked to hypoxia and glycolysis in one
of these studies [referred to as TAN-1 (56)]
strongly matched the T3 state in the murine
tumor, indicative that our T3 classification can
capture subsets independently identified to be
protumoral (fig. S15A). Similarly, tumor neu-
trophil gene signatures derived from our mouse
dataset (fig. S15B) were conserved in their
ability to distinguish between human tumor
neutrophils and showed strong specificity in
identifying T1, T2, and T3 neutrophils within
the tumor and not in the adjacent normal
the two groups (table S6). When T1 and T2
signatures were considered, high T2 signature
expression correlated with worse OS only in
the TCGA dataset (fig. S15, C and D), but this
was confounded by gender (table S5), whereas
high T1 signature expression was associated
with lower DFS only in the PACA-AU dataset
(fig. S15, E and F). Therefore, only the T3
signature was consistently associated with
poorer OS and DFS across both datasets. We
next considered whether the T3 neutrophil
signature was also associated with poorer OS
in other solid tumors. Within the TCGA pan-
cancer database (57), higher expression of the
T3 signature was associated with a significant-
ly higher risk of death across a subset of solid
,
types, it is thus unclear how neutrophils exert
a focused local effect in the tumor. Here, we
examined how neutrophils decouple their ini-
tial maturation phenotype from their eventual
protumoral function by undergoing conver-
gent reprograming within the tumor.
Our study demonstrates that tumor-infiltrating
immature and mature neutrophils acquire dis-
tinct epigenetic, transcriptomic, and proteomic
phenotypes while retaining features of their ini-
tial maturation status. We found that both im-
mature (T1) and mature (T2) tumor neutrophils
within the tumor converged toward a third pop-
ulation (T3) expressing dcTRAIl-R1. T3 neutro-
phils thus show intermediate maturation scores
when quantified at the population level, and have

Ng et al., Science 383, eadf6493 (2024) 12 January 2024 12 of 16

RES EARCH | R E S E A R C H A R T I C L E
both toroidal (immature) and hypersegmented
(mature) neutrophil nuclei, representative of
their being an admixture of reprogrammed
neutrophils of T1 and T2 origin. Our data re-
veal the capacity of neutrophils to simultane-
ously integrate different types of signals from
the environment, allowing them to layer a new
functional phenotype onto their pre-existing
differentiation stage. We demonstrated this
plasticity with immature and mature neutro-
phils from tumor-naïve mice, which are both
capable of acquiring phenotypic (dcTRAIL-R1
expression) and transcriptional traits of the T3
state when cultured in tumor-conditioned me-
dium or upon entering the tumor in vivo. By
circumventing neutrophil maturation as a rate-
limiting step, this adaptability allows imma-
ture neutrophils to be equally mobilized and
reprogrammed within the tumor in a shorter
time frame. Subsequently, this intrinsic ability
embedded in neutrophils consolidates the var-
ious functional neutrophil states into one ter-
minal neutrophil phenotype as directed by the
tissue, in this case, a tumor.
Circulating neutrophils have a predicted half-
life of ~10 hours (58), whereas tissue-resident
neutrophils persist for up to 1 day (49). Although
studies have hinted that neutrophils persist far
longer within the tumor (10, 59, 60), whether
this coincides with terminal differentiation in
the tumor remains an open question. Using a
BrdU pulse-labeling approach, we found that
up to 5% of the originally labeled neutrophils
could remain within the tumors for as long as
5.625 days upon entry. In addition, whereas
our curve fit was accurate for the first labeling
time points, there was a noticeable underfit-
ting at 12 and 15 days after labeling, represent-
ing a possible underestimation of the full
tumor neutrophil life span and indicating that
a small population of neutrophils could persist
even longer within the tumor. Long-lived neu-
trophils were predominantly of the dcTRAIL-R1hi
T3 phenotype, which indicates a correlation
between their ability to survive within chal-
lenging hypoxic and glycolytic environments
and their continued persistence within the
tumor. Spatial mapping at the transcriptome
and protein level placed T3 neutrophils pre-
dominantly within a hypoxic-glycolytic niche
nearer to the tumor core. By contrast, T1 and
T2 neutrophils were positioned at the stromal
and tumor parenchyma, where a large vessel
network exists, supporting the notion that
these cell states are in the process of reprogram-
ming after tumor entry. Given that hypoxia
is not required to trigger T3 reprogramming,
which can occur in normoxic conditions, we
propose that migration toward the hypoxic-
glycolytic niche occurs after acquisition of
T3 epigenetic and transcriptional programs
is fully complete to ensure neutrophil survival.
Consistent with this possibility, upstream tran-
scription factors regulating metabolic and oxida-
Ng et al., Science 383, eadf6493 (2024) 12 January 2024
tive stress are switched on in T3 neutrophils
compared with T1 or T2 populations. Deletion or
inhibiting these transcription factors to trace the
timeline of T3 reprogramming thus represent im-
portant goals for future studies within the field.
T3 neutrophils accelerated early tumor growth
experimentally, and although we still observed
tumor growth with other co-injected neutro-
phil subsets, this occurred at a slower rate,
likely due to recruitment and reprogramming
of endogenous T3 neutrophils at the later time
points. T1 and T2 neutrophils did not show sta-
tistically significant enhancement or inhibition
of tumor growth compared with other control
neutrophil populations or phosphate-buffered
saline (PBS), reflecting their transitional nature
as opposed to a fully anti- or protumoral popula-
tion. By contrast, the sustained growth advantage
conferred by T3 neutrophils stemmed, at least
in part, from T3-dependent remodeling of the
vasculature toward the core of the tumor. Given
their localization within the glycolytic-hypoxic
niche, an intriguing scenario is that T3 neutro-
phils serve as possible guide rails to direct angio-
genesis to relieve hypoxic and nutrient stress
in areas that would most require it. Consistent
with this, resistance toward anti-angiogenic
therapies in human cancer has been associated
with neutrophil infiltration (61), and neutrophil
depletion reduces tumor vascularization and
growth (51, 52, 62, 63). Studies facilitating further
understanding of T3-mediated vascular remodel-
ing would reveal new therapeutic targets against
pathological angiogenesis within the tumor.
A particularly interesting finding is the con-
servation of this differentiation program in
tumor-infiltrating neutrophils across tumor
type and species. When scRNAseq datasets
from other tumor models were examined, T3
annotation could identify protumoral clus-
ters found both in mouse [i.e., mN5 (11)] and in
human [i.e., TAN-1 (56)]. T3 neutrophils pro-
moted PDAC tumor growth in mice, whereas
ablation of T3 neutrophils or their pro-
angiogenic function removed this growth advan-
tage. In parallel, the T3 signature consistently
predicted poorer patient outcomes in two indepen-
dent human PDAC cohorts, as well as in a subset
of other solid tumors. Thus, different studies
converge upon our identification of a terminally
differentiated neutrophil state, the signature
of which might be used to better understand
neutrophil function in cancer and possibly to
predict tumor progression. We propose that the
reported heterogeneity of neutrophils across
tumors more likely reflects transitional states
derived from populations at different stages of
maturation and/or reprogramming.
Collectively, by ordering neutrophils through
the lens of their ontogeny, we have assessed
global neutrophil phenotypic heterogeneity
through maturation stages, extramedullary
sources, and the specialization of each neutro-
phil to each tissue (49). Although all of these
states coexist within the tumor-bearing mouse,
the neutrophil maturation trajectory remains
unchanged, consistent with findings in other
proinflammatory settings (19, 20, 64, 65). There-
fore, whereas our findings do not fully exclude
the possibility of neutrophil reprogramming
outside of the tumor, they do suggest that it
is unlikely that upstream changes in neutro-
phil progenitors specifically drive the forma-
tion of a protumoral neutrophil population.
Instead, it is the intrinsic capacity of recruited
neutrophils to adapt in response to the tumor
environment regardless of their initial pheno-
type that allows them to adopt convergent tra-
jectories to settle upon a final, protumoral state.
This feed-forward loop thus ensures the con-
tinued supply and differentiation of long-lived,
pro-angiogenic neutrophils that support tumor
growth. Expanding from the results of our study,
we propose that it is advantageous for tissues to
induce functional homogeneity in neutrophils
p
at the local scale to support tissue growth and
function. This process is then hijacked by the
tumor to favor a functional neutrophil state that
promotes aberrant tumor growth. Our findings
thus suggest a general mechanism by which
short-lived effector cells such as neutrophils
g
efficiently adjust their functions to meet the de-
mands of a tissue, and that local neutrophil re-
sponses can be therapeutically targeted.
y
Methods summary
scRNAseq
Sorted total neutrophils (Lin–CD45+CD115–
Ly6ClowSiglec-F–Gr1+CD11b+Ly6G+) from the
BM, spleen, blood, and tumor of mice bearing
orthotopic pancreatic tumors [generated as
described in (66)] were incubated with 0.5 mg
of Totalseq-A anti-mouse Hashtag antibodies
(BioLegend) per 100,000 cells for 30 min at 4°C.
Cells were washed with fluorescence-activated
y g
cell sorting (FACS) buffer, spun down, and re-
suspended in PBS with 1% bovine serum albumin.
Cells were pooled accordingly for 10X Genomics
3’ (v3) sequencing on NovaSeq (Illumina) follow-
ing the manufacturer’s protocol. Sequencing
reads were evaluated by FastQC and MultiQC.
y
High-quality reads were aligned to the GRCm38
,
mm10-2020-A genome assembly and quantified
using CellRanger (version 2.2.0, 10X Genomics).
The gene expression matrix was analyzed in R
using the Seurat package (4.0.5) (68). Hashtags
were demultiplexed using CITE-seq-Count. Dou-
blets and multiplets were filtered out, as well as
unique molecular identifiers (UMIs) with two or
more hashtags associated with them. UMIs
with excess mitochondrial reads (>5%), number
of features <200 (low read counts) and >4200
(outliers) were also removed. Normalization,
scaling, and clustering were performed with
the default Seurat pipeline. The destiny pack-
age (3.4.0.) was used for diffusion mapping (24).
Velocity analysis was performed with scVelo
(2.1.0) (26) using the default stochastic model
13 of 16
RES EARCH | R E S E A R C H A R T I C L E
and velocity vectors were projected onto the
diffusion map embedding. Regulatory net-
work analysis was carried out with PySCENIC
(0.10.4) (27) in Python and exported for visu-
alization in R.
High-parameter flow cytometry staining
and analysis
Pancreatic tumor single-cell suspensions pooled
from five tumor-bearing mice were stained with
a backbone panel of fluorophore-conjugated
antibodies. The stained cell suspension was
then distributed equally across all antibody wells
in the LEGENDScreenTM kit (BioLegend).
Staining and washing were performed accord-
ing to the manufacturer’s instructions. Cells were
finally stained with 1 mM DAPI, and 1 million
events per well were acquired. Flow cytometry
standard (.fcs) files corresponding to each unique
PE marker were compensated and analyzed
using FlowJo software (BD Biosciences).
Compensated live, CD45+ singlets were analyzed
using the InfinityFlow (1.4.0) in R (28, 29). The
analysis was transferred back to FlowJo, where
clusters were manually annotated and the im-
puted PE intensity values for each cluster were
exported. The geometric mean was calculated
for each cluster for heatmap plotting and
comparison.
Spatial transcriptomics for orthotopic
pancreatic tumors
Frozen tumor samples were quartered, em-
bedded in optimal cutting temperature filled
molds, and sectioned to a thickness of 10 mm at
−20 °C. To ensure capture of tumor-infiltrating
neutrophils, cryosections were screened for
Ly6G immunofluorescence before mounting
on Visium Spatial Gene Expression Slides (10X
Genomics). Tissue sections were fixed with
methanol; stained for cytokeratin, Ly6G, and
DAPI; and imaged (EVOS Auto FL2, Thermo
Scientific). After imaging, Visium Spatial Gene
Expression libraries were prepared according
to the manufacturer’s protocol. Libraries were
sequenced using HiSeq X or NovaSeq 6000 S4
(Illumina) at PE150 with 50,000 read pairs per
tissue-covered spot. Fastq data were processed
with SpaceRanger (version 1.2.2, 10X Genomics)
and mapped to the GRCm38 mm10-2020-A
genome assembly. Spots were filtered to re-
move DAPI-low/negative necrotic areas with
Loupe browser 6 (10X Genomics). Downstream
analysis was performed with Seurat (v3.2.3)
with default parameters, and joint clustering
was performed with BayesSpace (36). The
cell2location package (38) was used for de-
convolution. The UCell package (80) was used
to score spots independently for enrichment
in various GO processes.
MICS sample preparation and data analysis
Tumor cryosections were fixed in 4% para-
formaldehyde for 10 min, washed in PBS, and
Ng et al., Science 383, eadf6493 (2024) 12 January 2024
permeabilized for 20 min in blocking buffer
(5% donkey serum, 0.3% Triton X-100, and DAPI).
Sections were iteratively stained with fluorescein
isothiocyanate (FITC)-, PE- and allophycocyanin
(APC)-conjugated anti-mouse antibodies, after
which image acquisition and processing was
performed on the MACSima instrument as
described previously (39). Images were prepro-
cessed with MACS iQView software. Stitched
and registered TIFF files of individual chan-
nels were then imported into Imaris (Bitplane).
Staining artifacts were masked using the surface/
spot tool, and the masked channel was exported
back to MACS iQView. Image segmentation
based on DAPI-stained nuclei was performed
using the Advanced Tissue Morphology meth-
od with the donut algorithm and confirmed by
visual inspection. Identified cells and related
features containing marker intensities for all
channels were imported into FlowJo to gate
relevant populations. Seurat V4 was used for
further visualization and cell-type annotations,
and SPIAT (48) was used for distance-based
and colocalization metrics determining neutro-
phil subset localization with the tumor.
In vitro cell culture
TCM was collected from the culture of FC1242L
PDAC cell lines, spun down at 1000g for 10 min at
4°C to pellet down dead cells and debris, after
which the supernatant was aliquoted and stored
at –20°C. A total of 2.5 × 105 sorted neutrophils
per well were cultured in cDMEM (Gibco) or
TCM in normoxia or in a hypoxia incubator
at 5% O2. After 1 and 3 days, dcTRAIL-R1 up-
regulation was analyzed by flow cytometry
or wells were pooled, counted, and lysed for
analysis of RNA expression.
BrdU pulse-chase assay
Mice were injected intraperitoneally with 2 mg
of BrdU (Sigma-Aldrich) at the indicated time
points. To detect BrdU incorporation, cells from
the BM, spleen, blood, and tumor in pancreatic
tumor-bearing mice were stained with fixable
vitality dye (LIVE/DEAD Fixable Blue Dead
Cell Stain Kit, Invitrogen) and surface makers.
Cells were then fixed, permeabilized, and fi-
nally stained intracellularly with FITC-conjugated
anti-BrdU antibody according to the manu-
facturer’s protocol before analysis with flow
cytometry.
Subcutaneous pancreatic tumor model
A total of 1 × 105 tumor cells were resuspended
with or without with 1 × 105 neutrophils (spe-
cified in text) in 1× PBS, mixed in a 1:3 ratio of
Matrigel, and injected subcutaneously into the
right flanks of mice using a 30-gauge insulin
needle. At 3 and 7 days after the initial injec-
tion, 1 × 105 neutrophils were resuspended in
30 mL of PBS and injected directly into the ob-
servable Matrigel-tumor plug. If antibodies
were co-injected alongside neutrophils, 10 mg
of antibody was added to 1 × 105 neutrophils
and incubated for 10 to 20 min on ice before
the addition of 1 × 105 tumor cells. From day 14
after the injection of the tumor, tumors were
measured weekly with Vernier calipers and
tumor volumes were calculated using the fol-
lowing formula: 0.5 × length × width2. Mice
were euthanized at day 28 after injection and
tumor presence and weights were recorded.
Tumors were marked as rejected if there was
complete absence of tumor formation or only
a clear Matrigel plug left.
To evaluate vessel network formation in co-
injected tumors, tumors were excised, fixed with
4% paraformaldehyde overnight at 4°C, and
then washed with PBS. Samples were permea-
bilized with the SHANEL method, blocked,
and finally stained for CD31 and propidium
iodide (PI). Samples were then subjected to
solvent-based optical clearing and transferred
into ethyl cinnamate for imaging. All tumors
p
were fully imaged from one lateral edge to the
other at 2.0× magnification. Images were ac-
quired using a lightsheet LaVision Ultramicro-
scope II and captured using the PCO edge 4.2
sCMOS camera and then analyzed using Imaris
9.5.0 (Bitplane). Total tumor volume imaged
g
was first determined by surfacing positive PI
signal using the surface function. To standardize
the tumor volume evaluated, all tumor images
were processed to obtain the same percentage
y
volume (~35%) from the tumor midpoint, which
was then used for the downstream analysis. The
same image thresholds were applied consistent-
ly across tumors within the same experiment.
Surfaces based on CD31+ vessel staining were
then created with a standardized seed point val-
ue of 16.2, and their distribution from tumor
edge to the tumor midpoint was calculated
using the shortest distance function. Finally,
to quantify the vessels present, the total tumor
y g
length was then divided into percentile bins
(10%), after which CD31+ surface intensity was
quantified and normalized against the total
tumor volume intensity for each bin.
y
REFERENCES AND NOTES
1. L. G. Ng, R. Ostuni, A. Hidalgo, Heterogeneity of neutrophils.
,
Nat. Rev. Immunol. 19, 255–265 (2019). doi: 10.1038/
s41577-019-0141-8; pmid: 30816340
2. A. J. Gentles et al., The prognostic landscape of genes and
infiltrating immune cells across human cancers. Nat. Med. 21,
938–945 (2015). doi: 10.1038/nm.3909; pmid: 26193342
3. V. Bronte et al., Recommendations for myeloid-derived
suppressor cell nomenclature and characterization standards.
Nat. Commun. 7, 12150 (2016). doi: 10.1038/ncomms12150;
pmid: 27381735
4. J. Y. Sagiv et al., Phenotypic diversity and plasticity in
circulating neutrophil subpopulations in cancer. Cell Rep. 10,
562–573 (2015). doi: 10.1016/j.celrep.2014.12.039;
pmid: 25620698
5. B. E. Hsu et al., Immature low-density neutrophils exhibit
metabolic flexibility that facilitates breast cancer liver
metastasis. Cell Rep. 27, 3902–3915.e6 (2019). doi: 10.1016/
j.celrep.2019.05.091; pmid: 31242422
6. J.-I. Youn, S. Nagaraj, M. Collazo, D. I. Gabrilovich, Subsets of
myeloid-derived suppressor cells in tumor-bearing mice.
J. Immunol. 181, 5791–5802 (2008). doi: 10.4049/
jimmunol.181.8.5791; pmid: 18832739
14 of 16
RES EARCH | R E S E A R C H A R T I C L E
7. T. Condamine et al., Lectin-type oxidized LDL receptor-
1 distinguishes population of human polymorphonuclear
myeloid-derived suppressor cells in cancer patients.
Sci. Immunol. 1, aaf8943 (2016). doi: 10.1126/sciimmunol.
aaf8943; pmid: 28417112
8. C. Engblom et al., Osteoblasts remotely supply lung tumors
with cancer-promoting SiglecFhigh neutrophils. Science 358,
eaal5081 (2017). doi: 10.1126/science.aal5081; pmid: 29191879
9. F. Veglia et al., Fatty acid transport protein 2 reprograms
neutrophils in cancer. Nature 569, 73–78 (2019). doi: 10.1038/
s41586-019-1118-2; pmid: 30996346
10. C. Pfirschke et al., Tumor-promoting Ly-6G+ SiglecFhigh cells
are mature and long-lived neutrophils. Cell Rep. 32, 108164
(2020). doi: 10.1016/j.celrep.2020.108164; pmid: 32966785
11. R. Zilionis et al., Single-cell transcriptomics of human and mouse
lung cancers reveals conserved myeloid populations across
individuals and species. Immunity 50, 1317–1334.e10 (2019).
doi: 10.1016/j.immuni.2019.03.009; pmid: 30979687
12. H. Alshetaiwi et al., Defining the emergence of myeloid-
derived suppressor cells in breast cancer using single-cell
transcriptomics. Sci. Immunol. 5, eaay6017 (2020).
doi: 10.1126/sciimmunol.aay6017; pmid: 32086381
13. F. Veglia et al., Analysis of classical neutrophils and
polymorphonuclear myeloid-derived suppressor cells in cancer
patients and tumor-bearing mice. J. Exp. Med. 218, e20201803
(2021). doi: 10.1084/jem.20201803; pmid: 33566112
14. M. Evrard et al., Developmental analysis of bone marrow
neutrophils reveals populations specialized in expansion,
trafficking, and effector functions. Immunity 48, 364–379.e8
(2018). doi: 10.1016/j.immuni.2018.02.002; pmid: 29466759
15. Y. P. Zhu et al., Identification of an early unipotent neutrophil
progenitor with pro-tumoral activity in mouse and human bone
marrow. Cell Rep. 24, 2329–2341.e8 (2018). doi: 10.1016/
j.celrep.2018.07.097; pmid: 30157427
16. I. Kwok et al., Combinatorial single-cell analyses of granulocyte-
monocyte progenitor heterogeneity reveals an early uni-potent
neutrophil progenitor. Immunity 53, 303–318.e5 (2020).
doi: 10.1016/j.immuni.2020.06.005; pmid: 32579887
17. H. Q. Dinh et al., Coexpression of CD71 and CD117 identifies an
early unipotent neutrophil progenitor population in human
bone marrow. Immunity 53, 319–334.e6 (2020). doi: 10.1016/
j.immuni.2020.07.017; pmid: 32814027
18. D. E. Muench et al., Mouse models of neutropenia reveal
progenitor-stage-specific defects. Nature 582, 109–114 (2020).
doi: 10.1038/s41586-020-2227-7; pmid: 32494068
19. X. Xie et al., Single-cell transcriptome profiling reveals
neutrophil heterogeneity in homeostasis and infection.
Nat. Immunol. 21, 1119–1133 (2020). doi: 10.1038/
s41590-020-0736-z; pmid: 32719519
20. T. E. Khoyratty et al., Distinct transcription factor networks
control neutrophil-driven inflammation. Nat. Immunol. 22,
1093–1106 (2021). doi: 10.1038/s41590-021-00968-4;
pmid: 34282331
21. C. Wu et al., Spleen mediates a distinct hematopoietic
progenitor response supporting tumor-promoting
myelopoiesis. J. Clin. Invest. 128, 3425–3438 (2018).
doi: 10.1172/JCI97973; pmid: 29771686
22. M. Liu et al., PERK reprograms hematopoietic progenitor cells
to direct tumor-promoting myelopoiesis in the spleen. J. Exp.
Med. 219, e20211498 (2022). doi: 10.1084/jem.20211498;
pmid: 35266960
23. S. R. Hingorani et al., Trp53R172H and KrasG12D cooperate to
promote chromosomal instability and widely metastatic
pancreatic ductal adenocarcinoma in mice. Cancer Cell 7,
469–483 (2005). doi: 10.1016/j.ccr.2005.04.023;
pmid: 15894267
24. P. Angerer et al., destiny: Diffusion maps for large-scale
single-cell data in R. Bioinformatics 32, 1241–1243 (2016).
doi: 10.1093/bioinformatics/btv715; pmid: 26668002
25. L. Haghverdi, M. Büttner, F. A. Wolf, F. Buettner, F. J. Theis,
Diffusion pseudotime robustly reconstructs lineage branching.
Nat. Methods 13, 845–848 (2016). doi: 10.1038/nmeth.3971;
pmid: 27571553
26. G. La Manno et al., RNA velocity of single cells. Nature 560,
494–498 (2018). doi: 10.1038/s41586-018-0414-6;
pmid: 30089906
27. B. Van de Sande et al., A scalable SCENIC workflow for single-cell
gene regulatory network analysis. Nat. Protoc. 15, 2247–2276
(2020). doi: 10.1038/s41596-020-0336-2; pmid: 32561888
28. C.-A. Dutertre et al., Single-cell analysis of human mononuclear
phagocytes reveals subset-defining markers and identifies
circulating inflammatory dendritic cells. Immunity 51, 573–589.e8
(2019). doi: 10.1016/j.immuni.2019.08.008; pmid: 31474513
Ng et al., Science 383, eadf6493 (2024) 12 January 2024
29. E. Becht et al., High-throughput single-cell quantification of
hundreds of proteins using conventional flow cytometry and
machine learning. Sci. Adv. 7, eabg0505 (2021). doi: 10.1126/
sciadv.abg0505; pmid: 34550730
30. K. Neumann et al., Clec12a is an inhibitory receptor for uric
acid crystals that regulates inflammation in response
to cell death. Immunity 40, 389–399 (2014). doi: 10.1016/
j.immuni.2013.12.015; pmid: 24631154
31. W. Xu et al., Immune-checkpoint protein VISTA regulates
antitumor immunity by controlling myeloid cell–mediated
inflammation and immunosuppression. Cancer Immunol. Res.
7, 1497–1510 (2019). doi: 10.1158/2326-6066.CIR-18-0489;
pmid: 31340983
32. S. V. Ryzhov et al., Role of TGF-b signaling in generation of
CD39+CD73+ myeloid cells in tumors. J. Immunol. 193,
3155–3164 (2014). doi: 10.4049/jimmunol.1400578;
pmid: 25127858
33. M. Casanova-Acebes et al., Tissue-resident macrophages
provide a pro-tumorigenic niche to early NSCLC cells. Nature
595, 578–584 (2021). doi: 10.1038/s41586-021-03651-8;
pmid: 34135508
34. S. Kaur et al., Functions of thrombospondin-1 in the tumor
microenvironment. Int. J. Mol. Sci. 22, 4570 (2021).
doi: 10.3390/ijms22094570; pmid: 33925464
35. A. I. Markowska, F.-T. Liu, N. Panjwani, Galectin-3 is an
important mediator of VEGF- and bFGF-mediated angiogenic
response. J. Exp. Med. 207, 1981–1993 (2010). doi: 10.1084/
jem.20090121; pmid: 20713592
36. E. Zhao et al., Spatial transcriptomics at subspot resolution
with BayesSpace. Nat. Biotechnol. 39, 1375–1384 (2021).
doi: 10.1038/s41587-021-00935-2; pmid: 34083791
37. W. Lu, Y. Kang, Epithelial-mesenchymal plasticity in cancer
progression and metastasis. Dev. Cell 49, 361–374 (2019).
doi: 10.1016/j.devcel.2019.04.010; pmid: 31063755
38. V. Kleshchevnikov et al., Cell2location maps fine-grained cell
types in spatial transcriptomics. Nat. Biotechnol. 40, 661–671
(2022). doi: 10.1038/s41587-021-01139-4; pmid: 35027729
39. A. Kinkhabwala et al., MACSima imaging cyclic staining (MICS)
technology reveals combinatorial target pairs for CAR
T cell treatment of solid tumors. Sci. Rep. 12, 1911 (2022).
doi: 10.1038/s41598-022-05841-4; pmid: 35115587
40. Y. Kinugasa, T. Matsui, N. Takakura, CD44 expressed on
cancer-associated fibroblasts is a functional molecule
supporting the stemness and drug resistance of malignant
cancer cells in the tumor microenvironment. Stem Cells 32,
145–156 (2014). doi: 10.1002/stem.1556; pmid: 24395741
41. A. Costa et al., Fibroblast heterogeneity and
immunosuppressive environment in human breast cancer.
Cancer Cell 33, 463–479.e10 (2018). doi: 10.1016/j.
ccell.2018.01.011; pmid: 29455927
42. R. J. Slack, R. Mills, A. C. Mackinnon, The therapeutic potential
of galectin-3 inhibition in fibrotic disease. Int. J. Biochem.
Cell Biol. 130, 105881 (2021). doi: 10.1016/
j.biocel.2020.105881; pmid: 33181315
43. E. Fonsatti, M. Altomonte, M. R. Nicotra, P. G. Natali, M. Maio,
Endoglin (CD105): A powerful therapeutic target on
tumor-associated angiogenetic blood vessels. Oncogene 22,
6557–6563 (2003). doi: 10.1038/sj.onc.1206813;
pmid: 14528280
44. A. Giatromanolaki et al., Ectonucleotidase CD73 and CD39
expression in non-small cell lung cancer relates to hypoxia and
immunosuppressive pathways. Life Sci. 259, 118389 (2020).
doi: 10.1016/j.lfs.2020.118389; pmid: 32898522
45. K. Losenkova et al., Compartmentalization of adenosine
metabolism in cancer cells and its modulation during acute
hypoxia. J. Cell Sci. 133, jcs241463 (2020). doi: 10.1242/
jcs.241463; pmid: 32317394
46. N. Petruk et al., CD73 facilitates EMT progression and
promotes lung metastases in triple-negative breast cancer.
Sci. Rep. 11, 6035 (2021). doi: 10.1038/s41598-021-85379-z;
pmid: 33727591
47. Z. Fu et al., Proteolytic regulation of CD73 by TRIM21
orchestrates tumor immunogenicity. Sci. Adv. 9, eadd6626
(2023). doi: 10.1126/sciadv.add6626; pmid: 36608132
48. Y. Feng et al., Spatial analysis with SPIAT and spaSim to
characterize and simulate tissue microenvironments. Nat.
Commun. 14, 2697 (2023). doi: 10.1038/s41467-023-37822-0;
pmid: 37188662
49. I. Ballesteros et al., Cooption of neutrophil fates by tissue
environments. Cell 183, 1282–1297.e18 (2020). doi: 10.1016/
j.cell.2020.10.003; pmid: 33098771
50. L. Yang et al., Expansion of myeloid immune suppressor
Gr+CD11b+ cells in tumor-bearing host directly promotes tumor
angiogenesis. Cancer Cell 6, 13 (2004). pmid: 15488763
51. H. Nozawa, C. Chiu, D. Hanahan, Infiltrating neutrophils
mediate the initial angiogenic switch in a mouse model of
multistage carcinogenesis. Proc. Natl. Acad. Sci. U.S.A. 103,
12493–12498 (2006). doi: 10.1073/pnas.0601807103;
pmid: 16891410
52. F. Shojaei et al., Bv8 regulates myeloid-cell-dependent tumour
angiogenesis. Nature 450, 825–831 (2007). doi: 10.1038/
nature06348; pmid: 18064003
53. P. Kastana et al., “Matrigel plug assay for in vivo evaluation of
angiogenesis,” in The Extracellular Matrix, D. Vigetti,
A. D. Theocharis, Eds. (Springer, 2019), vol. 1952 of Methods in
Molecular Biology, pp. 219–232; http://link.springer.com/
10.1007/978-1-4939-9133-4_18.
54. V. Baeriswyl, G. Christofori, The angiogenic switch in
carcinogenesis. Semin. Cancer Biol. 19, 329–337 (2009).
doi: 10.1016/j.semcancer.2009.05.003; pmid: 19482086
55. N. G. Steele et al., Multimodal mapping of the tumor and
peripheral blood immune landscape in human pancreatic
cancer. Nat. Cancer 1, 1097–1112 (2020). doi: 10.1038/
s43018-020-00121-4; pmid: 34296197
56. L. Wang et al., Single-cell RNA-seq analysis reveals BHLHE40-
driven pro-tumour neutrophils with hyperactivated glycolysis in
pancreatic tumour microenvironment. Gut 72, 958–971 (2023).
doi: 10.1136/gutjnl-2021-326070; pmid: 35688610
57. J. Liu et al., An integrated TCGA pan-cancer clinical data
resource to drive high-quality survival outcome analytics. Cell
173, 400–416 (2018).
p
58. J. Pillay et al., In vivo labeling with 2H2O reveals a human
neutrophil lifespan of 5.4 days. Blood 116, 625–627 (2010).
doi: 10.1182/blood-2010-01-259028; pmid: 20410504
59. Y. Sawanobori et al., Chemokine-mediated rapid turnover of
myeloid-derived suppressor cells in tumor-bearing mice. Blood
111, 5457–5466 (2008). doi: 10.1182/blood-2008-01-136895;
pmid: 18375791
60. P.-B. Ancey et al., GLUT1 expression in tumor-associated
g
neutrophils promotes lung cancer growth and resistance to
radiotherapy. Cancer Res. 81, 2345–2357 (2021). doi: 10.1158/
0008-5472.CAN-20-2870; pmid: 33753374
61. L. M. Schiffmann et al., Tumour-infiltrating neutrophils
counteract anti-VEGF therapy in metastatic colorectal
y
cancer. Br. J. Cancer 120, 69–78 (2019). doi: 10.1038/
s41416-018-0198-3; pmid: 30377339
62. J. Jablonska, S. Leschner, K. Westphal, S. Lienenklaus,
S. Weiss, Neutrophils responsive to endogenous IFN-b regulate
tumor angiogenesis and growth in a mouse tumor model.
J. Clin. Invest. 120, 1151–1164 (2010). doi: 10.1172/JCI37223;
pmid: 20237412
63. Y. Itatani et al., Suppressing neutrophil-dependent
angiogenesis abrogates resistance to anti-VEGF antibody
in a genetic model of colorectal cancer. Proc. Natl. Acad.
Sci. U.S.A. 117, 21598–21608 (2020). doi: 10.1073/
pnas.2008112117; pmid: 32817421
64. R. Grieshaber-Bouyer et al., The neutrotime transcriptional
y g
signature defines a single continuum of neutrophils across
biological compartments. Nat. Commun. 12, 2856 (2021).
doi: 10.1038/s41467-021-22973-9; pmid: 34001893
65. E. Montaldo et al., Cellular and transcriptional dynamics of
human neutrophils at steady state and upon stress.
Nat. Immunol. 23, 1470–1483 (2022). doi: 10.1038/
s41590-022-01311-1; pmid: 36138183
y
AC KNOWLED GME NTS
,
We acknowledge and thank all members of the L.G.N. laboratory
for helpful discussion and feedback on the manuscript, the SIgN
Flow Cytometry team for sorting and flow cytometry assistance,
the SIgN Immunogenomics team for assistance with generating
and running the ATACseq and scRNAseq libraries, and the SIgN
mouse core facility for technical help and support. Funding: This
research was funded by Singapore Immunology Network (SIgN)
core funding and A*STAR, Singapore. L.G.N. is supported by
National Natural Science Foundation of China (grant 32270956)
and Shanghai Science and Technology Commission (grant
20JC1410100). M.S.F. is supported by the A*STAR Career
Development Fund (grant 202D8150). I.W.K. is supported by the
A*STAR Career Development Fund (grant 202D8197). S.Z.C. is
supported by the A*STAR Career Development Award (192D8043)
and core funding from SIgN. Y.T. and H.L.T. are supported by
Clinician Scientist Awards (NMRC/CSA-INV/0023/2017 and
CSAINV20nov-0003) from the National Medical Research Council
of Singapore. The SIgN Flow Cytometry facility is supported by
National Research Foundation (NRF) Singapore under Shared
Infrastructure Support (SIS) (NRF2017_SISFP09). For the use of
the Ultramicroscope II, we would like to thank A*STAR core funds,
15 of 16
RES EARCH | R E S E A R C H A R T I C L E

the HMBS IAF-PP grant (H1701a0004), and the National
Research Foundation’s Shared Infrastructure Support grant for
SingaScope, a Singapore-wide microscopy infrastructure network
(NRF2017_SISFP10), for continued support of the A*STAR
Microscopy Platform. D.C.-W. is supported by a CRI Irvington
postdoctoral fellowship (CRI3511). G.F.C. is supported by grant
PID2022-142341OB-I00 funded by the Spanish MCIN/AEI/
10.13039/501100011033. I.B. is supported by a Ramon y Cajal
fellowship (RYC2021-033511-I) from MICINN and by a Leonardo
fellowship (LEO22-2-2596) from the BBVA Foundation. A.V.
and Y.Z. are supported by NMRC OF-IRG grant to VA (OFIRG19nov-0112).
M.I. is supported by the European Research Council (ERC)
Consolidator Grant 725038, ERC Proof of Concept Grants 957502
and 101138728, Italian Association for Cancer Research (AIRC)
Grants 19891 and 22737, Italian Ministry for University and
Research Grants PE00000007 (INF-ACT) and PRIN (2022FMESXL),
and by a Funded Research Agreement from Asher Biotherapeutics
and VIR Biotechnology. A.H. is supported by grant R01AI165661 from
NIH/NIAD, RTI2018-095497- B-I00 from MCIN, HR17_00527 from
Fundación La Caixa, the Transatlantic Network of Excellence
(TNE-18CVD04) from the Leducq Foundation, and FET-OPEN
(grant 861878) from the European Commission. The CNIC is supported
by the MCIN and the Pro CNIC Foundation and is a Severo Ochoa
Center of Excellence (CEX2020-001041-S). F.G. is supported by
Singapore Immunology UIBR award from the Agency for Science,
Technology and Research (A*STAR), a Singapore NRF Senior
Investigatorship (NRFI2017-02), the Fondation Gustave Roussy, and
the ARC Foundation. Author contributions: Conceptualization: L.G.N.,
M.S.F., L.T., I.W.K., A.H., F.G.; Formal analysis: K.B.D., V.N., J.M.C.,
M.W.L., L.X.J., B.Y.S., G.F.C.; Funding acquisition: L.G.N., M.S.F., I.W.K.,
Y.H.L.; Investigation: M.S.F., L.T., I.W.K., Y.R., C.M.S., K.L., Y.N.Z., J.J.,
K.H.L., D.H.L., Y.C.T., D.C.-W., K.Y., C.B., C.N.; Methodology: M.S.F.,
L.T., I.W.K., Y.R., C.M.S., D.C.-W., K.Y.; Project administration: L.G.N.;
Supervision: L.G.N.; Visualization: M.S.F., L.T., I.W.K., Y.T., C.M.S., K.L.,
D.C.-W., K.Y.; Writing – original draft: L.G.N., M.S.F., L.T., I.W.K., A.H.,
F.G., D.C.-W.; Writing – review & editing: All authors Competing
interests: M.S.F., L.T., F.G., and L.G.N. are inventors on patent
WO2022132041 submitted by A*STAR that covers a method of
characterizing an immunosuppressive neutrophils in cancer. M.I.
participates in advisory boards and consults for Gilead Sciences,
Third Rock Ventures, Antios Therapeutics, Asher Biotherapeutics,
GentiBio, Clexio Biosciences, Sibylla Biotech, BlueJay Therapeutics,
and Aligos Therapeutics. H.L.T. is a cofounder and medical
adviser for RNAscence Pte. Ltd. F.G. participates in advisory
boards for iTheos and Genoskin. The remaining authors declare no
competing interests. Data and materials availability: All data
needed to evaluate the conclusions in the paper are present in
the paper and/or in the supplementary materials. scRNAseq
(GSE243466), ATACseq (GSE244531), and spatial transcriptomics
data (GSE244534) have been deposited in the NCBI Gene
Expression Omnibus. All three datasets have been linked under
into one SuperSeries GSE244536 for ease of access. License
information: Copyright © 2024 the authors, some rights reserved;
exclusive licensee American Association for the Advancement of
Science. No claim to original US government works. https://www.
science.org/about/science-licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.adf6493
Materials and Methods
Figs. S1 to S16
Tables S1 to S7
References (66–81)
MDAR Reproducibility Checklist
Movies S1 and S2
Submitted 9 November 2022; resubmitted 23 August 2023
Accepted 27 November 2023
10.1126/science.adf6493

p
g
y
y g
y
,

Ng et al., Science 383, eadf6493 (2024) 12 January 2024 16 of 16

RES EARCH

◥ environmental factors such as individual and

RESEARCH ARTICLE family-related stress, drug abuse, homeless-
ness, and social isolation. Depending on the
NEUROSCIENCE clinical outcome definition, up to 20 to 30% of
first-episode individuals (10) and more than
Illusory generalizability of clinical prediction models 50% with a relapse do not respond sufficiently
to antipsychotic medications (11).
Adam M. Chekroud1,2*, Matt Hawrilenko1, Hieronimus Loho2, Julia Bondar1, Ralitza Gueorguieva3, We examined the generalizability of clinical
Alkomiet Hasan4, Joseph Kambeitz5, Philip R. Corlett2, Nikolaos Koutsouleris6, Harlan M. Krumholz7, prediction models across multiple clinical trials
John H. Krystal2, Martin Paulus8 using the case study of antipsychotic treat-
ments for schizophrenia. Critically, this study
It is widely hoped that statistical models can improve decision-making related to medical treatments. directly evaluated the performance of a model
Because of the cost and scarcity of medical outcomes data, this hope is typically based on investigators on its initial training sample as well as how
observing a model’s success in one or two datasets or clinical contexts. We scrutinized this optimism the same model performed on truly inde-
by examining how well a machine learning model performed across several independent clinical trials of pendent clinical trial samples. This allowed us
antipsychotic medication for schizophrenia. Models predicted patient outcomes with high accuracy to assess two key risks: First, models may “overfit”
within the trial in which the model was developed but performed no better than chance when applied the data by fitting the random noise of one
out-of-sample. Pooling data across trials to predict outcomes in the trial left out did not improve particular dataset rather than a true signal
predictions. These results suggest that models predicting treatment outcomes in schizophrenia are likely to generalize across samples, leading to
highly context-dependent and may have limited generalizability. good predictions in the training data that do
not generalize to the testing data. The second

O
ne fundamental problem in medicine is
that despite similar treatments some pa-
tients get better whereas others show
no improvement. One goal of precision
medicine is to use machine learning to
p
key risk is poor model transportability. Models
the potential for statistical models to improve may lack external validity due to patients,
clinical practice (7–9). providers, or implementation characteristics
varying across trials (12).
Open data opens possibilities
As efforts toward mandatory randomized con- Data sources

find models that will help predict who will
respond to what type of treatment (1). For
precision medicine to affect clinical practice
and improve outcomes, the models that we
g
trolled trial (RCT) data deposition, archival We used treatment data from five international,
data sharing, and open science continue to multisite RCTs (NCT00518323, NCT00334126,
advance, opportunities arise to more rigor- NCT00085748, NCT00078039, and NCT00083668)
ously examine how well treatment predic- obtained through the YODA Project (https://

develop must robustly predict outcomes for
unseen, future patients (2–5).
However, models are not usually tested on
new patients in a different context because
data—especially data from controlled designs—
are scarce and expensive (6). Instead, research-
ers typically split a study’s participants into
two or more random groups, build a model
using the data from one of the groups, and test
its predictions on the other group (e.g., k-fold
y
tion models will fare in different contexts. The yoda.yale.edu/). These trials were selected be-
Yale Open Data Access (YODA) Project is one cause of their comparability and consistency.
such effort, which now includes a data archive All patients had a current DSM-IV diagnosis of
of over 246 clinical trials from all medical schizophrenia at the start of the trial; all trials
fields. randomized patients to an antipsychotic med-
The YODA project included several RCTs ication or placebo; all trials used the same
evaluating the comparative efficacy of anti- scale to measure treatment outcomes (the Posi-
psychotic medications for treating schizophrenia. tive and Negative Syndrome Scale, PANSS); all
Predicting treatment outcomes in schizophrenia trials included a 4-week timepoint to measure
could be especially advantageous because the outcomes; and all trials collected similar data

y g
cross-validation) (3, 4). When we use this kind clinical response to pharmacological interven- about the patients at baseline. Combined, the
of approximation based on one data set or tions is heterogeneous and depends on many trials also provide a heterogeneous patient
clinical sample, we have a fundamentally lim-
ited insight into the true potential for a model
to improve outcomes in the future. Validating
clinical prediction models in different clinical Table 1. Treatment outcomes across trials.

y
samples is an essential step in the model de-

,
velopment process. It generally results in pre- Adults first Adults - Adults - Older
dictive performance measures that are lower episode Chronic #1 Chronic #2 adults Teens Total
but allows for a more realistic assessment of Outcome definition (n = 321) (n = 430) (n = 481) (n = 99) (n = 182) (n = 1513)
264 208 266 32 47 816
25% Reduction PANSS
1 2 (82.2%) (48.4%) (55.3%) (32.3%) (25.8%) (54.0%)
Spring Health, New York City, NY 10010, USA. Department of .....................................................................................................................................................................................................................
Psychiatry, Yale University School of Medicine, New Haven, 50% Reduction 119 85 82 7 12 306
CT 06520, USA. 3Department of Biostatistics, Yale University, PANSS (37.1%) (19.8%) (17.0%) (7.1%) (6.6%) (20.3%)
.....................................................................................................................................................................................................................
New Haven, CT 06520, USA. 4Department of Psychiatry, RSWG remission 152 129 153 24 58
Psychotherapy and Psychosomatics, University Augsburg, 517 (34.2%)
86159 Augsburg, Germany. 5Department of Psychiatry and criteria (47.4%) (30.0%) (31.8%) (24.2%) (31.9%)
.....................................................................................................................................................................................................................
Psychotherapy, University of Cologne, Faculty of Medicine Percentage change
and University Hospital of Cologne, Cologne, Germany. 6Department -44.1 -26.9 -28.4 -18.0 -13.7 -28.8
in PANSS total
of Psychiatry and Psychotherapy, Ludwig-Maximilians- (23.1) (28.2) (25.3) (21.8) (21.5) (26.7)
University, Munich, Germany. 7Center for Outcomes Research score (SD)
.....................................................................................................................................................................................................................
and Evaluation, Yale New Haven Hospital, New Haven, CT Baseline total 103.0 92.4 92.9 91.1 90.0 94.4
06520, USA. 8Laureate Institute for Brain Research, Tulsa, PANSS (SD) (14.3) (13.0) (10.9) (8.8) (13.1) (13.2)
OK 74136, USA. .....................................................................................................................................................................................................................
*Corresponding author. Email: adam.chekroud@yale.edu

Chekroud et al., Science 383, 164–167 (2024) 12 January 2024 1 of 4

RES EARCH | R E S E A R C H A R T I C L E

population, with patients recruited from 194
sites across 4 continents, a pediatric trial, an
older adult trial, and a trial of individuals with
a first episode (see SM for more details). The
study design, outcome measure, and cross-
validation approach were preregistered on
2 August 2016 (YODA 2016-1005). Minor up-
dates to the preregistration were submitted on
2 May 2023 (included in the SM).
Patients and outcomes
From 29 March 2004 to 30 March 2009, 1962
total patients aged 12 to 81 years were enrolled
across five randomized controlled trials at
194 sites in North America, Asia, Europe, and
Africa. We assessed symptomatic outcomes
based on the PANSS (13) at week 4 for the
1513 participants with baseline and 4-week
follow-up data. Different definitions of re-
sponse, remission, and recovery are used in
schizophrenia research, which makes com-
duction. Table 1 reports treatment outcomes
for all definitions across the five trials.
We extracted all information available at
baseline across all trials and retained it as a
predictor variable if it was available for more
than 80% of patients. We also computed con-
dition (control versus treatment) X predictor
interaction terms. Drug dose was standar-
dized to paliperidone dose equivalents using
the defined daily dose method (18). Together,
this yielded 217 predictor variables that in-
cluded basic demographic features, psychiat-
ric history (DSM-IV diagnosis category, age of
diagnosis, psychiatric hospitalizations), clin-
ical data (PANSS, Clinical Global Impression)
(17), extrapyramidal symptom scales (Abnor-
mal Involuntary Movement Scale) (19) and
Simpson Angus Scale (20), biometric data
(blood chemistry panel, hematology, urinaly-
sis), and treatment randomization. The de-
tailed list of predictors, selection criteria, and
correlated with one another and predictors
may only be sparsely endorsed. It has been
successful in research predicting psychiatric
treatment outcomes (5, 23–25).
The elastic net model uses two penalty pa-
rameters, lambda and alpha, which balance
stability with parsimony. We examined 400
combinations of alpha and lambda penalties
(see supplement) and selected the optimal pe-
nalties using repeated 10-fold cross-validation.
The cross-validation part of this procedure se-
parates the data set into 10 random folds and
uses 9 of the subsets for training, repeating the
process such that each subset is left out once
for testing. The repeated part of this procedure
re-splits the data ten times to reduce the im-
pact of the random data split; in aggregate,
100 total models were fit to the 10 folds by 10
repeats. Model performance was calculated
by averaging the performance metric across
all 100 models. This entire procedure was run

paring and applying results in clinical practice
difficult (14–16). The primary outcome re-
ported here is the Remission in Schizophrenia
Working Group criteria (RSWG) (17). To ensure
that our findings were not driven by idiosyn-
crasies in how we defined treatment response,
missing data approach is provided in the SM.
Machine learning approach
We applied machine learning methods using
baseline data to predict whether a patient
would achieve clinically significant improve-
p
for each of the 400 combinations of alpha and
lambda values, and the final values were
chosen as the combination of alpha and
lambda values that optimized the model
performance metric. We used the metrics
of area under the receiver operating curve

we included three other definitions commonly
used in the field, including percentage change
with baseline correction (15, 16), and two bi-
nary definitions of 25 and 50% symptom re-
ments in symptoms over four weeks of anti-
psychotic treatment. We used the elastic net
algorithm (21, 22), a penalized regression meth-
od that is appropriate when covariates are
g
for binary outcomes and root mean square
error for continuous outcomes. The final
alpha and lambda values were applied to the
aggregate sample to estimate the prediction

y
Within-trial Within-trial Leave-one-
no validation cross-validation Paired-trial trial-out
1.0

0.9
Balanced accuracy

0.8

y g
0.7

0.6

y
,
0.5
s

E
E

ts

E
#1

s
s

#2

FE

E
#1

#2

E
#1

#2

lts

ts

s
FE

#1

#2

1
#2

lts

s
en

en
ult

ult

ult
en

en

en

en

en
ic #

ic #

ic #

ic #
sF

sF
sF

sF

sF
l

l
du

du

du

du
2 → ronic

ic

s → ronic

ic
ic

E → ronic

ic

nic

ic
ic
d

lts

Ad
lts
Te

Te

Te

Te

Te

Te

Te
ult

ult
ult

ult

ult
ron

ron
ron

rA

rA

ron

rA

2 → hron

rA

ron

ron
E → hron

rA

ron

rA
ron
du

du
hro
s→

s→

2→

er
E→

→
Ad

Ad
Ad

Ad

Ad
de

de
de

de

de

de
Ch

Ch

Ch

Ch

Ch

Ch

Ch

Ch

Ch
Ch

Ch
A

Old
C
C

C
#1

lts

ers
Ol

Ol
Ol

Ol

Ol

Ol
1→

2→

s→

→
E→

→
en

en

ic #
sF
→

1→

2→

s→

→
E→

→
→
du

oth
ic
1→

s→
→

→
E

lts

ers
Te

Te

ult
#1

#1

lts

lts

en

ers

ers
ic #

ron

ic #

ron

rA
sF

sF

du

en

en
ic #

ic #

ic #

c#
lts

lts

sF

sF

ers
oth
du

du

All
Ad

Te

en
ic #

ic #
sF

de
nic

ic

oth

oth
Ch

Ch
ult

ult

ron

ron

rA

Te

Te
du

du

i
ult

ult

oth
ron

ron

ron

ron

ron

rA

rA

Te
Ol

All
Ad

Ad
ro

ult
rA

rA

ron

ron

de
Ch

Ch

All

All
Ad

Ad

de

de
Ch

Ch

Ch

Ch

Ch

Ch

All
Ad

Ol
de

de

Ch

Ch

Ol

Ol
Ol

Ol

Study pair
Balanced accuracy is higher than chance no yes
FE = first episode

Fig. 1. Balanced accuracy for models predicting treatment outcome (Remission in Schizophrenia Working Group criteria) across all modeling scenarios.
Gray intervals represent 95% confidence intervals, not adjusted for multiple comparisons. Red markers denote statistical significance after applying the Benjamini-
Hochberg adjustment with the false discovery rate set to 5%. Repeated 10-fold cross-validation; FE, first episode.

Chekroud et al., Science 383, 164–167 (2024) 12 January 2024 2 of 4

RES EARCH | R E S E A R C H A R T I C L E
model coefficients. To interpret differential
performance across samples, we report a
metric known as balanced accuracy [(sensi-
tivity + specificity) / 2] whose null distribution
is centered on 50% (26, 27). To determine
whether balanced accuracy was statistically
significantly above chance, we bootstrapped
confidence intervals and adjusted for multiple
comparisons across all 35 comparisons using
the Benjamini-Hochberg adjustment with the
false discovery rate set to 5% (28). All analysis
was conducted using R version 4.1 (29), with
machine learning models fit using the caret
package (30).
Exploring the generalizability of machine
learning models
We evaluated the applicability of machine
learning models across four distinct scenarios
to gain insights into their generalizability:
First, we assessed the predictive accuracy of
the model within the trial, without any ex-
ternal validation beyond the training data.
Second, we also focused on within-trial pre-
diction accuracy but this time estimated using
the data excluded from the training set in a
repeated tenfold cross-validation process. Third,
we conducted a paired-across-trial prediction
accuracy assessment. In this case, models
trained on one trial were applied to all other
trials to evaluate their performance. Finally, in
the fourth scenario, we implemented a leave-
one-trial-out prediction accuracy assessment.
Models were trained using data aggregated
from four trials and their predictive accuracy
was tested on the fifth trial (Fig. 1). Balanced
accuracy for the RSWG criteria are shown in
Fig. 1, and data for alternative outcome de-
finitions and additional outcome metrics are
shown in the supplement.
No validation
In the scenario where we assessed within-trial
performance without any external validation,
the final prediction model created for a spe-
cific trial was applied to the entire sample
from that same trial. The balanced accuracy
was high and significantly above random
chance for all models, with an average of 0.72
(range: 0.66 to 0.77) across all five prediction
models. However, because the model was eval-
uated on the same sample used to develop it,
there is a risk of overfitting, making these re-
sults less likely to generalize.
Cross-validation
To estimate more generalizable prediction ac-
curacy, we employed within-trial cross-validation.
Performance characteristics of the optimal
alpha and lambda values were averaged across
the 100 left out folds (10 folds * 10 repeats)
from the repeated cross-validation procedure.
Each trial’s data were divided into 10 subsets,
with coefficients developed on 9 subsets and
Chekroud et al., Science 383, 164–167 (2024) 12 January 2024
then tested on the remaining subset. In this
scenario, balanced accuracy was lower in each
dataset, averaging 0.60 (range: 0.56 to 0.67)
across all five prediction models. Only three
out of five models performed above chance.
Paired-trial validation
Next, we directly assessed out-of-sample per-
formance in the paired-trial validation (16).
We applied the prediction models developed
using within-trial models across each of the
other trials, for a total of 20 trial pairs. Model
performance was low (mean across all trial
pairs was 0.54, range 0.48 to 0.61) with only
three trial pairs performing above chance.
Leave-one-trial-out validation
Given the availability of multiple archival trials
for developing a prediction model, a natural
extension of the paired-trial validation would
be a leave-one-trial-out approach. This approach
might enhance generalizability by allowing the
algorithm to be exposed to more information
through between-trial variability in baseline
phenotypes. We aggregated data across four
trials, leaving the fifth out for testing, and
repeated the process 5 times so that each trial
was left out once. Performance was once again
poor with low balanced accuracy in all con-
ditions (mean across all left out trials was 0.54
with range 0.50 to 0.58) and performance was
significantly above chance in only two of the
five testing sets.
Sensitivity analyses
The pattern of results observed was not due
to idiosyncrasies of how we measured treat-
ment response. We found the same pattern of
results when we reproduced all four modeling
scenarios using other binary and continuous
definitions of treatment response (see SM).
This lack of model generalizability to un-
seen patients was also observed for another
machine learning algorithm. When we applied
random forest models, which can detect com-
plex patterns of interactions amongst predic-
tor variables, we observed the same pattern of
results except that excessive overfitting occur-
red for no-validation conditions (see SM).
Discussion
Machine learning prediction of treatment out-
comes in medicine is exciting but challenging.
Our modeling scenarios using antipsychotic
treatment outcome prediction in schizophrenia
suggest that predictive models are fragile and
that excellent performance in one clinical con-
text is not a strong indicator of performance
on future patients. This is highly concerning as
most predictive studies today rely on internal
samples for testing and validation. When
models were tested on the same sample on
which they were developed, models routinely
produced strong predictions. Cross-validation
tempered these performance estimates but
even the models that performed well in cross-
validation were little better than chance when
predicting outside of the sample in which they
were developed—even when the unseen sam-
ples were well-phenotyped. In a world where
we hope that predictive models might eventu-
ally improve clinical practice, the ability to
generalize to other carefully controlled clin-
ical contexts is only the first step to generalize
to settings with more heterogeneity in patient
presentations and methods of care delivery.
Why model generalizability is challenging
There are three key reasons why predictive
models might not generalize across trials.
First, patient groups may be too different
across trials. The umbrella category of schiz-
ophrenia is useful for clinical practice but also
means that patients with different disease
stages are coerced into the same diagnostic
p
category in clinical trials. If key information
that differentiates patients is not captured in
the data or if the range of that information is
more restricted in the dataset used to develop
the model compared with the target trial, pre-
dictions may be inaccurate. Thus, patient pop-
g
ulations may differ considerably between trials
within the same diagnostic category. However,
the current study found little evidence that
results would generalize across even the most
y
similar trials. The three cross-trial pairs with
predictions slightly greater than chance were
amongst the three studies of adults aged 18
and over but this pattern of results did not
consistently replicate across other outcome
definitions.
Second, these trials may not have collected
the type or volume of data needed to make
good predictions. This study used clinical,
sociodemographic, and simple biomarker data-
y g
based on almost 2000 patients. However,
additional data types may have been more
relevant to treatment outcomes. Psychosocial
information and social determinants of health
were not included in this study but have pre-
viously been found to predict treatment out-
y
comes in first episode psychosis (27). Preliminary
,
research suggests that longitudinal patterns
of symptom co-occurrence—either before or
during treatment—can be specifically relevant
to how a patient will respond to treatment
although it may delay care to collect this data
(31–34). Some have suggested the use of
neuroimaging and genetic data but there
is currently little evidence to suggest that
such data would improve predictions; further,
collecting these data would pose additional
barriers for routine implementation (35–37).
Finally, having data from more participants
may allow for more nuanced modeling of in-
dividual differences.
A third reason why predictive models may
not generalize is that patient outcomes may be
3 of 4
RES EARCH | R E S E A R C H A R T I C L E
too context-dependent. Trials may have subtly
important differences in recruiting procedures,
inclusion criteria, or treatment protocols. Be-
cause these characteristics do not vary across
patients within a trial, they cannot be modeled
as predictors within a single trial. However,
this study used multinational RCTs conducted
by large pharmaceutical companies and con-
tract research organizations, minimizing non-
specific concerns especially in comparison to
the variability we would expect in real clinical
practice going from one site or provider to the
next. Of course, different antipsychotic drugs
may differ from one another in ways that af-
fect outcome prediction, and the D2 dopamine
receptor blockade intended to correct overstim-
ulation of D2 receptors by endogenous dopa-
mine may be too far downstream from the
primary pathology of schizophrenia or the symp-
tom severity criteria used to measure it (38).
Improving model generalizability
It is worth considering how we might improve
the situation in the future. From a statistical
modeling perspective, capturing important
heterogeneity through phenotyping or strat-
ification procedures might help improve the
generalizability of models. Identifying trial-
level characteristics that relate to patient out-
comes may provide information to better
equip prediction models to generalize across
settings. Such trial-level variation can be studied
using Bayesian approaches or recent techni-
ques that incorporate replicability across con-
texts or populations into the algorithm training
process (39). From a population perspective,
there may be some patients for whom the
choice of treatment has no impact on their
clinical course, which represents an inherent
limitation of predicting treatment outcomes.
However, this could also be an opportunity for
further improvement in identifying which
patients have a wider range of potential out-
comes and for whom selecting the optimal
treatment would provide clinical benefit (40).
Longitudinal validation methods, in which
a validation sample is drawn from the same
population at a later point in time, may pro-
vide a limited but pragmatic path to avoid
generalizing from one clinical setting to an-
other. The growth of large mental health care
delivery systems provides the opportunity to
collect large amounts of data and deploy pre-
diction models in the same setting in which
they were developed (41). This strategy can
reduce challenges associated with patient het-
erogeneity and context-dependence, and also
help identify temporal or geographic trends
that affect a model’s predictions. However,
when a model is trained and validated on
samples from the same population, it may
perform well in that specific context but fail
when applied to a different population with
different characteristics.
Chekroud et al., Science 383, 164–167 (2024) 12 January 2024
Conclusions
The present study offers an underwhelming
but realistic picture of our current ability to
develop truly useful predictive models for schiz-
ophrenia treatment outcomes. Models that
performed with excellent accuracy in one sam-
ple routinely failed to generalize to unseen
patients. These findings suggest that approx-
imations based on a single data set are a
fundamentally limited insight into future per-
formance and represent a potential concern
for prediction models throughout medicine.
The field as a whole—present authors included—
hope that machine learning approaches can
eventually improve the allocation of treatments
in medicine; however, we should a priori re-
main skeptical of any predictive model find-
ings that lack an independent sample for
validation.
RE FERENCES AND NOTES
1. F. S. Collins, H. Varmus, N. Engl. J. Med. 372, 793–795 (2015).
2. D. G. Altman, P. Royston, Stat. Med. 19, 453–473 (2000).
3. D. G. Altman, Y. Vergouwe, P. Royston, K. G. M. Moons,
BMJ 338, 1432 (2009).
4. E. W. Steyerberg, in Clinical Prediction Models (Springer,
2009), pp. 11–31.
5. A. M. Chekroud et al., Lancet Psychiatry 3, 243–250 (2016).
6. G. C. M. Siontis, I. Tzoulaki, P. J. Castaldi, J. P. A. Ioannidis,
J. Clin. Epidemiol. 68, 25–34 (2015).
7. E. W. Steyerberg, Y. Vergouwe, Eur. Heart J. 35, 1925–1931
(2014).
8. R. D. Riley et al., Stat. Med. 40, 4230–4251 (2021).
9. M. Pavlou et al., Stat. Methods Med. Res. 30, 2187–2206
(2021).
10. Y. Zhu et al., Eur. Neuropsychopharmacol. 27, 835–844 (2017).
11. S. Leucht et al., Am. J. Psychiatry 174, 927–942 (2017).
12. B. Li, C. Gatsonis, I. J. Dahabreh, J. A. Steingrimsson,
Biometrics 79, 2382–2393 (2023).
13. S. R. Kay, A. Fiszbein, L. A. Opler, Schizophr. Bull. 13, 261–276
(1987).
14. M. Lambert, A. Karow, S. Leucht, B. G. Schimmelmann,
D. Naber, Dialogues Clin. Neurosci. 12, 393–407 (2010).
15. S. Leucht, J. Clin. Psychiatry 75, 8–14 (2014).
16. M. Obermeier et al., BMC Psychiatry 11, 113 (2011).
17. J. Busner, S. D. Targum, Psychiatry 4, 28–37 (2007).
18. S. Leucht, M. Samara, S. Heres, J. M. Davis, Schizophr. Bull. 42,
S90–S94 (2016).
19. W. Guy, Ecdeu Assessment Manual for Psychopharmacology
(The George Washington University, 1976).
20. G. M. Simpson, J. W. Angus, Acta Psychiatr Scand Suppl. 45,
11–19 (1970).
21. H. Zou, T. Hastie, J. R. Stat. Soc. Series B Stat. Methodol. 67,
301–320 (2005).
22. J. Friedman, T. Hastie, R. Tibshirani, J. Stat. Softw. 33, 1–22 (2010).
23. A. M. Chekroud et al., Psychiatr. Serv. 69, 927–934 (2018).
24. A. M. Chekroud et al., JAMA Psychiatry 74, 370–378 (2017).
25. Z. D. Cohen, R. J. DeRubeis, Annu Rev Clin Psychol. 14,
209–236 (2018).
26. K. H. Brodersen, C. S. Ong, K. E. Stephan, J. M. Buhmann,
in 2010 20th International Conference on Pattern Recognition
(2010), pp. 3121–3124.
27. N. Koutsouleris et al., Lancet Psychiatry 3, 935–946 (2016).
28. Y. Benjamini, Y. Hochberg, J. R. Stat. Soc. B 57, 289–300
(1995).
29. R Core Team, R: A language and environment for statistical
computing (R Foundation for Statistical Computing, 2021);
www.R-project.org/.
30. M. Kuhn, J. Stat. Softw. 28, 1 (2008).
31. A. J. Fisher, J. D. Medaglia, B. F. Jeronimus, Proc. Natl. Acad.
Sci. U.S.A. 115, E6106–E6115 (2018).
32. A. J. Fisher, P. Soyster, Generating Accurate Personalized
Predictions of Future Behavior: A Smoking Exemplar. PsyArXiv
e24v6 [Preprint] (2019); doi:10.31234/osf.io/e24v6
33. A. J. Fisher, J. F. Boswell, Assessment 23, 496–506
(2016).
34. A. J. Fisher et al., Behav. Res. Ther. 116, 69–79 (2019).
35. A. M. Chekroud, N. Koutsouleris, Mol. Psychiatry 23, 24–25
(2018).
36. A. M. Chekroud, JAMA Psychiatry 74, 1183–1184
(2017).
37. M. P. Paulus, JAMA Psychiatry 74, 1185–1186 (2017).
38. R. A. McCutcheon et al., Biol. Psychiatry 94, 561–568 (2023).
39. P. Patil, G. Parmigiani, Proc. Natl. Acad. Sci. U.S.A. 115,
2578–2583 (2018).
40. R. J. DeRubeis et al., PLOS ONE 9, e83875 (2014).
41. J. Bondar et al., JAMA Netw. Open 5, e2216349 (2022).
AC KNOWLED GME NTS
Funding: No funding source had any role in the study design, data
collection, data analysis, data interpretation, writing, or submission
of this report. All trials were originally funded by Janssen Research and
Development. The study design, outcome measure, and cross-
validation approach were preregistered on 1 Aug 2016 (YODA #2016-
1005). The study was approved on 15 December 2016 (IRB/HSC#
1610018521) by the Yale University Institutional Review Board. Access
was granted on 5 January 2017, and data were analyzed until October,
2023. Author contributions: Conceptualization: A.C. Methodology:
A.C., M.H., R.G., H.L., J.B., A.H., N.K., J.H.K., H.K. Data Acquisition: H.K.,
on behalf of the Yale Open Data Access Initiative. Data Analysis: A.C.,
M.H., H.L., J.B. Visualization: M.H., H.L., R.G. Supervision: A.C., M.P.,
P.C., and J.H.K. Writing – original draft: A.C. Writing – substantial
review and editing: A.C., M.H., A.H., N.K., P.C., H.K., J.H.K., and M.P.
Competing interests: A.C. holds equity in Spring Care and is the lead
p
inventor on 3 patent submissions relating to treatment for major
depressive disorder (US Patent and Trademark Office number
Y0087.70116US00 and provisional application numbers 62/491 660
and 62/629 041). M.H. and J.B. are employed by and hold equity
in Spring Care. RG received royalties from the book Statistical
Methods in Psychiatry and Related Fields published by CRC Press and
is an inventor on US patent application 20200143922. A.H. was a
member of advisory boards and received paid speakership by
g
Boehringer-Ingelheim, Lundbeck, Otsuka, Rovi, and Recordati. He
received paid speakership by AbbVie and Advanz. He is editor
of the AWMF German guidelines for schizophrenia. J.H.K. has been a
consultant and/or advisor to or has received honoraria from Janssen/
J&J, Lundbeck and Boehringer Ingelheim. P.C. is co-founder and
y
Board Member of Tetricus Labs and reported holding stock and stock
options in Tetricus Labs Inc. H.K. received funding from Johnson
and Johnson through Yale University. J.Kr. reported holding patents
licensed to Johnson and Johnson and Freedom Biosciences; co-
founder of Freedom Biosciences, stock in Spring Health, Biohaven
Pharmaceuticals, Neumora Pharmaceuticals; Consultant to Biogen,
Bionomics Limited, Boehringer Ingelheim International, Cerevel
Therapeutics, Jazz Pharmaceuticals, Otsuka American Pharmaceutical
Inc., Perception Neuroscience, Sumitomo America, Taishisho, Takeda,
BioXcel, Psychogenics Inc. Data and materials availability: The
study design, outcome measure, and cross-validation approach was
preregistered on 1 August 2016 (YODA #2016-1005). All data are
available via the Yale University Open Data Access (YODA) platform
(https://yoda.yale.edu/request). Data accession numbers for the
y g
five trials analyzed here are: R076477-PSZ-3001 (Teens trial),
R076477-SCH-3015 (Adults First Episode), R076477-SCH-302
(Older Adults), R076477-SCH-305 (Adults Chronic #1), R076477-
SCH-303 (Adults Chronic #2). The code to reproduce all results in
this manuscript and the supplement is available at Zenodo (42). This
study, carried out under YODA Project #2016-1005, used data
y
obtained from the Yale University Open Data Access Project, which
has an agreement with JANSSEN RESEARCH & DEVELOPMENT,
,
L.L.C. The interpretation and reporting of research using this data are
solely the responsibility of the authors and does not necessarily
represent the official views of the Yale University Open Data Access
Project or JANSSEN RESEARCH & DEVELOPMENT, L.L.C. License
information: Copyright © 2024 the authors, some rights reserved;
exclusive licensee American Association for the Advancement of
Science. No claim to original US government works. https://www.
sciencemag.org/about/science-licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.adg8538
Methods and Results
Figs. S1 to S15
Tables S1 to S20
References (42)
MDAR Reproducibility Checklist
Data S1
Submitted 27 January 2023; resubmitted 20 July 2023
Accepted 10 November 2023
10.1126/science.adg8538
4 of 4
RES EARCH

OPTICS correspond to the spontaneous formation of

diverse spatiotemporal patterns (Fig. 1D).
Free-electron interaction with nonlinear optical Electrons traversing the optical evanescent
near field of the air-cladded microresonator
states in microresonators undergo inelastic electron-light scattering (IELS)
by absorbing and emitting photons, leading to
Yujia Yang1,2*†, Jan-Wilke Henke3,4†, Arslan S. Raja1,2†, F. Jasmin Kappert3,4†, Guanhao Huang1,2, “photon sidebands” in the electron spectrum
Germaine Arend3,4, Zheru Qiu1,2, Armin Feist3,4, Rui Ning Wang1,2, Aleksandr Tusnin1,2, spaced by the photon energy (10). The side-
Alexey Tikan1,2, Claus Ropers3,4*, Tobias J. Kippenberg1,2* band amplitude and the overall spectral width
are determined by a coupling parameter g (33),
The short de Broglie wavelength and strong interaction empower free electrons to probe structures which is sensitive to the electric field along the
and excitations in materials and biomolecules. Recently, electron-photon interactions have e-beam direction
enabled new optical manipulation schemes for electron beams. In this work, we demonstrate the e þ∞ w0

interaction of electrons with nonlinear optical states inside a photonic chip–based microresonator. g ðx; yÞ ¼  ∫ Ez ðx; y; z′Þei v z′ dz′ ð1Þ
2ℏw0 ∞
Optical parametric processes give rise to spatiotemporal pattern formation corresponding to
coherent or incoherent optical frequency combs. We couple such “microcombs” to electron where z is the coordinate in the e-beam direc-
beams, demonstrate their fingerprints in the electron spectra, and achieve ultrafast temporal gating tion, (x, y) is the transverse coordinate, e is the
of the electron beam. Our work demonstrates the ability to access solitons inside an electron electron charge, ℏ is Planck’s constant h di-
microscope and extends the use of microcombs to spatiotemporal control of electrons for imaging vided by 2p, w0 is the angular frequency of
and spectroscopy. the optical carrier wave, and v is the electron
velocity. In a simplified picture, a continuous

N
onlinear optical phenomena are preva-
lent in science and technology, allowing
broadband supercontinuum for optical
frequency metrology (1), squeezed light
for gravitational-wave astronomy (2),
taneous parametric generation of electron-
photon pairs (25). However, only the linear
cavity response was exploited to enhance the
intracavity field and the electron-light coupling,
whereas nonlinear optical responses have been
p
e-beam samples the electric field of the in-
tracavity waveforms at random arrival times
of the electrons within the beam. This results
in a characteristic electron spectral shape
for each nonlinear optical intracavity state
(Fig. 1D).

and entangled photons for quantum informa-
tion science (3). Over the past decade, triggered
by advances in ultralow-loss (ultrahigh quality
factor Q) photonic microresonators, continuous-
predicted to endow electron-light interaction
with new capabilities (26, 27). Kerr microreso-
nators support diverse intracavity nonlinear
dissipative structures, including DKSs (5), Turing
g
The experimental setup is illustrated in Fig. 2A.
We transiently generated nonlinear optical
states in the transverse-electric mode family
by continuously scanning the pump laser fre-

wave driven microresonators with Kerr non-
linearity (c(3)) have given rise to a host of
“dissipative structures”—namely, self-organizing
spatiotemporal light patterns in nonlinear op-
tical systems (4). In particular, dissipative Kerr
solitons (DKSs) (5) lead to coherent optical fre-
quency combs with widespread applications,
from atomic clocks (6) and telecommunications
(7) to photonic computing (8) and astrophysical
spectroscopy (9).
patterns (28), chaotic modulation instability
(MI) (29), breathing solitons (30), and soliton
crystals (31). We studied the in situ coupling
of electron beams with such spatiotemporal
patterns.
Transient observation of electron spectral
modulation by nonlinear optical states
We studied and harnessed nonlinear optical
effects in a transmission electron microscope
y
quency across a cavity resonance near 1551.5 nm
and recorded electron energy spectra (200 keV
center energy) in parallel. Several oscilloscope
traces of the nonlinearly generated light and
the detuning-dependent electron spectra are
shown in Fig. 2B. The oscilloscope traces bear
the characteristics of a detuning scan for a
Kerr nonlinear cavity with anomalous disper-
sion, exhibiting a typical “soliton step” (5). We
identified five regions corresponding to dis-

Photonic structures mediating electron-photon
interactions have brought new optical manip-
ulations to electron beams, enabling high space/
time/energy–resolution electron microscopy
(10–12), quantum-coherent optical modulation
and probing (13–17), longitudinal and trans-
(TEM) using ultralow-loss Si3N4 photonic chip–
based microresonators (24, 25). The fiber-
packaged photonic chip was placed in the TEM
with the quasi-monochromatic electron beam
(e-beam) passing over the microresonator sur-
face in an aloof geometry (Fig. 1A). The micro-
y g
tinct intracavity states, agreeing with the sim-
ulated stability chart (Fig. 2C) in terms of the
pump power and detuning: (i) At low power,
nonlinear optical responses are negligible, lead-
ing to a monochromatic CW intracavity field;
(ii) with an increased detuning, and thus a higher

y
verse electron beam shaping (18–20), dielectric resonator (Fig. 1B) features a high Q of ~3.0 × intracavity power, cascaded FWM generates

laser acceleration (21), and electron-driven light
sources (22, 23). Recently, high-Q silicon nitride
(Si3N4) photonic chip–based microresonators
have been used to demonstrate continuous-
beam electron phase modulation (24) and spon-
1
Institute of Physics, Swiss Federal Institute of Technology
Lausanne (EPFL), CH-1015 Lausanne, Switzerland. 2Center
for Quantum Science and Engineering, EPFL, CH-1015
Lausanne, Switzerland. 3Department of Ultrafast Dynamics,
Max Planck Institute for Multidisciplinary Sciences, D-37077
Göttingen, Germany. 4Georg-August-Universität Göttingen,
D-37077 Göttingen, Germany.
*Corresponding author. Email: yujia.yang@epfl.ch (Y.Y.);
claus.ropers@mpinat.mpg.de (C.R.); tobias.kippenberg@epfl.ch (T.J.K.)
†These authors contributed equally to this work.
106 and anomalous group velocity dispersion
to resonantly enhance the intracavity field and
facilitate nonlinear frequency mixing. When
pumped by a continuous-wave (CW) laser above
the parametric oscillation threshold (32), cas-
caded four-wave mixing (FWM) leads to the
generation of incoherent or coherent optical
frequency combs corresponding to various dis-
sipative structures (4) attainable by scanning the
pump laser frequency from the blue to the red
side of the resonance—changing the frequency
difference or detuning between the pump laser
and the cavity resonance (Fig. 1C). In particular,
DKSs exist in the bistability regime of the reso-
nance tilt caused by Kerr nonlinearity. In the
time domain, nonlinear dissipative structures
,
new frequency components and an amplitude
modulation of the intracavity field, known as a
Turing pattern; (iii) when further red-detuning
the pump, the intracavity state enters the
chaotic MI regime, corresponding to a disor-
dered, rapidly varying waveform and an inco-
herent Kerr frequency comb; (iv) when tuning
the pump to the soliton step, localized struc-
tures are spontaneously formed, which at small
detunings consist of breathing solitons with
periodically oscillating temporal and spec-
tral shapes; and (v) with an increased detun-
ing on the soliton step, shape-invariant stable
DKSs are generated, featuring one or multiple
femtosecond temporal pulses and a coherent
frequency comb.

Yang et al., Science 383, 168–173 (2024) 12 January 2024 1 of 6

RES EARCH | R E S E A R C H A R T I C L E

A B C

Normalized intracavity power (arb. u.)

Pump laser tuning direction

g
1

n in
Electron gun

tu
de
ue
-
e 0.8

bl
e
Fiber

tiv
c
e- energy

fe
Ef

ble
0.6

sta
Power

Un
Microresonator 0.4
Time

0.2
Photonic Light output
chip-based 0.5 mm Effective red
detuning
micro- 0
-5 0 5 10
resonator Normalized detuning (arb. u.)
Power

D Continuous Chaotic modulation Dissipative

Time wave instability Kerr soliton

e- energy Light input

p
Electron
Spectrum (log.)

Spectrum (log.)

Spectrum (log.)
detector

Magnetic
prism

g
Electron energy Electron energy Electron energy

Fig. 1. Free-electron interaction with nonlinear optical states. (A) detuning showing a resonance tilt and bistability from the Kerr effect.

Schematic of the experiment. Electrons in a TEM pass by a photonic chip–based
nonlinear microresonator. Stimulated inelastic scattering between electrons
and nonlinear optical states induces electron spectral broadening. (B) Photograph of
the fiber-packaged Si3N4 photonic chip. (C) Intracavity power versus laser
y
(D) Illustration of intracavity waveforms and post-interaction electron spectra
for CW, chaotic MI, and DKS states. The electron spectral broadening originates
from incoherent summation of electrons that interact with the intracavity field at
different times.

For electrons interacting with these states, The simulated intracavity waveforms and ing in constructive or destructive interfer-
the coupling parameter (Eq. 1) becomes optical spectra are shown in Fig. 2D, as well ence owing to a phase delay between the two

 as the measured optical and electron spectra interactions (24, 35). We changed this phase
e þ∞ z′ for four intracavity states. Electron spectra in delay by scanning the e-beam position along
2ℏw0 ∫∞
g ðx; yÞ ¼  A f þ f0 þ D 1 ; t

y g
v Fig. 2, B and D, demonstrate disparate charac- the chip surface and acquired electron spec-
  teristic shapes for different states. For the CW tra while maintaining a given intracavity state
 ef ðx; yÞcos f  er ðx; yÞsin f
w0 w0
state, the electron spectrum has the typical (Fig. 3). For dissipative structures, the evolu-
 ein c Rf ei v z′ dz′ ð2Þ double-peak shape from coherent phase mod- tion of both the optical carrier and envelope
ulation. For the Turing pattern and chaotic between the two interaction regions provides
where A(f, t) is the intracavity waveform in MI, the electron spectrum broadens because more complex free-electron modulations than

y
the frame rotating at the optical group velocity of increasing intracavity power. A Gaussian- the coherent phase modulation at a uniform

with the microresonator angular coordinate f
and the slow time t (4); D1/2p is the microreso-
nator free spectral range (FSR); ef and er are
the tangential and radial optical modal fields,
respectively; n is the linear refractive index; c
is the speed of light; R is the microresonator
ring radius; and f0 is the angular offset of the
rotating frame from the laboratory frame. The
coupling parameter further determines the elec-
tron spectrum through the Nth photon side-
h iN
band amplitude JN ½2jg ðx; yÞj jgg ððx;y Þ
x;yÞj , where
JN is the Bessel function of the first kind. We
obtained continuous e-beam spectra by means
of an incoherent averaging over the angular
offset f0 and the slow time t (34).
,
shaped background appears, and the double field strength.
peaks reduce in intensity. We attribute these For the monochromatic CW field (Fig. 3B),
observations to the intracavity field strength the electron spectral width oscillation along
no longer being uniform. In the stable DKS the radial position has the typical pattern of
state stochastically generated by the detuning previously reported Ramsey-type interferences
scan, the electron spectrum features a strong, (Fig. 3D) (24). The oscillation nodes or destruc-
narrow central peak on a weak, broad plateau. tive interference arises from a p-phase shift of
the optical field at the two interactions.
Fingerprints of dissipative structures For a Turing pattern, the generated frequency
in Ramsey-type interference components and the accompanying amplitude
We gained further insights into the interaction modulation (Fig. 3C) alter the electron spec-
of electrons with nonlinear dissipative struc- trum. The electron spectrum (Fig. 3E) has a cen-
tures by investigating the Ramsey-type inter- tral, low-energy-change part of the spectrum
ferences from two sequential interactions with that resembles the double-peak shape from a
the microresonator. The e-beam intersects the CW interaction but with a narrower spectral
waveguide near field twice (Fig. 3A), result- width caused by a reduced intracavity pump

Yang et al., Science 383, 168–173 (2024) 12 January 2024 2 of 6

RES EARCH | R E S E A R C H A R T I C L E

A B C 0.3

Electron energy shift (eV) Generated light (arb. u.)

2 3
4

OSA TEM 1.5 5

e-
soliton
2
Light 1 step
ESA output

Pump power (W)

0.2
0.5
BS
PD 1
OSC FBG Si3N4 0
1 2 3 4 5
-30 105
PD CIRC -20

e- counts (arb. u.)

VNA 104
-10 0.1
Light
input 103 CW
FPC 0
Turing pattern
EDFA 10 102 Chaotic MI
CW pump
EOM
20 Breather
101
Stable DKS
Trigger 30
AFG 2 2.5 3 3.5 4 0 0.4 0.8 1.2
Time (µs) Detuning (GHz)
D Simulation Simulation Experiment Experiment
Simulated intracavity waveform (arb. units)

Measured electron spectrum (50 dB/div)

Simulated optical spectrum (80 dB/div)

Measured optical spectrum (80 dB/div)

1

p
3

g
- -100 100 1490 1620 -32 32
Resonator angular coordinate Mode number Wavelength (nm) Electron energy change (eV)

Fig. 2. Electron spectra imprinted by nonlinear dissipative structures. indicate regions corresponding to different intracavity states. (C) Simulated
(A) Experimental setup. OSA, optical spectrum analyzer; ESA, electronic stability chart of the microresonator showing the two-dimensional parameter

y
spectrum analyzer; OSC, oscilloscope; VNA, vector network analyzer; space of pump power and detuning. Colored regions indicate the existence
AFG, arbitrary waveform generator; PD, photodetector; EOM, electro-optic ranges for the labeled states. The dashed arrow depicts the detuning scan in (B).
modulator; EDFA, erbium-doped fiber amplifier; FBG, fiber Bragg grating; (D) Simulated optical waveforms and spectra, as well as measured optical
CIRC, optical circulator; FPC, fiber polarization controller; and BS, beam splitter. and electron spectra for (top to bottom, respectively) CW, Turing pattern, chaotic
(B) (Top) Oscilloscope traces of the generated light and (bottom) the electron MI, and DKS states. The numbering and color-coding are equivalent to that
spectrum when scanning the pump laser frequency across a resonance. in (B) and (C). Measured electron spectra are vertical slices of the electron
Electron spectra are summed over multiple scans. Dashed lines and numbers spectrum in (B), indicated with the colored arrows.

power due to FWM. Additionally, the electron and absence of nodes in the position-dependent D2 is the second-order dispersion (5). The ob-

y g
spectrum has a broad shoulder that is absent scan (Fig. 3H), resulting from random intracav- served corresponding electron spectra have a
in the CW case. This shoulder arises from the ity field strengths that prohibit well-defined distinct shape featuring a strong, narrow, low-
interaction of electrons with the intensity peaks interference conditions. The e-beam modulated energy-change region and a weak, broad pla-
of the intracavity waveform (Fig. 3E, inset). For by the chaotic MI state possesses a Gaussian teau (Fig. 3I). The former exhibits the hallmark
the chosen e-beam position, the temporal pe- spectrum similar to that previously reported of a coherent phase modulation produced by
riodicity of the Turing pattern amplitude mod- for thermal states (17). Rather than invoking the CW background. The latter has a low total

y
ulation is approximately commensurate with photon statistics, our analysis attributes the spectral weight and a broad width originating

,
the travel time difference of the electron and spectral shape to the statistical fluctuations of from the interaction between electrons and the
the optical envelope between the two interac- the intracavity optical intensity, stochastically high-peak-power femtosecond soliton pulse.
tion regions (DTtravel ≈ NTTuring, N ∈ ℤ) (34). sampled by the electrons and averaged on the Because the DKS pulse duration (~100 fs) is
Hence, the electron experiences nearly equal detector (34). much shorter than the round-trip time (~10 ps),
optical intensities in the two interactions be- In the stable DKS state, a single temporal only a small fraction of the electrons in the con-
cause of the time-translation invariance of the soliton pulse circulates in the cavity on a weak tinuous e-beam interact with the pulse. Hence,
periodic Turing pattern; this leads to similar CW background (Figs. 2D and 3A), forming a the plateau is much lower than the central region.
modulation strengths just like the CW case. coherent, low-noise optical frequency comb In addition, the plateau has a moderate, position-
Therefore, the spectral widths of the low-energy- with a sech2 spectral shape (Fig. 3G). The an- dependent spectral width oscillation, arising from
change part and the broad shoulder jointly in- alytical waveform is the Ramsey interference of electrons interacting
crease and decrease. rffiffiffiffiffiffi
rffiffiffiffiffiffi  with the soliton pulse and the CW background
For the chaotic MI (incoherent Kerr comb) 4D iq0 2D one time each. This oscillation has the same pe-
Y ≃ Y0 þ e sech f ð3Þ
(Fig. 3F), the characteristic double peaks from k D2 riod as that of the CW Ramsey pattern but with
uniform phase modulation disappear, and the an offset that encodes the soliton phase (34).
electron spectrum has a smooth Gaussian shape. where D denotes the pump detuning, k is the For all four intracavity states, experimental
We further found a reduced spatial dependence cavity decay rate, q0 is the soliton phase, and results were well reproduced with simulations

Yang et al., Science 383, 168–173 (2024) 12 January 2024 3 of 6

RES EARCH | R E S E A R C H A R T I C L E

A e- B 0
C 0
CW Turing pattern

Optical spectrum (dBm)

Optical spectrum (dBm)

CW -40 -40
pump

e -e
ne -80 -80
rgy 1500 1550 1600 1450 1500 1550 1600 1650
sh sition Wavelength (nm) Wavelength (nm)
ift Radial po
D 1.38 Exp. Intensity (dB) Sim.
E 1.38 Exp. Sim.
-50 0 interaction
e-
Radial position (µm)

Radial position (µm)

second
interaction

0 0 e-
e- spectrum (dB)

e- spectrum (dB)
0 0

p
-50 -50
-30 -15 0 15 30 -30 -15 0 15 30 -30 -15 0 15 30 -30 -15 0 15 30
Electron energy change (eV) Electron energy change (eV) Electron energy change (eV) Electron energy change (eV)
F G
0 Chaotic MI 0 Single DKS
Optical spectrum (dBm)

Optical spectrum (dBm)

g
-40 -40
sech2

y
-80 -80
1450 1500 1550 1600 1650 1450 1500 1550 1600 1650
Wavelength (nm) Wavelength (nm)
H 1.38 I 1.38
Exp. Sim. Exp. Sim.
Radial position (µm)

Radial position (µm)

0 0

y g
e- spectrum (dB)

e- spectrum (dB)

0 0

-50 -50
-30 -15 0 15 30 -30 -15 0 15 30 -30 -15 0 15 30 -30 -15 0 15 30
Electron energy change (eV) Electron energy change (eV) Electron energy change (eV) Electron energy change (eV)

y
Fig. 3. Ramsey-type interference of nonlinear intracavity states. (A) Illustration (Sim.) Ramsey interference patterns for these states. For each interference pattern,

,
of the double interactions between the e-beam and the microresonator supporting two line cuts (green and blue dashed lines) are plotted at bottom. The schematic
an exemplary DKS state. Position-dependent electron spectra are shown at bottom. in (E) depicts snapshots of the Turing pattern at the two interactions for one
(B, C, F, and G) Measured optical spectra for (B) CW, (C) Turing pattern, (F) chaotic exemplary electron arrival time. The sech2 fitting of the optical spectrum in
MI, (G) and single-DKS states. (D, E, H, and I) Measured (Exp.) and simulated (G) corresponds to a DKS pulse duration of ~98.5 fs.

by using the reduced time-dependent Schrödinger of the Kerr microresonator (Fig. 4A). Adjusting the detuning-dependent DKS frequency comb
equation for electrons and the Lugiato-Lefever the detuning D alters the DKS pulse duration, width, with increasing detuning leading to a
equation for nonlinear optics (Fig. 3). We also peak field, and background CW amplitude. broader spectrum—that is, a shorter DKS pulse
pffiffi qffiffiffiffi
observed additional electron spectral features Shown in Fig. 4B is a DKS state characterized
duration tFWHM ≃ 2 arcosh
D1
2
 D2D2 .
caused by multisoliton states, incommensurate by scanning the electro-optic-modulation side-
Turing patterns, soliton pulse edges, and breath- bands of the pump by means of a vector net- When measuring the electron spectrum, an
ing solitons (supplementary materials) (34). work analyzer (VNA) (Fig. 2A) for three different increasing detuning D leads to a wider pla-
detunings (D1, D2, and D3). The VNA traces il- teau, induced
pffiffiffiffiffiffiffiffiffiffi
ffi by an increasing DKS peak field
Probing soliton dynamics with free electrons lustrate the soliton (S) and cavity (C) resonances, (∼ 4D=k ) (Fig. 4D). Meanwhile, the plateau
We used free electrons to probe basic proper- with the latter revealing the detuning (36). The height slightly decreases because the DKS pulse
ties of DKSs existing in the bistability regime measured optical spectra (Fig. 4C) demonstrate duration tFWHM decreases, and fewer electrons

Yang et al., Science 383, 168–173 (2024) 12 January 2024 4 of 6

RES EARCH | R E S E A R C H A R T I C L E

A B C -30

ch
Laser
Δ1

ran
fit

pe W
-20
up C
(pump)
Δ2
rb
-40

Power (dBm)
power (a. u.)

Δ3

Power (dB)
Intracavity

-50
Δ Soliton
-40
branch
-60
CW-lower -60
branch Δ1 Δ2 Δ3 -70

λ 0 400 800 1200 1500 1550 1600 1650

Detuning Frequency (MHz) Wavelength (nm)
D 0 E
Δ1 400

e- counts
Δ2
e- spectrum (dB)

Pulse width (fs)

200 300 110
Δ3

2
|g max |
0
-2 0 2
Electron energy change (eV) 100
-25 200
90
100
80
-50
-30 -20 -10 0 10 20 30 400 500 600 700 800
Electron energy change (eV) Detuning (MHz)

p
Fig. 4. Electron probing of DKSs. (A) Resonance bistability caused by the Kerr effect. DKS is formed when the high-power CW branch (black) decays to the soliton
branch (green). In the soliton state, the pump is effectively red-detuned, and both soliton branch and low-power CW branch (blue) exist. (B) Double-resonance
response of soliton (S) and cavity (C) observed with a vector network analyzer. (C and D) The (C) optical and (D) electron spectra acquired at three different detunings
(D1, D2, and D3) indicated in (B). (E) Detuning dependence of the dimensionless coupling parameter jgmax j2 and soliton pulse duration. The error bars indicate maximal
and minimal values obtained at 100 e-beam positions.

g
interact with the pulse. A larger detuning more- by use of DKSs instead of mode-locked lasers. 21. R. J. England et al., Rev. Mod. Phys. 86, 1337–1389 (2014).
over leads to a weaker CW background Y0, Proper spectral filtering will lead to a train of 22. G. Adamo et al., Phys. Rev. Lett. 103, 113901 (2009).
23. M. Taleb, M. Hentschel, K. Rossnagel, H. Giessen, N. Talebi,
which is reflected by the weaker modulation electron pulses with <100-fs duration and a Nat. Phys. 19, 869–876 (2023).

y
of the central region of the electron spectra gigahertz-to-terahertz repetition rate, which is 24. J.-W. Henke et al., Nature 600, 653–658 (2021).
(Fig. 4D, inset). We illustrate in Fig. 4E the de- orders of magnitude greater than that of state-of- 25. A. Feist et al., Science 377, 777–780 (2022).
26. A. Konečná, V. Di Giulio, V. Mkhitaryan, C. Ropers,
tuning dependence of the plateau width [ap- the-art ultrafast TEMs, allowing a much greater F. J. García de Abajo, ACS Photonics 7, 1290–1296 (2020).
proximated by 4jgmax jℏw0 , where gmax is the beam current. Moreover, by accurately phase- 27. F. J. García de Abajo, E. J. C. Dias, V. Di Giulio, Phys. Rev. Lett.
maximum coupling parameter (13)] extracted matching the electrons with the DKS pulses, a 129, 093401 (2022).
28. S.-W. Huang et al., Phys. Rev. X 7, 041002 (2017).
from electron spectra, and the DKS pulse du- scheme of small-footprint and high repetition- 29. T. Hansson, D. Modotto, S. Wabnitz, Phys. Rev. A 88, 023819
ration extracted from optical spectra. Because rate dielectric laser particle accelerators could (2013).
jgp j isffi proportional to the DKS peak field
ffiffiffiffiffiffiffiffiffiffi
max be foreseen. Our work thus presents avenues 30. M. Yu et al., Nat. Commun. 8, 14569 (2017).
∼ 4D=k, its square scales linearly with the de- for ultrafast electron-light-matter interactions 31. D. C. Cole, E. S. Lamb, P. Del’Haye, S. A. Diddams, S. B. Papp,
Nat. Photonics 11, 671–676 (2017).
tuning D. A continuous detuning scan shows a driven by chip-based temporal solitons. 32. T. J. Kippenberg, S. M. Spillane, K. J. Vahala, Phys. Rev. Lett.

y g
similar scaling (34). 93, 083904 (2004).
33. S. T. Park, M. Lin, A. H. Zewail, New J. Phys. 12, 123028 (2010).
RE FERENCES AND NOTES 34. Materials and methods are available as supplementary
Discussion and outlook
1. J. M. Dudley, G. Genty, S. Coen, Rev. Mod. Phys. 78, 1135–1184 materials.
Optical manipulation of free electrons is extended (2006). 35. K. E. Echternkamp, A. Feist, S. Schäfer, C. Ropers, Nat. Phys.
2. H. Grote et al., Phys. Rev. Lett. 110, 181101 (2013). 12, 1000–1004 (2016).
beyond the regimes of pulsed or continuous-
3. P. G. Kwiat et al., Phys. Rev. Lett. 75, 4337–4341 (1995). 36. H. Guo et al., Nat. Phys. 13, 94–102 (2017).
wave lasers. Bringing free electron–light inter-

y
4. T. J. Kippenberg, A. L. Gaeta, M. Lipson, M. L. Gorodetsky, 37. M. T. Hassan, H. Liu, J. S. Baskin, A. H. Zewail, Proc. Natl. Acad.
action to the nonlinear optics regime opens up Science 361, eaan8083 (2018). Sci. U.S.A. 112, 12944–12949 (2015).

5. T. Herr et al., Nat. Photonics 8, 145–152 (2014).
the potential to noninvasively probe nonlinear
6. S. B. Papp et al., Optica 1, 10 (2014).
optical dynamics and devices with nanometer- 7. P. Marin-Palomo et al., Nature 546, 274–279 (2017).
femtosecond spatiotemporal resolution and 8. J. Feldmann et al., Nature 589, 52–58 (2021).
direct access to the intracavity field. The in- 9. E. Obrzud et al., Nat. Photonics 13, 31–35 (2019).
10. B. Barwick, D. J. Flannigan, A. H. Zewail, Nature 462, 902–906
tegrated photonics toolbox provides diversity (2009).
and flexibility for on-chip arbitrary optical wave- 11. D. Nabben, J. Kuttruff, L. Stolz, A. Ryabov, P. Baum, Nature
form generation and frequency conversion, prom- 619, 63–67 (2023).
12. A. Polman, M. Kociak, F. J. García de Abajo, Nat. Mater. 18,
ising advanced electron control schemes with 1158–1171 (2019).
flexible and programmable sideband amplitudes. 13. A. Feist et al., Nature 521, 200–203 (2015).
Furthermore, we achieved ultrafast electron- 14. K. Wang et al., Nature 582, 50–54 (2020).
15. O. Kfir et al., Nature 582, 46–49 (2020).
light interaction with chip-based femtosecond
16. V. Di Giulio, M. Kociak, F. J. G. de Abajo, Optica 6, 1524
temporal solitons in the absence of pulsed lasers (2019).
or electron sources. This approach facilitates 17. R. Dahan et al., Science 373, eabj7128 (2021).
18. K. E. Priebe et al., Nat. Photonics 11, 793–797 (2017).
ultrafast electron microscopy in a conventional
19. G. M. Vanacore et al., Nat. Mater. 18, 573–579 (2019).
TEM equipped with a photonic chip and a CW 20. A. Konečná, F. J. G. de Abajo, Phys. Rev. Lett. 125, 030801
laser, using temporal photon-gating (37, 38) (2020).
,
38. X. Fu et al., Nat. Commun. 11, 5770 (2020).
39. Y. Yang et al., Figure data for: Free-electron interaction with
nonlinear optical states in microresonators. Zenodo (2023);
https://doi.org/10.5281/zenodo.10104801.
AC KNOWLED GME NTS
Photonic chips were fabricated in the Center of
MicroNanoTechnology (CMi) and the Institute of Physics
cleanroom at EPFL. We thank J. Liu for helping with the fabrication.
Funding: This material is based on work supported by the Air
Force Office of Scientific Research under award FA9550-19-1-0250
and by the Swiss National Science Foundation under grant
agreement 185870 (Ambizione). Y.Y. acknowledges support from
the EU H2020 research and innovation program under the
Marie Skłodowska-Curie IF grant agreement 101033593 (SEPhIM).
The experiments were conducted at the Göttingen UTEM Lab,
funded by the Deutsche Forschungsgemeinschaft (DFG;
German Research Foundation) through grant 432680300/SFB
1456 (project C01) and the Gottfried Wilhelm Leibniz program,
and the EU H2020 research and innovation program under
grant agreement 101017720 (FET-Proactive EBEAM).

Yang et al., Science 383, 168–173 (2024) 12 January 2024 5 of 6

RES EARCH | R E S E A R C H A R T I C L E

Author contributions: Conceptualization, supervision, project
administration, and funding acquisition: T.J.K. and C.R. Methodology:
A.S.R., Y.Y., J.-W.H., F.J.K., G.H., Z.Q., A.F., R.N.W., A.Tu., and A.Ti.
Software: F.J.K., A.Ti., J.-W.H., and A.Tu. Validation: A.S.R., J.-W.H.,
Y.Y., and F.J.K. Formal analysis: Y.Y., J.-W.H., A.S.R., F.J.K., and
G.H. Investigation: J.-W.H., Y.Y., F.J.K., A.S.R., G.H., G.A., and
A.F. Resources: F.J.K., Z.Q., R.N.W., A.F., and G.A. Data curation:
J.-W.H., F.J.K., Y.Y., and A.S.R. Visualization: Y.Y., F.J.K., A.S.R.,
and J.-W.H. Writing (original draft preparation): Y.Y., J.-W.H.,
A.S.R., F.J.K., C.R., and T.J.K. Writing (review and editing): all SUPPLEMENTARY MATERIALS
authors. Competing interests: The authors declare no competing science.org/doi/10.1126/science.adk2489
financial interests. Data and materials availability: The code and Materials and Methods
data used to produce the plots within this work are available in Supplementary Text
Zenodo (39). License information: Copyright © 2024 the authors, Figs. S1 to S17
some rights reserved; exclusive licensee American Association References (40–84)
for the Advancement of Science. No claim to original US
government works. https://www.science.org/about/science- Submitted 10 August 2023; accepted 17 November 2023
licenses-journal-article-reuse 10.1126/science.adk2489

p
g
y
y g
y
,

Yang et al., Science 383, 168–173 (2024) 12 January 2024 6 of 6

RES EARCH

ELECTROCHEMISTRY ylporphyrin) has been studied recently as a

molecular ORR electrocatalyst and was selected
Synthetic dioxygenase reactivity by pairing as the catalyst for the present investigation
(Fig. 2A) (23). Cyclic voltammetry (CV) of 1-Cl
electrochemical oxygen reduction and water oxidation in acetonitrile showed a reversible redox wave
at –0.75 V versus Fc+/Fc (where Fc is ferrocene),
Md. Asmaul Hoque, James B. Gerken, Shannon S. Stahl* corresponding to the MnIII/II redox couple.
Substantial enhancement of the cathodic cur-
The reactivity of molecular oxygen is crucial to clean energy technologies and green chemical synthesis, rent was observed at this potential under 1 atm
but kinetic barriers complicate both applications. In synthesis, dioxygen should be able to undergo O2 in the presence of acetic acid (AcOH) (Fig.
oxygen atom transfer to two organic molecules with perfect atom economy, but such reactivity is rare. 2B, blue trace a), resembling studies of ORR
Monooxygenase enzymes commonly reductively activate dioxygen by sacrificing one of the oxygen atoms electrocatalysis with molecular Mn complexes
to generate a more reactive oxidant. Here, we used a manganese-tetraphenylporphyrin catalyst to pair (23–26). The addition of Ph2S to this reaction
electrochemical oxygen reduction and water oxidation, generating a reactive manganese-oxo at mixture suppressed the catalytic current (Fig.
both electrodes. This process supports dioxygen atom transfer to two thioether substrate molecules, 2B, red trace b). This current suppression was
generating two equivalents of sulfoxide with a single equivalent of dioxygen. This net dioxygenase also evident in constant-potential electrolysis
reactivity consumes no electrons but uses electrochemical energy to overcome kinetic barriers. experiments performed under the optimized
reaction conditions (see below), with an ap-
plied potential of –0.8 V versus Fc+/Fc (Fig. 2C).

C
ontrolling the chemistry of molecular the kinetic limitations of OAT by generating a In the presence of the Ph2S substrate, a stable
oxygen (O2) is notoriously challenging. stronger oxidant, which is reflected by the current of ~–5 mA was recorded. A slow decay

Reactive oxygen species (ROS) derived
from O2 play a major role in human
disease and aging. The energy ineffici-
ency of the oxygen reduction and evolution reac-
tions (ORR and OER, respectively) in fuel
cells and electrolyzers impedes the broader
higher reduction potential of H2O2 relative to
O2: E°H2O2/H2O = 1.76 V versus E°O2/H2O = 1.23 V
(Fig. 1A). The challenge of using O2 as a four-
electron oxidant in two OAT reactions may be
compared with the four-electron electrochemical
ORR and OER processes. Studies of molecular
p
in the current was evident in the absence of
the substrate, but the initial current of –8.9 mA
was nearly twice that observed in the presence
of Ph2S.
The reduction in current evident in the pres-
ence of substrate suggests that Ph2S can inter-

adoption of these clean energy technologies.
O2 finds limited use as an environmentally
benign chemical oxidant because of its poor
selectivity in reactions with most organic mol-
catalysts for the latter reactions commonly
invoke metal-oxo intermediates that undergo
further reduction to H2O (ORR) (14) or that
react with H2O or another metal-oxo species
g
cept the proposed MnV=O ORR intermediate
(Fig. 2D, path b), enabling OAT rather than full
reduction of O2 to H2O (Fig. 2D, path a). Bulk
electrolysis data suggest that this alternate

ecules. These examples highlight the multitude
of redox steps available to O2: single-electron
steps associated with ROS, 4 e–/4 H+ processes
in the ORR and OER, and two-electron oxida-
tion reactions of organic molecules. Oxygen-
atom transfer (OAT) reactions are prominent
two-electron oxidation reactions that play a
central role in the synthesis of organic chem-
icals, ranging from pharmaceuticals to polymer
building blocks. Energetic considerations indi-
to form the O–O bond (OER) (15). These mech-
anisms inspired consideration of an electro-
chemical strategy to achieve net dioxygenase
reactivity by pairing monooxygenase and de-
hydrogenase reactivity at the cathode and
anode, respectively (Fig. 1B). Parallel pathways
for reductive activation of O2 and oxidation of
H2O will generate the same metal-oxo species
(Fig. 1C), enabling OAT from O2 to two sub-
strate molecules. Here, we demonstrate this
y
pathway can be quite efficient. The Mn-catalyzed
OAT reaction with Ph2S proceeded with an
excellent FE (87%) during constant current
electrolysis experiments, with a >30:1 ratio
of sulfoxide to sulfone (Ph2SO2) (5 mA; Fig. 2E,
entry 1). The nonunity FE likely arises from
electrochemical reduction of the MnV=O inter-
mediate, reflecting 4 e– ORR (23). The relative
rate of these pathways changed under different
conditions. When AcOH was replaced with tri-

cate that O2 should be able to add oxygen to
two organic molecules with perfect atom econ-
omy. Such reactions are favorable (1), but
thermodynamically challenging oxidations, such
as C–H hydroxylation, thioether oxygenation,
and alkene epoxidation (Fig. 1A), have rela-
concept by using a manganese porphyrin cat-
alyst to support double OAT with a single
equivalent of O2, with thioether oxidation to
sulfoxide as a representative high-potential oxi-
dation reaction. This specific example of linear
paired electrolysis (16–19) shows how electro-
y g
fluoroacetic acid (TFAH), then negligible Ph2S
oxidation was observed (Fig. 2E, entry 2). This
observation and results obtained with other
carboxylic acids suggest that stronger acids
favor proton-coupled reduction of the MnV=O
intermediate over OAT (see section 6 of the

y
tively small potential difference between O2 chemical energy may be used to overcome supplementary materials for additional data).

reduction and substrate oxidation. Such reac-
tions have little kinetic margin for error, and
successful examples are exceedingly rare (2–8).
Monooxygenase enzymes such as cytochrome
P450 and methane monooxygenase generate a
reactive metal-oxo species by using a sacrificial
reductant to activate O2 (9–11). Dioxygenase
enzymes transfer both oxygen atoms from O2
to substrate(s), but these enzymes typically use
a-ketoglutarate or another sacrificial substrate
to promote two-electron oxidation reactions
(8, 12, 13). Reductive activation of O2 overcomes
Department of Chemistry, University of Wisconsin–Madison,
Madison, WI 53706, USA.
*Corresponding author. Email: stahl@chem.wisc.edu
kinetic barriers in the generation of reactive
OAT species from O2 without the net con-
sumption of electrons (Fig. 1C).
Cathodic monooxygenase activity
This study was initiated by independently ex-
ploring the cathodic monooxygenase and anodic
dehydrogenase reactions using the oxidation of
diphenyl sulfide (Ph2S) to the corresponding
sulfoxide (Ph2SO) as a benchmark OAT reaction.
Prior studies have explored the use of Fe- and Mn-
porphyrins in electrochemical monooxygenase
reactions (20–22). These studies did not report
synthetic yields, and only moderate Faradaic
efficiency (FE) was observed (typically <50%
FE). Mn(TPP)Cl (1-Cl, where TPP is tetraphen-
,
Even better OAT results were observed upon
adding N-methyl imidazole (NMI) as an axial
ligand to the conditions with AcOH, resulting
in 99% FE (Fig. 2E, entry 3). High FE (94%) was
retained even when the catalyst loading was
reduced to 0.1 mol% (Fig. 2E, entry 4). Col-
lectively, these results show that catalytic ORR
reactivity can be diverted to support efficient
monooxygenase OAT reactivity through electro-
chemical reductive activation of O2.
Anodic dehydrogenase activity
The same Mn(TPP) complex was then eval-
uated as an anodic OAT electrocatalyst (27–32),
with the goal of generating a MnV=O species
by proton-coupled oxidation of a MnIII–OH2

Hoque et al., Science 383, 173–178 (2024) 12 January 2024 1 of 6

RES EARCH | R E S E A R C H A R T I C L E

p
g
y
Fig. 1. Adaptation of O2 reduction and H2O oxidation for organic oxidation reaction. (A) Energetic analysis for O2 reduction, substrate oxidation, and net
dioxygenase reactivity (gray box). (B) Electrochemical O2 reduction and evolution reactions may be adapted to support oxygen atom–transfer reactions through
reductive activation of O2 and H2O oxidation. (C) Linear paired electrolysis strategy to support synthetic dioxygenase reactivity using a common catalyst.

species (Fig. 3A) (33). This reactivity resembles
key steps in the electrocatalytic OER. Cofacial
bridged Mn-porphyrins and mononuclear
Mn-macrocycles have been studied previous-
appeared just beyond the MnIV/III wave. Con-
stant current electrolysis experiments revealed
nearly quantitative FE in the presence of 0.1 or
1 M H2O, forming only Ph2SO (Fig. 3C, entries
y g
(1:1 acid:base) with pKa values ranging from
8.8 to 14.7. The resulting data provide the basis
for a nonaqueous Pourbaix diagram (Fig. 3D)
(37). Several features of the Pourbaix diagram

y
ly as electrochemical OER electrocatalysts 1 and 2). No sulfone by-product was observed merit attention. First, the MnIII/II redox process

(34–36). The proposed role of MnV=O inter-
mediates in these reactions suggested that it
might be possible to intercept this species
with an organic substrate, enabling OAT
rather than O2 evolution. CV analysis of 1-Cl
in MeCN in the presence of 1 M H2O showed
two quasireversible redox features at 0.59 and
0.82 V versus Fc+/Fc, corresponding to MnIV/III
and MnV/IV redox events, respectively (Fig. 3B,
black trace). The lack of a catalytic wave at
either potential indicates that reaction of
these species with H2O is slow on the CV
time scale. By contrast, substantial catalytic
current was observed when 100 mM Ph2S was
added to this solution (Fig. 3B, red trace). The
catalytic onset potential of 0.64 V versus Fc+/Fc
under these conditions, and inclusion of NMI
as an axial ligand had minimal impact on the
reactivity (95% FE; Fig. 3C, entry 3). Once
again, excellent performance was observed upon
lowering the catalyst loading to 0.1 mol%
(97% FE).
Reactions with other substrates revealed that
the catalytic onset potential can vary. For exam-
ple, the electron-rich substrate thioanisole fea-
tured an onset potential aligned with the MnIV/III
potential, whereas the electron-deficient sub-
strate 4-cyano-thioanisole aligned with the MnV/IV
potential (fig. S12). To gain more insights into
Mn-based redox processes, redox potentials
were recorded for [Mn(TPP)(OH2)]SbF6, 1-OH2,
with a series of different buffering electrolytes
,
is independent of the buffer pKa, implicating
a [MnIII-OH2]+/[MnII-OH2] redox couple that
involves no proton transfer to or from the
bound H2O molecule. Second, the MnIV/III
potential varies with the buffer and exhibits
a change in slope near the middle of the buf-
fer pKa range. These results suggest that
the Mn IV/III potential corresponds to [MnIII
-OH2]+/[MnIV-OH]+ at lower electrolyte pKa and
[MnIII-OH2]+/[MnIV=O] at higher pKa. Fitting
of the data to reflect theoretical slopes of –59
and –118 mV/pKa units across the full range
yields a pKa value of 12.4 for the [MnIV-OH]+
species (see section 8.3 in the supplementary
materials for fitting procedure). Third, the MnV/IV
potential exhibits a –59 mV/pKa slope at pKa <

Hoque et al., Science 383, 173–178 (2024) 12 January 2024 2 of 6

RES EARCH | R E S E A R C H A R T I C L E

p
g
y
y g
y
Fig. 2. Analysis of electrochemical monooxygenase reactivity. (A) Electrochemical monooxygenase reactions. The intermediate MnV=O species can undergo oxygen

reductive activation of O2 for monooxygenase-type oxygenation of Ph2S. (B) Cyclic
voltammograms of 1 mM 1-Cl in acetonitrile under N2 (black), 1 atm O2/200 mM
AcOH (blue), and 1 atm O2/200 mM AcOH/100 mM Ph2S (red), all with 0.1 M
n-Bu4NPF6. Scan rate = 50 mV/s. (C) Constant potential electrolysis of 1 mM 1-Cl
(1 mol%) in a divided cell, 1 atm O2/200 mM AcOH/100 mM Ph2S in MeCN (red),
and 1 atm O2/200 mM AcOH in MeCN (blue). (D) Proposed mechanism for electrochemical
,
atom transfer or proton-coupled reduction to H2O. (E) Influence of acid source and NMI
on Ph2S oxidation during constant current electrolysis (5 mA) in an H-type divided cell.
A 38.6 C charge was passed to attain 20% theoretical conversion. Faradaic yield is the
charge used to form product (2 e−/mol Ph2SO, 4 e−/mol Ph2SO2)/total charge passed;
TON is the turnover number (mmol product/mmol 1-Cl). FEc, FE in the cathode (sum
of the Faradaic yields); RVC, reticulated vitreous carbon. *0.1 mM 1-Cl (0.1 mol%).

12.4, consistent with a [MnIV-OH]+/[MnV=O]+
redox couple. For buffers with pKa > 12.4, the
MnV/IV potential was less well defined because
of the redox instability of the electrolytes;
however, the data fit reasonably well with a
pKa-independent redox couple, corresponding
to [MnV=O]+/[MnIV=O]. These data, together
with the observed onset potentials for catalytic
substrate oxidation, provide the basis for the
catalytic mechanism shown in Fig. 3E. The
substrate-dependent variations in catalytic
onset potentials suggest that OAT can be ini-
tiated by either MnIV-(hydr)oxo or MnV-oxo
species. Both pathways enable OAT to the or-
ganic substrate with no competition from H2O
oxidation to O2.
Net dioxygenase activity
The results presented in Figs. 2 and 3 show
that the same Mn(TPP) catalyst supports both
monooxygenase reactivity through reductive

Hoque et al., Science 383, 173–178 (2024) 12 January 2024 3 of 6

RES EARCH | R E S E A R C H A R T I C L E

p
g
y
y g
y
,
Fig. 3. Analysis of electrochemical dehydrogenase reactivity. (A) Electrochemical current electrolysis (5 mA) in an H-type divided cell. A 38.6 C charge was passed
oxidative activation of H2O for dehydrogenase-type oxygenation of Ph2S. (B) Cyclic to attain 20% theoretical conversion. FEa, FE in the anode (sum of the Faradaic yields).
voltammograms of 1 mM 1-Cl in acetonitrile with 1 M H2O in the absence of 100 mM *0.1 mM 1-Cl (0.1 mol%). (D) Nonaqueous Pourbaix diagram with various organic
Ph2S (black) and in the presence of 100 mM Ph2S (red), all with 0.1 M n-Bu4NPF6. buffer solutions (see fitting details in section 8.3 of the supplementary materials).
Scan rate = 50 mV/s. (C) Influence of H2O and NMI on Ph2S oxidation during constant (E) Proposed mechanism for electrochemical dehydrogenase reactions.

activation of O2 at the cathode and dehydro-
genase reactivity through H2O oxidation at
the anode. Because H2O is a stoichiometric by-
product in the cathodic reaction and a sub-
strate in the anodic reaction, pairing of these
two reactions within a single electrochemical
cell would support net dioxygenase reactivity
using one molecule of O2 to support two OAT
reactions (Fig. 4A). Paired electrolysis perfor-
mance was evaluated by using a “combined FE”
metric that accounts for the charge associated
with the formation of sulfoxide (2 e–) and sul-
fone (4 e–) at either electrode, divided by the
total charge passed in the electrochemical cir-
cuit. FEs (i.e., Faradaic yields) of 200% are
theoretically possible according to this metric
because passing two electrons through the
circuit can generate up to two equivalents of
sulfoxide. Although this metric is used in the lit-
erature (16, 18) and is helpful for “bookkeeping,”
there is no consumption or generation of elec-
trons in the net dioxygenase reaction (compare
Fig. 1B and see further discussion below).
To test this possibility, the Mn(TPP) catalyst
1-Cl and NMI were combined with the Ph2S
substrate, H2O (1 M), and AcOH (200 mM) in
an undivided cell and placed under 1 atm of O2.

Hoque et al., Science 383, 173–178 (2024) 12 January 2024 4 of 6

RES EARCH | R E S E A R C H A R T I C L E

Fig. 4. Analysis of electrochemical dioxygenase reactivity. (A) Electrochemical was charged with 1 atm O2. FEtotal, FE of the cell (sum of the Faradaic yields p
g
y
y g
y
,
reductive activation of O2 and oxidative activation of H2O for oxygenation of Ph2S. in both the cathode and anode). *Reaction performed in an undivided
(B) Linear paired electrolysis test for Ph2S oxidation during constant current cell under identical reaction conditions to entry 2. (C) Preparative scale
electrolysis (5 mA) in an H-type divided cell. A 38.6 C charge was passed to attain demonstration of linear paired electrolysis. (D) Energetic analysis of dioxygenase
20% theoretical conversion in each compartment. The cathode compartment reactivity.

The AcOH was included to support O2 reduc-
tion at the cathode (compare Fig. 2); control
experiments showed that AcOH does not in-
terfere with the anodic OAT reaction (see
section 7.2 of the supplementary materials).
Conducting the reaction with a constant cur-
rent of 5 mA led to the formation of Ph2SO as
the major product with small amounts of
Ph2SO2; 140% FE was observed (138:2%
Faradaic yield of Ph2SO:Ph2SO2) (Fig. 4B,
entry 1). The lack of side product formation
suggests that the FE is lowered by redox cycl-
ing of the Mn catalyst at the cathode and
anode, passing charge without oxidizing the
substrate. This problem was avoided by using
a divided cell, because 195% FE was observed
when conducting the same experiment in an
H cell equipped with a glass frit divider (Fig. 4B,
entry 2, and see photograph of the reaction
cell). Further improvement was achieved when
NMI was not added to the anode compartment
(compare Fig. 3): 200% FE was observed
(197:3% for Ph2SO:Ph2SO2), reflecting perfect
utilization of electrons in the double-OAT

Hoque et al., Science 383, 173–178 (2024) 12 January 2024 5 of 6

RES EARCH | R E S E A R C H A R T I C L E
process. Excellent FE was also achieved on a
preparative (1 mmol) scale, evident from the
189 to 196% FE observed for oxidation of a diaryl,
a dialkyl, and an aryl alkyl sulfide (Fig. 4C).
The net stoichiometry of this electrochemical
process corresponds to a dioxygenase reaction
in which O2 adds an oxygen atom to two equiv-
alents of organic substrate. Electrons are neither
generated nor consumed in this net reaction,
but the electrochemical potential provides the
energy input needed to overcome the kinetic
barriers that limit thermal reactivity. This
energy input may be quantified by analyzing
the cathodic and anodic redox potentials (Fig.
4D). CV analysis reveals an onset potential
at –0.59 V versus Fc+/Fc for O2 reductive activa-
tion and at +0.65 V versus Fc+/Fc for H2O
oxidation (the onset potential was defined
as Icat = 20 mA). The difference of 1.24 V
corresponds to a free energy of +28.6 kcal/mol
used to promote the dioxygenase reaction. The
Pourbaix diagram in Fig. 3D suggests that it
might be possible to reduce the energy gap by
using an electrolyte with a higher pKa. With an
acetic acid/acetate (2:1) electrolyte, the cathod-
ic and anodic onset potentials shifted to –0.69
and +0.29 V versus Fc+/Fc, respectively, while
retaining quantitative current efficiency (Fig.
4D). This 0.98 V gap reflects a reduced energy
input of 22.6 kcal/mol (for similar analysis
with buffering electrolytes with intermediate
pKa values, see figs. S17 to S20). The thermo-
dynamic potential for Me2SO/Me2S under
these buffered conditions is E°′ = –0.15 V. Thus,
the overpotentials for thioether oxidation are
approximately balanced between the two half
reactions, with a slightly higher value for the
cathodic monooxygenase step (DE ~ 0.54 V)
relative to the anodic dehydrogenase step
(DE ~ 0.44 V) (Fig. 4D). This analysis permits
comparison to photochemical methods that have
been used to promote dioxygenase-like reactivity.
An oxo-bridged cofacial bis-FeIII-porphyrin has
been shown to support oxidation of dimethyl
sulfide by O2 in a 2:1 ratio under irradiation with
365 to 425 nm light, corresponding to an energy
input of 66.9 to 78.4 kcal/mol (4, 38). In addition
to the three- to fourfold higher energy input, this
process features a quantum yield of only 10−4 to
10−8, which may be compared to the quanti-
tative FE of the electrochemical process. Another
method for photochemically promoted aerobic
oxidation of thioethers features a Ru-salen cat-
alyst under 467 nm irradiation (61.2 kcal/mol),
although the lack of a reported quantum effi-
ciency prevents quantitative comparison (39).
Hoque et al., Science 383, 173–178 (2024) 12 January 2024
The analysis outlined above provides a cru-
cial bridge between the reactivity of molecular
oxygen in synthetic and energy conversion
applications. Electrochemical metrics such as
overpotential and 2- versus 4-e– reaction se-
lectivity are widely used in the development of
ORR and OER electrocatalysts for fuel cells
and electrolyzers. The present study shows how
similar metrics may be used to guide the devel-
opment of (electro)catalysts for chemical
synthesis, enabling a rare demonstration of
dioxygenase reactivity. Sulfide oxidation was
chosen as a prototypical example of a “high-
potential” substrate, but many other oxidation
reactions, including C–H hydroxylation and
alkene epoxidation, feature similar challenges
(compare Fig. 1A). Dioxygenase reactivity with
each of these substrate classes is thermody-
namically possible but faces substantial kinetic
barriers associated with O2 activation and sub-
sequent OAT. The analysis introduced here
provides a foundation for the systematic design
and optimization of new catalysts that achieve
better efficiency and access effective dioxy-
genase reactivity with diverse substrates.
RE FERENCES AND NOTES
1. A. G. Stamoulis, D. L. Bruns, S. S. Stahl, J. Am. Chem. Soc. 145,
17515–17526 (2023).
2. T. Pu, H. Tian, M. E. Ford, S. Rangarajan, I. E. Wachs,
ACS Catal. 9, 10727–10750 (2019).
3. J. T. Groves, R. Quinn, J. Am. Chem. Soc. 107, 5790–5792
(1985).
4. B. J. Pistorio, C. J. Chang, D. G. Nocera, J. Am. Chem. Soc. 124,
7884–7885 (2002).
5. S. Koya, Y. Nishioka, H. Mizoguchi, T. Uchida, T. Katsuki,
Angew. Chem. Int. Ed. 51, 8243–8246 (2012).
6. T. D. H. Bugg, Tetrahedron 59, 7075–7101 (2003).
7. E. G. Kovaleva, J. D. Lipscomb, Nat. Chem. Biol. 4, 186–193
(2008).
8. F. P. Guengerich, F. K. Yoshimoto, Chem. Rev. 118, 6573–6655
(2018).
9. A. L. Feig, S. J. Lippard, Chem. Rev. 94, 759–805 (1994).
10. I. G. Denisov, T. M. Makris, S. G. Sligar, I. Schlichting,
Chem. Rev. 105, 2253–2278 (2005).
11. D. E. Torres Pazmiño, M. Winkler, A. Glieder, M. W. Fraaije,
J. Biotechnol. 146, 9–24 (2010).
12. S.-S. Gao, N. Naowarojna, R. Cheng, X. Liu, P. Liu, Nat. Prod. Rep.
35, 792–837 (2018).
13. S. M. Barry, G. L. Challis, ACS Catal. 3, 2362–2370 (2013).
14. M. L. Pegis, C. F. Wise, D. J. Martin, J. M. Mayer, Chem. Rev.
118, 2340–2391 (2018).
15. R. Matheu et al., Nat. Rev. Chem. 3, 331–341 (2019).
16. W. Li, T. Nonaka, J. Electrochem. Soc. 146, 592–599 (1999).
17. J. G. Ibanez, B. A. Frontana-Uribe, R. Vasquez-Medrano, J. Mex.
Chem. Soc. 60, 247–260 (2016).
18. J. Strehl, M. L. Abraham, G. Hilt, Angew. Chem. Int. Ed. 60,
9996–10000 (2021).
19. H. Sheng et al., Nat. Catal. 5, 716–725 (2022).
20. S. E. Creager, S. A. Raybuck, R. W. Murray, J. Am. Chem. Soc.
108, 4225–4227 (1986).
21. P. Leduc, P. Battioni, J. F. Bartoli, D. Mansuy, Tetrahedron Lett.
29, 205–208 (1988).
22. M. Mukherjee, A. Dey, ACS Cent. Sci. 5, 671–682 (2019).
23. G. Passard, D. K. Dogutan, M. Qiu, C. Costentin, D. G. Nocera,
ACS Catal. 8, 8671–8679 (2018).
24. N. Kobayashi, H. Saiki, T. Osa, Chem. Lett. 14, 1917–1920
(1985).
25. R. L. Shook et al., J. Am. Chem. Soc. 133, 5810–5817
(2011).
26. E. N. Cook, C. W. Machan, Chem. Commun. 58, 11746–11761
(2022).
27. K.-C. Cheung, W.-L. Wong, D.-L. Ma, T.-S. Lai, K.-Y. Wong,
Coord. Chem. Rev. 251, 2367–2385 (2007).
28. A. K. Vannucci, Z. Chen, J. J. Concepcion, T. J. Meyer,
ACS Catal. 2, 716–719 (2012).
29. A. Das, J. E. Nutting, S. S. Stahl, Chem. Sci. 10, 7542–7548
(2019).
30. K. Jin et al., J. Am. Chem. Soc. 141, 6413–6418 (2019).
31. B. Chandra, K. M. Hellan, S. Pattanayak, S. S. Gupta,
Chem. Sci. 11, 11877–11885 (2020).
32. S. G. Robinson, J. B. C. Mack, S. N. Alektiar, J. Du Bois,
M. S. Sigman, Org. Lett. 22, 7060–7063 (2020).
33. C. J. Gagliardi, A. K. Vannucci, J. J. Concepcion, Z. Chen,
T. J. Meyer, Energy Environ. Sci. 5, 7704–7717 (2012).
34. Y. Naruta, M. Sasayama, T. Sasaki, Angew. Chem. Int. Ed. 33,
1839–1841 (1994).
35. X. Li et al., J. Am. Chem. Soc. 143, 14613–14621 (2021).
36. D. Wang, J. T. Groves, Inorg. Chem. 61, 13667–13672
(2022).
37. B. D. McCarthy, J. L. Dempsey, Inorg. Chem. 56, 1225–1231
p
(2017).
38. J. Rosenthal, B. J. Pistorio, L. L. Chng, D. G. Nocera, J. Org. Chem.
70, 1885–1888 (2005).
39. H. Tanaka, H. Nishikawa, T. Uchida, T. Katsuki, J. Am. Chem. Soc.
132, 12034–12041 (2010).
AC KNOWLED GME NTS
We thank I. A. Guzei and K. M. Sanders for assistance with x-ray
g
crystallographic characterization and D. T. Hofsommer for helpful
discussions. Funding: This work was initially supported by the
Center for Molecular Electrocatalysis, an Energy Frontier Research
Center funded by the US Department of Energy, Office of Science,
Office of Basic Energy Sciences, and continued with support
y
from the National Institutes of Health (NIH grant R35 GM134929).
The spectrometers were supported by the National Science
Foundation (NSF grant CHE-1048642) and by a generous gift
from Paul J. and Margaret M. Bender. The mass spectrometer
was supported by the NIH (grant S10 OD020022). The x-ray
diffractometer was partially supported by the NSF (grant CHE-
1919350 to the Univeristy of Wisconsin−Madison Department of
Chemistry). Author contributions: This project was conceived by
S.S.S. in collaboration with Md.A.H. and J.B.G. All the work was
done by Md.A.H. in consultation with J.B.G. and S.S.S. All authors
contributed to the preparation of the manuscript. Competing
interests: The authors declare no competing interests. Data and
materials availability: X-ray structural data with access code
CCDC 2289621 for 1-Cl and 2289622 for 1-OH2 are available free
y g
of charge from the Cambridge Crystallographic Data Centre (www.
ccdc.cam.ac.uk/data_request/cif; Email: data_request@ccdc.cam.
ac.uk). All other data are available in the main text or the
supplementary materials. License information: Copyright © 2024
the authors, some rights reserved; exclusive licensee American
Association for the Advancement of Science. No claim to original
y
US government works. https://www.science.org/about/science-
licenses-journal-article-reuse
,
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.adk5097
Materials and Methods
Figs. S1 to S20
Tables S1 to S15
NMR Spectra
References (40–59)
Submitted 2 September 2023; accepted 4 December 2023
10.1126/science.adk5097
6 of 6
RES EARCH

MEMBRANES fusion of adsorbed molecules (Fig. 1A). Relative

to the whole narrow sieving channel, the local
A molecular sieve with ultrafast adsorption kinetics sieving channel shortens the restricted diffusion
path of adsorbed molecules. Porous crystalline
for propylene separation materials that are self-assembled by secondary
building blocks (SBUs) offer the possibility of
Jiyu Cui1, Zhaoqiang Zhang3, Lifeng Yang1*, Jianbo Hu2, Anye Jin1, Zhenglu Yang1, Yue Zhao4, tailor-made porous materials with designed
Biao Meng5, Yu Zhou5, Jun Wang5, Yun Su6, Jun Wang6, Xili Cui1,2, Huabin Xing1,2* pore size (39, 40), shape (41, 42), and pore
chemistry (43–45). Rational modification using
The design of molecular sieves is vital for gas separation, but it suffers from a long-standing issue of anions based on the original two-dimensional
slow adsorption kinetics due to the intrinsic contradiction between molecular sieving and diffusion within (2D) structure provides the possibility of pore
restricted nanopores. We report a molecular sieve ZU-609 with local sieving channels that feature shape and pore size control, from which mo-
molecular sieving gates and rapid diffusion channels. The precise cross-sectional cutoff of molecular lecular sieve ZU-609 was designed.
sieving gates enables the exclusion of propane from propylene. The coexisting large channels constituted
Propylene molecular sieve design
by sulfonic anions and helically arranged metal-organic architectures allow the fast adsorption kinetics
of propylene, and the measured propylene diffusion coefficient in ZU-609 is one to two orders of ZU-609 is self-assembled with inorganic
magnitude higher than previous molecular sieves. Propylene with 99.9% purity is obtained through metal node and organic linkers (EDS2−, 1,2-
breakthrough experiments with a productivity of 32.2 L kg−1. ethanedisulfonate; dps, 4,4’-dipyridyl sulfide)
(Fig. 1B). It consists of a 2D network that is

A
coordinated by copper metal nodes and dps
s a key feedstock, the production of pro- in molecular diffusion rates (18–21). However, ligands, and it extends in a helical-like con-

pylene (C3H6) exceeded 100 metric tonnes
(Mt) per year in 2020 and its demand is
expected to exceed 150 Mt per year in
2050 (1). Propane (C3H8) is a common
byproduct of propylene manufacturing and
separation of propylene and propane is essen-
as a result of low to moderate selectivity, these
methods inevitably suffer from co-adsorption
of similar alkanes, and consecutive adsorption-
desorption cycles are needed to further im-
prove the purity of the produced olefins (22–24).
Molecular sieving avoids co-adsorption as
p
formation along the direction that is vertical
to the network (Fig. 1B and fig. S1). The EDS2−
sulfonic anions coordinate to the bare Cu2+ on
the adjacent 2D networks to form 3D coordi-
nation networks (Fig. 1C, figs. S1 and S2, and
table S1) (46, 47). It possesses large one-

tial for production of polymer-grade propylene
(2–4). Conventional cryogenic distillation is
energy intensive as a result of the close rela-
tive volatility and the production of olefin and
it only adsorbs molecules within the molec-
ular size or shape cutoff (25–27) and is ad-
vantageous in the production of high-purity
olefins (28–30). However, the design of ideal
g
dimensional channels with a size of 7.5 × 8.1 ×
11.1 Å3, connected by the local contraction
resulting from the tilted pyridine rings (Fig.
1D). The cross-section size of the contraction

paraffin accounts for nearly 1% of global car-
bon emissions (5–7). It is estimated that the
development of nonthermally driven alterna-
tives could make the separation ten times more
energy efficient (8, 9).
Physisorption based on porous materials is
recognized as an alternative to cryogenic dis-
tillation owing to its moderate energy con-
sumption (10–12). The development of advanced
porous materials such as metal-organic frame-
molecular sieves with optimal adsorption ther-
modynamics and kinetics for gas separation
remains a daunting task because of two im-
portant challenges: (i) Pore size control for
molecular size sieving (31) and (ii) pore shape
control for fast adsorption kinetics (32). Con-
sidering the difficulties in fine-tuning the pore
size in 0.2 to 0.4 Å increments within the 3
to 5 Å range, the full exclusion of propylene
from propane remains a challenge as a result
y
is 4.2 Å × 5.1 Å, which falls just between the
size of C3H6 (4.1 Å × 5.1 Å) and C3H8 (5.1 Å ×
5.3 Å) molecules, suggesting the size/shape
sieving potential of ZU-609 (Fig. 1E). The sur-
face area of ZU-609 is explored using both N2
and CO2 as probes. Because of the obvious
diffusion barrier of N2 in ZU-609 at 77 K, the
CO2 probe is selected and the measured ap-
parent surface area and pore volume of ZU-609
are 380 m2 g−1 and 0.15 cm3 g−1, respectively

works, zeolites, and others has attracted con-
siderable interest (13–15). Adsorbents with open
metal sites and cations can selectively interact
with propylene through p complexation but
show low selectivity (16, 17). Additionally, kinetic-
driven adsorbents with restricted pore struc-
of their small size difference (< 0.4 Å) (33–35).
Adsorption kinetics is scarified in molecular
sieves as ultrahigh selectivity is achieved based
on restricted channels (2, 36). Thus, a high ad-
sorption temperature is required to reach the
diffusion rate threshold; however, this approach
y g
(fig. S3). Additionally, the thermal stability of
ZU-609 is around 200°C and the activated
structure remains stable after different treat-
ments, such as air, water, and solutions of dif-
ferent pH (pH 3 to pH 11) (figs. S4 and S5).

y
tures realize separation based on the difference sacrifices capacity and energy efficiency. Equilibrium propylene/propane

1
Key Laboratory of Biomass Chemical Engineering of Ministry of
Education, College of Chemical and Biological Engineering,
Zhejiang University, Hangzhou 310012, P.R. China. 2Engineering
Research Center of Functional Materials Intelligent Manufacturing
of Zhejiang Province, ZJU-Hangzhou Global Scientific and
Technological Innovation Center, Hangzhou 311215, P.R. China.
3
Department of Chemical and Biomolecular Engineering, National
University of Singapore, Singapore 117585, Singapore.
4
Coordination Chemistry Institute, State Key Laboratory of
Coordination Chemistry, School of Chemistry and Chemical
Engineering, Nanjing University, Nanjing 210023, P.R. China.
5 pore space within molecular sieves, a sieving
State Key Laboratory of Materials-Oriented Chemical
Engineering, College of Chemical Engineering, Nanjing Tech
University, Nanjing 211816, P.R. China. 6School of Chemistry and
Chemical Engineering, Nanchang University, Nanchang 330031,
P.R. China.
*Corresponding author. Email: xinghb@zju.edu.cn (H.X.);
lifeng_yang@zju.edu.cn (L.Y.)
The dilemma is ascribed to the continuous
diffusion limitation path along the whole nar-
row sieving channel (Fig. 1A), a typical pore
structure in molecular sieves, which allows
only “single-file diffusion” of adsorbates. The
development of common cage-type structures
with narrow windows and large voids breaks
the tradeoff between separation selectivity and
capacity, but still suffers from diffusion limita-
tions (37, 38). To rationally utilize the limited
channel that features a local limited diffusion
path is proposed to solve the challenge. The
local contraction along the diffusion path func-
tions as a sieve to exclude large molecules and
the coexisting large channel allows the rapid dif-
,
separation performance
Pure-component equilibrium adsorption iso-
therms for C3H6 and C3H8 were measured,
and ZU-609 showed sieving performance for
C3H6 and C3H8 with an uptake ratio of 22.3 at
298 K, 1 bar (Fig. 2A, fig. S6, and table S2). The
time-dependent adsorption profile of C3H8 was
measured and its uptake was almost invariable
in 24 hours, indicating that the low C3H8 up-
take of ZU-609 is attributed to the size-sieving
effect (fig. S7). The C3H6 adsorption isotherm
approaches linear type with a moderate iso-
steric heat of adsorption (Qst) of 43 kJ mol−1
(figs. S8 and S9), providing a considerable
C3H6 equilibrium working capacity of 2.0 mmol
g−1 between 0.1 and 1 bar (Fig. 2A). The value is

Cui et al., Science 383, 179–183 (2024) 12 January 2024 1 of 5

RES EARCH | R E S E A R C H A R T I C L E

p
g
y
y g
Fig. 1. Channel types and crystal structure of ZU-609. (A) The schematic diagrams of different kinds of sieving channels. (B) The secondary building blocks of ZU-609.
(C) The assembled crystal structure of ZU-609 and its channel structure (color code: C, gray; H, white; N, blue; Cu, pink; O, red; S, yellow). (D) Measurements of the
large channel. (E) Measurements of the sieving gate and the molecular sizes of propylene and propane (the carbon atoms of both are highlighted in black).

higher than reported C3H6/C3H8 sieving mate- bution of the tested materials was measured by (2.35 × 10−10 cm2 s−1) (fig. S14 and table S4).

y
rials such as Co-gallate (1.3 mmol g−1) (36), laser diffraction (fig. S16). The calculated C3H6 Such a high diffusion coefficient of ZU-609

KAUST-7 (1.2 mmol g−1) (2), Y-abtc (0.58 mmol g−1)
(48) and common zeolites 5A (0.41 mmol g−1),
4A (0.15 mmol g−1), ZSM-5 (0.25 mmol g−1)
(figs. S10 and S11). The adsorption behavior is
attributed to the embedded large pore space
within the sieving channel and is ideal for the
pressure swing adsorption process (fig. S12).
Propylene diffusion performance
In addition to the thermodynamic equilib-
rium capacity, we also evaluated the C3H6
adsorption kinetics of ZU-609 and the other
reported materials. Time-dependent C3H6 up-
take profiles were assessed (figs. S13 to S15).
Considering the influence of particle size on
gas diffusion behavior, the particle size distri-
diffusivity coefficients of ZU-609 with different
particle size are close, around 1.00 × 10−9 cm2 s−1
(4.8 mm), 1.36 × 10−9 cm2 s−1 (26.8 mm), based
on time-dependent adsorption profiles mea-
sured volumetrically (Fig. 2B, fig. S13, and
table S3). This value is nearly an order of mag-
nitude larger than the C3H6/C3H8 sieving mate-
rials Co-gallate (6.47 × 10−11 cm2 s−1), KAUST-7
(9.49 × 10 −11 cm2 s −1 ) and is several orders
of magnitude higher than robust Zeolite-4A
(1.01 × 10−12 cm2 s−1) (Fig. 2B). The consistent
trend is observed based on the concentration-
swing frequency response method for micro-
pore diffusion measurement. The value of
ZU-609 (2.38 × 10−9 cm2 s−1) is around 10 times
that of Co-gallate (2.54 × 10−10 cm2 s−1), KAUST-7
,
enables the possible implementation of the
adsorption process under near ambient con-
ditions (fig. S12) without the necessity of in-
creasing temperature to enhance diffusion, as
in the case of Zeolite-4A (49).
Molecular-level understanding about
adsorption and diffusion behavior
An in-situ powder x-ray diffraction experiment
was conducted on ZU-609·C3H6 to study its
C3H6 adsorption behavior (figs. S17 and S18).
Each unit cell could accommodate 5.6 C3H6
molecules, which is close to its adsorption
capacity at 298 K (table S5). The adsorbed
C3H6 molecules are mainly distributed around
the anions and bounded by EDS2− anions

Cui et al., Science 383, 179–183 (2024) 12 January 2024 2 of 5

RES EARCH | R E S E A R C H A R T I C L E

Fig. 2. Gas sorption

properties of ZU-609 and
molecular-level understand-
ing of C3H6 adsorption
and diffusion behavior.
(A) Gas sorption isotherms of
propylene (red) and propane
(blue) at 298 K for ZU-609.
(B) Comparison of propylene
diffusivity coefficient of ZU-609
with reported porous materials
using time-dependent adsorp-
tion profiles measured
volumetrically. Crystal sizes
of Zeolite 4A (1.5 mm), Co-
gallate (1.7 mm), KAUST-7
(1.9 mm), ZU-609 (4.8 mm),
and ZU-609 crystal (26.8 mm).
(C) The adsorption configura-
tions of C3H6 in ZU-609
obtained by in situ PXRD

p
experiments. (D) The represent-
ative minimum energy path
(MEP) profile and snapshots
for C3H6 in ZU-609. The energy
relative to the lowest energy
configuration is shown as a

g
function of the position of the
central carbon on the C3H6
guest molecule. (Color code:
C, gray; H, white; N, blue; Cu,

y
pink; O, red; S, yellow; the carbon
of propylene is highlighted in
black.)

through multiple hydrogen-bond interactions
C-H···O (2.37 to 2.70 Å) (Fig. 2C). The C3H6
binding energy (DE) revealed by the dispersion-
corrected density functional theory (DFT-D)
trapped for 18.5 min (24.6 min g−1), affording a
dynamic C3H6 capacity of approximately
1.64 mmol g−1 (Fig. 3, A and B). The corre-
sponding C3H6 productivity of ZU-609 is about
y g
alkane impurities (figs. S26 and S27). The rapid
C3H6 diffusion behavior of ZU-609 enables the
adsorption process to be implemented under
high gas velocity without obvious dynamic

y
calculation is around 47 kJ mol−1 (fig. S19). The 32.2 L kg−1 (99.9%), higher than that of 21.9 L capacity loss (figs. S28 to S30). Even under a

minimum energy path profile using DFT-D cal-
culations indicates the local diffusion energy
barriers as the propylene molecules diffuse
along the channel (Fig. 2D and fig. S20). These
studies provide a molecular-level understand-
ing about confined C3H6 adsorption and dif-
fusion behavior.
Propylene mixture separation performance
and pressure swing adsorption simulation
The C3H6/C3H8 (50/50) mixture separation
performance of ZU-609 was evaluated by break-
through experiments. C3H8 breaks out imme-
diately along with the inert gas He, indicating
the C3H8 sieving effect exhibited by ZU-609
(Fig. 3A and fig. S21). C3H6 is continuously
kg−1 (97.7%) for Co-gallate (36) and 16.6 L kg−1
(93.2%) for KAUST-7 (table S6). The purity of
the eluted C3H6 is calculated to be 99.9% with
a yield of 87.9% on average (Fig. 3C and figs.
S22 and S23), which is higher than those of
other materials (for JNU-3, purity was 99.5%
and recovery was 51%; for KAUST-7, purity was
93.2% and recovery was 83%) (figs. S24 and
S25) (50). Even for complex mixtures mimick-
ing the composition of propane dehydrogena-
tion CH4(2%)/C2H6(5%)/C2H4(5%)/C3H6(44%)/
C3H8(44%)/H2O (2000 ppm), ZU-609 also shows
good separation performance with high olefin
(C3H6+C2H4) purity (99.72%) and recovery
(91.90%) at 283 K, and its separation ability is
least influenced by the coexistence of C1-C3
,
C3H6/C3H8 gas flow rate of 9 NmL min−1, the
C3H6 dynamic capacity of ZU-609 preserves
94.1% of the equilibrium capacity; by contrast,
the corresponding value on KAUST-7 is only
57.4% (Fig. 3D and fig. S31). In addition, the
breakthrough performance of ZU-609 is invar-
iable over the consecutive adsorption-desorption
cycles (Fig. 3E and fig. S32). Even when ZU-609
is presaturated with high water content (500 to
3000 ppm) the C3H6 purity is not influenced
(>99.9%) (figs. S27 and S33).
The pressure swing adsorption (PSA) proc-
ess is used to evaluate the actual capability of
ZU-609 toward C3H6/C3H8 separation, as well
as Co-gallate and KAUST-7 (Fig. 3F and figs.
S34 to S36). ZU-609 is able to recover 89.7%

Cui et al., Science 383, 179–183 (2024) 12 January 2024 3 of 5

RES EARCH | R E S E A R C H A R T I C L E

Fig. 3. Experimental breakthrough results of ZU-609 and simulated pres-
sure swing adsorption process results. (A) Breakthrough curves of ZU-609
for an equimolar C3H6/C3H8 mixture at 298 K and 1 bar. The breakthrough
experiments were carried out in a packed column with 0.75 g sample at a flow
p
g
process with 3 NmL min−1 of N2 at 298 K. (D) Ratios of C3H6 dynamic uptake
(Udynamic)/equilibrium uptake (Uequilibrium) of ZU-609 (n = 3 per group and
are presented as means ± SD) and KAUST-7 under different superficial gas
velocity for C3H6/C3H8 (50/50) mixture. (E) Dynamic uptake of C3H6 in cycling

rate of 3 NmL min−1 [N represents the standard temperature (273 K) and
pressure (1 bar)]. (B) Comparison of C3H6 dynamic uptake of ZU-609 with other
reported C3H6/C3H8 sieving materials. (C) Flow rate curve of the desorbed C3H6
and C3H8 from ZU-609 and the purity of eluted C3H6 during the regeneration
y
tests on ZU-609 for C3H6/C3H8 (50/50) mixture. (F) Schematic model for 2-bed PSA
process. (G) Simulation results of C3H6 recovery for 2-bed PSA process. (H) Simulation
results of C3H6 production efficiency for 2-bed PSA process. (I) Simulation results of
energy consumption for 2-bed PSA process.

C3H6 with purity 99.7% (Fig. 3G). The C3H6 adsorption kinetics and provides opportuni- 23. Y. Yang et al., Nat. Chem. 13, 933–939 (2021).
production efficiency of ZU-609 is up to 2.28 mol ties to address other challenging gas separa- 24. M. Chang et al., AIChE J. 68, e17794 (2022).

kg−1 hour−1, 345 and 316% more than that of

25. Z. Liu, W. Qiu, W. Quan, W. J. Koros, Nat. Mater. 22, 109–116
tion processes. (2023).

y g
Co-gallate and KAUST-7, respectively (Fig. 3H). 26. D. D. Zhou et al., Nat. Mater. 18, 994–998 (2019).
Meanwhile, the corresponding consumed en- RE FERENCES AND NOTES 27. L. Li et al., Science 377, 335–339 (2022).
ergy of ZU-609 is the least, at ~2.35 MJ kg 1. IEA, The future of petrochemicals: towards more sustainable 28. R. B. Lin et al., Nat. Mater. 17, 1128–1133 (2018).

(C3H6)−1 (Fig. 3I). plastics and fertilisers. IEA Publications; http://www.iea.org 29. L. Wang et al., Angew. Chem. Int. Ed. 62, e202218596
(IEA, 2020). (2023).
2. A. Cadiau, K. Adil, P. M. Bhatt, Y. Belmabkhout, M. Eddaoudi, 30. Q. Wang et al., Nat. Commun. 13, 2955 (2022).
Conclusions 31. X. Zhao, Y. Wang, D. S. Li, X. Bu, P. Feng, Adv. Mater. 30,

y
Science 353, 137–140 (2016).
3. J. J. H. B. Sattler, J. Ruiz-Martinez, E. Santillan-Jimenez, e1705189 (2018).
Progress has been made in the separation per-

formance of molecular sieves in challenging
gas separations, as demonstrated by our de-
signed sulfonate microporous material ZU-609
(fig. S37). The local-limited diffusion path en-
sures the sieving effect and the coexisting
relatively unrestricted diffusion path enables
fast adsorption kinetics. Such a strategy offers
solutions to the contradictory qualities of dif-
fusion, capacity, and selectivity of common
molecular sieves and exhibits attractive olefin/
paraffin separation performance in industrial
operation processes, which highlights the im-
portance of requisite dual control of pore size
and pore shape in molecular sieve designs (figs.
S38 to S40). Our work offers proof of the
high efficiency of molecular sieves with fast
B. M. Weckhuysen, Chem. Rev. 114, 10613–10653 (2014).
4. W. Wang et al., Science 381, 886–890 (2023).
5. L. Li et al., Science 362, 443–446 (2018).
6. K. J. Chen et al., Science 366, 241–246 (2019).
7. A. Tullo, C&EN 99, i9 (2021).
8. D. S. Sholl, R. P. Lively, Nature 532, 435–437 (2016).
9. G. Liu et al., Nat. Mater. 17, 283–289 (2018).
10. X. Han, S. Yang, M. Schröder, Nat. Rev. Chem. 3, 108–118 (2019).
11. C. Gu et al., Science 363, 387–391 (2019).
12. P. G. Boyd et al., Nature 576, 253–256 (2019).
13. P. Q. Liao, N. Y. Huang, W. X. Zhang, J. P. Zhang, X. M. Chen,
Science 356, 1193–1196 (2017).
14. X. Cui et al., Science 353, 141–144 (2016).
15. S. J. Datta et al., Science 350, 302–306 (2015).
16. E. D. Bloch et al., Science 335, 1606–1610 (2012).
17. Y. Chai et al., Science 368, 1002–1006 (2020).
18. P. J. Bereciartua et al., Science 358, 1068–1071 (2017).
19. Q. Liu et al., Angew. Chem. Int. Ed. 62, e202218854 (2023).
20. M. Yang et al., Nat. Commun. 13, 4792 (2022).
21. Y. Su et al., Nature 611, 289–294 (2022).
22. L. Yang et al., Chem. Soc. Rev. 49, 5359–5406 (2020).
,
32. C. Zhang et al., J. Phys. Chem. Lett. 3, 2130–2134 (2012).
33. Q. Ding, S. Zhang, Energy Fuels 36, 7337–7361 (2022).
34. Y. Xiao et al., J. Am. Chem. Soc. 145, 10980–10986
(2023).
35. D. Liu et al., Angew. Chem. Int. Ed. 62, e202218590 (2023).
36. B. Liang et al., J. Am. Chem. Soc. 142, 17795–17801
(2020).
37. P. Nugent et al., Nature 495, 80–84 (2013).
38. B. Li et al., Adv. Mater. 29, 1704210 (2017).
39. Z. Chen et al., Science 368, 297–303 (2020).
40. T. He et al., Nat. Mater. 21, 689–695 (2022).
41. X. Han et al., Nat. Mater. 17, 691–696 (2018).
42. S. H. Lo et al., Nat. Chem. 12, 90–97 (2020).
43. W. Xu et al., Nat. Rev. Mater. 5, 764–779 (2020).
44. K. Jayaramulu et al., Chem. Rev. 122, 17241–17338 (2022).
45. R. Freund et al., Angew. Chem. Int. Ed. 60, 23946–23974
(2021).
46. G. Y. Zhang, G. F. Wei, Z. P. Liu, R. J. Oliver, H. H. Fei,
Chem. Mater. 28, 6276–6281 (2016).
47. J. Wang et al., J. Am. Chem. Soc. 142, 9744–9751 (2020).
48. H. Wang et al., Adv. Mater. 30, e1805088 (2018).

Cui et al., Science 383, 179–183 (2024) 12 January 2024 4 of 5

RES EARCH | R E S E A R C H A R T I C L E

49. C. A. Grande, F. Poplow, A. E. Rodrigues, Sep. Sci. Technol. 45,
1252–1259 (2010).
50. H. Zeng et al., Nature 595, 542–548 (2021).
ACKN OW LEDG MEN TS
Funding: This work was supported by the National Natural Science
Foundation of China (22122811, 22227812, and 22108240)
and Zhejiang Provincial Natural Science Foundation of China
(LR20B060001). Author contributions: H.X. and L.Y. initiated
and supervised the research. J.C., Z.Z., and A.J. performed the
adsorbents preparation. J.C., Z.Z., Y.S., J.W., Y.Z., B.M., Y.Z.,
and J.W. performed the characterization. Z.Y., J.C., L.Y., and
J.H. performed the simulated calculations. J.C., L.Y., H.X., and X.C.
analyzed the data and wrote the paper. Competing interests: SUPPLEMENTARY MATERIALS
The authors declare no competing financial interests. The authors science.org/doi/10.1126/science.abn8418
and their affiliated institutions have a patent pending related to Materials and Methods
the results presented here. Data and materials availability: All Supplementary Text
data are available in the manuscript or the supplementary Figs. S1 to S41
materials. The related CIF of ZU-609, ZU-609 after refinement and Tables S1 to S14
ZU-609 loaded with C3H6 have been uploaded to the Cambridge References (51–55)
crystallographic data center (CCDC 2266913, 2277294, and Data S1 and S2
2277295). License information: Copyright © 2024 the authors,
some rights reserved; exclusive licensee American Association Submitted 23 December 2021; resubmitted 5 July 2023
for the Advancement of Science. No claim to original US Accepted 1 December 2023
government works. https://www.science.org/about/science- Published online 14 December 2023
licenses-journal-article-reuse 10.1126/science.abn8418

p
g
y
y g
y
,

Cui et al., Science 383, 179–183 (2024) 12 January 2024 5 of 5

RES EARCH

ANTHROPOLOGY and archaeobotanists to investigate the entire

valley. The scope of the fieldwork results,
Two thousand years of garden urbanism in the which have led to a better understanding of the
valley’s earthworks, has recently been broad-
Upper Amazon ened by LIDAR survey.
Our fieldwork led to results concerning the
Stéphen Rostain1*, Antoine Dorison2, Geoffroy de Saulieu3, Heiko Prümers4, Jean-Luc Le Pennec5, pre-Hispanic cultural sequence, habitat and
Fernando Mejía Mejía6, Ana Maritza Freire7, Jaime R. Pagán-Jiménez8, Philippe Descola9 diet, and the ancient volcanic activity and mor-
phogenesis of the valley. Archaeological work
A dense system of pre-Hispanic urban centers has been found in the Upano Valley of Amazonian focused on mounds, central plazas, and roads,
Ecuador, in the eastern foothills of the Andes. Fieldwork and light detection and ranging (LIDAR) analysis as well as sites without built features (supple-
have revealed an anthropized landscape with clusters of monumental platforms, plazas, and streets mentary text S1, figs. S1 to S6, and tables S1
following a specific pattern intertwined with extensive agricultural drainages and terraces as well as wide and S2). Large-scale excavations in platforms and
straight roads running over great distances. Archaeological excavations date the occupation from plazas at two major settlements (Sangay and
around 500 BCE to between 300 and 600 CE. The most notable landscape feature is the complex road Kilamope) revealed domestic floors, with post-
system extending over tens of kilometers, connecting the different urban centers, thus creating a holes, caches, pits, hearths, large jars, grinding
regional-scale network. Such extensive early development in the Upper Amazon is comparable to similar stones, and burnt seeds. The construction
Maya urban systems recently highlighted in Mexico and Guatemala. methods consisted of cutting the natural slope
to form a base on which the mound was built

U
(27). Intentional artifact deposits suggest that
rbanism and Amazonia are rarely asso- Located in the Upper Amazon on the eastern the building process was accompanied by

ciated when we evoke pre-Hispanic times.
However, Francisco de Orellana, leading
an expedition down the great river in
1541–42, witnessed large cities along its
banks (1) yet was called a fabulist upon his
return. More than four centuries later, in the
slopes of the Ecuadorian Andes, its dense pop-
ulation extended across the high alluvial ter-
races that border the Upano River (17). The
data of this study stem from more than two
decades of interdisciplinary investigations in
the region, the scope of which was recently
p
ritual activities. The newly established dating
sequence indicates the succession of at least
five cultural ensembles. The original building
of earthen platforms and roads took place be-
tween approximately 500 BCE and 300 to
600 CE and was carried out by groups from

1980s, great archaeological sites with thick
layers of Amazonian dark earth were discov-
ered in the middle Amazon, confirming the ex-
istence of extensive pre-Hispanic settlements
broadened by light detection and ranging
(LIDAR) mapping in a 300-km2 area. This re-
search revealed the largest urban network
of erected and excavated features known in
g
the Kilamope and later Upano cultures. Some
mounds were then reoccupied, after a hiatus,
by groups of the Huapula culture between 800
and 1200 CE (supplementary text S2 and fig.

along the river (2). Orellana did not lie.
Recently, various monumental archaeologi-
cal sites were brought to light in Amazonia
(3, 4). Some of them display earthen platforms
of various sizes with a variety of features, in-
cluding causeways, mounds, canals, and/or forti-
fications [as in the Barinas llanos in Venezuela
(5), the Llanos de Mojos in Bolivia (6), and the
Upper Xingu and central Amazon in Brazil
(7–9)] or roads [as in southwestern Amazonia
Amazonia, whose beginnings date back to
2500 years ago (18). This discovery raises many
questions, among them the following: What
types of features were built by the pre-Hispanic
inhabitants? Were the settlements contempo-
rary and connected to each other? Where did
the inhabitants cultivate the large quantity of
plants needed for their subsistence?
Context
y
S7). All the evidence indicates a local cultural
evolution process. There is no a priori reason
to think that the pre-Hispanic inhabitants origi-
nated from the Andes and not from the Amazon.
The Kilamope and Upano people were sed-
entary agrarian societies that densely occu-
pied the valley, where even today, according to
local farmers, fertile volcanic soils still allow
up to three harvests per year. Analyses of starch
grains from potteries revealed the consump-

(10, 11)]. Pre-Hispanic inhabitants of the Amazon
were indeed remarkable land builders who
intensively reworked their environment, thus
changing the morphology of their territories
(12–14) and its vegetation cover (15, 16).
Here we present an extensive case of a pre-
Stretching along the foothills of the Andes in
southern Ecuador, enclosed between the Andes
to the west and the Cutucú range to the east,
the Upano Valley is a region where Amazonian
and Andean ecosystems meet, with an impor-
tant seismic risk—the earthquake of 1995 reached
y g
tion of maize (Zea mays), beans (Phaseolus sp.),
manioc (Manihot esculenta), and sweet potato
(Ipomoea batatas). The microtraces on a maize
starch grain are identical to those left today by
the manufacture of the traditional “chewed”
chicha (sweet beer), suggesting that this bev-

y
Hispanic urban system in the Amazon region. a moment magnitude of 7.0 (19). Towering erage may have been served in the Upano

1
Archaeology of the Americas (UMR8096), French National
Center for Scientific Research (CNRS), Paris, France.
2
Archaeology of the Americas (UMR8096), Paris-1 University,
Paris, France. 3Local Heritage, Environment and Globalization
(UMR 208), French National Research Institute for
Sustainable Development, Paris, France. 4Kommission für
Archäologie Außereuropäischer Kulturen, Bonn, Germany.
5
Geo-Ocean, Brest University, CNRS, Ifremer, Institut
Universitaire Européen de la Mer (UMR6538), French
National Research Institute for Sustainable Development,
Plouzané, France. 6Anthropology and Archaeology
Department, Pontifical Catholic University of Ecuador, Quito,
Ecuador. 7Museum of the Banco Central of Ecuador,
Guayaquil, Ecuador. 8Center for Social Research, University
of Puerto Rico, San Juan, Puerto Rico. 9Social Anthropology
Laboratory, Collège de France, Paris Sciences et Lettres,
Paris, France.
*Corresponding author. Email: stephen.rostain@cnrs.fr
above it, the active Sangay stratovolcano, a
steep-sided, cone-shaped, snow-capped moun-
tain, peaks at 5230 m above sea level. After
descending the Andean slopes, the Upano River
goes straight south along the sub-Andean
Range (20).
Groups of earthen platforms have been ar-
chaeologically explored: They form large settle-
ments extending over the 70- to 100-m-high
alluvial terraces along the river (Fig. 1). Few
sites have been excavated (21–24). One of the
largest settlements, Sangay, was discovered
in the late 1970s and excavated by three suc-
cessive teams (17, 25, 26). Started in the mid-
1990s, the Upano interdisciplinary project
brought together archaeologists, geoscientists,
,
vessel (28). Upano pottery is well made and
comes in many decorated types, the most com-
mon being the red-banded incised type with
straight or curved painted and incised lines (29).
Upano bowls and jars were exported up to the
Andes, near the modern town of Cuenca (30).
Sangay volcano’s intense activity had a def-
inite impact on pre-Hispanic communities
(31). Previous studies demonstrated contin-
uous cone growth and two major flank failures
during the Pleistocene (32). The resulting de-
bris avalanches (33) spread as far as 60 km from
the volcano (34) and left thick hummocky de-
posits, on which human settlements devel-
oped in late Holocene times. Analyses also
indicate that large explosive eruptions occurred

Rostain et al., Science 383, 183–189 (2024) 12 January 2024 1 of 7

RES EARCH | R E S E A R C H A R T I C L E
Fig. 1. Map of archaeological Upano Valley. (Left) Map of the Middle Upano with five major settlements and 10 secondary sites. m a.s.l., meters above sea level.
(Right) Map of pre-Hispanic dug roads.
in the nearer past (32). The hypothesis that the
Upano culture came to an abrupt end after a
massive eruption ~400 to 600 CE has been
raised (18) but has been called into question by
the disparity of radiocarbon dates obtained
recently for these levels.
LIDAR results
In 2015, the Ecuadorian National Institute for
Cultural Heritage commissioned the LIDAR
survey of a 600-km2 area stretching from the
Upper Upano to the Pastaza River to highlight
anthropogenic features hidden under the can-
opy (35, 36) (Fig. 1). After a 1-m-resolution
digital elevation model had been derived from
the point cloud, the Upano project team studied
the 300-km2 area constituting the southern part
of this landscape (supplementary text S3, figs.
S8 to S11, and table S3). The number and size
of anthropogenic features led us to target only
specific scientific goals. Given that previous
fieldwork had already provided information on
the internal organization of residential sites
(37), we focused on the “gaps” between settle-
ments and the road network rather than on
Rostain et al., Science 383, 183–189 (2024) 12 January 2024
intrasite features. This objective was also mo-
tivated by the identification in the field of
various road starts (38). We set up a remote
sensing methodology based on multiple LIDAR
visualizations (39, 40) (Fig. 2). Pre-Hispanic
features were distinguished from modern ele-
ments according to their orientation and spa-
tial coherence.
The analysis revealed an elaborate anthro-
pogenic landscape. Settlements with standard-
ized archaeological features evidencing a
shared cultural background are intercon-
nected by short-distance (intrasite) and long-
distance (intersite) roads, which strongly
suggests contemporaneity.
Platforms and plazas
The most common features are earthen plat-
forms. Going far beyond the field identification
of just some individual platforms, the LIDAR
data enabled the detection of >6000 platforms
within the 300-km2 area. The standard shape
is rectangular (although a few were circular)
and about 20 m by 10 m with a preserved ele-
vation of 2 to 3 m (Fig. 3).
p
g
y
Rarely isolated, the platforms typically occur
in groups—or complexes—of three to six units
y g
around a plaza, often with a central platform.
The most common complexes measure 40 m
by 40 m (~1600 m2) and are interpreted as
1residential (37). However, LIDAR imagery
also highlighted monumental complexes prob-
ably bearing a civic-ceremonial function, with
y
much larger platforms, but of similar height to
,
the smaller ones. The largest complex, at Kila-
mope, covers 10 ha and includes a 140 m by
40 m monumental platform (4.5 m high).
This pattern is repeated almost evenly
throughout the area. Average density is
16.6 platforms/km2, but some agglomerated
areas have densities of >100 features/km2.
Because of their ubiquity and close relation-
ship with the road network, we consider the
complexes to be the elementary built features
of the pre-Hispanic landscape.
Settlements
The distribution of elements in the study
area reflects a settlement pattern in which the
buffer zones between residential architectures
2 of 7
RES EARCH | R E S E A R C H A R T I C L E

p
g
y
y g
y
,

Fig. 2. Kilamope site. Anthropogenic features in the center of the Kilamope site, including residential platforms, dug footpaths, and agricultural structures. The
four images on the right side of the figure illustrate different LIDAR visualizations used to interpret the digital elevation model in the same area (dotted rectangle) in
order to highlight the drained-fields pattern. From top to bottom: hillshade from multiple directions, simple local relief models with 10- and 50-pixel radii overlaid,
slopes reclassified according to geomorphological models (40), and color classification of the elevation.

played an important role. Moreover, cluster- ceremonial platforms, and connections with centers) and 10 secondary sites (Fig. 1). Among
ing trends occur. Clusters of complexes have other complexes. the first group, Sangay stands out for its higher
been identified as settlements on the basis of Fifteen settlements were classified into two density (>125 platforms/km2) and its ostenta-
three criteria: feature density, size of the civic- categories: five major sites (large and/or dense tious core. Accessed by a 2.5-km-long straight

Rostain et al., Science 383, 183–189 (2024) 12 January 2024 3 of 7

RES EARCH | R E S E A R C H A R T I C L E
Fig. 3. Earth platforms complex. Earth platforms complex of Nijiamanch on the right bank of the Middle
Upano. Drained fields are visible around the artificial mounds on the LIDAR image (bottom).
road, it lies at the highest point of the cliff
that borders the Upano’s northern bank, thus
dominating the valley. Junguna and Kunguints
seem to be chiefly residential sites with many
small complexes adjoining intrasite circulation
axes, whereas Kilamope and Copueno are char-
acterized by large civic-ceremonial complexes.
Despite the apparent architectural and spa-
tial homogeneity among these sites, several
elements suggest that the settlements were
exposed to threats. These include peripheral
ditches blocking access to some settlements
(e.g., east of Sangay) and obstructed roads near
some large complexes (e.g., Copueno). We in-
terpreted these elements as the result of ten-
sions between groups or reinforcement of the
sites against external threats.
Rostain et al., Science 383, 183–189 (2024) 12 January 2024
Modified hills
Numerous “truncated” hills—natural reliefs
with flat summits—were detected. While their
general morphology is probably of volcanic
origin (hummocks), a plausible hypothesis,
supported by the accesses built on their slopes,
is that their flattening is artificial, so that they
constitute an integral component of the pre-
Hispanic landscape (41).
Roads and pathways
Perhaps the most notable elements of the
landscape are the intrasite pathways and a
regional-scale intersite road network (Fig. 4).
Roads are considered as such because they
systematically link complexes (thus distinguish-
ing them from other features, such as canals,
described below). Some of these complexes,
unusually isolated in the landscape, adjoin major
roads as if they were layover points along the way.
Four types of roads or pathways have been
identified. Although we cannot totally dis-
count the possibility that a few paths could
be the result of usage (sunken lanes) (42, 43),
the morphology of the features detected (e.g.,
rectilinearity, depth, curbs) strongly suggests
that these are excavated roads. Therefore, the
generic term “dug roads” is used here.
The most widespread are the straight-dug
footpaths and roads (thus classified on the
basis of their size and length) whose morphol-
ogy is documented by fieldwork (27). Essen-
tially straight and about 2 to 3 meters deep on
average, they probably result from digging
and accumulating earth on either side of the
path, creating a U-shaped profile bordered
by curbs. The width is 4 to 6 m for the smallest
paths and up to >15 m for the largest ones,
p
creating a walkable surface that is 2 to 5 m
wide in the middle. Despite discontinuities in
their layout owing to the heterogeneity of the
LIDAR coverage, we can reasonably estimate
that the longest roads—Uyunts-Jurumbuno and
Kilamope-Kunguints—run for more than 14 and
g
25 km, respectively. Both are likely to continue
beyond the boundaries of the study area. We
believe that dug roads were designed to be as
straight as possible despite the natural irregu-
y
larities of the terrain.
The second type includes roads running
along the interfluve in hilly terrain. Occasion-
ally also dug, they follow topographic relief
and connect open spaces (e.g., plateaus, river-
beds) through steep areas, such as the Andean
footslopes or the valley cliffs. The third cate-
gory includes possible elevated causeways with
parallel ditches on either side of the road rem-
iniscent of “causeways-canals” in the Llanos de
y g
Mojos of Bolivia (44). Finally, dug footpaths
regularly end with a descent into one of the
gullies (or quebradas) that spread on the Upano
alluvial terrace and come out of the same gully
a few hundred meters further on. This seems to
constitute a fourth type of path, which takes ad-
y
vantage of the natural layout of the quebradas.
,
In line with this, we argue that, beyond con-
necting spaces, most pathways were intimately
related to surface water management and agri-
cultural practices.
Drained fields and terraces
The LIDAR survey highlighted numerous agrar-
ian features of two main types: drained fields
and terraces. Their strong spatial coherence
with the rest of the detected remains and, con-
versely, their location under forested areas
support the inference that they were an in-
tegral part of the pre-Hispanic anthropogenic
landscape.
Drained fields extend over hundreds of hect-
ares into orthogonal and continuous plot systems
4 of 7
RES EARCH | R E S E A R C H A R T I C L E
Fig. 4. Dug roads. Wide and deep dug roads connecting Upano monumental settlements.
(Fig. 2). The elementary unit is a rectangular
field 10 to 40 m wide and several dozens of
meters long. The field limits are ditches of 4-m
width and 40-cm depth on average. They are
connected to drainage canals, slightly wider
and deeper, and mitigate waterlogging in this
climate with daily rainfall. These canals then
flow into the hydrographic network of the
quebradas, whose course has sometimes been
modified.
Drained fields are linked to the network of
dug footpaths often surrounding them, mak-
ing it sometimes difficult to distinguish be-
tween a road and a canal. In such cases, a clear
connection to a complex was the discriminat-
ing criterion. However, it is very likely that
pathways had a dual function of circulation
and water management. This agricultural tech-
nique resembles those documented for the
Rostain et al., Science 383, 183–189 (2024) 12 January 2024
Llanos de Mojos in Bolivia and is still used
today, for example, by the Karinya of the Orinoco
llanos (45). Less widespread, terraces are
found occasionally along the edges of quebra-
das, along the alluvial terrace of the Upano,
perpendicular to concave hillslopes or on the
lower Andean slopes, where they are associ-
ated with drains parallel to the slope.
Agrarian features fill the “gaps” between com-
plexes and settlements noted during the first
reading of the LIDAR data. Their ubiquity, their
close connection with residential and ceremo-
nial areas, and the variety of geomorphological
contexts exploited demonstrate the importance
of agricultural activity in the settlement pattern.
Discussion and conclusions
The results of fieldwork and LIDAR analysis
demonstrate that the Upano Valley was dense-
ly populated around the beginning of the com-
mon era. The intimate link between residential
and agricultural areas brings to mind the “gar-
den cities” and “green urbanism” described or
theorized by other researchers (13, 46–52). Far
from the utopia that these terms imply, how-
ever, the garden urbanism of the Upano valley
constitutes a concrete, dynamic, and probably
contested landscape and provides further proof
that Amazonia is not the pristine forest once
depicted.
The settlement pattern is composed of dense
sites with standardized domestic groups of
platforms around plazas and monumental civic
architecture connected by streets. Establish-
ments are linked to each other over great dis-
tances by a vast network of roads intertwined
with intensive agricultural layouts. The or-
ganizational and architectural homogeneity,
as well as the consistent interweaving of
monumental-ceremonial features, domestic
p
spaces, and economic areas, strongly suggests
that the whole network was at least partly
contemporary.
Despite variations in the settlement density,
it must be emphasized that few areas of the
valley are devoid of remains. Apparent empty
g
buffer zones between complexes of platforms
were in fact dedicated to agriculture. Two main
patterning strategies have been recognized
and were seemingly induced by the valley’s
y
geomorphology. In the southern part of the
study area, extensive parcels of drained fields
are spread within a low-density distribution of
complexes in the wetlands of the Upano al-
luvial terraces. Meanwhile, the hilly terrain of
the northern zone is more prone to the cluster-
ing of complexes, with agricultural terraces
intertwined with this denser grid in a more
opportunistic manner. However, these two
contrasting “environment-forced” sectors are
y g
all connected by the road network and its
orthogonal layout, independent of all geo-
morphic constraints and constituting per-
haps the most distinctive characteristic of
this built landscape. Straight roads cross at
right or nearly right angles without deviating
y
in front of hills or ravines. If these roads fa-
,
cilitated exchanges, it is very likely that they
also had a marked symbolic and powerful rit-
ual function and participated in the construc-
tion of a cultural landscape. The complexity
and dynamic nature of the latter are further
demonstrated by the presence of defensive
features that raise the question of alliances
and tensions or possibly even episodic wars.
However, it would be imprudent to infer
that these cities were organized around a cen-
tralized authority capable of mobilizing labor
in a more or less coercive manner. The inter-
locking of groups of filiation and segmental
solidarities, regularly reinforced by ceremo-
nial exchanges, are sufficient to ensure the co-
hesion and coordination necessary for the
5 of 7
RES EARCH | R E S E A R C H A R T I C L E
Fig. 5. Comparison of site sizes. Comparison at the same scale of the core areas of major sites of the Upano, ancient Egypt, and ancient Mesoamerica (for
comparison with low-density urbanism sites in Amazonia, see supplementary text S4 and fig. S12).
organization of a highly structured settlement.
Contemporary ethnology shows that exchanges
are not so much based on economic logic as on
the circulation of specialized production within
ethnic confederations, each defining a local
identity (53–55). Indeed, the overall organiza-
tion and standardization of architectural fea-
tures, the road patterns, or the defense systems,
suggests, for instance, the existence of ad-
vanced engineering (mature architectural
tradition relying on sighting and surveying
instruments and/or expertise).
The Upano sites are quite different from
other monumental sites of Amazonia, which
are all more recent, considerably less dense in
terms of features, and, until proven otherwise,
not embedded in such a vast and dense com-
munication network (fig. S12) (13, 14). This
original 2000-year-old society of the Upano
valley constitutes the earliest and largest low-
density agrarian urbanism ever documented
in the Amazon so far. At a supraregional level,
beyond exchanges with the Cuenca area, the
relations and potential bilateral influence
with the contemporary Andean world, such
as Chavín de Huántar’s sphere of influence
(56, 57), remain difficult to grasp (58, 59). Yet,
at present, there is no reason to believe that
this is not an endogenous development.
The Upano’s LIDAR coverage confirms the
exceptional archaeological potential of Amazonia,
recently recalled by extrapolative modeling in
Brazil (60). However, whereas the latter article
focuses on isolated earthworks (i.e., monumen-
tal ditches), the Upano materialized another
Rostain et al., Science 383, 183–189 (2024) 12 January 2024
scale of landscape anthropization, where
urbanism covers hundreds of square kilo-
meters. It echoes more clearly the Upper
Xingu, where road networks suggest a low-
density pre-Hispanic urbanism (52). This sug-
gests that, being denser and less accessible,
northern Amazonia may have been under-
estimated in terms of archaeological potential
and promises numerous discoveries with fu-
ture exploratory work and LIDAR coverage.
Broadening perspectives, considering the region-
scale density of features, the population den-
sity they must have supported, or the highly
anthropized landscape—and although the size
of the valley does not rival that of the Yucatán
biosphere reserves—the human investment in
the Upano Valley is comparable to that of con-
temporary Central Maya Lowlands (61–65). Going
further, the major ceremonial cores, with mo-
numental platforms, plazas, and causeways, are
comparable in size to those of other great cul-
tures of the past, such as Mexican Teotihuacan
or the Egyptian Giza Plateau (Fig. 5).
What is very noteworthy is that this example
of garden urbanism appears to be associated
with a mound-building tradition. One wonders
whether such architecture might echo other
regions of the Americas and a pre-Hispanic
ideology wherein fertility, origin of humanity,
and ancestors are key aspects. Although the
design of roads and canals is not radial in the
Upano Valley, the grid it establishes is rem-
iniscent of the ceque system of Cuzco (66)
and its connections between kin-based and
segmented territories, ritual places of worship,
p
g
and human-built hydraulic features. Indeed, it
is tempting to see the Upano valley’s geomet-
y
rical layout, which cuts across the topography,
as a cosmological design rather than a com-
mon and practical system of communication.
This strong emphasis on abstract lineal con-
nections between sites is more reminiscent of
the sacred topography of Andean polities as
“analogist” systems than of contemporary
“animist” cultures of the Upper Amazon [in
the sense of Descola (67)]. For lack of ethno-
graphic analogies in settlement patterns in the
y g
contemporary lowlands, one should probably
look for current highlands autochthonous pol-
ities whose sociocosmic structure is embed-
ded in a topographical grid [for instance, the
Chipaya of Bolivia (68)]. Such a discovery is
another vivid example of the underestimation
y
of Amazonia’s twofold heritage: environmental
,
but also cultural, and therefore Indigenous.
Like many others (69–71), we believe that it
is crucial to thoroughly revise our preconcep-
tions of the Amazonian world and, in doing so,
to reinterpret contexts and concepts in the
necessary light of an inclusive and participa-
tory science.
REFERENCES AND NOTES
1. G. de Carvajal, Descubrimiento del río de la Amazonas
(Estudios ediciones y medios, Sevilla, [1542] 1992).
2. U. Lombardo et al., Nat. Commun. 13, 3444 (2022).
3. S. Rostain, Islands in the Rainforest: Landscape
Management in Pre-Columbian Amazonia (Left Coast Press,
2012).
4. S. Rostain, C. Jaimes Betancourt, Eds., Las Siete Maravillas
de la Amazonia precolombina (IV Encuentro Internacional de
6 of 7
RES EARCH | R E S E A R C H A R T I C L E
Arqueología Amazónica, Bonner Altamerika-Sammlung und
Studien, Plural Editores, 2017).
5. E. M. Redmond, C. S. Spencer, Anthropol. Pap. Am. Mus.
Nat. Hist. 2007, 86 (2007).
6. U. Lombardo, H. Prümers, J. Archaeol. Sci. 37, 1875–1885 (2010).
7. M. J. Heckenberger et al., Science 301, 1710–1714 (2003).
8. M. Heckenberger, The Ecology of Power: Culture, Place, and
Personhood in the Southern Amazon, AD 1000-2000
(Routledge, 2005).
9. E. Neves, Sob os tempos do equinócio: Oito mil anos de história
na Amazônia central (Ubu Editora, 2022).
10. J. Iriarte et al., J. Comput. Appl. Archaeol. 3, 151–169 (2020).
11. S. Saunaluoma, J. Moat, F. Pugliese, E. G. Neves,
Lat. Am. Antiq. 32, 1–15 (2020).
12. W. Denevan, Cultivated Landscapes of Native Amazonia and the
Andes (Oxford Univ. Press, 2001).
13. M. J. Heckenberger et al., Science 321, 1214–1217 (2008).
14. H. Prümers, C. J. Betancourt, J. Iriarte, M. Robinson,
M. Schaich, Nature 606, 325–328 (2022).
15. C. R. Clement et al., Proc. R. Soc. London Ser. B 282, 20150813
(2015).
16. W. Balée, T. Swanson, M. G. Zurita-Benavides,
J. C. Ruiz Macedo, Lat. Am. Antiq. 34, 1–15 (2023).
17. P. Porras, Natl. Geogr. Res. 5, 374–381 (1989).
18. S. Rostain, Bull. Inst. Fr. Études Andines 28, 53–89 (1999).
19. D. Legrand et al., in La cuenca oriente: Geología y petróleo,
P. Baby, M. Rivadeneira, R. Barragán, Eds. (IFEA, IRD,
Petroecuador, 2004), pp. 168–182.
20. S. Bès de Berc et al., in La cuenca oriente: Geología y petróleo,
P. Baby, M. Rivadeneira, R. Barragán, Eds. (IFEA, IRD,
Petroecuador, 2004), pp. 153–167.
21. L. Rampón, in Cuadernos de Investigaciones Científicas
No. 1 (Arqueología, Misiones Católicas de la Amazonía, 1959).
22. P. Porras, Natl. Geogr. Res. 5, 374–381 (1989).
23. P. Moncayo Echeverría, Revista Sarance 20, 147–154 (1994).
24. A. Rostoker, thesis, City University of New York (2005).
25. P. Porras, Investigaciones arqueológicas a las faldas del Sangay,
Tradición Upano (Centro de Investigaciones Arqueológicas,
Universidad Católica del Ecuador, 1987).
26. E. Salazar, in Handbook of South American Archaeology,
H. Silverman, W. Isbell, Eds. (Springer, Kluwer, Plenum, 2008),
pp. 263–278.
27. S. Rostain, STRATA 1, e2 (2023).
28. J. R. Pagán Jiménez, S. Rostain, “Uso de plantas
económicas y rituales (médicinales o energizantes)
en dos comunidades precolombinas de la Alta Amazonía
ecuatoriana: Sangay (Huapula) y Colina Moravia (400 a.C.-
1200 d.C.)” in Antes de Orellana. Actas del 3er Encuentro
Internacional de Arqueología Amazónica, S. Rostain, Ed.,
IFEA, FLACSO, Embajada de los EEUU, Quito (2014),
pp. 313–322.
29. S. Rostain, in Cerâmicas Arqueológicas da Amazônia: Rumo a
uma nova síntese, C. Barreto, H. Lima, C. Jaimes Betancourt,
Eds. (MPEG, Belém-do-Pará, 2016), pp. 573–585.
30. K. Bruhns, J. H. Burton, A. Rostoker, in En Tecnología y
organización de la producción de cerámica prehispánica en los
Andes, I. Shimada, Ed. (PUCP, Fondo Editorial, 1994),
pp. 53–66.
31. C. Von Hillebrandt, in El paisaje volcánico de la sierra
ecuatoriana, P. Mothes, Ed., vol. 4 of Estudios de Geografía
(Corporación Editora Nacional, Colegio de Geógrafos del
Ecuador, 1991), pp. 39–53.
32. M. Monzier et al., J. Volcanol. Geotherm. Res. 90, 49–79 (1999).
33. L. Siebert, J. Volcanol. Geotherm. Res. 22, 163–197 (1984).
34. V. Valverde, P. A. Mothes, B. Beate, J. Bernard, J. Volcanol.
Geotherm. Res. 411, 107172 (2021).
Rostain et al., Science 383, 183–189 (2024) 12 January 2024
35. Technoproject, “Informe del levantamiento aéreo y del análisis
preliminar e identificación de las anomalías del relieve
relacionadas con las estructuras arqueológicas. Superficie
Upano–Mera,” field report, INPC, Quito (2015).
36. A. Sánchez-Polo, R. Á. Litben, STRATA 1, e3 (2023).
37. S. Rostain, in Communities in Contact. Essays in Archaeology,
Ethnohistory and Ethnography of the Amerindian Circum-
Caribbean, C. L. Hofman, A. van Duijvenbode, Eds.
(Sidestone Press, 2011), pp. 455–475.
38. S. Rostain, Quat. Int. 249, 31–42 (2012b).
39. Z. Kokalj, R. Hesse, Airborne Laser Scanning Raster Data
Visualization: A Guide To Good Practice, vol. 14 of Prostor,
kraj, čas (Založba ZRC, 2017).
40. A. Dorison, C. Siebe Grabach, M. Elliott, G. Pereira,
Boletín de la Sociedad de Geología Mexicana 74, A180622 (2022).
41. G. de Saulieu, S. Rostain, J.-L. Le Pennec, “El Formativo del
Alto Pastaza (Ecuador), entre arqueología y vulcanología” in
Antes de Orellana. Actas del 3er Encuentro Internacional de
Arqueología Amazónica, S. Rostain, Ed., IFEA, FLACSO,
Embajada de los EEUU, Quito (2014), pp. 199–205.
42. M. Cardale de Schrimpff, Caminos prehispánicos en Calima
(Fundación de Investigaciones Arqueológicas Nacionales,
Santafé de Bogotá, 1996).
43. N. Nir et al., Catena 215, 106312 (2022).
44. C. L. Erickson, in Landscapes of Movement: Trails, Paths,
and Roads in Anthropological Perspective, J. E. Snead,
C. L. Erickson, J. A. Darling, Eds. (Univ. of Pennsylvania Press,
2011), pp. 204–231.
45. W. M. Denevan, K. H. Schwerin, Antropológica 50, 3–91 (1978).
46. E. Howard, Garden Cities of Tomorrow (2007).
47. C. M. Arensberg, in Urban Life: Readings in Urban Anthropology,
G. Gmelch, W. P. Zenner, Eds. (St. Martin’s Press, 1980),
pp. 37–47.
48. W. J. Eggeling, J. Ecol. 34, 20–87 (1947).
49. E. Graham, Archeol. Pap. Am. Anthropol. Assoc. 9, 185–194 (1999).
50. C. Isendahl, M. E. Smith, Cities 31, 132–143 (2013).
51. D. Marken, M. C. Arnauld, Eds., Building an Archaeology of
Maya Urbanism: Planning and Flexibility in the American Tropics
(Univ. Press of Colorado, 2023).
52. M. Heckenberger, in Landscapes of Preindustrial Urbanism,
G. Farhat, Ed. (Dumbarton Oaks Research Library and
Collection, 2020), pp. 225–262.
53. J. Jackson, The Fish People: Linguistic Exogamy and Tukanoan
Identity in the Northwest Amazon (Cambridge Univ. Press,
1983).
54. P. Descola, In the Society of Nature (Cambridge University
Press, 1994).
55. B. Franchetto, M. Heckenberger, Eds., Os povos do Alto
Xingu - história e cultura (UFRJ, 2001).
56. R. L. Burger, Chavin and the Origins of Andean Civilization
(Thames & Hudson, 1992).
57. R. L. Burger, in Handbook of South American Archaeology,
H. Silverman, W. H. Isbell, Eds. (Springer, 2008), pp. 681–703.
58. A. J. Pearce, D. G. Beresford-Jones, P. Heggarty, Eds.,
Rethinking the Andes-Amazonia Divide: A Cross-Disciplinary
Exploration (UCL Press, 2020).
59. R. Clasby, J. Nesbitt, Eds., The Archaeology of the Upper
Amazon: Complexity and Interaction in the Andean Tropical
Forest (Univ. Press of Florida, 2021).
60. V. Peripato et al., Science 382, 103–109 (2023).
61. R. Fletcher, Insights 2, 1–19 (2009).
62. A. F. Chase, D. Z. Chase, Annu. Rev. Anthropol. 45, 361–376
(2016).
63. E. Lemonnier, in Settlement Ecology of the Ancient Americas,
L. C. Kellett, E. E. Jones, Eds. (Routledge, 2017),
pp. 167–194.
64. M. A. Canuto et al., Science 361, eaau0137 (2018).
65. T. Beach et al., Proc. Natl. Acad. Sci. U.S.A. 116, 21469–21477
(2019).
66. T. Zuidema, The Ceque System of Cuzco: The Social
Organization of the Capital of the Inca (Brill, 1964).
67. P. Descola, Beyond Nature and Culture (Univ. of Chicago Press,
2013).
68. N. Wachtel, Le retour des ancêtres. Les Indiens Urus de Bolivie,
XX°-XVI° siècle. Essai d’histoire régressive, Bibliothèque des
sciences humaines (Éditions Gallimard, 1990).
69. C. Gnecco, P. Ayala, Eds., Indigenous Peoples and Archaeology
in Latin America (Left Coast Press, 2011).
70. B. C. da Rocha, thesis, Institute of Archaeology, University
College London (2017).
71. P. Erikson, V. Vapnarsky, Eds., Living Ruins: Native
Engagements with Past Materialities in Contemporary
Mesoamerica, Amazonia, and the Andes (Univ. Press of
Colorado, 2022).
AC KNOWLED GME NTS
LIDAR images were provided thanks to an agreement elaborated
by J. F. Moscoso Novillo (director of the Ecuadorian National
Institute for Cultural Heritage). The following persons helped
during the fieldwork conducted by S.R.: I. Cangas, P. Baby,
R. Bartone, S. Bes de Berc, J. Burgos, T. Bray, J. Burton, A.M.F.,
F. Fuentes, Y. Graber, J.-L.L.P., K. Olsen Bruhns, L. Pintér,
p
P. Podwojewski, H.P., A. Rostoker, G.d.S., L. Zlonay, and several
Ecuadorian and French students. Local people also contributed,
especially J. Vega, Juan Vele, Julio Vele, and several workers.
A. Gómez de la Peña, K. Leonard, and J.R.P.-J. did the
archaeobotanical analyses. Thanks to D. McKey and F. Savatier for
their reading and corrections and to the reviewer for their
suggestions. Funding: This work was supported by the French
National Center for Scientific Research (CNRS), Laboratory
g
“Archaeology of the Americas” (UMR8096), French Ministry for
Europe and Foreign Affairs (MAEE), French National Research
Institute for Sustainable Development (IRD), German Kommission
für Archäologie Außereuropäischer Kulturen (KAAK), Ecuadorian
National Institute for Cultural Heritage (INPC), Banco Central del
Ecuador, and National Secretary of Higher Education, Science,
y
Technology and Innovation (SENESCyT). Author contributions:
Conceptualization: S.R. and A.D. Methodology: S.R., A.D., H.P., and
J.-L.L.P. Investigation: S.R., A.D., G.d.S., H.P., J.-L.L.P., F.M.M.,
A.M.F., J.R.P.-J., and P.D. Visualization: S.R., A.D., and H.P. Funding
acquisition: S.R., G.d.S., H.P., and J.-L.L.P. Project administration:
S.R. Supervision: S.R. Writing – original draft: S.R., A.D., G.d.S.,
H.P., J.-L.L.P., and P.D. Writing – review & editing: S.R., A.D.,
G.d.S., H.P., J.-L.L.P., and P.D. Competing interests: The authors
declare that they have no competing interests. Data and
materials availability: All data are available in the main text or the
supplementary materials. License information: Copyright © 2024
the authors, some rights reserved; exclusive licensee American
Association for the Advancement of Science. No claim to original
y g
US government works. https://www.science.org/about/science-
licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.adi6317
Materials and Methods
y
Figs. S1 to S13
Tables S1 to S3
,
References
Submitted 10 May 2023; accepted 15 November 2023
10.1126/science.adi6317
7 of 7
RES EARCH

CANCER IMMUNOLOGY mately, we tested the proof-of-principle of

pSAP targeting to DCs as a possible mode of
Hyperglycosylation of prosaposin in tumor dendritic cancer immunotherapy.

cells drives immune escape Results

Saposins promote cross-presentation of
membrane-associated tumor antigen
Pankaj Sharma1†, Xiaolong Zhang1†, Kevin Ly1†, Ji Hyung Kim1, Qi Wan1, Jessica Kim1, Mumeng Lou1,
Lisa Kain2, Luc Teyton2, Florian Winau1* First, we investigated the effect of saposins on
the integrity of apoptotic bodies derived from
Tumors develop strategies to evade immunity by suppressing antigen presentation. In this work, we tumor cells. For this purpose, we exposed mu-
show that prosaposin (pSAP) drives CD8 T cell–mediated tumor immunity and that its rine MCA fibrosarcoma cells to g irradiation
hyperglycosylation in tumor dendritic cells (DCs) leads to cancer immune escape. We found that [100 Gy (unit of absorbed dose of ionizing
lysosomal pSAP and its single-saposin cognates mediated disintegration of tumor cell–derived apoptotic radiation); 1 Gy = 100 rads] to trigger apoptotic
bodies to facilitate presentation of membrane-associated antigen and T cell activation. In the tumor cell death (Fig. 1A). Successful induction of
microenvironment, transforming growth factor–b (TGF-b) induced hyperglycosylation of pSAP and its apoptosis was controlled by measuring PS
subsequent secretion, which ultimately caused depletion of lysosomal saposins. pSAP hyperglycosylation exposure with AnnexinV staining. Subsequent-
was also observed in tumor-associated DCs from melanoma patients, and reconstitution with pSAP ly, we purified apoptotic vesicles from cell
rescued activation of tumor-infiltrating T cells. Targeting DCs with recombinant pSAP triggered tumor culture supernatants using differential ultra-
protection and enhanced immune checkpoint therapy. Our studies demonstrate a critical function of centrifugation (100,000g pellets) and visual-
pSAP in tumor immunity and may support its role in immunotherapy. ized them by transmission electron microscopy
(Fig. 1B). We then loaded the fluorescent dye

A
ntigen-specific T cell responses are cen-
tral to protection against cancer. Cyto-
toxic CD8 T lymphocytes recognize tumor
antigens presented by major histo-
compatibility complex I (MHC-I) molecules
tumors often develop mechanisms to evade
immune responses; for example, by producing
immunosuppressive cytokines such as trans-
forming growth factor–b (TGF-b) (14, 15). Thus,
the goal of cancer therapy is to target and
p
calcein into those apoptotic vesicles using a
liposome extruder with a 100-nm pore size.
After incubation with different recombinant
saposins, we measured calcein release and
found that saposins disintegrate tumor cell–
derived apoptotic vesicles when compared with

and subsequently deploy their effector func-
tions, such as target-cell killing and produc-
tion of inflammatory cytokines (1). However,
tumor cells often fail to directly activate T cells
overcome immune evasion to restore protec-
tive immunity.
Prosaposin (pSAP) is a precursor protein
that is transported from the Golgi apparatus
g
control bovine serum albumin (BSA) (Fig. 1B).
In addition to leakage of the small molecule
calcein, we also tested whether saposins facil-
itate vesicular release of larger proteins and

because of down-regulation of their MHC-I
pathway (2, 3). Therefore, other antigen-
presenting cells, such as dendritic cells (DCs),
are critical to engulf tumor antigens for sub-
sequent processing and display to MHC-I–
restricted CD8 T cells in a process called
cross-presentation (4). On a cellular level, this
mechanism can be broadly divided into a cyto-
solic and vacuolar pathway (5, 6). According to
cytosolic processing, endosomal antigens are
to the lysosome, assisted by its chaperone,
sortilin (16, 17). In the lysosome, cathepsins
cleave pSAP into the single saposins A to D.
Saposins are also called sphingolipid activa-
tor proteins because they function as small,
nonenzymatic cofactors for lysosomal hydro-
lases that are required for sphingolipid degra-
dation (18). Moreover, at the low acidic pH of
the endolysosomal compartment, saposins are
able to interact with anionic phospholipids,
y
incubated liposomes enclosing fluorescein iso-
thiocyanate (FITC)–labeled ovalbumin (OVA-
FITC) with single saposins. In contrast to
incubation with BSA, saposins triggered the
release of soluble OVA protein, with various sapo-
sins exhibiting different membrane-perturbing
potency (fig. S1A).
We next explored the impact of saposins on
the processing of apoptotic bodies in DCs. To
this end, we first assessed the subcellular lo-

retrotranslocated into the cytosol for degrada-
tion by proteasomes and subsequent reimport
into the endosome for MHC-I loading. By con-
trast, in the vacuolar pathway, the endosome
is more autonomous and relies on its pro-
teases for antigen processing (7–9). Among
such as phosphatidylserine (PS), exposed on
intralysosomal vesicles (19). These membrane-
perturbing properties facilitate vesicle dis-
integration and pertain to apoptotic vesicles
that characteristically contain PS in their lipid
bilayers (20, 21). In this context, tumors, owing
y g
calization of pSAP in the endosomal or lyso-
somal compartments using DC2.4 cells (murine
DC line) transfected with a pSAP expression
vector containing a C-terminal FLAG tag. We
found maximum colocalization of pSAP with
LAMP-1 protein, indicating its predominant

y
the different DC subsets, classical DC1s are to their uncontrolled growth kinetics, pro- distribution in lysosomes (fig. S1B). Next, we

especially efficient in cross-presentation and
fulfill this function in tumor-draining lymph
nodes for T cell priming as well as in the tumor
microenvironment to activate tumor-infiltrating
T lymphocytes (10, 11). Abundance and proper
function of immune cells, including antigen-
presenting cells and lymphocytes, at the tumor
site are vital for effective immunity and con-
trol of cancer growth (12, 13). Unfortunately,
1 antigen and activation of CD8 T cell responses
Program in Cellular and Molecular Medicine, Boston
Children's Hospital, Department of Pediatrics, Harvard
Medical School, Boston, MA 02115, USA. 2Department of
Immunology and Microbiology, Scripps Research Institute,
La Jolla, CA 92037, USA.
*Corresponding author. Email: florian.winau@childrens.harvard.edu
†These authors contributed equally to this work.
duce a substantial amount of dying cells and
apoptotic bodies that contain tumor antigens
to potentially trigger the immune system. Not-
ably, membrane-associated particulate anti-
gen is more immunogenic than soluble protein,
and thus, antigen presentation pathways based
on vesicular processing might be central to the
induction of protective T cell immunity. In this
study, we explored the impact of saposins on
presentation of membrane-associated tumor
that protect against cancer growth. This work
also describes a mechanism by which the
tumor counteracts saposin-mediated process-
ing by triggering pSAP hyperglycosylation and
secretion from tumor-associated DCs. Ulti-
,
pulsed bone marrow–derived DCs with car-
boxyfluorescein diacetate succinimidyl ester
(CFSE)–labeled apoptotic cells derived from
g-irradiated MCA101 tumor cells and followed
their fate along the endolysosomal compart-
ment. Confocal microscopy revealed colocali-
zation of apoptotic bodies with LAMP-1, which
indicated trafficking to saposin-containing
lysosomes (Fig. 1C). When comparing the kinetic
digestion of apoptotic cells in pSAP-deficient
or wild-type (WT) DCs, we found that early
uptake of apoptotic material was similar, sug-
gesting that phagocytosis is not affected by
saposin deficiency. However, at later time points
after endocytosis, pSAP-deficient DCs accu-
mulated CFSE-labeled cells, demonstrating

Sharma et al., Science 383, 190–200 (2024) 12 January 2024 1 of 11

RES EARCH | R E S E A R C H A R T I C L E

A Apoptotic B
tumor cells

% Calcein Leakage
100
80
Irradiate DCs DC:CD8 T cell 60
tumor cells coculture
40
Confocal microscopy 72 hours
EM and calcein FACS 20
Antigen presentation
leakage assay assay 0

SAP-B
SAP-C
SAP-D
SAP-A

BSA
on vesicles

C D E

Number of ApoBDs/cell
ns
12
15
ns
10

MFI (x1000)

p
8 10
6
4 5

2
0
0

m A
VA
2h 4h 8h

V
sO

g
F G
ns

y
% CFSElow cells
*** ***
% CFSElow cells

80 80

60 60
40 40
20 20
0
0

Fig. 1. Saposins promote cross-presentation of membrane-associated tumor microscopy images showing the kinetics of apoptotic cell disintegration in WT and
antigens. (A) Diagram depicting the experimental read-outs used in Fig. 1. MCA101 pSAP-KO DCs. DCs were pulsed with CFSE-labeled, g-irradiated apoptotic MCA101

fibrosarcoma cells were g-irradiated (100 Gy) prior to the collection of apoptotic
vesicles from supernatant and analysis with electron microscopy (EM) and calcein
leakage assay. Apoptotic MCA101 or MCA101-OVA cells were used to pulse bone
marrow–derived DCs from WT or pSAP-KO mice prior to the analysis of digestion of
apoptotic cells with confocal microscopy and antigen processing and T cell
y g
tumor cells for 2 hours, and the numbers of apoptotic bodies (ApoBD) were
quantified at the indicated time points with ImageJ software (DAPI, blue; LAMP-1,
red; ApoBD, green). (E) Representative histogram overlays and bar graph showing
flow cytometry staining and mean fluorescence intensity (MFI) of MHC-I-SIINFEKL
peptide on the surface of WT or pSAP-KO DCs after incubation with either soluble

y
activation with FACS. (B) Calcein leakage assay to quantify the effect of saposins on OVA (sOVA) or irradiated MCA101-OVA tumor cells (mOVA, membrane-associated

disintegration of apoptotic bodies. Apoptotic vesicles were prepared with differential
ultracentrifugation (100,000g) from the supernatant of irradiated MCA101 cells
and visualized with transmission electron microscopy (left). Scale bar, 200 nm.
Apoptotic bodies were further loaded with calcein dye prior to incubation with
indicated saposins or BSA (negative control), and calcein release was quantified in
the supernatant with fluorimetry. (Right) Depicted is percent leakage compared with
100% lysis induced by Triton X-100. SAP, saposin. (C) Representative confocal
microscopy image showing colocalization of apoptotic bodies (green) with LAMP-
1 (red). WT DCs were pulsed with CFSE-labeled, g-irradiated apoptotic MCA101
tumor cells for 2 hours. ApoBD, apoptotic body. (D) Representative confocal
,
OVA) for 4 hours. (F) Histograms and bar graph showing frequencies of proliferating
CFSElow CD8 T cells after a three-day coculture with WT or pSAP-KO DCs pulsed
with soluble OVA. (G) Histograms and bar graph depicting the frequencies of
CFSElow CD8 T cells after a three-day coculture with WT or pSAP-KO DCs pulsed
with irradiated MCA101-OVA cells. Additionally, pSAP-KO DCs were reconstituted
with 10 mg/ml of recombinant pSAP prior to the T cell assay. Data shown in all
graphs represent mean ± SD from three to five independent replicates. P values
were determined with one-way analysis of variance (ANOVA) [(B) and (G)] or
unpaired Student’s t test [(D), (E) and (F)]. *P < 0.05; **P < 0.01; ***P < 0.001;
ns, not significant.

the importance of saposins for processing of
apoptotic bodies in DCs (Fig. 1D).
To test for saposin-dependent cross-presentation
and CD8 T cell activation, we pulsed DCs from
pSAP-knockout (KO) or WT mice with apo-
ptotic MCA101 cells expressing a membrane-
associated form of the antigen OVA prior
to coculture with OVA-specific CD8 T cells.
MCA101-OVA cells express the model antigen
coupled to the C1C2 domain of lactadherin,
which allows for targeting of OVA to PS-
expressing vesicles (22). We analyzed produc-
tive antigen processing by staining for the
processed OVA epitope in complex with MHC-I
(H-2Kb-SIINFEKL) on the DC surface. Flow

Sharma et al., Science 383, 190–200 (2024) 12 January 2024 2 of 11

RES EARCH | R E S E A R C H A R T I C L E
Fig. 2. pSAP is required for tumor immunity and boosts T cells derived
from melanoma patient samples. (A) Experimental scheme of tumor challenge
studies. WT and pSAP-KO BM chimeric mice were primed with 4 × 105 g-irradiated
MCA101-OVA cells subcutaneously (s.c.) and subsequently inoculated with 1 × 106
live MCA101-OVA cells (s.c.) 7 days post priming (D0). (B) Comparison of tumor
sizes between WT and pSAP-KO mice on day 17 (left) and the kinetics of the tumors’
growth (right). (C) Representative histogram overlay and bar graph depicting the
staining and MFI of MHC-I-SIINFEKL peptide on the surface of tumor DCs from
pSAP-KO or WT animals. (D) FACS plots and bar graphs showing frequencies of
MHC-I (Kb-SIINFEKL) tetramer- and IFN-g–positive tumor-infiltrating CD8 T cells in
pSAP-KO or WT mice. MHC-I tetramer specifically detects CD8 T cells that are
reactive with SIINFEKL peptide. (E) Experimental setup for the coculture of myeloid
and CD8 T cells isolated from human melanoma. Single-cell suspensions from
human melanoma samples were FACS-sorted for CD146+ melanoma cells, CD8+ T cells,
Sharma et al., Science 383, 190–200 (2024) 12 January 2024
p
g
y
y g
y
,
and CD11c/b+ myeloid cells. CD146+ cells were g-irradiated and incubated with DCs,
which were further cocultured with CD8 T cells in the presence or absence of
recombinant pSAP. (F) FACS plots and bar graph showing the frequencies of IFN-g–
positive CD8 T cells following the indicated culture conditions. (G) Representative
histogram overlay and bar graph demonstrating surface staining and MFI of LAMP-1 on
CD8 T cells according to the indicated culture conditions. Graph colors represent the
same as in (F). (H) Flow cytometry analysis and summarizing bar graph depicting the
frequencies of antigen-specific CD8 T cells reactive with HLA-A*0201 tetramers loaded
with epitopes from gp100, MART-1, tyrosinase, and NY-ESO-1 following the indicated
culture setups. Amino acid residues are depicted on top, and percentages of gated cells
are shown as mean ± SD in the dot plots. Data shown in (B) to (D) are representative of
three independent experiments, whereas (F) to (H) depict mean ± SD from seven
independent subjects. P values were determined by unpaired Student’s t test [(B) to (D)]
or one-way ANOVA [(F) to (H)]. **P < 0.01; ***P < 0.001; ****P < 0.0001.
3 of 11
RES EARCH | R E S E A R C H A R T I C L E
cytometry demonstrated that processing of
soluble OVA was equally efficient in pSAP-KO
and WT DCs (Fig. 1E). However, processing
and MHC-I loading of membrane-associated
antigen derived from tumor cells was signifi-
cantly hampered in the absence of pSAP (Fig. 1E).
These findings were in accordance with our T cell
data, as pSAP-deficient DCs efficiently activated
CD8 T cells in response to soluble antigen as
well as OVA-coated latex beads (Fig. 1F and
fig. S1C). In sharp contrast, when DCs were
pulsed with tumor cells containing membrane-
associated antigen, CD8 T cell activation by
pSAP-KO DCs was greatly reduced, as high-
lighted by their truncated CFSE dilution profile
in flow cytometry (Fig. 1G). Notably, incubation
of pSAP-deficient DCs with recombinant pSAP
fully reconstituted CD8 T cell activation (Fig.
1G). Similarly, we pulsed pSAP-KO or WT DCs
with irradiated B16F10 melanoma cells prior
to coculture with antigen-specific CD8 T cells
that recognize the integral membrane protein
Pmel-1 (gp100). Impaired T cell activation by
pSAP-deficient DCs revealed that saposins are
equally important for processing of membrane-
associated melanoma antigens (fig. S1D). In ad-
dition, pSAP deficiency also compromised
MHC-II–restricted presentation of tumor anti-
gens derived from apoptotic bodies, as indicated
by diminished stimulation of antigen-specific
CD4 T cells (fig. S1E). Taken together, saposins
disintegrate apoptotic vesicles and process
membrane-associated antigens for presenta-
tion to CD4 and CD8 T cells.
pSAP is required for tumor immunity
Next, we investigated how pSAP function af-
fects T cell activation in vivo and protection
against cancer. To assess T cell priming, we
transferred naïve CFSE-labeled, OVA-specific
CD8 T cells into pSAP-deficient or WT re-
cipients prior to subcutaneous administra-
tion of apoptotic MCA101-OVA cells (fig. S2A).
Four days after tumor cell injection, we iso-
lated DCs from skin-draining lymph nodes
and analyzed antigen processing. Expression
of H-2Kb-SIINFEKL proved to be reduced in
DCs from pSAP-KO when compared with that
of WT mice (fig. S2B). Moreover, antigen-
specific proliferation and interferon (IFN)–g
production by CD8 T cells from lymph nodes
were severely hampered in the absence of
pSAP (fig. S2C). Thus, pSAP facilitates CD8
T cell priming in response to particulate anti-
gen. Because straight pSAP-KO mice have a
reduced life span, we generated chimeric mice
by transferring pSAP-KO or WT bone marrow
to WT recipients to allow for tumor challenge
experiments. In this context, we immunized
mice with irradiated tumor cells and chal-
lenged them with a higher number of live
MCA101-OVA cells one week later (Fig. 2A).
Subsequently, we monitored tumor growth in
the skin and found a drastic expansion of
Sharma et al., Science 383, 190–200 (2024) 12 January 2024
cancer in pSAP deficiency (Fig. 2B). Flow
cytometry analysis of isolated DCs from the
tumor site showed a pSAP-dependent decrease
in antigen processing and presentation (Fig.
2C). Furthermore, MHC-I tetramer–mediated
detection of antigen-specific CD8 T cells showed
reduced frequency of tumor-infiltrating T cells
as well as cytokine production in pSAP-deficient
mice (Fig. 2D). Additionally, we also challenged
bone marrow–chimeric mice with live tumor cells
without prior vaccination (fig. S3A). As a result,
cancer protection, antigen processing in tumor
DCs, frequency of tumor-infiltrating, antigen-
specific T cells, and cytokine production and
cytotoxicity were all greatly reduced when pSAP
was lacking (fig. S3, B to E). To exclude that
pSAP-dependent tumor immunity was influ-
enced by differential DC migration, we exam-
ined migratory DCs in tumor-draining lymph
nodes and found no alteration in DC frequen-
cies and expression of the homing marker
CCR7 in pSAP deficiency (fig. S3, F and G).
Moreover, we analyzed DC-attracting chemo-
kines at the tumor site and observed compara-
ble expression of CCL19 and CCL21 in pSAP-KO
and WT mice (fig. S3H). Because pSAP has
been associated with macrophage function
and inflammation (23), we controlled our mouse
model for a possible impact on tumor-associated
macrophages. However, pSAP deficiency neither
affected the number of macrophages nor their
expression of inflammatory genes in the tumor
microenvironment (TME) (fig. S3, I and J).
Altogether, these findings demonstrate that
tumor immunity critically depends on pSAP
function.
T lymphocytes from melanoma patients are
boosted by pSAP
To study pSAP in the context of human cancer,
we used dissociated tumor cell samples from
melanoma patients. A detailed description of
the patient samples can be found in table S1.
Briefly, the majority of specimens were isolated
from primary melanoma lesions of the skin,
assigned as clinical stage III, including those
from white female and male patients older
than 50 years before treatment (table S1). To
purify antigen-presenting cells, responder T
cells, and tumor cells as source of antigen, we
used fluorescence-activated cell sorting (FACS)
on CD146+ melanoma cells, CD11b/c+ myeloid
cells, and CD8+ T cells (fig. S4). We then irra-
diated CD146+ melanoma cells and pulsed them
onto sorted myeloid cells prior to coculture with
autologous CD8 T cells (Fig. 2E). In parallel, we
also treated DCs and T cells with human re-
combinant pSAP. Five days after culture, we
analyzed effector functions of CD8 T cells and
found that recombinant pSAP was able to boost
IFN-g production (Fig. 2F) and cytolytic activity,
indicated by surface LAMP-1 staining as a sign of
cytotoxic degranulation (Fig. 2G). Furthermore,
we measured the frequencies of tumor antigen–
specific CD8 T cells by staining with MHC-I
tetramers loaded with dominant melanoma
antigens, including MART, gp100, tyrosinase,
and NY-ESO-1. The abundance of melanoma-
specific CD8 T cells was greatly increased when
tumor DCs were treated with pSAP (Fig. 2H).
Notably, the patient samples were human leu-
kocyte antigen (HLA)–typed by flow cytometry
beforehand to select the proper haplotype
for MHC-I tetramer analysis (HLA-A02). Over-
all, the impact of pSAP on tumor DCs is able
to rescue T cell activation from melanoma
patients.
Hyperglycosylation of pSAP in tumor DCs leads
to its secretion
Beyond the use of pSAP deficiency in a mouse
model, we aimed at understanding the regu-
lation of pSAP in tumor-associated DCs in a
pathophysiological context. To this end, we
inoculated WT mice with live MCA101-OVA
p
cells subcutaneously, and after tumor out-
growth, we isolated DCs from the TME, tumor-
draining lymph nodes, and spleen (fig. S5A).
We FACS-purified the two main classical DC
subsets on the basis of their established
markers as cDC1 (XCR1+ or CD103+) and cDC2
g
(SIRP1a+ or CD11b+) prior to performing an
array of antigen processing and presentation
assays. Accordingly, pulsing with FITC-dextran
showed that the phagocytosis rate of tumor
y
DCs was not altered when compared with
lymph node and spleen (fig. S5B). Incubation
with a self-quenched antigen conjugate (DQ-
OVA), which exhibits fluorescence upon pro-
teolytic degradation, demonstrated that mainly
cDC2s in the tumor are compromised to pro-
cess soluble antigen (fig. S5C). These findings
were in line with the ultimate epitope expres-
sion on surface MHC-I following pulsing with
soluble OVA, which revealed hampered anti-
y g
gen presentation by tumor cDC2 (fig. S5D). In
sharp contrast, the presentation capacity of
cDC1 in the TME was only affected after in-
cubation with irradiated MCA101-OVA cells,
which contain antigen in membrane-associated
form (fig. S5D). This phenomenon was reflected
y
in functional T cell experiments because DCs
,
isolated from tumors were severely perturbed
to induce T cell responses reactive to membrane-
associated antigen (fig. S5E).
Because we found that saposins are critical
for presentation of particulate antigen, we then
hypothesized that pSAP function might be
modulated in tumor DCs as a basis for poor
T cell induction in the TME. Indeed, analysis
by immunoblot revealed the expression of a
75-kDa high molecular weight form of pSAP
in tumor DCs when compared with pSAP-65,
which was predominant in DCs from spleen
(Fig. 3A). Moreover, the small, single saposins
were severely depleted in DCs from the TME
(Fig. 3A). When we cultured the respective DC
subsets ex vivo, we observed secretion of pSAP
4 of 11
RES EARCH | R E S E A R C H A R T I C L E

p
g
y
y g
y
,

Fig. 3. Hyperglycosylation of pSAP in tumor DCs leads to its secretion. WT inoculation, cDC1 and cDC2 populations were FACS-sorted from tumor and
mice were inoculated with 1 × 106 live MCA101 cells, and 18 days post tumor spleen. (A) Immunoblot showing the abundance of pSAP and saposins in

Sharma et al., Science 383, 190–200 (2024) 12 January 2024 5 of 11

RES EARCH | R E S E A R C H A R T I C L E

DCs as well as pSAP secretion. The left blot shows the expression of pSAP and
saposins in FACS-sorted DC subsets from tumor and spleen, whereas the right
blot shows secreted pSAP in the culture supernatant. pSAP-75, hyperglycosy-
lated pSAP; pSAP-65, glycosylated pSAP; SAPs, saposins. (B) Quantification of
pSAP secreted by DCs. FACS-sorted splenic and tumor DC subsets were
cultured in cRPMI for 48 hours, and pSAP in culture supernatant was quantified
with ELISA. (C) Immunoblot of Endo H–treated pSAP. (Left) Mechanism of Endo
H that leads to cleavage of high-mannose, but not complex, glycans. (Right)
FACS-sorted DCs from tumor and spleen were lysed in radioimmunoprecipitation
assay (RIPA) buffer and cell lysates were treated with Endo H for 12 hours at
37°C prior to analysis with immunoblot. (D) Matrix-assisted laser desorption/
ionization–time-of-flight (MALDI-TOF) mass spectrometry analysis of permethy-
lated N-linked glycans of pSAP immunoprecipitated from FACS-sorted CD11c+
DCs. Enzymatically released N-glycans from pSAP of splenic (top) and tumor
(bottom) DCs were analyzed. Glycan compositions were assigned based on m/z
values. x axis, mass to charge ratio (m/z). y axis, signal intensity of the ions.
green circle, mannose; yellow circle, galactose; red triangle, fucose; blue square,
N-acetylglucosamine; magenta diamond, sialic acid. (E) Heat map of differen-
tially expressed genes in tumor DCs involved in glycosylation, as analyzed by
real-time RT2 profiler PCR array. Splenic DCs were used as the control to
splenic DC1. (G) PLA of pSAP and sortilin. Confocal microscopy images of tumor
and splenic DC subsets reveal PLA signal between pSAP and sortilin. Blue
indicates cell nucleus (DAPI) and magenta represents ligation signal. The violin
plot shows quantification of PLA signal, where 200 cells from each sample
were analyzed for statistics. (H) Coimmunoprecipitation of sortilin and pSAP in
tumor and splenic DCs. (Top) Blot of sortilin pulled down by anti-pSAP antibody.
(Bottom) Immunoblot of total sortilin in corresponding DC populations. Bar
graphs depict the densitometric and statistical analysis of sortilin abundance
in the immunoblots. Red, tumor; blue, spleen. IP-pSAP, immunoprecipitation of
pSAP. (I) PLA of pSAP and sortilin in human melanoma and monocyte-derived
DCs (MoDCs). Melanoma DCs were sorted as CD11c+ cells from viable CD45+
cells isolated from human melanoma samples, whereas MoDCs were generated
by culturing monocytes with interleukin 4 (IL-4) and granulocyte-macrophage
colony-stimulating factor (GM-CSF) for 4 days. Blue indicates cell nucleus (DAPI),
and magenta represents ligation signal. The violin plot shows quantification of
PLA signal, where 200 cells from each sample were analyzed for statistics.
(J) Immunoblot of pSAP in human melanoma DCs and MoDCs. (K) Illustration
visualizing glycosylation mechanisms that control pSAP trafficking in tumor DCs.
Hyperglycosylation of pSAP compromises its interaction with sortilin and reroutes it
to the secretory pathway. Data shown in all graphs are representative of three

p
calculate fold change in gene expression. (F) Bar graph depicting glycosyl- independent experiments, and P values were determined with unpaired Student’s
transferase and glycosidase gene expression in tumor compared with that of t test. *P < 0.05; **P < 0.01; ***P < 0.001; ****P < 0.0001; ns, not significant.

into the cell culture supernatant mainly by cient lysosomal delivery of pSAP. For this pur- the compromised antigen presentation capac-
tumor DCs (Fig. 3, A and B). To demonstrate pose, we subjected tumor and splenic DCs to a ity in the tumor microenvironment.

that the occurrence of pSAP-75 was due to
glycosylation, we tested endoglycosidase H
(Endo H) sensitivity of pSAP. Endo H cleaves
N-linked glycans between the two proximal
proximity ligation assay (PLA), in which anti-
bodies against pSAP and sortilin are coupled
to oligonucleotide probes that allow for sub-
sequent ligation and amplification in case the
g
TGF-b induces pSAP hyperglycosylation in tumor
DCs and drives immune evasion
We mechanistically addressed the question

N-acetylglucosamine residues only in high-
mannose carbohydrate chains, but not in com-
plex glycans. After treatment of protein lysates
with Endo H, pSAP-65 from splenic DCs was
cleaved to lower molecular weight forms,
whereas pSAP-75 from tumor DCs proved
to be Endo H–resistant, suggesting that it
contains complex glycans (Fig. 3C). To corrob-
orate this finding, we performed deeper mo-
lecular analysis using mass spectrometry of
two target proteins are in close vicinity (10 to
80 nm). Confocal microscopy revealed these
PLA events as discrete spots and showed
that the spatial relationship between pSAP
and sortilin was perturbed in DCs from the
TME (Fig. 3G). To assess the direct physical
interaction between pSAP and sortilin through
biochemistry, we performed immunoprecipi-
tation of pSAP and subsequently resolved
sortilin using immunoblot. As a result, the
y
how hyperglycosylation of pSAP is regulated.
To this end, we screened a panel of recombi-
nant cytokines upon incubation of a murine
DC line (DC2.4) and found that TGF-b can
induce pSAP secretion (fig. S6A). Treatment
of DCs with TGF-b triggered dose-dependent
induction of pSAP-75 and subsequent secre-
tion into the culture supernatant, as detected
by immunoblot and enzyme-linked immuno-
sorbent assay (ELISA), without affecting the

sugar structures based on purified pSAP bands
derived from splenic (pSAP-65) or tumor DCs
(pSAP-75). Tandem mass spectrometry showed
that the glycan of pSAP-65 mainly consisted
of mannose residues, whereas pSAP-75 ex-
hibited complex glycans involving additions
abundance of sortilin recovered from the pSAP
precipitate was clearly reduced in tumor DCs,
indicating that the interaction of pSAP with its
chaperone sortilin is hampered in the TME (Fig.
3H). To translate these findings to the human
system, we explored pSAP-sortilin interaction in
y g
expression of the pSAP chaperone, sortilin
(Fig. 4, A to D). Moreover, we measured the
gene expression of a panel of enzymes involved
in the glycosylation pathway in DC2.4 cells
treated with TGF-b (fig. S6B). We observed
that the up-regulated genes in TGF-b–treated

y
of N-acetylglucosamine, galactose, and sialic DCs isolated from the tumor site of melanoma DCs correlated well with the enzyme signature

acid (Fig. 3D). Because glycan structures are
synthesized by a diverse set of glycosyltrans-
ferases, we compared the expression of glyco-
syltransferases between tumor and splenic
DCs using a quantitative reverse transcription
polymerase chain reaction (qRT-PCR) array
(Fig. 3E). In tumor DCs, we found up-regulation
of several enzymes that facilitate the attach-
ment of complex glycan residues, such as
N-acetylglucosaminyltransferases, galactosyl-
transferases, and sialyltransferases (Fig. 3F).
Because hyperglycosylation of pSAP leads to
its secretion and therefore, the reduced gen-
eration of single saposins, we next investigated
the interaction of pSAP with its chaperone
sortilin, which is normally required for effi-
patients compared with that of monocyte-
derived DCs. The resulting PLA signals dem-
onstrated that the spatial interaction of pSAP
with sortilin was reduced in melanoma DCs
(Fig. 3I). Furthermore, we examined the abun-
dance of the different molecular weight forms
of pSAP and found that tumor DCs from mela-
noma patients exclusively expressed hypergly-
cosylated pSAP-75 (Fig. 3J).
Taken together, our results demonstrate that
pSAP is hyperglycosylated in tumor DCs, fails
to interact with its chaperone sortilin, and fol-
lows instead a secretory route (Fig. 3K). This
mechanism of pSAP hyperglycosylation leads
to a depletion of intracellular saposins availa-
ble for antigen processing, which might explain
,
detected in tumor DCs (Fig. 4E), suggesting
that TGF-b is responsible for triggering the
respective glycosylation program. To test
whether TGF-b signaling is indeed required
for pSAP hyperglycosylation in vivo, we used
mice that lacked TGF-b receptor II specifically
in DCs (CD11c-Cre x Tgfbr2flox/flox) for chal-
lenge with live MCA101-OVA tumor cells (Fig. 4F).
As anticipated, the lack of TGF-b downstream
signaling in DCs caused better tumor protec-
tion, increased antigen presentation in tumor
DCs, and stronger IFN-g production by tumor-
infiltrating CD8 T cells (Fig. 4, G to I). More
notably, when we isolated DCs, macrophages,
and other CD45+ leukocytes from the tumor
site for analysis by immunoblot, we found that

Sharma et al., Science 383, 190–200 (2024) 12 January 2024 6 of 11

RES EARCH | R E S E A R C H A R T I C L E

A B C D E

mRNA fold change in TGF- β

103
β R2=0.56 B3gnt8
2.0

pSAP (ng/ml)

stimulated DC2.4 cells

B3gnt4
102 St8sia6 St8sia2
1.5 Mgat4a&5
Neu1 B3gnt3 Galnt3&9
1.0
101 Galnt6
0.5
B4galt1 B4galt5
0.0 100

Unst.
2.5 ng/ml
10 ng/ml
β 10-1
0.25 1 4 16
mRNA fold change in tumor DCs
TGF- β compared to splenic DCs
F G Tumor volume (mm 3)
H
2000 3

MFI (x1000)
1500 ***
*** 2
1000
*** 1
500 **

p
0 0
0
6
8
10
12
14
16
18
20 Days

I J K

g
% IFN-γ+ CD8+ T cells

20
15

y
10
5
0
γ

L
mRNA fold change in tumor DC1

10

y g
compared to splenic DC1

Fig. 4. TGF-b induces pSAP hyperglycosylation in DCs and com- 5

promises tumor immunity. (A) Experimental setup of murine DC
(DC2.4) cells treated with TGF-b and read-outs shown in (B) to 0

(E). (B) Immunoblot showing dose-dependent induction of pSAP -5

hyperglycosylation and secretion by DC2.4 cell line incubated with

y
recombinant TGF-b for 48 hours. (C) Quantification of pSAP in -10

,
the culture supernatant of DC2.4 cells after incubation with
-15
recombinant TGF-b for 2 days as measured with ELISA. Unst., xxx.
Man1b1

Man2b1
Man1c1
B4galt 1

Uggt1

Fuca 1
St3Gal1
St8sia2
St8sia6
St8sia4
Galnt6
B3gnt 8
B3gnt 3
B3gnt 4
B3gnt 2
B4galt 5

Mgat4a

Mgat5b

Man1a2

Galnt13
Galnt14
Galnt7

Fuca 2
Mgat5

(D) Immunoblot of sortilin in DC2.4 cells after incubation with

recombinant TGF-b for 2 days. (E) Scatter plot showing the
correlation of gene expression of glycosyltransferases and glyco-
sidases between tumor DCs and TGF-b–stimulated DC2.4 cells. mRNA fold (J) Immunoblot of pSAP and saposins (SAPs) in tumor DCs, macrophages,
changes were quantified by real-time RT2 profiler PCR array. CD11c+ tumor DCs and other CD45+ leukocytes from Tgfbr2DDC or Tgfbr2f/f mice on day 20
were analyzed with splenic DCs as the control, whereas TGF-bstimulated after tumor cell inoculation. (K) Immunoblot of sortilin in tumor DCs
DC2.4 cells were compared with sham-treated DC2.4 cells. (F) Experimental from Tgfbr2DDC or Tgfbr2f/f mice on day 20 after tumor cell inoculation.
scheme of tumor cell challenge. Tgfbr2f/f and CD11c–Cre x Tgfbr2f/f (Tgfbr2DDC) (L) Differentially expressed glycosyltransferase and glycosidase genes in
BM chimeric mice were inoculated with 1 × 106 live MCA101-OVA cells (s.c.). tumor DCs isolated from Tgfbr2DDC or Tgfbr2f/f mice on day 20 after tumor
(G) Kinetics of tumor growth in Tgfbr2f/f and Tgfbr2DDC mice. (H) Histogram challenge. Splenic DCs from Tgfbr2f/f animals were used as control to
overlay and bar graph depicting H-2Kb-SIINFEKL staining and MFI on tumor DCs calculate mRNA fold change. Data shown in all graphs are representative
from Tgfbr2DDC mice or Tgfbr2f/f controls on day 20 after tumor cell injection. of three independent experiments, and P values were determined with
(I) FACS plots and bar graph showing frequencies of IFN-g+ tumor-infiltrating one-way ANOVA (C) or unpaired Student’s t test [(G) to (I)]. *P < 0.05;
CD8 T cells in Tgfbr2DDC or Tgfbr2f/f animals on day 20 post tumor challenge. **P < 0.01; ***P < 0.001; ****P < 0.0001.

Sharma et al., Science 383, 190–200 (2024) 12 January 2024 7 of 11

RES EARCH | R E S E A R C H A R T I C L E
pSAP-75 was specifically absent in DCs that
lacked TGF-b signaling, whereas the expres-
sion of sortilin in tumor DCs was not depen-
dent on TGF-b (Fig. 4, J and K). Additionally,
tumor DCs lacking TGF-b receptor exhibited
an abundance of single saposins, which was in
great contrast to the saposin depletion and ex-
clusive expression of hyperglycosylated pSAP
(pSAP-75) in WT DCs as well as tumor-associated
macrophages and other CD45+ leukocytes
(Fig. 4J). Furthermore, the enzyme signature
involved in glycosylation proved to be reduced
when TGF-b signaling was deficient in tumor
DCs (Fig. 4L).
Next, we investigated how TGF-b mechanis-
tically mediates the glycan complexity of pSAP
and its effects on sortilin interaction and pSAP
secretion. Our mass spectrometry data revealed
that the complex glycan of pSAP-75 is generated
through the addition of N-acetylglucosamine,
galactose, and sialic acid to the ternary man-
nose core (Fig. 3D and fig. S7A). On the basis of
this observation, we hypothesized that TGF-b
controls pSAP hyperglycosylation by modu-
lating the expression of specific enzymes in-
volved in glycan synthesis. Analyzing our gene
expression array, we selected TGF-b–induced
N-acetylglucosaminyltransferases (Mgat4a
and Mgat5) and galactosyltransferase (B4galt1)
based on their relevance in the corresponding
pSAP glycan pathway identified by mass spec-
trometry (fig. S7A). In addition to the TGF-b–
regulated enzymes, the pSAP glycan pathway
is complemented by constitutive glycosyltrans-
ferases, such as Mgat1/2, which synthesize
precursor molecules for Mgat4a and Mgat5,
as well as the sialyltransferases St6Gal1 and
St6Gal2 that add terminal sialic acid to the
carbohydrate chain generated by B4galt1 (fig.
S7A). Subsequently, we performed small inter-
fering RNA (siRNA)–mediated knockdown of
those enzyme candidates in TGF-b–treated
DC2.4 cells (fig. S7, B and C) prior to testing
the impact on pSAP interaction with sortilin
using PLA, pSAP secretion based on ELISA,
and T cell activation using antigen presentation
assays. As a negative control, we chose the TGF-
b–regulated enzyme Mgat3 that produces a bi-
secting N-acetylglucosamine on the trimannosyl
core, which represents a carbohydrate structure
absent from our mass spectrometry data on
pSAP-75 and therefore is not expected to play a
role in the pSAP glycan pathway (fig. S7B).
Notably, the silencing of all pSAP-75–
associated glycosyltransferases restored the in-
teraction of pSAP with sortilin, as indicated by
drastically amplified PLA signal (fig. S7D). By
contrast, knockdown of negative control Mgat3
had no effect on pSAP-sortilin engagement
(fig. S7D). In a complementary manner, down-
regulation of the respective enzyme candi-
dates prevented the release of pSAP into the
supernatant of DCs and facilitated proper
CD8 T cell activation, whereas Mgat3 modu-
Sharma et al., Science 383, 190–200 (2024) 12 January 2024
lation failed to affect pSAP secretion and T cell
induction (fig. S7, E and F). Notably, gene
silencing of St6Gal enzymes was sufficient to
reconstitute the pSAP-sortilin PLA signal, im-
plicating terminal sialic acid residues as crit-
ical blockers of this interaction. To confirm
these findings, we used an established inhib-
itor of sialyltransferase (3F-NeuAc) upon incu-
bation of TGF-b–treated DC2.4 cells (fig. S7G),
which completely abrogated the addition of
sialic acid to the pSAP glycan, as indicated
by ELISA-based binding of a lectin specific
for sialic acid (fig. S7H). Consistent with the
siRNA results, inhibition of pSAP sialylation
reinstated its interaction with sortilin and
prevented eventual secretion, thereby promot-
ing activation of CD8 T cells (fig. S7, I to K).
These data indicate that TGF-b–regulated Mgat
and B4galt1 enzymes are crucial for the se-
quential synthesis of the complex pSAP glycan
and that terminal sialylation by St6Gal im-
pedes the pSAP and sortilin interaction.
Furthermore, we examined the molecular
mechanism of TGF-b downstream signaling
and its regulation of signature enzymes in-
volved in pSAP hyperglycosylation. For this
purpose, we performed chromatin immuno-
precipitation of the signal transducers Smad2/3
using genomic DNA derived from TGF-b–
treated DC2.4 cells (fig. S7L). When compared
with isotype control immunoglobulin G (IgG)
and the housekeeping gene b-actin, subsequent
amplification by PCR revealed significant en-
richment of promoter regions containing Smad-
responsive elements in the genes coding for
Mgat4a, Mgat5, and B4galt1 (fig. S7L).
Taken together, these findings demon-
strate (i) the constitutive and TGF-b–regulated
enzyme pathway required for the generation
of hyperglycosylated pSAP, (ii) the critical func-
tion of terminal sialic acid to hamper inter-
action with sortilin, and (iii) the presence of
TGF-b–responsive gene promoter sequences
in signature enzymes involved in pSAP glycan
synthesis. Overall, TGF-b is essential for hyper-
glycosylation of pSAP in tumor DCs, a mech-
anism associated with immune escape.
Immunotherapeutic targeting of tumor DCs with
recombinant pSAP
Based on the importance of pSAP for antigen
presentation in the tumor microenvironment,
we then aimed at targeting recombinant pSAP
to tumor DCs. For this purpose, we coupled
pSAP to anti-DEC205, an antibody well estab-
lished to target the endocytic receptor DEC205
on DCs, using chemical conjugation as previ-
ously described (24, 25). Briefly, we incubated
recombinant pSAP with tris (2-carboxyethyl)
phosphine hydrochloride to expose its sulfhy-
dryl groups, and in parallel, the anti-DEC205
antibody or isotype control IgG were activated
for chemical conjugation by sulfosuccini-
midyl 4-[N-maleimidomethyl] cyclohexane-1-
carboxylate. Following overnight incubation,
the pSAP-antibody conjugates were concen-
trated, and successful coupling was analyzed
by immunoblot (fig. S8A). In addition, we con-
trolled that the amount of pSAP coupled to
anti-DEC205 or isotype control was compara-
ble (fig. S8B). Moreover, we verified that pSAP
conjugation still preserved the fine specificity
of the DEC205 antibody by showing that the
pSAP–anti-DEC205 conjugate stained a simi-
lar percentage of DCs when compared to sepa-
rate anti-DEC205 detection by flow cytometry
(fig. S8C). Furthermore, we incubated pSAP-
KO DCs with the pSAP-antibody conjugates
and tested OVA epitope expression and CD8
T cell activation after the pulsing of DCs with
apoptotic MCA101-OVA cells (fig. S8D). Ac-
cordingly, we observed reconstitution of anti-
gen presentation (H-2Kb-SIINFEKL) and CD8
T cell priming (CD69) specifically by pSAP
coupled to anti-DEC205 in a dose-dependent
p
manner (fig. S8, E and F). To assess the impact
of glycosylation on pSAP’s biological activity,
we purified pSAP-65 and pSAP-75 from DC2.4
cells. Treatment of pSAP-65 with peptide
N-glycosidase (PNGase) F generated nongly-
cosylated pSAP-55 (fig. S8G). Subsequently,
g
we coupled the differentially glycosylated pSAP
forms with anti-DEC205 and incubated pSAP-KO
DCs with the conjugates prior to performing
functional T cell assays. In contrast to pSAP-
y
65, hyperglycosylated and nonglycosylated pSAP
mediated significantly lower CD8 T cell activ-
ation (fig. S8H), which highlights that optimal
glycosylation of pSAP is required for deploying
its functions in the context of endosomal de-
livery by pSAP-based therapeutics.
After validation of the targeting tool in vitro
and ex vivo, we inoculated pSAP-KO bone
marrow–chimeric mice with MCA101-OVA
tumor cells and injected 100 mg of pSAP
y g
coupled with anti-DEC205 or isotype IgG on
days 9 and 13 after tumor challenge (Fig. 5A).
On day 14, we isolated and sorted intratu-
moral DCs, focusing on the two major clas-
sical DC subsets, cDC1 and cDC2. Staining
for intracellular pSAP in the respective DC
y
subsets revealed effective delivery of pSAP tar-
,
geted through DEC205 when compared with
the isotype control (Fig. 5B). Additionally,
immunoblot analysis revealed that pSAP-
deficient tumor DCs took up recombinant
pSAP-65 and generated massive amounts of
single saposins upon lysosomal cleavage (Fig. 5C).
This demonstrated successful reconstitution
of pSAP-deficient DCs in the TME, especially in
the cDC1 subtype that is central to cross-
priming of CD8 T cells. We then followed a
similar experimental schedule using pSAP tar-
geting of tumor-inoculated WT animals (Fig.
5D). Treatment with pSAP coupled to anti-
DEC205 greatly reduced tumor burden in WT
mice (Fig. 5E). Correspondingly, antigen pre-
sentation of OVA peptide by tumor DCs as
8 of 11
RES EARCH | R E S E A R C H A R T I C L E

A B C

50

% pSAPpos DCs
40
30
20
10
0

2
C

C
D

D
D E F

Tumor volume (mm3)

2000 15

MFI (x1000)
1500 ***
*** 10
1000
***
500 5

p
0 0

0
6
8
10
12
14
16
18
20
Days

G H

% MHC-I-Tet+ cells % MHC-I-Tet+ cells

% IFN-γ+ T cells

g
10 10
8 8
6 6
4 4
2 2
0 0

y
% IFN-γ+ T cells

80 10
60
5
γ

40
20
0 0

I 2500 Anti-DEC205/pSAP + anti-PD-L1

y g
Anti-DEC205/pSAP
Tumor volume (mm 3)

2500
Tumor volume (mm 3)

2000 Anti-PD-L1
IgG/pSAP 2000
1500
1500

y
1000
1000

,
500 500

0 0
Day 32
20

28
16

24

32
12
8
0

Days

Fig. 5. Reconstitution of tumor DCs with recombinant pSAP drives
protection from cancer. (A) Regime of pSAP targeting to tumor DCs. pSAP-
KO BM-chimeric mice were inoculated with 1 × 106 live MCA101-OVA cells.
On days 9 and 13 after tumor cell injection, mice were intravenously treated
with pSAP coupled with either anti-DEC205 or isotype control antibodies.
(B) FACS plots and bar graph showing the amount of pSAP uptake by
DCs at the tumor site as analyzed on day 14 after tumor challenge.
(C) Immunoblot of delivered pSAP and single saposins in pSAP-deficient
tumor DCs on day 20 after tumor cell inoculation. (D) Experimental setup
depicting tumor cell inoculation and pSAP targeting through DEC205 in WT
mice. (E) Comparison of tumor sizes on day 20 (left) and kinetics of tumor
growth (right). (F) Histogram overlay and bar graph showing H-2Kb-SIINFEKL
peptide surface staining and MFI on tumor DCs on day 20 post tumor
challenge. (G) FACS plots and bar graphs showing percentages of IFN-g–
positive CD8 T cells in tumors and dLNs in mice treated with pSAP coupled
to anti-DEC205 or isotype control. (H) Flow cytometry and bar graphs
showing MHC-I (Kb-SIINFEKL) tetramer (MHC-I-Tet)–positive CD8 T cells in
tumors and dLNs. (I) Experimental setup depicting B16F10 melanoma cell
inoculation, pSAP targeting through DEC205, and tumor growth kinetics.
WT mice were inoculated with 3 × 106 live melanoma cells and were treated

Sharma et al., Science 383, 190–200 (2024) 12 January 2024 9 of 11

RES EARCH | R E S E A R C H A R T I C L E
with pSAP coupled with anti-DEC205 or isotype control antibodies, either
alone or in combination with anti–PD-L1 antibodies. Statistical analysis of tumor
volume across all treatment groups is shown on day 32 after tumor challenge.
Percentages of gated cells are shown as mean ± SD in the dot plots in (B), (G),
well as frequency of IFN-g-producing, antigen-
specific T cells at the tumor site and in drain-
ing lymph nodes (dLNs) were significantly
amplified upon pSAP treatment (Fig. 5, F
to H).
Moreover, we investigated whether pSAP
therapy is able to amplify antitumor memory. To
this end, we first subcutaneously administered
1 × 106 MCA101-OVA tumor cells into the left
flank of WT mice prior to treatment with 100 mg
of pSAP coupled to anti-DEC205 or isotype
IgG (fig. S8I). When tumors reached a volume
of ~600 mm3, mice were anesthetized and
tumors were surgically removed, followed by
tumor cell rechallenge into the contralateral
side (fig. S8I). Monitoring tumor volume after
rechallenge revealed that DC-targeted pSAP
treatment significantly reduced tumor growth
(fig. S8J). This increased protection was re-
flected in a drastic expansion of antigen-specific
CD8 T cells at the tumor site (fig. S8K). These
data suggest that pSAP-based immunotherapy
drives immunological memory toward tumor
protection.
Because pSAP delivery to tumor DCs pro-
moted an immunologically active TME, we
next asked whether pSAP could also rescue
immune suppression in immunologically cold
tumors. For this purpose, we used the B16F10
melanoma model that displays limited T cell
infiltration and low susceptibility to treatment
with immune checkpoint inhibitors such as
anti–PD-1 despite strong expression of PD-L1
(26, 27). When compared with mice treated
with anti–PD-L1 alone, the tumor growth kinet-
ics revealed that pSAP combination therapy
was able to overcome the resistance of B16F10
melanoma to immune checkpoint blockade in
order to enable protection (Fig. 5I). Further-
more, pSAP combination therapy led to sig-
nificantly increased frequencies of total CD4
and CD8 T cells among various subsets of
tumor-infiltrating immune cells (fig. S9, A and
B). Notably, treatment with recombinant pSAP
and immune checkpoint blockade activated
CD8 T cells, amplified the frequencies of T
lymphocytes specific for the melanoma anti-
gen Pmel-1 (gp100), and facilitated robust
cytokine production (fig. S9, C and D).
On the basis of our results that genetic
ablation of TGF-b signaling in DCs fully abro-
gates pSAP hyperglycosylation and secretion,
we investigated whether restored antigen pre-
sentation could increase the susceptibility
of Tgfbr2DDC mice to anti–PD-L1 treatment
in the B16F10 melanoma model. Our tumor
challenge data show that the intrinsic loss
of TGF-b signaling in DCs indeed enhances the
Sharma et al., Science 383, 190–200 (2024) 12 January 2024
and (H). Data shown in all graphs are representative of three independent
experiments, and P values were determined with unpaired Student’s t test
[(B) and (E) to (H)] or one-way ANOVA (I). **P < 0.01; ***P < 0.001;
****P < 0.0001.
responsiveness to immune checkpoint blockade
(fig. S9E).
Taken together, these results highlight the
crucial impact of pSAP on antigen presenta-
tion by tumor DCs to trigger powerful intra-
tumoral T cell responses and point to a viable
future strategy for immunotherapy of cancer.
Discussion
In this work, we demonstrated the critical func-
tion of saposins in the processing of corpus-
cular, membrane-associated antigens required
for efficient CD8 T cell activation. Notably, sap-
osins were not essential for cross-presentation
of soluble antigen, which mechanistically in-
volves a rather early endosomal or phagosomal
compartment (6). In cancer immunity, the
provision of particulate antigen derived from
dying tumor cells represents a physiologic and
potent route of antigen delivery. In this con-
text, it has been shown that only DCs that
contain tumor-derived vesicles are able to in-
duce T cell responses (28). Thus, our findings
highlight the importance of the lysosome in
the cross-priming of T cells and underscore
the impact of saposins on the immunogenic
pathway of vesicular processing.
In tumor biology, previous reports suggested
a trophic function of pSAP stimulating the
proliferation of cancer cells (29, 30). However,
pSAP and a pSAP-derived synthetic cyclic
peptide were able to prevent tumor metastasis
in a mouse model (31, 32). Our work estab-
lished a genuine immunological function of
pSAP in tumor DCs. Thus, pSAP-driven anti-
gen processing and presentation at the tumor
site amplified abundance and functionality of
tumor-infiltrating T lymphocytes, ultimately
leading to cancer control.
TGF-b is a pleiotropic cytokine with a diverse
set of immunosuppressive functions and is
also produced by most cell types. Therefore,
tumors themselves, as well as the infiltrating
cells of the TME, can serve as the source and
target of TGF-b. For example, tumor-derived
TGF-b is capable of restricting T cell infiltra-
tion or functionally blocking differentiation
of protective T lymphocyte populations (15, 33).
Our results revealed that TGF-b acts on tumor
DCs to trigger hyperglycosylation of pSAP and
its subsequent secretion, depleting the lyso-
somal pool of saposins required for proper
antigen processing and presentation (fig. S10).
The secretion of pSAP has been shown to be
caused by pSAP oligomerization in a cell line
(34). However, in primary DCs in the tumor
microenvironment, we did not observe large
oligomers and instead found perturbed inter-
action between sortilin and hyperglycosylated
pSAP. An additional lysosomal player called
progranulin might be involved in this mechan-
ism, as it has been shown to bridge the inter-
action between pSAP and sortilin (35). Overall,
tumors are prone to immune escape, and our
work describes a further cancerous strategy
to manipulate an antigen processing molecule
through glycosylation.
Because hyperglycosylation of pSAP occurs
along the secretory pathway, an approach to
overcome tumor-induced saposin deficiency
entails the feeding of recombinant pSAP into
the endocytic route of DCs (fig. S10). In this
context, our targeting experiments using pSAP
p
coupled to anti-DEC205 demonstrated that
reconstitution with fully functional pSAP can
restore antigen presentation in tumor DCs,
leading to amplified T cell responses and
eventual tumor protection. DCs are not only
crucial for T cell priming in draining lymph
g
nodes; a growing body of evidence indicates
that tumor-associated DCs have vital func-
tions regarding recruitment of effector T cells
and stimulation of tumor-infiltrating T lym-
y
phocytes, which are overall required for ef-
fective immunity to cancer (36–39). Current
therapeutic options, such as immune check-
point blockade, aim at reinvigorating ex-
hausted T cells to augment antitumor responses
(40). However, these boosted T lymphocytes
need to encounter their respective antigens
to deploy their functions in a tumor-specific
manner. Consistent with that notion, we showed
that pSAP combination treatment can over-
y g
ride the resistance of immunologically cold
tumors to immune checkpoint inhibitors. Ta-
ken together, a pSAP-based therapy might
help to restore powerful antigen presenta-
tion by tumor DCs with the goal of driving
protective immune responses at the site of
y
the cancer.
,
REFERENCES AND NOTES
1. M. D. Vesely, M. H. Kershaw, R. D. Schreiber, M. J. Smyth,
Annu. Rev. Immunol. 29, 235–271 (2011).
2. M. S. Rooney, S. A. Shukla, C. J. Wu, G. Getz, N. Hacohen, Cell
160, 48–61 (2015).
3. M. Sade-Feldman et al., Nat. Commun. 8, 1136 (2017).
4. A. Alloatti et al., J. Exp. Med. 214, 2231–2241 (2017).
5. J. S. Blum, P. A. Wearsch, P. Cresswell, Annu. Rev. Immunol.
31, 443–473 (2013).
6. J. M. Blander, Annu. Rev. Immunol. 36, 717–753
(2018).
7. L. Shen, L. J. Sigal, M. Boes, K. L. Rock, Immunity 21, 155–165
(2004).
8. L. Saveanu et al., Science 325, 213–217 (2009).
9. T. L. Tang-Huau et al., Nat. Commun. 9, 2570 (2018).
10. K. Hildner et al., Science 322, 1097–1100 (2008).
11. S. T. Ferris et al., Nature 584, 624–629 (2020).
12. T. F. Gajewski, H. Schreiber, Y. X. Fu, Nat. Immunol. 14,
1014–1022 (2013).
10 of 11
RES EARCH | R E S E A R C H A R T I C L E

13. S. K. Wculek et al., Nat. Rev. Immunol. 20, 7–24 (2020).
14. R. A. Flavell, S. Sanjabi, S. H. Wrzesinski, P. Licona-Limón,
Nat. Rev. Immunol. 10, 554–567 (2010).
15. S. Mariathasan et al., Nature 554, 544–548 (2018).
16. S. Lefrancois, J. Zeng, A. J. Hassan, M. Canuel, C. R. Morales,
EMBO J. 22, 6430–6437 (2003).
17. L. Yuan, C. R. Morales, J. Histochem. Cytochem. 58, 287–300
(2010).
18. A. Darmoise, P. Maschmeyer, F. Winau, Adv. Immunol. 105,
25–62 (2010).
19. T. Kolter, K. Sandhoff, Annu. Rev. Cell Dev. Biol. 21, 81–103 (2005).
20. F. Ciaffoni et al., J. Biol. Chem. 276, 31583–31589 (2001).
21. F. Winau et al., Immunity 24, 105–117 (2006).
22. I. S. Zeelenberg et al., Cancer Res. 68, 1228–1235 (2008).
23. M. M. T. van Leent et al., Sci. Transl. Med. 13, eabe1433 (2021).
24. W. Jiang et al., Nature 375, 151–155 (1995).
25. J. Volckmar et al., J. Vis. Exp. (168): (2021).
26. I. J. Fidler, Cancer Res. 35, 218–224 (1975).
27. S. A. Quezada et al., J. Exp. Med. 205, 2125–2138 (2008).
28. M. K. Ruhland et al., Cancer Cell 37, 786–799.e5 (2020).
29. Y. Jiang et al., J. Pathol. 249, 26–38 (2019).
30. S. Ishihara, D. R. Inman, W. J. Li, S. M. Ponik, P. J. Keely,
Cancer Res. 77, 6179–6189 (2017).
31. S. Y. Kang et al., Proc. Natl. Acad. Sci. U.S.A. 106, 12115–12120
(2009).
32. S. Wang et al., Sci. Transl. Med. 8, 329ra34 (2016).
33. M. Liu et al., Nature 587, 115–120 (2020).
34. L. Yuan, C. R. Morales, Exp. Cell Res. 317, 2456–2467 (2011).
35. X. Zhou et al., Nat. Commun. 8, 15277 (2017).
36. H. Salmon et al., Immunity 44, 924–938 (2016).
37. S. Spranger, D. Dai, B. Horton, T. F. Gajewski, Cancer Cell 31,
711–723.e4 (2017).
38. J. E. Jang et al., Cell Rep. 20, 558–571 (2017).
39. K. C. Barry et al., Nat. Med. 24, 1178–1191 (2018).
40. P. Sharma et al., Cancer Discov. 11, 838–857 (2021).
ACKN OWLED GMEN TS
We thank C. Théry for providing us with OVA-expressing MCA101
fibrosarcoma cells. We are grateful to R. Cummings, S. Lehoux, and all
members of the National Center for Functional Glycomics (NCFG) at
Harvard Medical School for performing glycan mass spectrometry.
Illustrations used in the manuscript were created using BioRender.com.
Funding: This work was supported by National Institutes of Health
grant R01 AI136939 to F.W. Author contributions: P.S. and
F.W. designed the study and performed data analysis. P.S., X.Z.,
K.L., J.H.K., Q.W., J.K., M.L., and L.K. performed experiments.
L.T. produced recombinant pSAP and saposins. P.S. and F.W.
wrote the manuscript. Competing interests: F.W. and P.S. are
listed as inventors on a patent application filed by Boston
Children's Hospital (provisional application no. 63/350,734),
which covers the use of pSAP in cancer therapy. The authors
declare no conflict of financial interests. Data and material
availability: Gene expression data have been deposited to Gene
Expression Omnibus and are available under accession number
GSE248278. Mass spectrometry data are available from
GlycoPOST (ID: GPST000381). All data are available in the
manuscript or the supplementary materials. License
information: Copyright © 2024 the authors, some rights
reserved; exclusive licensee American Association for the
Advancement of Science. No claim to original US government
works. https://www.science.org/about/science-licenses-
journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.adg1955
Materials and Methods
Figs. S1 to S10
Tables S1 and S2
References (41–46)
MDAR Reproducibility Checklist
Submitted 7 December 2022; resubmitted 29 August 2023
Accepted 27 November 2023
10.1126/science.adg1955

p
g
y
y g
y
,

Sharma et al., Science 383, 190–200 (2024) 12 January 2024 11 of 11

RES EARCH

TRIBOLOGY or surface contamination. Second, a suitable

asperity-based friction model, able to predict
Designing metainterfaces with specified friction laws the global frictional behavior as the collective
response of the population of asperities, is iden-
tified. Depending on the expected relevant phys-
Antoine Aymard, Emilie Delplanque, Davy Dalmas, Julien Scheibert* ics, the model can range from analytical (20–22)
to numerical (23) through artificial intelligence
Many devices, including touchscreens and robotic hands, involve frictional contacts. Optimizing these [e.g., by extending the approaches of previous
devices requires fine control of the interface’s friction law. We lack systematic methods to create work (24, 25) to asperity-based descriptions of
dry contact interfaces whose frictional behavior satisfies preset specifications. We propose a generic surface topography]. Based on the inverted
surface design strategy to prepare dry rough interfaces that have predefined relationships between asperities’ geometries, a corresponding sam-
normal and friction forces. Such metainterfaces circumvent the usual multiscale challenge of tribology ple can be manufactured. The material and
by considering simplified surface topographies as assemblies of spherical asperities. Optimizing the characteristic size of the prescribed asperities
individual asperities’ heights enables specific friction laws to be targeted. Through various centimeter- contribute to the selection of the relevant man-
scaled elastomer-glass metainterfaces, we illustrate three types of achievable friction laws, including ufacturing method. Finally, shearing tests against
linear laws with a specified friction coefficient and unusual nonlinear laws. This design strategy the smooth counter surface enable determina-
represents a scale- and material-independent, chemical-free pathway toward energy-saving and tion of the resulting friction law, F(P), and,
adaptable smart interfaces. through direct comparison with Ftarget(P), as-
sessment of the overall reliability of the work-

D
flow. Discrepancies with respect to the target
espite centuries of investigation, a com- Results friction law may arise from an incomplete cal-

prehensive understanding of friction is
still lacking. For instance, predicting from
first principles the value of the friction
force, F, of a given dry contact interface
remains out of reach, mainly because of the
multiscale character of surfaces and the multi-
General interface design strategy
To bypass the multiscale and multiphysics chal-
lenges of friction along rough interfaces, we
propose a compromise between simplicity and
richness in the surface description, by con-
sidering flat-flat contact interfaces between a
p
ibration of the single asperity behavior, incor-
rect assumptions in the friction model, or
manufacturing imperfections.
Applying the strategy to centimetric
elastomer-glass interfaces

physics nature of contact interactions (1). Thus,
time- and resource-consuming experimental
tests remain necessary to calibrate the fric-
tional behavior of a contact interface as soon
smooth and a rough surface. Once the mate-
rial pair is fixed, the designable feature of the
interface is the topography of the rough sur-
face, which is built as an ensemble of indi-
g
Although the generic flowchart shown in Fig. 1 is
expected to be applicable to a large variety of
frictional interfaces, the following scientific
and technological choices represent only one

as any change is brought to the material pair,
shape of the solids, loading conditions, sur-
face finish, or environmental conditions. This
inability to master friction is a major obstacle
to the optimization of devices whose function
relies on dry contact interfaces. For soft inter-
faces, these devices include sport accessories
[e.g., racket coatings (2), shoe soles (3, 4)],
robotic grasping devices (5–7), haptic feedback
tools for virtual reality (8), and conveyor belts (9).
vidual microasperities, with both well-controlled
geometrical properties and calibrated contact
and friction behaviors against the smooth
counter surface. The richness of the emerging
macroscale behavior stems from the countless
possible combinations of geometrical proper-
ties of all individual asperities.
Just as the microstructures of the materials
can be engineered to provide metamaterials
with macroscale properties that are rarely
y
example of how our design strategy can be im-
plemented. First, we chose to work with con-
tacts between polydimethylsiloxane (PDMS)
and glass (26), a widely used pair of materials in
contact mechanics and friction studies (1, 11). The
behavior of sheared PDMS-glass, single sphere-
plane contacts has been extensively characterized
recently (27–30), which provides useful insights
into the single-asperity laws that should be used
in the friction model. In particular, the friction

At the present time, surface functionaliza-
tion is the main approach that is followed
worldwide to provide contact interfaces with
improved frictional capabilities (1, 10). It often
consists of creating a certain surface topog-
raphy at various length scales or adding a
found in nature [see (16–19) for mechanical
metamaterials], we propose a design strategy
(Fig. 1) to prepare contact interfaces with com-
plex predefined frictional behavior. We de-
note these as metainterfaces. The strategy starts
with a target friction law expressed as the
y g
force, F, is found to be proportional to the contact
area, A, through F = sA, where s is the frictional
strength of the PDMS-glass interface. Predicting
the macroscale friction force F is thus substan-
tially simplified because, for a given s, predicting
F reduces to predicting the total contact area

y
homogeneous or heterogeneous thin coating macroscopic relationship Ftarget(P) between the A of all microcontacts.

at the solid surface. Unfortunately, despite
many successes in a variety of specific cases
(11–15), this approach is still based on trial
and error. It does not offer a general, system-
atic design strategy to convert a set of fric-
tional specifications into an actual interface
that offers the expected behavior. By circum-
venting the main pitfalls that make it difficult
to understand friction in natural interfaces,
we propose and validate a general design strat-
egy to prepare multicontact interfaces with
on-demand frictional features.
Université de Lyon, École Centrale de Lyon, CNRS, ENTPE,
LTDS, UMR5513, 69130 Ecully, France.
*Corresponding author. Email: julien.scheibert@cnrs.fr
normal load P applied to the interface and
the target macroscopic friction force Ftarget.
This law is the input of an inversion step, the
output of which is a geometrical description
of the surface’s topography, including the num-
ber of necessary asperities and the list of their
individual properties (shape, size, height, po-
sition). The inversion is based on two main
ingredients. First, the indentation and fric-
tional behaviors of a single microcontact are
obtained through a preliminary calibration (top-
right illustration in Fig. 1), which may, but need
not, be captured by an existing tribological
model. Crucially, those calibrated behaviors
contain any specific effect due to the man-
ufacturing process, interface physicochemistry,
,
Second, as illustrated in Fig. 1, we chose to
use a square lattice of 64 asperities as sphe-
rical caps, with a common radius of curvature
R = 526 ± 5 mm and distributed summit
heights hi. We obtained them by molding a
PDMS slab onto an aluminum mold prepared
with deterministic spherical holes using a sphere-
ended cutting tool in a micromilling machine
(26) (for a typical image of a PDMS textured sam-
ple, see Fig. 1). This method of preparing popu-
lations of spherical PDMS asperities has been
successfully applied in the literature (31–33) but
not used to design interfaces with predefined
friction laws, as is done in this work. We have
calibrated the indentation and shear behavior of
single such microcontacts (26). Their normal

Aymard et al., Science 383, 200–204 (2024) 12 January 2024 1 of 5

RES EARCH | R E S E A R C H A R T I C L E

indentation is well captured by the classical
Hertz model of a linear-elastic sphere–rigid plane
contact (34) (Fig. 2B), with a composite elastic
modulus E* (E
¼ 1n E
2 , where E and n are the
Young’s modulus and Poisson’s ratio of the
PDMS, respectively). Consistent with previous
work (27–30), the contact region is initially
circular with an area a0 but decreases aniso-
tropically under shear, by ~10 to 15%, down to
an area af (Fig. 2A) at the onset of sliding (sta-
tic friction). The ratio B = af /a0 thus represents
Target macroscale friction law: Ftarget(P)
DESIGN STRATEGY
Definition of relevant Calibration of
the fraction of the initial contact area that re-
mains when the contact is sheared and just
about to slide. We found B to be independent of
a0 (Fig. 2C). We also confirmed the propor-
tionality between the contact area and static
friction force, f = saf (Fig. 2D).
The third choice we made was the type of
friction model to be inverted. For the present
proof of concept, we found that a very simple
linear-elastic asperity-based model was suffi-
cient. We considered N spherical linear-elastic
f p
asperities with common curvature radius, R,
and composite elastic modulus, E*, but dis-
tributed summit heights, hi. This population
of asperities is brought into contact with a
rigid smooth plane, forming a number of mi-
crocontacts, which are each assumed to obey
both Hertz’s model (34) under pure compres-
sion and the calibrated friction behavior under
shear. Microcontacts are assumed to be inde-
pendent [see (26) for justification] such that
their individual in-plane locations become ir-
relevant. The area of the ith microcontact is
expressed as a0,i = pR(hi – d) if hi ≥ d and a0,i =
0 if hi < d, where d is the separation between
the rigid indenting plane and the base plane
that supports the asperities. The macroscopic
XN
quantities are simple sums: A0 ¼ a ,
i¼1 0;i
XN XN
3=2
Af ¼ a ,P ¼
i¼1 f;i 3=2 a0;i (Hertz’s
4E

asperity-based indentation & friction of XN i¼1 3p R

single microcontacts P model), and F ¼ f ¼ sAf , where Af =
i¼1 i
friction model

p
BA0. The model parameter values were exper-
F imentally calibrated (26). Inverting the above
Inversion of the friction model consists in determining a suit-
friction model able list of heights hi such that the predicted
R friction law satisfies specifications set a priori.
Number of asperities
hi Interestingly, in the model, the friction law

g
F
+ list of individual F(P) can be rescaled as sB versus EP
, where the
F P
geometrical function relating sB and E
depends only on the
properties geometrical parameters of the asperities, R
and hi, and not on the material parameters, s,

y
B, and E*. Thus, in the following three exam-
Surface manufacturing
& friction test ples, we illustrate the success of our design
1cm
strategy using
 the material-independent rela-
F P
tionship sB E
. This rescaled friction law char-
acterizes solely the effect of surface topography,
Actual macroscale friction law: F(P) which is the actual interfacial quantity on
which the design is performed.
Fig. 1. Flowchart of the design strategy. Shown at the top right is an illustration of a single microcontact
submitted to a normal force p and a friction force f. An illustration of a metainterface submitted to Building a friction law from operating points
normal and friction forces, P and F, is shown at the middle right. Dark gray ellipses represent real contact For our first example, we consider specifica-

y g
regions. The topography is made of N asperities, each with specifically designed geometrical properties
tions
Pi Fi
in
 terms of a list of operating points
 
F P
E
; sB through which the friction law sB E

(here, spherical caps of height hi and curvature radius R). At the bottom right is an image of a typical
centimetric elastomer-based realization of such a textured sample. Photographs of metainterfaces can be must pass. An infinity of suitable lists of as-
found in Figs. 3 to 5. perity heights hi exist as the solution to this

y
A B 0.2 C D
,
a0 af 1 0.4
a0 (mm2 )

af (mm2 )

f (N)

0.1 B
0.5 0.2

0 0 0
0 0.05 0.1 0.15 0 0.5 1 0 0.5 1
0.5 mm p 2/3
(N 2/3
) a0 (mm2) af (mm2)

Fig. 2. Tribological calibration of single microcontacts. Illustrations for PDMS
batch 1. (A) Images of a microcontact under normal load p = 0.061 N with no
shear, initial area a0 (dark central region) (left) and at the onset of sliding, area af
(tangential force applied upward in the image plane) (right). (B) a0 versus p2/3,
for three indentation experiments using three different samples (data points). The
 2=3
solid line is Hertz’s prediction a0 ¼ p 3Rp
4E
for R = 526 mm and E* = 1.36 MPa
[see (26) for information about how those values were determined]. (C) Typical
evolution of af as a function of a0 (data points). The solid line is the linear fit
showing the existence of a constant area reduction ratio B = af /a0. (D) Typical
friction force f versus af (data points). The solid line is the linear fit showing the
existence of a constant friction strength s = f/af. Error bars are ±5 mN for p,
pffiffiffiffiffi
±(1 mN + 0.01f) for f, and ±2:8 pa  106 m2 for a0 and af.

Aymard et al., Science 383, 200–204 (2024) 12 January 2024 2 of 5

RES EARCH | R E S E A R C H A R T I C L E

5 problem, each giving a different shape of the

h1 > h 2 > h3 friction law between operating points. For
pedagogical purposes, we adopted an inversion
strategy (26) in which each operating point
4 is reached using a single level of asperity height
(fig. S1). To pass from the first (trivial) operating
point (0, 0) to the next, one evaluates the max-
h2 imum number of asperities with identical heights
h1 &
F/ B (mm2)

3 (first height level) required to approach the op-

ly
h 1 on erating point from below and adds a single
adjustment asperity whose height is selected to
reach exactly the desired operating point. The
2 procedure is repeated as many times as the
number of additional operating points. As a re-
sult, upon normal loading, a prescribed number
of new asperities enter the contact as soon as
1 the previous operating point is reached. We
applied this procedure to design a metainter-
Real contact
face with a predicted three-branched friction
behavior passing through three nonaligned op-
0 erating points (Fig. 3). Our measurement points

p
0 0. 2 0.4 0. 6 0.8
P/E* (mm2 ) delineate a friction law that closely approaches
the three target friction forces to better than 5%
Fig. 3. Designing an interface that reaches three preset frictional operating points. The green, blue, and (2.5% for the first two points). This agreement
red disks represent targeted operating points. The black points represent measured friction law sB F
versus EP
. Error shows that the design strategy can be used suc-
bars are calculated as described in (26). The solid line is the model prediction. Dashed lines are continuations cessfully to prepare real-life metainterfaces that
of the two first branches of the solid line if the subsequent height levels of the asperities had not been populated. target a nontrivial list of friction forces.

g
The top-left inset shows a full image of the unsheared contact under P = 0.92 N. Asperities at each of the The latter design strategy can be extended
three height levels (h1 > h2 > h3) are circled with the same color as their associated operating point (main panel). to an arbitrary number of operating points
Scale bar is 1.5 mm. The bottom-right inset shows a magnified view of the real contact of nine microcontacts (although sufficiently smaller than the number
(the region outlined by the dashed line in the top-left inset). N of available asperities), opening the way to the

y
design of interfaces with friction laws defined by
many operating points. We expect that inver-
3.5
sion procedures that are more versatile than
B the pedagogical one we present will be devel-
µ~
~ m = 2.41
E* f oped, for instance, those that invert not only
3 mf
hi but also other geometrical quantities such
as the curvature radius of each asperity.
2.5
Specifying the friction coefficient at a constant
material pair
mt

y g
2 For our second example, we acknowledged
that the most classical descriptor of the fric-
F/ B (mm2)

mf tional properties of an interface is the value

1.5 of its friction coefficient, m. Commonly con-
sidered to be a characteristic of a pair of
mt materials in contact, the friction coefficient

y
1 does not relate to a particular operating point

,
but rather to the global shape of the friction
law. Indeed, unlike the curve shown in Fig. 3, in
0.5 B most natural or human-made interfaces, the
µ~
~ m = 1.51
E* f friction law F(P) is found to be linear [Amontons-
Coulomb friction (1, 35)], with m being the slope
0
0 0.1 0.2 0.3 0.4 of the law. Hence, we targeted linear friction
P/E* (mm2 ) laws with tunable slopes.
For surface topographies made of spherical
Fig. 4. Specifying the friction coefficient at a constant material pair. Experimental friction laws (data points) asperities that obey Hertz’s indentation law,
of two metainterfaces with exponential-like distributions of asperity heights (table S1). Error bars are as described as in our metainterfaces, a statistical frame-
in Fig. 3. Lines are linear fits of the data [over the last 12 (7) points for the blue (red) data]. The red and work relating the contact area A0 to the normal
blue shaded regions around the lines represent the 68% confidence interval on the linear regression. The fitted load P exists [see Greenwood and Williamson’s
slopes mf of the rescaled laws, 6.73 ± 0.22 (red) and 9.53 ± 0.25 (blue) [which correspond to friction coefficients model (20)]. This framework predicts a linear
E
mf ¼ 1:51 and 2.41 that are typical of PDMS-glass rough contacts (27, 29)] match the targeted slopes
m ≈ sB A0(P) relationship when the asperities’ heights,
(dashed lines), mt = 6.87 (red) and 9.43 (blue), to better than 2.1%. The insets show typical photographs of the hi, follow a probability distribution that is ex-
two metainterfaces under pure normal force P = 0.39 N (blue) or 0.93 N (red). Scale bars are 1.5 mm. ponential (e−h/l/l, where l is the scale parameter

Aymard et al., Science 383, 200–204 (2024) 12 January 2024 3 of 5

RES EARCH | R E S E A R C H A R T I C L E

of the exponential). We adapted this model to
account for the existence of a maximum asperity
height hm due to the finite number of asperities
and extended it to predict linear-like friction laws
[details in (26)]. In particular,
  we derived the
F P
value of the slope of sB
qffiffiffiffi E
for large P: m ¼
pffiffi
x
pffiffi
l erf ð xÞp2ffip xex , where x = hm/l and erf is
pR 1e
the error function. m is a rescaled estimator for
E

the friction coefficient of the interface (m ≈ sB m).
For illustrative purposes, we fixed R and hm
and targeted two different slopes m. Using the
above expression of m, we identified the two l
that were suitable to generate two metainter-
faces, each with one of the target slopes (26). The
two friction laws that we obtained are linear,
and their slopes match the predictions to better
than 2.1% (Fig. 4). This agreement validates that
our design strategy quantitatively enables the
prescription of the slope of a linear friction law
[after a nonlinear part at small P that is due to
for centuries. Thus, controlling friction based on
generic principles is often considered unreach-
able. Our work demonstrates that the friction
of dry elastomer-based metainterfaces can ac-
tually be finely tuned when the surface topog-
raphy is designed according to our proposed
strategy (Fig. 1). We emphasize that our three
examples far from exhaust the myriad possible
variations around our design strategy, which
provides a simple framework that enables
friction-control opportunities, starting from,
but not limited to, the universally used friction
coefficient. Such topography-based control con-
firms, in particular, that the static friction co-
efficient is not a constant for a given pair of
materials, as predicted by models (22, 23, 36)
and observed in stick-slip experiments (37).
The potentially accessible friction laws F(P)
are countless, in terms of both their nonlinear
shape and the way specifications are expressed
(forces, slopes, etc.). Those alternative, richer
approaches proposed here as pedagogical
illustrations.
Static friction is not the only possible tar-
get functionality, because the design strategy
can presumably be adapted to control other
topography-related quantities, including inter-
facial stiffnesses and adhesion forces. Such
extensions require different types of calibra-
tions of individual microcontacts that are ded-
icated to those target properties. In this respect,
note that in our experiments, the friction de-
sign strategy can be applied indifferently to
the static friction force (as in Figs. 3 to 5) or
sliding friction force [see (26)].
Our design strategy should be applicable to
other material pairs than PDMS and glass. For
materials whose contact remains elastic [see
(26) for a criterion], the approach developed
in our examples should hold, with material
changes being accounted for by the rescaled,
material-independent
  version of the friction

the finite value of hm/l and is responsible for a
nonvanishing intercept; see (26)]. The slopes m
that
 we  achieved for the rescaled friction law
F P
sB E
differ by a ratio of 1.42 (1.60 for the fric-
tion coefficients m≈ sB E
m). For a discussion about
the range of achievable friction coefficients, see
prescriptions likely require the inversion of
more topographic features than the individ-
ual asperities’ heights, including their individ-
ual radii of curvature and in-plane locations.
Such additional degrees of freedom should
prompt the development and use of advanced
p
F P
law, sB E
. For nonelastic materials, qualita-
tive changes might need to be brought to the
calibrated microcontact behavior laws, friction
model, and related inversions. For instance, for
elastoplastic contacts, enriched asperity-based
friction models are required (40, 41). In addi-

(26). We emphasize that the results provide a
practical way to tune the friction coefficient of a
contact interface by manipulating the surface
topography only, that is, without bringing any
g
friction models (22, 38) and inversion tools tion, the plasticity-induced irreversible mod-
(25, 39) that are not limited to the analytical ifications of the topography will likely affect all
2

y
change to the bulk materials or physicochem-
istry of their surfaces.

Building friction laws with multiple specified m 2,f

m 2,t
linear branches 1.5
For our third example, we show that unnatu-
ral, nonlinear-shaped friction laws can also be m 1,f
designed, enabling one to go beyond Amontons-
Coulomb–like friction. As an illustration, we
F/ B (mm2)

targeted a succession of two linear branches 1

y g
with specified slopes, m1 and m2, and a cross-
over at a specified critical normal load, Pc/E*.
We first describe how such a friction law can
be obtained [see (26)] using a weighted expo-
nential distribution of heights, which gener- 0.5
ates two subpopulations of asperities. We then Pc,f /E*

y
explain how to invert the law and find a suit- m 1,t

able list of asperity heights, hi [see (26)].
We demonstrate that such bilinear-shaped
friction laws can indeed be achieved (Fig. 5)
through two different realizations. As targeted,
both realizations share the same two slopes,
which are different by a ratio of 1.75, but have
two different crossover normal loads, Pc/E*
(ratio 1.94). The slopes values are fulfilled by
better than 5.2% (10% for the ratio of Pc/E*),
further demonstrating experimentally the suc-
cess of our design strategy in satisfying quan-
titatively nontrivial friction prescriptions.
Discussion
Understanding friction in its generality re-
mains a formidable challenge and has been
,
Pc,t /E*
0
0 0.05 0.1 0.15 0.2
2
P/E* (mm )
Fig. 5. Friction law made of two specified linear branches. Two experimental realizations of bilinear friction
laws (data points) based on height distributions given by eq. S6 (table S2). Error bars are as described in Figs. 3
and 4. Solid lines are linear fits of each branch [six (four) first data points for the first branch of the blue (red)
curve; eight (seven) last data points for the second branch of the blue (red) curve]. The red and blue shaded
regions around the lines are as in Fig. 4. The fitted slopes mf of both linear-like branches in each curve match the target
slopes mt (dashed black lines), 7.15 and 12.52, to better than 5.2%. Vertical lines show target (Pc,t/E*, dashed
lines) and observed (Pc,f/E*, solid lines) crossover normal loads (in the same colors as the corresponding curves).
Insets show typical photographs of the metainterfaces that underlie both friction laws under pure normal
force P = 0.23 N (red) or 0.34 N (blue). The frame color refers to that of the corresponding friction law. Scale bars
are 1.5 mm. See movie S1 for a qualitative illustration experiment on analogous metainterfaces.

Aymard et al., Science 383, 200–204 (2024) 12 January 2024 4 of 5

RES EARCH | R E S E A R C H A R T I C L E
successive uses of the metainterface such that
the friction law is no longer a constant but
rather a loading history–dependent function.
Irreversible changes of the topography may
also arise as a result of wear-induced loss of
material. Upon sliding, various wear mech-
anisms may occur, including progressive mate-
rial removal that leads to blunted asperities
and asperity-sized debris formation due to frac-
ture processes (42). With such alterations of
the surface topography, the behavior of meta-
interfaces will progressively deviate from the
target one. In this context, evaluating the
lifetime of metainterfaces as a function of
the materials, roughness, and loading condi-
tions will be an important step before use in
any specific engineering application.
We expect our design strategy to be appli-
cable over a large variety of length scales. We
used asperities with a submillimetric radius of
curvature that was obtained using micromil-
ling. Larger asperities can be obtained in the
same way or, for example, through 3D printing
(43). For smaller scales, other well-established
methods might be suitable, from laser ablation
for asperities down to the micrometer scale
(44) to micro- or nanolithography techniques
at submicrometer scales (45). In all cases, ex-
perimental challenges include the reproducibil-
ity of the shapes of asperities (for microcontact
calibration to be relevant), the level of accuracy
of the manufactured asperity heights with re-
spect to their prescribed values, and the po-
tential misalignment between the two surfaces
during the friction measurements [see (26) for
a discussion of the latter two issues].
Many different types of frictional metainter-
faces can be developed by using our design strat-
egy, constituting a useful toolbox for designers of
friction-based devices. By operating in optimized
conditions, these devices should benefit from in-
creased energy efficiency and lifetime. If further
equipped with suitable sensors or actuators that
Aymard et al., Science 383, 200–204 (2024) 12 January 2024
bring relevant real-time changes to the topog- 34. J. Barber, Contact Mechanics, Solid Mechanics and Its
raphy, these metainterfaces could also become Applications Series, vol. 250 (Springer, 2018).
35. E. Popova, V. L. Popov, Friction 3, 183–190 (2015).
an asset in the development of smart systems 36. J. Scheibert, D. K. Dysthe, Europhys. Lett. 92, 54001 (2010).
(46) that incorporate functional solid contacts. 37. O. Ben-David, J. Fineberg, Phys. Rev. Lett. 106, 254301
(2011).
38. Y. Xu, J. Scheibert, N. Gadegaard, D. M. Mulvihill, J. Mech.
RE FERENCES AND NOTES Phys. Solids 164, 104878 (2022).
1. A. I. Vakis et al., Tribol. Int. 125, 169–199 (2018). 39. T. Djourachkovitch, N. Blal, N. Hamila, A. Gravouil,
2. M. Varenberg, A. Varenberg, Tribol. Lett. 47, 51–56 (2012). Comput. Struc. 255, 106574 (2021).
3. L. Damm et al., Footwear Sci. 6, 155–164 (2014). 40. W. R. Chang, I. Etsion, D. B. Bogy, J. Tribol. 109, 257–263
(1987).
4. J. Worobets, J. W. Wannop, Sports Biomech. 14, 351–360 (2015).
5. M. R. Cutkosky, P. K. Wright, Int. J. Robot. Res. 5, 20–37 (1986). 41. H. Ghaednia et al., Appl. Mech. Rev. 69, 060804 (2017).
6. A. J. Spiers, B. Calli, A. M. Dollar, IEEE Robot. Autom. Lett. 3, 42. R. Aghababaei, D. H. Warner, J.-F. Molinari, Nat. Commun. 7,
4116–4123 (2018). 11816 (2016).
7. Y. Golan, A. Shapiro, E. Rimon, IEEE Robot. Autom. Lett. 5, 43. R. Sahli et al., Sci. Rep. 10, 15800 (2020).
2889–2896 (2020). 44. B. Mao, A. Siddaiah, Y. Liao, P. L. Menezes, J. Manuf. Process.
8. G. Millet, M. Otis, D. Horodniczy, J. R. Cooperstock, 53, 153–173 (2020).
Mechatronics 46, 115–125 (2017). 45. Z. Zhang, C. Geng, Z. Hao, T. Wei, Q. Yan, Adv. Colloid Interface
9. H.-W. Zhu et al., Powder Technol. 367, 421–426 (2020). Sci. 228, 105–122 (2016).
10. T. Ibatan, M. S. Uddin, M. A. K. Chowdhury, Surf. Coat. Tech. 46. M. A. McEvoy, N. Correll, Science 347, 1261689 (2015).
272, 102–120 (2015).
11. J. Scheibert, S. Leurent, A. Prevost, G. Debrégeas, Science AC KNOWLED GME NTS
323, 1503–1506 (2009).
We thank A. Malthe-Sørenssen, K. Thøgersen, H. A Sveinsson, C. Oliver,
12. B. Murarash, Y. Itovich, M. Varenberg, Soft Matter 7, 5553–5557 (2011).
M. Guibert, D. Bonamy, and G. Pallares for discussions. We thank
13. M. J. Baum, L. Heepe, E. Fadeeva, S. N. Gorb, Beilstein J.
D. Roux and N. Morgado for their help with the experimental device
Nanotechnol. 5, 1091–1103 (2014).
used in movie S1 and N. Morgado for the sketches in Fig. 1. We thank
p
14. N. Li, E. Xu, Z. Liu, X. Wang, L. Liu, Sci. Rep. 6, 39388 (2016).
Z. Lin and X. Qi for their help in developing the inversion scheme
15. J. Dufils et al., Surf. Coat. Tech. 329, 29–41 (2017).
for friction laws defined by operating points. Funding: This work was
16. M. Kadic, T. Bückmann, R. Schittny, M. Wegener, Rep. Prog. Phys.
funded by LABEX iMUST (grant ANR-10-LABX-0064) of the Université
76, 126501 (2013).
17. B. Florijn, C. Coulais, M. van Hecke, Phys. Rev. Lett. 113, 175503 (2014).
de Lyon within the program “Investissements d’Avenir” (grant
ANR-11-IDEX-0007) operated by the French National Research Agency
18. K. Bertoldi, V. Vitelli, J. Christensen, M. van Hecke, Nat. Rev. Mater.
(ANR) (J.S.), the ANR through grant ANR-18-CE08-0011 (PROMETAF
2, 17066 (2017).
project) (J.S.), and the Institut Carnot Ingénierie@Lyon (PREGLISS project)
19. J. U. Surjadi et al., Adv. Eng. Mater. 21, 1800864 (2019).
(D.D.). Author contributions: Conceptualization: J.S.; Methodology:
g
20. J. A. Greenwood, J. B. P. Williamson, Proc. R. Soc. London Ser. A
A.A., J.S., D.D.; Investigation: A.A., J.S., E.D.; Visualization: A.A., J.S.,
295, 300–319 (1966).
D.D.; Writing – original draft: J.S.; Writing – review and editing:
21. O. M. Braun, M. Peyrard, Phys. Rev. Lett. 100, 125501 (2008).
A.A., D.D., J.S.; Funding acquisition: J.S., D.D. Competing interests:
22. K. Thøgersen, J. K. Trømborg, H. A. Sveinsson, A. Malthe-Sørenssen,
The authors declare that they have no competing interests. Data and
J. Scheibert, Phys. Rev. E Stat. Nonlin. Soft Matter Phys. 89,
052401 (2014).
materials availability: All data are available in the main text or the
y
supplementary materials. License information: Copyright © 2024
23. J. K. Trømborg et al., Proc. Natl. Acad. Sci. U.S.A. 111, 8764–8769 (2014).
the authors, some rights reserved; exclusive licensee American
24. G. Yang, Q. J. Li, Y. Zhan, Y. Fei, A. Zhang, J. Comput. Civ. Eng.
Association for the Advancement of Science. No claim to original US
32, 04018052 (2018).
25. P. Cinat, G. Gnecco, M. Paggi, Front. Mech. Eng. 6, 29 (2020).
government works. https://www.science.org/about/science-licenses-
journal-article-reuse
26. Materials and methods are available as supplementary materials.
27. R. Sahli et al., Proc. Natl. Acad. Sci. U.S.A. 115, 471–476 (2018).
SUPPLEMENTARY MATERIALS
28. J. C. Mergel, R. Sahli, J. Scheibert, R. A. Sauer, J. Adhes. 95,
1101–1133 (2019). science.org/doi/10.1126/science.adk4234
29. R. Sahli et al., Phys. Rev. Lett. 122, 214301 (2019). Materials and Methods
Figs. S1 to S3
30. J. Lengiewicz et al., J. Mech. Phys. Solids 143, 104056 (2020).
Tables S1 and S2
31. V. Romero, E. Wandersman, G. Debrégeas, A. Prevost,
References (47–51)
Phys. Rev. Lett. 112, 094301 (2014).
Movie S1
32. S. Yashima et al., Soft Matter 11, 871–881 (2015).
y g
33. V. Acito, M. Ciavarella, A. M. Prevost, A. Chateauminois, Submitted 21 August 2023; accepted 22 November 2023
Tribol. Lett. 67, 54 (2019). 10.1126/science.adk4234
y
,
5 of 5
RES EARCH

IMMUNOLOGY We next interrogated the usage of each codon

within IGHM relative to the human protein-
Antibody production relies on the tRNA inosine coding genome by calculating the probability
of finding a CDS with a lower usage frequency.
wobble modification to meet biased codon demand Overall, IGHM used few codons at a frequency
comparable with that of the average human
Sophie Giguère1*, Xuesong Wang1, Sabrina Huber2†, Liling Xu1, John Warner1, Stephanie R. Weldon1, gene: P(Usage) of 0.5 indicates standard usage;
Jennifer Hu2, Quynh Anh Phan1, Katie Tumang1, Thavaleak Prum1, Duanduan Ma3, Kathrin H. Kirsch1, 1.0 indicates a higher usage in IGHM than in all
Usha Nair1, Peter Dedon2,4, Facundo D. Batista1,5,6,7* other CDS (Fig. 1C). Thus, the pattern of codon
usage in IGHM is distinct from the general
Antibodies are produced at high rates to provide immunoprotection, which puts pressure on the B cell biases of human protein-coding genes.
translational machinery. Here, we identified a pattern of codon usage conserved across antibody genes. One All members of the IgH family have high SCUO
feature thereof is the hyperutilization of codons that lack genome-encoded Watson-Crick transfer RNAs scores (fig. S1D). We found the pattern identified
(tRNAs), instead relying on the posttranscriptional tRNA modification inosine (I34), which expands the in immunoglobulin M (IgM) throughout: IgH
decoding capacity of specific tRNAs through wobbling. Antibody-secreting cells had increased I34 levels isotype P(Usage) scores were highly correla-
and were more reliant on I34 for protein production than naïve B cells. Furthermore, antibody ted, indicating a conserved pattern of codon
I34-dependent codon usage may influence B cell passage through regulatory checkpoints. Our work bias within the family (Fig. 1, C and D).
elucidates the interface between the tRNA pool and protein production in the immune system and has To interrogate whether this pattern was asso-
implications for the design and selection of antibodies for vaccines and therapeutics. ciated with protein structure, we examined the
T cell receptor (TCR) constant-region family.

P
Like IgHs, TCRs are antigen receptors with a

rotein production is among the most ener-
getically costly and stressful processes that
cells undertake (1). Efficient translation
is critical in cell fitness and function, par-
ticularly for highly expressed proteins (2).
Antibodies must be synthesized in substan-
proteotoxic stress and energetic costs (7–10).
Although a correlation between codon usage
of abundant proteins and tRNA gene copy
numbers—often used as a proxy for abundance
but weakly correlated—is well documented
in single-celled organisms, the interactions be-
p
cell-specific variable region and universal con-
stant region and structurally resemble the Fab
portion of immunoglobulins (18). The codon
usage patterns of human TCR constant re-
gions did not correlate with those of IgH iso-
types, except for TCRb, which was still less

tial quantities to provide systemic protection
against infection (3). Initially expressed in the
hundreds of thousands per cell as membrane-
bound immunoglobulin B cell receptors (BCRs),
tween tRNA supply and codon demand may be
less straightforward in multicellular eukaryotes
(11–14). The tRNA pool of metazoan cells is
dynamic, and changes in tRNA expression and
g
similar to any IgH isotype than the IgH iso-
types were to each other (Fig. 1, D and E).
To determine whether this specific pattern
of codon bias was conserved, we compared

antibodies are secreted by plasma cells (PCs)
at rates of up to 100 million molecules per
hour, per cell, for years (4, 5). Antibody pro-
duction in PCs is tightly regulated, requiring
dramatic alterations in cellular biology. Key
processes include the coordinated activity of
multiple transcription factors, metabolic changes,
and expansion of the endoplasmic reticulum
(ER) and Golgi complex because cell survival
requires the up-regulation of unfolded pro-
modifications affect cell differentiation ver-
sus proliferation, as well as tumor metastasis
(12, 15–17). Given the importance of antibody
production to immunity, we investigated whe-
ther antibody-secreting cells have tRNA pools
tailored to immunoglobulin production.
Antibody constant-region genes demonstrate
a biased codon pattern that does not predict
the corresponding tRNA
y
human and murine IgH constant regions by
using hierarchical clustering analysis of codon
P(Usage) for all IgH isotypes (Fig. 1F). Across all
isotypes, human and murine IgH sequences
shared a set of 12 hyperutilized codons, as well
as 11 hypoutilized relative to the correspond-
ing genome. Most notably, ACC (Thr) had high
P(Usage) in both mice and humans. ACC (Thr)
appeared, on average, more frequently in IgH
sequences than in nearly any human [P(usage) =

tein response factors and ER chaperones to
prevent toxic antibody aggregation (6). How-
ever, little is known about how the transla-
tional machinery adapts to antibody production.
tRNAs, the primary interpreters of the gen-
etic code, are central and dynamic compo-
The human immunoglobulin heavy chain (IgH)
constant region comprises the bulk of the anti-
body protein and, consequently, its production
demand. Unlike the antigen-recognizing vari-
able region, IgH has not evolved for hyper-
mutability and remains relatively consistent
y g
0.995] or murine (0.96) CDS. Murine and hu-
man IgH genes shared a characteristic codon
bias pattern.
To assess the ability of the tRNA pool to match
immunoglobulin codon patterns, we examined
gene copy numbers for tRNAs corresponding

y
nents of translation. The adaptation of codon across cells and individuals. to codons hyperutilized in IgH genes in both

usage to the abundance of corresponding
tRNAs can affect translational efficiency and
accuracy. Perturbations of this balance induce
1
The Ragon Institute of Mass General, Massachusetts
Institute of Technology (MIT), and Harvard, Cambridge, MA
02139, USA. 2Department of Biological Engineering, MIT,
Cambridge, MA 02139, USA. 3BioMicro Center, MIT,
Cambridge, MA 02139, USA. 4Singapore-MIT Alliance for
Research and Technology, Singapore 138602. 5Department
of Immunology, Harvard Medical School, Boston, MA 02115,
USA. 6Department of Microbiology, Harvard Medical School,
Boston, MA 02115, USA. 7Department of Biology, MIT,
Cambridge, MA 02139, USA.
*Corresponding author. Email: giguere.sophie@gmail.com (S.G.);
fbatista1@mgh.harvard.edu (F.D.B.)
†Present Address: ETH Zürich, Laboratory of Toxicology, 8092 Zürich,
Switzerland.
In human IGHM—the first isotype expressed
and, at 49 kDa, the largest—the frequencies of
the 61 sense codons varied substantially, in-
cluding within synonymous codon sets (Fig.
1A and fig. S1A). Seven codons were not used,
despite the presence of all standard amino
acids (Fig. 1A and fig. S1, A and B). Each of the
10 most common codons in IGHM encoded a
different amino acid and was far more fre-
quent than its synonyms, suggesting strong
bias (Fig. 1A and fig. S1C). IGHM synonymous
codon usage order (SCUO), which quantifies
gene-wide codon bias, was within the top 10%
of human coding sequences (CDS) (Fig. 1B),
suggesting potential selective pressure on IGHM
codon use.
,
species. Eight had multiple directly correspond-
ing tRNA genes for the anticodon (Fig. 1, G
and H, and fig. S1F). However, four hyperutil-
ized codons lacked corresponding tRNA genes
according to classic Watson-Crick base pairing
rules: GTC (Val), TCC (Ser), CCC (Pro), and ACC
(Thr) (Fig. 1, G and H, and fig. S1F). Genome-
encoded tRNAs were incongruous with the IgH
pattern of codon bias. Thus, to translate anti-
body genes, cells must reprogram available tRNAs
to meet codon demand.
tRNA modifications in antibody-secreting cell
lines expand decoding capacity
tRNA decoding capacity is determined by the
genome-encoded primary sequences of the

Giguère et al., Science 383, 205–211 (2024) 12 January 2024 1 of 7

RES EARCH | R E S E A R C H A R T I C L E

A B C IGHM R2 = 0.90
P < 2e-16
1.00
0.06
0.6

P(Usage) Score
Frequency

0.75

SCUO
0.04
0.4
0.50

0.02
0.2
0.25

0.00
0.0
0.00
CCG
GCG

TCG
TTG

ATG
AGG
CGG
TGG

ACG
GGG

CTG

GTG
CGC

CTC

TGC

ATC
GGC
GTC
AGC
CCC
TTC

GCC
TCC
ACC
CGA

GGA

AGA
CAA

TCA

GAA

AAG
ACA

AAA

CAG
GAG
CCA
CGT

GTT
TTT
ACT
ATT

TGT
AGT

GCA
GGT
CTT

GCT
CCT
TCT

TAC

CAC

AAC

GAC
ATA

CTA
GTA
TTA

CAT

TAT

AAT

GAT

M
om
Codon

H
IG
en
G
Codon
D IGHA R2 = 0.82
P < 2e-16 IGHD R2 = 0.59
P < 3e-13 IGHE R2 = 0.71
P < 2e-16 IGHG R2 = 0.90
P < 2e-16
1.00
P(Usage) Score

0.75

0.50

0.25

0.00

p
Codon Codon Codon Codon
E TCR alpha R2 = 0.0091
TCR beta R2 = 0.41
TCR delta R2 = 0.020 TCR gamma R2 = 0.00070
P = 0.21 P < 7e-9 P = 0.14 P = 0.31
1.00
P(Usage) Score

0.75

0.50

g
0.25

0.00
Codon Codon Codon Codon

y
GGG

TAG
CGG

GCG

CCG
AGG

GAG
GGC
ATG

CAG

TGG

GTG
TAC
CGC

GCC

A CG
CTG

TCG
TAT

CCC
AGC
GAC
ATC

TGC

GTC
TTG
GAT

GTA

ATA
GGT

CAC

ACC
CTC

TCC
GCT

CAT

CTA

CGT

AAG
GGA

CCT
AGT

TTC
GCA

CGA
ATT

GTT

TGT
TTA
A CT

CCA
CTT

TCT
AGA

AAC
AAT

TAA

TGA
A CA
TCA
TTT

GAA

CA A

AAA

Codon Order:
G
F 20 N

IgG2
tRNA Gene Count

IgG4 15 Average
IgG1 P(Usage)
IgG3 Score
IgA2 V 1.00
10 F
IgA1 VSPT 0.75
S 0.50
IgE W P
IgD T 0.25
P(Usage) 5 V T 0.00
S S P
Score
1.0
0.8 0 F S N S P V T
IgH Isotype

y g
0.6
AGT

TGG
CCA
CCG

ACG
CCC
GTC
AGC
ACC
TTT

GTT
TCT
CCT
ACT
TCA
ACA
TCG

TTC
TCC
GTG
AAC
GTA
AAT

IgG2 0.4
0.2
Codon
H
0.0

IgG3 Species Asn (N) Val (V)

Human 20
IgE
tRNA Gene Count

Mouse Average
IgG1 15 P(Usage)
Score

y
1.00
10 0.75
IgA

,
0.50
5 0.25
IgD 0.00
IgE
GGG
CGG
GCG
AGG

GAG
CCG
TGG

GTG

CAG

ACG
CTG

AAG
TCG
ATG

TTG
GGC
CGC

GCC
AGC
CCC

GAC
GTC
ACC

TGC

CAC
AAC

TCC

CTC

TAC

ATC
TTC

GGA
GCA

CGA

GGT

0
CCA

CGT

GCT

GAA

AGA
ACA

CAA

GAT

CCT

GTA
AGT
GTT

CTA

CAT

TCA

ACT
TGT
AAA
CTT

AAT

ATA

TCT
TAT

ATT

TTA
TTT

AAT

AAC

GTA

GTT

GTG

GTC

Hyperutilized Hypoutilized
Codons Codons Codon

Fig. 1. Biased codon usage patterns across IgH constant regions. (A) Frequency
of codon usage in the human IGHM gene. Mean ± SD. (B) Quartile distribution
of SCUO scores, representing total codon usage bias across the human IGHM gene
compared with all genes from the human-genome Consensus Coding Sequence
(CCDS) dataset. (C) Mean of codon P(Usage) scores for human IGHM. The
codons on the x axis are ranked by the combined mean scores for all five human
IgH isotypes, using a linear correlation model. (D) Mean of codon P(Usage)
scores for all other human IgH isotypes. The x axis is ordered as in (C), and
codons were analyzed with a linear correlation model. (E) Same as (C) and (D)
but showing scores for each human TCR constant-region chain with the x axis
ordered as in (C), analyzed with a linear correlation model. (F) Hierarchical
clustering analysis of codon P(Usage) scores for human and mouse IgH
genes. Each row represents one gene isoform [as delimited by the International
Immunogenetics Information System (IMGT)]; each column represents a
codon. (G and H) tRNA gene copy number in the human genome for the
amino acids corresponding to hyperutilized codons bracketed in (F). Codons
and tRNAs are matched by Watson-Crick base pairing. Codons lacking genome-
encoded Watson-Crick complementary tRNAs are highlighted in red. Bars are
color-coded by the average codon probability score for all IgH isotypes across
humans. Source data for this figure can be found in data S1 and S2.

Giguère et al., Science 383, 205–211 (2024) 12 January 2024 2 of 7

RES EARCH | R E S E A R C H A R T I C L E
anticodon, overlaid by posttranscriptional mod-
ifications, some of which may expand the de-
coding capacity of that tRNA (19, 20). We
compared the tRNA modification profiles of
cells that secrete antibodies with those that
do not to determine if there were any differ-
ences suggesting adaptation to IgH production.
We isolated small RNA (sRNA) [<200 nucleo-
tides (nt)] from murine splenic naïve B cells
(NBCs), a murine NBC-like line (WEHI231), two
murine antibody-secreting PC lines (J558, MPC11),
and a human reference, human embryonic
kidney 293T (HEK293T), which is frequently
used to produce antibodies. The bulk of the
extracted RNA was 70 to 90 nt, the standard
size of full-length tRNA (fig. S2A). We quan-
tified a panel of 25 tRNA modifications with
liquid chromatography–mass spectrometry
(LC-MS) multiple reaction monitoring (Fig. 2A).
The most frequently found modifications were
methylations (Fig. 2B), which are often asso-
ciated with tRNA structure (21). There is no
evidence, however, that methylations can ex-
pand tRNA decoding capacity.
Modifications on position 34, or the so-called
wobble position, are the only ones known to
modulate decoding capacity (19–21). In our
samples, five modifications were found that
occur only at site 34, some of which alter
decoding (Fig. 2C) (21). The antibody-secreting
J558 and MPC11 clustered together in a prin-
cipal components analysis (PCA) of overall tRNA
modification profiles, as well as a comparison
focused to site 34 (Fig. 2, D and E, and fig.
S2B). To compare frequencies of site 34 tRNA
modifications across antibody-secreting PC
lines and NBC lines, we normalized modifi-
cations in all other lines relative to the NBC-like
WEHI231 and found that the two murine
antibody-secreting PC lines most resembled
one another (Fig. 2F). Moreover, the profile of
HEK293T—which does not naturally produce
antibodies but can be induced to express anti-
bodies efficiently—more closely resembled the
PC lines (Fig. 2F).
The most consistently up-regulated wobble
modification in PC lines is inosine (I34) (Fig.
2F). Introduced through the deamination of
adenosine in eight anticodons, I34 is a “super-
wobbler” able to decode cytidine, uridine, and
(within limits) adenine (Fig. 2G). I34 is neces-
sary for the decoding of eight C-ending codons
in most eukaryotes, given the absence of tRNA
genes with the corresponding Watson-Crick
base pairing (22–25). Because multiple codons
hyperutilized in IgH genes lack a directly decod-
ing tRNA and because I34 is up-regulated in cell
lines that express high levels of antibody, I34
could partially bridge the gap between IgH
codon demands and the tRNA supply. Both
human and mouse IgH sequences were highly
enriched for I34-dependent codons (Fig. 2I).
Indeed, four I34-dependent codons were consist-
ently hyperutilized across mouse and human
Giguère et al., Science 383, 205–211 (2024) 12 January 2024
IgH genes relative to their genomic distributions
(Fig. 2H). Thus, I34 modification is required to
decode IgH and is also found at elevated fre-
quencies in antibody-secreting cells.
To distinguish between structural constraints
and the demands of antibody production on I34
dependency, we compared the I34-dependency
score across antibody light-chain constant region
genes (IgL) and TCR constant regions, as well
as the broader set of immunoglobulin super-
family genes (IGSF), all of which contain struc-
turally similar immunoglobulin-like domains
(Fig. 2J). Of these, IgH had the highest I34-
dependency score, followed by IgL, another
antibody component. By contrast, TCRs, which
are produced at lower levels and in smaller
sizes than antibodies, have I34-dependency
scores near the species average. Histones, which
are structurally unrelated to antibodies but also
highly expressed, were extremely enriched for
I34-dependent codons. Similarly, I34-dependent
codons are enriched in highly expressed yeast
proteins (22). Across eukaryotes, I34-dependent
codons are enriched for in low-complexity pro-
tein domains: regions characterized by low
amino acid diversity and often difficult to
translate (26–28). Thus, antibody expression,
rather than structure, may underlie I34
dependency.
The tRNA pools of antibody-secreting cell
lines are enriched for inosine-containing
anticodons
Changes in tRNA modification abundance de-
tected with MS may have been caused by
differential expression of modifiable tRNAs,
the extent to which available tRNAs were
modified, or both. For a more comprehensive
comparison of tRNA expression in B cells, we
performed tRNA sequencing (tRNA-seq) (29)
on the sRNA extractions (<200 nt) of two
murine NBC lines (WEHI231 and Bcl 5b1b) and
two antibody-secreting PC lines (J558 and
MPC11). Multiple tRNAs, decoding a variety of
amino acids, were significantly differentially
expressed between these cell types (Fig. 3A).
Generally, tRNAs encoding serine, arginine, and
tryptophan were significantly up-regulated in
PC lines, whereas those encoding histidine and
lysine were down-regulated (fig. S2C), although
this did not map consistently to amino acid
frequency in the gene (fig. S2, D and E). tRNAs
for the same anticodon may have different
efficiencies or regulatory aspects (30). Indeed,
there was divergence within tRNA genes of
the same anticodon (Fig. 3A). For example, tRNA
Arg-ACG-3 was up-regulated in PC lines, whereas
Arg-ACG-2 was down-regulated. Thus, the tRNA
pool differs between antibody-secreting cells
and NBCs, potentially to meet different trans-
lational needs.
I34-modifiable tRNAs were up-regulated in
the PC lines (Fig. 3, A and B), reflecting the in-
creased demand from the high-level production
of I34-dependent immunoglobulin genes. This
was recapitulated in an independent sequenc-
ing run for J558, MPC11, and WEHI231 (fig. S2F).
Furthermore, when we grouped tRNA genes for
each anticodon and analyzed differential expres-
sion, anticodons up-regulated in PC lines were
enriched for I34-modifiable species (Fig. 3C).
To confirm that I34-modifiable tRNA species
in B cells were in fact modified, we examined
the anticodon read sequences. Inosine preferen-
tially matches with cytidine bases, whereas the
unmodified adenosine original preferentially
matches with uracil (20, 31). Thus, the presence
of I34 can be detected through an A-to-G mis-
match at the wobble position. Nearly all reads
overlapping the anticodon region for I34-
modifiable species were modified at the wobble
position (Fig. 3D). The fractions of modified
species were comparable across naïve and PC
lines. One caveat, however, is that multiple mu-
rine tRNA Val-AAC and Val-CAC gene sequen-
p
ces are highly similar. The A-to-C mismatches
that we observed may have been an artifact of
alignment parameters. However, a wobble posi-
tion G definitively indicated inosine introduc-
tion due to the absence of a tRNA Val-GAC
gene in the mouse genome. This was also true
g
for Leu, Ser, and Pro (Fig. 3D). Thus, the in-
crease in inosine observed in PC tRNA pools
can likely be attributed to differential expression
rather than variation in the rate of modification.
y
To determine whether the enrichment of
I34-modified tRNAs in PC lines represents an
adaptation to IgH translational demands, we
compared the expression of anticodons in PC
versus NBC lines, relative to the mean codon
P(Usage) of murine antibody IgH regions (Fig. 3,
E and F). If genome-encoded Watson-Crick base
pairing was used, there was no significant
correlation between codons and anticodons
(Fig. 3E). However, if tRNA wobbling was per-
y g
mitted, there was a strong increase in correla-
tion between codons enriched within IgH and
up-regulated anticodons in PC lines (Fig. 3F).
Thus, tRNA wobbling generally, but I34-based
translation particularly, has an important place
in PC lines. Moreover, I34-modifiable tRNA
y
expression rates may constitute an adaptation
,
to high-level antibody production.
I34-decoding efficiency is critical to effective
protein production, particularly in antibody-
secreting cells
To assess the functional relevance of I34-
modified tRNA decoding to protein production,
we designed a fluorescent reporter expression
assay (fig. S3A). We generated plasmids with
enhanced green fluorescent protein (GFPWT)—
which hyperutilized I34-dependent codons
compared with all murine CDSs—or I34 codon-
modified eGFP (GFPI34mod), which dramatically
underutilized these codons (fig. S3B), and we
used mCerulean as an internal control (fig. S3C).
In GFPI34mod, all I34-dependent codons (NNC)
3 of 7
RES EARCH | R E S E A R C H A R T I C L E

A m1A
B 1e8

1e7

Signal Intensity
Normalized
3e7

2e7
7.5e6
7
mG 1e7

m1G
0
Counts

m2G
m2,2G

m
m A

m5 A

Am
ac m

C
U

m
m1

m
U
Y

D
I

A
G
G
G
t6
t A6
t6

G
4
p3

m5

m6

m1
m2
m1
m7
U
U

,2
m6
s2
5e6

C
G
ac

m2
C 1.2e6 D
Cm
0.4
mcm5s2U

Signal Intensity

PC2 (16.45%)
2.5e6 Cell:

Normalized
Inosine Gm
8e5 0.2 WEHI231
m5U 4 6 6 Splenic
ac C mtA
D acp3U ncm5Um t6A x B Cells
m 6A ms2t6A 0.0 x
ncm5U 1 x J558
Y Um m I m5Um Am
4e5 MPC11
cm5U
0 tRNA -0.2 x HEK293T
0 5 10 position 34
Acquisition Time (min) 0 -0.4
-0.25 0.00 0.25 0.50

nc m U

cm U
In U
nc m

e
5

m m5
s5 2

in
U
PC1 (48.45%)

m5

os
c
E 0.4 F 0.4

p
0.2 x
PC2 (34.21%)

Cell: 0.2 x
0.0 WEHI231 Cell:
x x WEHI231
Log(FC)

x Splenic x
B Cells x Splenic
-0.2 J558 0.0 B Cells
MPC11 x J558
x HEK293T x x MPC11
-0.4
-0.2 x x HEK293T
x x

g
-0.6

-0.50 -0.25 0.00 0.25 0.50 -0.4

PC1 (46.22%)
Um

e
U
U

s5 2

in
m5

U5
m5

cm

os
cm
nc

In
nc

y
G H IgG3
IgE
I 1.00 J 1.00

IgM

IgG2
IgG1 Probability 0.75 0.75
Score
Probability Score

Probability Score
I34-Dependency

I34-Dependency

CAA
IgG2 1
GU U
AAAAA 0.8
IgG1 0.6 Human
H₂O 0.50 0.50 Human
0.4 Mouse
IgA 0.2 Mouse
NH₃
0
IgE Species
IgD

y g
IgE
Human 0.25 0.25
IgM Mouse
IgA
IgD
IgG3 0.00 0.00
IgG1
IgG4
CA I IgG3
IgG2
s
L

SF

R
H

ne
Ig

GU C
ACC
GTC
TCC
CCC
ATC
CTC
GCC
CGC

TC
Ig

IG

AAAAA
to

y
is
H

,
Fig. 2. The tRNA modification profiles of antibody-secreting plasma cells splenic B cells (dark blue), and HEK293T cells (gray), relative to the murine NBC-like
and NBCs differ. (A) Total ion chromatogram of LC-MS-MS analysis of line WEHI231 (blue line). Each dot represents the average of three replicates run
modified sRNA nucleosides performed using multiple reaction monitoring from concurrently. Mean ± SEM. (G) Schematic of tRNA deamination to introduce
WEHI231 cells. (B and C) Normalized signal intensities for modifications from inosine at the wobble position (I34). (H) Hierarchical clustering analysis of codon
WEHI231 sRNA. Each dot represents the average of three replicates run probability scores for I34-dependent codons in human and mouse IgH genes.
concurrently. n = 3 independent experiments. (B) shows modifications found Each row represents one gene isoform (as delimited by IMGT); each column
throughout tRNAs, and (C) shows modifications that occur exclusively on tRNA represents a codon. (I and J) Quartile distribution of combined I34-dependent
position 34 (wobble position). (D) PCA of MS-based quantification of 25 tRNA codon probability scores for various gene families in humans (gray) and mice
modifications in various cell types from one set of samples run concurrently. (brown). (I) shows IgH constant regions, and (J) shows IgL constant regions,
Each dot represents the average of three replicates run concurrently. (E) As IGSF, TCRs, and histones. In panels B through F, unfilled red circles and triangles
shown in (D), PCA of MS-based quantification of the five tRNA modifications denote antibody-secreting B cell lines, and filled blue diamonds and squares
detected that occur exclusively on position 34 of tRNAs. Each dot represents denote non–antibody secreting B cell lines. The non-B cell line is a gray X. Source
the average of three replicates run concurrently. (F) Log of fold-change (FC) values data for MS can be found in data S3, and source data for I34-dependent codon
for position 34 tRNA modifications in murine PC lines (J558, MPC11; red), murine probability scores can be found in data S4.

Giguère et al., Science 383, 205–211 (2024) 12 January 2024 4 of 7

RES EARCH | R E S E A R C H A R T I C L E

A tRNA:
Trp−CCA−2 B C
Arg−CCG−1 ACG
Met−CAT−3
0.50 8
Arg−TCG−3
Val−AAC−4
Val−CAC−4

Fraction of Significantly Differentially

Val−CAC−7

Expressed I34-Modifiable tRNAs

Trp−CCA−5

−Log10 adjusted p−value

Arg−ACG−3
Thr−AGT−4
Ile−AAT−1
Ile−AAT−3 6
Arg−ACG−1
Val−AAC−3
Val−CAC−3
Ser−GCT−3
Thr−AGT−2
Met−CAT−4 4 I34-
Arg−TCT−4 0.25
Modifiable
Ile−TAT−2 4
Gln−CTG−5 2 AAT Non I34-
Ile−AAT−2 Modifiable
Val−CAC−5 0
Ala−AGC−8
Asp−GTC−3
Glu−TTC−1 -2
Ile−TAT−1
Met−CAT−1
2
AGC
Arg−TCG−1
Leu−TAG−3
-4 AGA
G
Leu−CAG−2
Ala−TGC−2
Ala−TGC−3
Lys−CTT−2 AAC
C
Lys−CTT−3 0.00
AAG
Thr−AGT−1
Thr−AGT−3 0
His−GTG−2
AG AGT
AGG

in NBC Lines

Upregulated
in PC Lines
His−GTG−3

Upregulated
Arg−ACG−2 -0.50 -0.25 0.00 0.25
Asn−GTT−1
Ser−TGA−1
Arg−CCT−4 Log2 fold change
Leu−TAA−2
Arg−CCT−1
Phe−GAA−1
Cys−GCA−3
Cys−GCA−7
Ala−AGC−4
Arg−TCG−2

WEHI231 BCL J558 MPC11

5b1b
D E F

p
AGC (Ala) ACG (Arg) AAT (Ile) AGT (Thr) Watson-Crick Base Pairing Wobble Pairing
0.50 0.50
1.00
r 2 = 0.00617, p = 0.6 r 2 = 0.0871, p = 0.05
0.75 ATC
CGC GCC
0.50
Anticodon Log2(FC)

Anticodon Log2(FC)
TCC
Mismatch Fraction

0.25 0.25
0.25
Mismatch
0.00 A C GTC
ACC

g
AAG (Leu) AGG (Pro) AGA (Ser) AAC (Val) A G 0.00 0.00 CCC
1.00 CTC
A T
0.75

0.50 -0.25 -0.25

A/U I34/C
0.25 None None

y
G/C G/U
0.00
-0.50 -0.50
WEHI231

MPC11
WEHI231

MPC11

WEHI231

MPC11

WEHI231

MPC11

J558
Bcl 5b1b
J558

J558

J558
Bcl 5b1b

Bcl 5b1b

Bcl 5b1b

0.00

0.25

0.50

0.75

1.00

0.00

0.25

0.50

0.75

1.00
IgH Codon P(Usage) Score IgH Codon P(Usage) Score

Fig. 3. The tRNA pool of PC lines is enriched for inosine at the wobble expressed anticodons between PC lines and NBC-like lines. I34-modified
position as compared with NBC lines. (A) Heatmap of significantly anticodons are labeled and highlighted in orange. (D) Fraction of nucleotide
differentially expressed tRNA genes between murine PC lines (red) and mismatch at the wobble position of I34-modifiable tRNA species from tRNA-seq.
murine NBC-like lines (blue), showing normalized expression values (E and F) Correlation between mean P(Usage) codon probabilities for murine IgH
(rlog normalization, using DESeq2) that are centered and scaled by row. sequences and the differentially expressed anticodons between murine PC lines and

y g
I34-modifiable tRNAs are highlighted in orange. Each column represents NBC-like lines. (E) Anticodon-codon pairs are restricted to Watson-Crick
an individual sample. n = 1 independent experiment. (B) Fraction of (nonwobble) base-pairing interactions, where applicable. (F) Anticodon-codon
significantly up-regulated tRNAs in NBC lines versus up-regulated tRNAs pairs are restricted to wobble, where applicable. I34-preferred codons are
in PC lines that are I34-modifiable. (C) Volcano plot of differentially labeled. Raw tRNA sequencing data are available from Dryad (51).

were switched to the corresponding NNT codon
(in NNC and NNT, N is any nucleotide), which
can be decoded by the unmodified tRNA, and,
at low efficiency, by I34-modified tRNAs (20, 31).
We transfected the two murine PC lines, J558
and MPC11, and the murine NBC line WEHI231,
with one of the two plasmids. GFP production—
both baseline and normalized to mCerulean—
was reduced in all cell lines when I34-dependent
codon usage was reduced (Fig. 4, A and B, and
fig. S3D). A greater difference between normal-
ized GFPWT and GFPI34mod indicates a greater
contribution of I34-decoding efficiency to pro-
tein production. WEHI231 was least affected,
whereas normalized GFPI34mod showed signifi-
cantly greater decreases in both PC lines (Fig.
y
,
4C). Direct attribution to I34 modification is unable to generate viable Adat2-KO B cell
complicated by the concomitant change in GC- lines using CRISPR/Cas9, in line with lethality
content, which may affect RNA processing and observed broadly in eukaryotic cells and likely
expression (32, 33). Nonetheless, decreasing owing to the dependency of many genes on I34
I34-dependent codons was more detrimental modification (23, 25, 35–37). Furthermore, when
to protein production in PC lines. we generated a mouse in which ADAT2 was spe-
cifically ablated in the B cell compartment—
Translation of I34-dependent codons affects Adat2-flox crossed with Mb1-Cre (fig. S4, B to
B cell development and differentiation D)—homozygous Adat2-flox Mb1-Cre mice
I34 is introduced into tRNAs through deami- exhibited nearly complete loss of peripheral
nation, by a complex of adenosine deaminase blood B cells (Fig. 4, D and E). To study the
tRNA-specific 2 (ADAT2) and ADAT3. Although effects of I34 dependency, we instead sought
both enzymes are required, ADAT2 contains to regulate the decoding efficiency of I34-
the active deaminase domain (25, 34). Adat2 modified tRNAs by altering codon compo-
is expressed in mouse B cells and at especially sition. In contrast to the constant region, the
high levels in PC lines (fig. S4A). We were antibody variable region is specific to B cell

Giguère et al., Science 383, 205–211 (2024) 12 January 2024 5 of 7

RES EARCH | R E S E A R C H A R T I C L E

A WEHI231 J558 MPC11 C D WT

2.0
**
* Single Cells
1.5 0.6 105
Density

Log Normalized GFP Difference

wtGFP
1.0 104
GFPI34mod

0.5 103

B220
0
0.4
0.0
-10-3
-2 -1 0 1 -2 -1 0 1 -2 -1 0 1
-10-3 0 103 104 105
Log(Normalized GFP) Log(Normalized GFP) Log(Normalized GFP)
CD19

B WEHI231 J558 MPC11

0.2
Adat2 KOB cell Hz
4 4
4
Single Cells
Normalized GFP

3 3 105
3
wtGFP
0.0 104
2 2 GFPI34mod
2

WEHI231

MPC11
J558
103
1 1 1

B220
0

p
-10-3
0 0 0
-10-3 0 103 104 105
CD19

E **** F 100 G 100 **

60
% Peripheral B Cells
Expressing CLK19

% Peripheral B Cells
75

Expressing CLK19
75
% B220+ CD19+

g
40
50
50

20 25
25

y
0
0
0 IgHCLK19/WT IgHI34/WT

T
/W

W
4/
19
B H t2

H I3
T

LK
W

a
z

HC
KO Ad

Ig
Ig
H I J
** * * **

0.75
I34-Dependency

0.72

y g
Donor Donor
I34-Dependency

I34-Dependency
P(Usage)

01a 0.70 01a

0.70
P(Usage)

P(Usage)

02a 02a
04 04
07 07
0.68 10 10
13 13
0.65 20 20
22 22
0.65

y
0.64

,
0.60
C d
C e

C e

PC
C C

C ry
st
B ate
B aiv

a od

C ve
G

C ve
B mo
la

PC

C ry
l
l

l
l

l
el
el

el
el

el
BM

l
ab

l
m N

B mo
el
B Nai

el
N

B Nai

el
tiv

BM
as h
m

M
Ac

e
Pl mp
as

M
Ly
Pl

Fig. 4. I34 and I34-dependent codons play key roles at multiple stages of
B cell development. (A) Representational density plots of log scores for normalized
GFP (GFP/mCerulean), by cell. The red label denotes an antibody-secreting B cell line,
whereas the blue label denotes a non–antibody secreting B cell line. (B) Mean of
normalized GFP by cell type. n = 5 independent experiments. (C) Difference in log-
normalized GFP between GFPWT and GFPI34mod by cell type. Each dot represents the
average of technical duplicates. n = 5 independent experiments; mean ± SD; linear mixed
model with Tukey post hoc correction. *P < 0.05; **P < 0.01; ***P < 0.001. Filled blue
symbols indicate non–antibody secreting B cell line; open red circles and triangles
denote antibody-secreting B cells lines. Source data for (A) to (C) can be found in data
S5. (D) Representative flow cytometry plots of blood cells from (top) WT and (bottom)
Adat2-flox Mb1-Cre mice. (E) Quantification of percentage of B220+ cells in the blood.
Each dot represents a single mouse. n = 1 independent experiment; mean ± SEM; Welch’s
two-sample t test. Source data for (D) and (E) can be found in table S1. (F) Percentage
of peripheral B cells expressing full-length CLK19 variant from IghCLK19/WT or IghI34/WT
knockin mice, by single-cell RNA-seq. (G) Quantification of (F), with mean ± SEM. Welch’s
two-sample t test. Source data for (F) and (G) can be found in table S2. (H) Boxplot of
I34-dependency P(Usage) scores in variable-region antibody repertoires from eight human
donors, in various B cell compartments. (I and J) Same as (H) but showing individual
median donor scores in (I) NBCs versus long-lived BMPCs and (J) NBCs versus MBC
repertoires. Linear mixed model with Tukey post hoc correction. *P < 0.05; **P < 0.01;
***P < 0.001. Source data for (H) to (J) can be found in data S6.

Giguère et al., Science 383, 205–211 (2024) 12 January 2024 6 of 7

RES EARCH | R E S E A R C H A R T I C L E
clonal lineages. Variable-region differences
affect immunoglobulin binding target and
affinity, as well as B cell progression through a
developmental checkpoint in the bone marrow
and an antigen-dependent activation and differ-
entiation checkpoint in secondary lymphoid
tissues. We hypothesized that the codon iden-
tity of the variable region might affect B cell
progression through these checkpoints. To
this end, we modified our previously reported
humanized mouse model, IghCLK19/WT (38). The
amino acid sequences of IghI34/WT and IghCLK19/WT
are identical, but IghI34/WT uses I34-dependent
codons at much greater frequency (fig. S4, G
and H). In heterozygous IghI34/WT mice, ~80%
of mature NBCs expressed CLK19, in contrast
to ~38% in heterozygous IghCLK19/WT mice.
Thus, IghI34/WT was more likely than IghCLK19/WT
to outcompete the endogenous murine allele
(Fig. 4, F and G). As higher expression of an
immunoglobulin transgene better suppresses
endogenous WT variable regions (39), this may
indicate increased IgH expression, which could
also enhance tonic signaling and B cell survival
in the periphery (40, 41). Increased I34-dependent
codon usage appears to benefit selection, either
during intracellular competition between im-
munoglobulins in developing B cells or through
the persistence of B cells in the periphery.
Following allelic exclusion, B cells can be ac-
tivated and enter the germinal center (GC), in
which the BCR variable region undergoes muta-
tion and selection for improved antigen binding
(42). Successful B cells exit the GC to become
long-lived bone-marrow plasma cells (BMPCs)
or memory B cells (MBCs). Survival in and exit
from the GC requires sufficient immunoglobu-
lin production to capture antigen, and long-term
BMPC survival requires sustainable, high-level
antibody production. By examining a post-mRNA
vaccination human B cell repertoire RNA-seq
dataset (43), we found that I34-dependent codon
usage was increased in BMPC and MBC antibody
repertoires as compared with its usage in the
NBC repertoire (Fig. 4, H to J). Thus, increased
I34-dependent codon usage in the variable region
is favored in long-lived B cell compartments
that have undergone a GC reaction, suggest-
ing that the codon composition of immuno-
globulins plays a critical role in the progression
and functionality of B cells.
Discussion
The coordination between the codon-usage pat-
tern of antibody genes and the tRNA profile of
specialized antibody-secreting cells facilitates
both antibody production and selection. Anti-
body constant-region genes are unusually reliant
on codons that can only be translated through
I34 modification, and PCs demonstrate speci-
fic tRNA pool adaptations to the production of
proteins that are highly I34-dependent. The
balance in tRNA supply and codon demand is
critical to cell fitness and functionality (2, 44–47),
Giguère et al., Science 383, 205–211 (2024) 12 January 2024
and our results align with studies performed
in cancer, the nervous system, and stem cells,
among others (15–17, 48, 49). Human develop-
mental syndromes associated with ADAT com-
plex mutations, presumably with downstream
I34 effects, have been identified (50). Although
no related immune deficits have yet been iden-
tified, this may be an important area for future
investigation.
Our results also highlight a potential route
for improving protein expression, particularly
exogenous production of antibodies for phar-
maceutical purposes, by designing constructs
that consider tRNA expression and tRNA
modifications according to cell type. More-
over, variable-region alleles with increased
I34-dependent codon usage were selected for
during B cell development, and I34-dependent
codon usage was higher in B cell compart-
ments emerging from the selective GC reac-
tion. This discovery opens codon usage as a
category for the selection of target antibodies
in rational vaccine design. Our findings demon-
strate that tRNA remodeling and gene codon
usage are important regulatory mechanisms of
humoral immunity.
RE FERENCES AND NOTES
1. F. Buttgereit, M. D. Brand, Biochem. J. 312, 163–167
(1995).
2. H. Gingold, Y. Pilpel, Mol. Syst. Biol. 7, 481 (2011).
3. S. A. Plotkin, Clin. Vaccine Immunol. 17, 1055–1065 (2010).
4. T. Hibi, H.-M. Dosch, Eur. J. Immunol. 16, 139–145 (1986).
5. M. K. Slifka, R. Antia, J. K. Whitmire, R. Ahmed, Immunity 8,
363–372 (1998).
6. S. L. Nutt, P. D. Hodgkin, D. M. Tarlinton, L. M. Corcoran,
Nat. Rev. Immunol. 15, 160–171 (2015).
7. A. Dana, T. Tuller, Nucleic Acids Res. 42, 9171–9181 (2014).
8. T. Tuller et al., Cell 141, 344–354 (2010).
9. S. Varenne, J. Buc, R. Lloubes, C. Lazdunski, J. Mol. Biol. 180,
549–576 (1984).
10. E. B. Kramer, P. J. Farabaugh, RNA 13, 87–96 (2007).
11. T. Ikemura, J. Mol. Biol. 146, 1–21 (1981).
12. K. A. Dittmar, J. M. Goodenbour, T. Pan, PLOS Genet. 2, e221
(2006).
13. S. Kanaya, Y. Yamada, Y. Kudo, T. Ikemura, Gene 238, 143–155
(1999).
14. S. Kanaya, Y. Yamada, M. Kinouchi, Y. Kudo, T. Ikemura,
J. Mol. Evol. 53, 290–298 (2001).
15. H. Gingold et al., Cell 158, 1281–1292 (2014).
16. H. Goodarzi et al., Cell 165, 1416–1427 (2016).
17. S. Bornelöv, T. Selmi, S. Flad, S. Dietmann, M. Frye, Genome Biol.
20, 119 (2019).
18. J. H. Wang et al., EMBO J. 17, 10–26 (1998).
19. P. F. Agris, Biochimie 73, 1345–1349 (1991).
20. F. H. C. Crick, J. Mol. Biol. 19, 548–555 (1966).
21. T. Pan, Cell Res. 28, 395–404 (2018).
22. E. M. Novoa, M. Pavon-Eternod, T. Pan, L. Ribas de Pouplana,
Cell 149, 202–213 (2012).
23. S. Tsutsumi et al., J. Biol. Chem. 282, 33459–33465
(2007).
24. A. G. Torres et al., FEBS Lett. 588, 4279–4286 (2014).
25. A. P. Gerber, W. Keller, Science 286, 1146–1149 (1999).
26. Z. W. Dickson, G. B. Golding, Mol. Biol. Evol. 39, msac087
(2022).
27. À. Rafels-Ybern, C. Stephan-Otto Attolini, L. Ribas de Pouplana,
Int. J. Mol. Sci. 16, 17303–17314 (2015).
28. A. G. Torres et al., Nucleic Acids Res. 49, 7011–7034
(2021).
29. J. F. Hu et al., Nat. Biotechnol. 39, 978–988 (2021).
30. R. Geslain, T. Pan, J. Mol. Biol. 396, 821–831 (2010).
31. F. H. Martin, M. M. Castro, F. Aboul-ela, I. Tinoco Jr., Nucleic
Acids Res. 13, 8927–8938 (1985).
32. G. Kudla, L. Lipinski, F. Caffin, A. Helwak, M. Zylicz, PLOS Biol.
4, e180 (2006).
33. G. Kudla, A. W. Murray, D. Tollervey, J. B. Plotkin, Science 324,
255–258 (2009).
34. E. Ramos-Morales et al., Nucleic Acids Res. 49, 6529–6548
(2021).
35. W. Zhou, D. Karcher, R. Bock, Plant Physiol. 166, 1985–1997
(2014).
36. A. G. Torres et al., Nucleic Acids Res. 43, 5145–5157
(2015).
37. À. Rafels-Ybern et al., Mol. Biol. Evol. 36, 650–662 (2019).
38. X. Wang et al., EMBO J. 40, e105926 (2021).
39. N. Watanabe et al., J. Exp. Med. 190, 461–469 (1999).
40. K. P. Lam, R. Kühn, K. Rajewsky, Cell 90, 1073–1083
(1997).
41. T. T. T. Nguyen et al., Nat. Immunol. 18, 321–333 (2017).
42. G. D. Victora, M. C. Nussenzweig, Annu. Rev. Immunol. 40,
413–442 (2022).
43. W. Kim et al., Nature 604, 141–145 (2022).
44. P. C. Dedon, T. J. Begley, Chem. Res. Toxicol. 27, 330–337
(2014).
45. C. T. Y. Chan et al., Nat. Commun. 3, 937 (2012).
46. D. D. Nedialkova, S. A. Leidel, Cell 161, 1606–1618 (2015).
47. W. Qian, J.-R. Yang, N. M. Pearson, C. Maclean, J. Zhang, PLOS
Genet. 8, e1002603 (2012).
48. S. Martin et al., Nat. Commun. 13, 5003 (2022).
49. F. Rapino et al., Nature 558, 605–609 (2018).
50. A. M. Alazami et al., J. Med. Genet. 50, 425–430 (2013).
p
51. S. Giguere et al., Antibody production relies on the tRNA inosine
wobble modification to meet biased codon demand, Version 6,
Dryad, (2023); https://doi.org/10.5061/dryad.f4qrfj723.
AC KNOWLED GME NTS
We would like to thank all members of the Batista Lab for
assistance with experiments. We also thank A. K. Chakraborty,
P. M. Murugan, and K. G. Sprenger for help with the design of
g
the IghI34/WT mouse model; L. Wu and the Harvard Gene
Modification Facility for mouse injections; S. S. Levine and the
MIT BioMicroCenter for sequencing; M. T. Waring and the Ragon
Flow Core for assistance with cell sorting and flow cytometry;
J. Berube for statistical help; and A. E. Ringel, D. Lingwood, and
Y. Shulgina for scientific discussion of the manuscript. Funding:
y
This research was funded by the Collaboration for AIDS Vaccine
Discovery (CAVD) grants INV009585 and INV046626 (to F.D.B.)
through the Bill and Melinda Gates Foundation; the National
Institute of Allergy and Infectious Diseases (NIAID) UM1 AI144462
(Scripps Consortium for HIV/AIDS Vaccine Development) (to
F.D.B.); the National Science Foundation of Singapore through the
Singapore-MIT Alliance for Research and Technology Antimicrobial
Resistance Interdisciplinary Research Group (to P.D.); NIH grant
ES031529 (to P.D.); Swiss National Science Foundation Fellowship
grant P2SKP3_174681 (to S.H.); the Ragon Institute of Mass
General, MIT, and Harvard (to F.D.B.); and the Herchel Smith Fellowship
(to S.G.). Author contributions: Conceptualization: S.G. and F.D.B.
Data curation: J.W. and S.G. Formal Analysis: S.G., S.H., P.D.,
y g
and F.D.B. Funding acquisition: F.D.B. and P.D. Investigation: S.G.,
X.W., L.X., S.H., J.W., J.H., D.M., Q.A.P., K.T., and T.P. Methodology:
S.G., U.N., P.D., S.H., and J.H. Project administration: F.D.B.
Resources: K.H.K. Supervision: U.N., P.D., and F.D.B. Visualization:
S.G., F.D.B., and S.R.W. Writing – original draft: S.G., S.R.W., and
F.D.B. Writing – review and editing: S.G., S.R.W., P.D., and F.D.B.
Competing interests: J.H. is an employee of GentiBio. Data and
y
materials availability: Adat2-flox mice are available from F.D.B.
under a material agreement with Massachusetts General Hospital.
,
tRNA-seq datasets are available at Dryad (51). All other data
are available in the main text or the supplementary materials.
Requests for material and correspondence should be addressed to
S.G. and F.D.B. License information: Copyright © 2024 the
authors, some rights reserved; exclusive licensee American
Association for the Advancement of Science. No claim to original
US government works. https://www.science.org/about/science-
licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.adi1763
Materials and Methods
Figs. S1 to S4
Tables S1 to S10
References (52–70)
MDAR Reproducibility Checklist
Data S1 to S6
Submitted 13 April 2023; accepted 27 November 2023
10.1126/science.adi1763
7 of 7
RES EARCH

NANOMATERIALS ventional film deposition techniques is limited

by its poor wetting characteristics on many
Twisted epitaxy of gold nanodisks grown between surfaces, including MoS2. For thin layers (nom-
inal thickness of <1 nm), Au epitaxial nano-
twisted substrate layers of molybdenum disulfide particles (NPs) are formed; in effect, the higher
surface area favored NPs over a continuous
Yi Cui1, Jingyang Wang1,2, Yanbin Li1, Yecun Wu1, Emily Been1, Zewen Zhang1, Jiawei Zhou1, film. In the epitaxial MoS2-Au system, Au has
Wenbo Zhang1, Harold Y. Hwang1,3*, Robert Sinclair1*, Yi Cui1,4,5* an fcc crystal structure with a 3  symmetry
along the ½111 axis, whereas MoS2 has a hex-
We expand the concept of epitaxy to a regime of “twisted epitaxy” with the epilayer crystal agonal structure (P63/mmc) and is sixfold-
orientation between two substrates influenced by their relative orientation. We annealed symmetric along its [001] axis. A schematic
nanometer-thick gold (Au) nanoparticles between two substrates of exfoliated hexagonal molybdenum SAED pattern of epitaxial Au on MoS2 is
disulfide (MoS2) with varying orientation of their basal planes with a mutual twist angle ranging from shown in Fig. 1B, where the Au {220} electron-
0° to 60°. Transmission electron microscopy studies show that Au aligns midway between the top diffraction spots and the MoS2 {110} spots are
and bottom MoS2 when the twist angle of the bilayer is small (<~7°). For larger twist angles, Au has only a epitaxially aligned and are connected by a re-
small misorientation with the bottom MoS2 that varies approximately sinusoidally with twist angle of the ciprocal lattice vector Dga. The lattice mismatch
bilayer MoS2. Four-dimensional scanning transmission electron microscopy analysis further reveals a of the parallel Au {220} planes (0.144 nm) and
periodic strain variation (<|±0.5%|) in the Au nanodisks associated with the twisted epitaxy, consistent MoS2 {110} planes (0.158 nm) yields a Moiré
with the Moiré registry of the two MoS2 twisted layers. pattern with spacing |Dga|−1 equal to 1.63 nm
(Fig. 1C), which can be observed in bright-field

E
TEM images.

pitaxy is an important phenomenon in
materials science that describes the sit-
uation when one crystalline material
grows on a substrate with exact crystal-
lographic registry in an ordered manner
(1, 2). Epitaxy has been used in synthesizing
gaps between graphene vdW materials has
been reported, these studies did not examine
the influence of both substrates in the twisted
epitaxy regime (38–40). For example, a single
layer of b-CuI, which only exists in the 645
to 675 K temperature range, was stabilized
p
Design and synthesis
In this study, we attempted to synthesize ultra-
thin 2D Au nanodisks and examine their
orientation within physically prepared bilayer
MoS2. A schematic of the synthetic process

and fabrication of many materials, including
planar semiconductor heterostructures (3, 4)
and linear and core-shell nanocrystals (5–8)
and nanowires (9–11), with applications in-
at room temperature by graphene encapsu-
lation (38), but the orientation of the as-
synthesized b-CuI relative to twisted bilayers
of graphene was not fully investigated. To fab-
g
of the MoS2-Au-MoS2 sandwich structure is
shown in Fig. 1D. First, a thin flake of MoS2
~50 mm in size and ~3 nm in thickness was
exfoliated onto a 25-nm-thick amorphous SiNx

cluding transistors (12–15), light-emitting
diodes (16–18), lasers (19–22), and quantum de-
vices (23–26). The recent development of two-
dimensional (2D) van der Waals (vdW) materials
has expanded the scope of epitaxy to vdW
epitaxy (27–30), remote epitaxy (31, 32), and
confined epitaxy (33–35). Thus far, research
has been mainly based on the framework of
growing one crystal epitaxially onto or into
one substrate. Here, we explore a regime of
ricate 2D low-melting-point metals, Briggs et al.
used silicon carbide (SiC) (0001) as a tem-
plate, with epitaxial graphene (EG) on top
instead of bilayer graphene (35). Metal atoms
(Ga, In, and Sn) were intercalated into the in-
terface of EG and SiC to produce 2D meta-
materials of three-atomic-layer thickness. These
2D metals were thought not to be bonded
to the top graphene layer but rather cova-
lently bonded to the Si atoms below and thus
y
TEM grid using the “scotch tape” method.
Next, we deposited Au onto the exfoliated
MoS2 using electron-beam evaporation with a
nominal thickness set to 0.3 nm. Imaging with
TEM showed that, as expected, Au formed
epitaxial NPs ~3.5 nm in diameter. Afterward,
the Au was physically encapsulated by a sec-
ond exfoliated MoS2 layer with orientation
(qm) varied from 0° to 60° rotation. Optical
microscope images of the sample are shown in

epitaxy between two twisted substrates, or
“twisted epitaxy,” in which both substrates in-
teract with the grown crystal epilayer and in-
fluence its crystallographic registry. This
approach allows the relative orientation of
two substrates to be used as a structural con-
had epitaxial orientation only with the SiC
crystal (35).
We propose using 2D molybdenum disul-
fide (MoS2) as the two substrates and gold (Au)
as the twisted epilayer. The chemical interac-
tion between Au and sulfur (S) atoms is much
y g
fig. S1. Finally, we annealed the as-synthesized
MoS2-Au-MoS2 samples between 300° and
700°C in a tube furnace in an argon (Ar) atmo-
sphere to promote atomic rearrangement of
the Au NPs between the two MoS2 substrate
layers. By plan-view TEM examination, we

y
trol parameter. stronger than conventional vdW interaction found that the original epitaxial Au NPs be-

We explored 2D vdW materials as substrate
candidates because they can easily be stacked
mechanically to form nanoscale separations
for confining atoms and molecules (36, 37).
Although the growth of crystals in nanoscale
1 the hexagonal MoS2 planes ({001}) (30, 44).
Department of Materials Science and Engineering, Stanford
University, Stanford, CA 94305, USA. 2Materials Sciences
Division, Lawrence Berkeley National Laboratory, Berkeley,
CA 94305, USA. 3Department of Applied Physics, Stanford
University, Stanford, CA 94305, USA. 4Stanford Institute for
Materials and Energy Sciences, SLAC National Accelerator
Laboratory, Menlo Park, CA 94025, USA. 5Department of
Energy Science and Engineering, Stanford University,
Stanford, CA 94305, USA.
*Corresponding author. Email: hyhwang@stanford.edu (H.Y.H.);
bobsinc@stanford.edu (R.S.); yicui@stanford.edu (Y.C.)
but weaker than covalent bonding (41, 42, 43).
This configuration would not only stabilize 2D
Au but also tune its orientation. Previous studies
of Au through vacuum deposition onto single-
crystal MoS2 have shown that there is an
epitaxial relationship between the face-centered
cubic (fcc) close-packed Au planes ðf111gÞ and
The morphology of Au and the interfacial
orientation were determined by transmission
electron microscopy (TEM) and selected-area
electron diffraction (SAED) patterns. The as-
deposited Au exhibits a morphology of 3D
nanoislands with a typical size of 20 nm and
thickness of 8 nm (Fig. 1A) (30, 43). The ability
to synthesize and utilize ultrathin Au by con-
,
come nanodisks upon annealing and that
their original epitaxial orientation (qa) was
modified by the orientation of the upper MoS2
layer (Fig. 1E), as determined by electron dif-
fraction (Fig. 1F) and Moiré image analysis
(Fig. 1G and fig. S2).
Synthesis of twisted epitaxial Au nanodisks
Figure 2A shows a TEM image of the MoS2-
Au-MoS2 sample at room temperature (RT)
before annealing with a bilayer twist angle qm
equal to 19.1°. The edge of the bottom MoS2
flake is marked in blue, and the edge of the top
MoS2 flake is marked in green, dividing the
sample area into four regions: Au on the amor-
phous SiNx substrate; Au on the bottom MoS2

Cui et al., Science 383, 212–219 (2024) 12 January 2024 1 of 8

RES EARCH | R E S E A R C H A R T I C L E

Fig. 1. Twisted epitaxial Au

nanodisks. (A) Schematic
of epitaxially aligned Au on
MoS2. (B) Reciprocal space
model of the SAED pattern of
the MoS2-Au sample in (A).
(C) Plan-view atomic model of
3D Au on 2D MoS2 for Moiré
patterns generated by Dga in
(B), where |Dga|−1 notes the
3D Moiré fringe spacing. Note
that we use |Dga|−1 for theoret-
ical Moiré fringe spacing
calculated from SAED patterns
and S for Moiré fringe spacing
measured from TEM images.
(D) Schematic synthesis
procedure of Au nanodisks on an
amorphous silicon nitride sub-
strate showing the physical
twisting and annealing steps.

p
(E) Schematic of the twisted Au
nanodisks encapsulated in the
vdW spacing of the bilayer
MoS2. qm notes the arbitrary
rotation angle of the top MoS2
relative to the bottom one. qa

g
notes the twist angle of Au
relative to the bottom MoS2.
(F) Reciprocal space model of
the SAED pattern of the

y
MoS2-Au-MoS2 sample in (E).
(G) Plan-view atomic model for
the Moiré patterns generated
by Dga in (F).

(MoS2-Au, region I); Au below the top MoS2 nealed the sample at elevated temperatures the top MoS2 was weaker than that with the

(Au-MoS2, region II); and Au encapsulated
between the top and bottom MoS2 (MoS2-
Au-MoS2, region III). The diameter of Au
NPs in regions I and III was 3.5 nm with a
standard deviation of 1.2 nm, which is smaller
than the NPs in region II and on the SiNx
and used the results from 400° and 500°C to
illustrate the changes. The schematic mor-
phology and orientation changes of Au under
different annealing conditions are given in fig.
S5. The TEM image of the MoS2-Au-MoS2 sam-
ple after annealing in Ar at 400°C for 5 hours
y g
bottom one (Fig. 2H).
After annealing the sample in Ar at 500°C
for 2 hours (Fig. 2I), Au nanodisks with an
average diameter of ~50 nm were observed as
larger, less dark contrast regions around the
edge of the sandwiched region III. In regions I

y
(5.1 ± 2.6 nm). The crystal orientation of Au (Fig. 2E) shows that the Au NPs on SiNx and in and II, the Au NPs grew slightly larger, but no

relative to MoS2 was revealed by SAED. For
Au on amorphous SiNx without MoS2 top and
bottom layers, the NPs displayed random
orientations (fig. S3). The SAED of Au NPs in
region I (Fig. 2B) revealed epitaxial alignment
with the bottom MoS2, where {220} Au diffrac-
tion spots aligned with {110} MoS2, which was
consistent with previous studies (42, 43). In
region II, Au NPs on SiNx also displayed ran-
dom orientations and did not align with the
top MoS2 (Fig. 2C and fig. S4). The SAED pat-
terns of the MoS2-Au-MoS2 sandwich structure
(region III) (Fig. 2D) showed that Au aligned
with the bottom MoS2 layer.
To determine how the temperature affected
the morphology and orientation of Au, we an-
region I underwent Ostwald ripening and grew
larger (10.4 ± 4.5 nm), whereas Au NPs in re-
gions II and III remained small (3.7 ± 1.2 nm).
The SAED of the four regions showed that the
Au on SiNx remained randomly orientated, and
that Au in region I was still aligned with the
bottom MoS2 (Fig. 2F). The Au in region II epi-
taxially aligned with the top MoS2, where {220}
Au spots were aligned with {110} MoS2 spots
(Fig. 2G, inset). In region III, {220} planes of
Au epitaxially aligned with the {110} planes
of bottom MoS2 (Fig. 2H, bottom-right inset),
whereas no epitaxial diffraction spots of {220}
Au were observed aligned with the top {110}
MoS2 spot (Fig. 2H, top-right inset). These re-
sults indicate that the interaction of Au with
,
Au nanodisks were observed in these two re-
gions (fig. S6). The presence of Au was con-
firmed by energy-dispersive x-ray spectroscopy
(EDS; fig. S7), and the disk morphology was
verified by sample tilting (fig. S8). The SAED
patterns of regions I and II are shown in Fig.
2, J and K, respectively. The intensity of {220}
Au spots became stronger compared with those
after annealing at 400°C in both regions and
retained epitaxial alignment with the {110} MoS2
spots. The SAED pattern of Au nanodisks in
region III is shown in Fig. 2L, where the bi-
layer twist angle qm was determined to be un-
changed after annealing. The orientation of
Au {220} planes still matches closely with that
of the bottom MoS2 {110} planes. However,

Cui et al., Science 383, 212–219 (2024) 12 January 2024 2 of 8

RES EARCH | R E S E A R C H A R T I C L E
Fig. 2. Synthesis, alignment, and thickness of Au nanodisks. (A) Bright-field
TEM image of a freshly prepared MoS2-Au-MoS2 sample on a 25-nm-thick SiNx
membrane showing the Au nanoparticles as dark on a bright background. The edges
of the top and bottom MoS2 are marked in green and blue, respectively. (B to D)
The SAED patterns of regions I, II, and III in (A), respectively. The diameter of the
selected area was about 200 nm. The insets have the same orientation as the
smaller boxed regions, and they show the epitaxial alignment of Au {220} diffraction
spots with MoS2 {110} spots. Scale bars: 2 nm−1. (E) Bright-field TEM image of MoS2-Au-
MoS2 sample after annealing in Ar at 400°C for 5 hours. (F to H) The SAED patterns
of regions I, II, and III in (E), respectively, showing the epitaxial alignment of Au {220}
Cui et al., Science 383, 212–219 (2024) 12 January 2024
p
g
y
y g
y
,
diffraction spots with MoS2 {110} spots. Scale bars: 2 nm−1. (I) Bright-field TEM
image of MoS2-Au-MoS2 sample after annealing in Ar at 500°C for 2 hours.
(J to L) The SAED patterns of regions I, II, and III in (I), respectively. The yellow lines
in (H), (J), and (L), drawn radially, note the ideal alignment. Scale bars: 2 nm−1.
(M to P) Thickness measurement of Au nanodisks. (M) Bright-field TEM image of Au
nanodisks (gray particles) and Au nanoballs (dark particles). The edges of the top
and bottom MoS2 are marked in green and blue, respectively. (N) AFM image of the
same region as in (M). (O) Superposition of (M) and (N), indicating the position of
Au nanodisks relative to Au nanoballs. (P) Thickness profile of four NPs in (O), where
the thickness of nanodisk 2 is 3 nm, and the thickness of nanodisks 3 and 4 is 2 nm.
3 of 8
RES EARCH | R E S E A R C H A R T I C L E

p
g
y
y g
y
,

Fig. 3. MoS2 rotation angle–dependent twist of Au nanodisks. (A) Schematic
reciprocal lattice orientation of Au and the top and bottom MoS2 in (B) with a small
bilayer twist angle qm = 0.8°. The red arrow notes the reciprocal lattice vector of
the Au {220} planes, and the green and blue arrows are reciprocal lattice vectors of
the {110} planes of the top and bottom MoS2, respectively. (B) Bright-field TEM
image of MoS2-Au-MoS2. A white dashed line denotes the edge of the top MoS2.
(C) Bright-field TEM image of a typical Au nanodisk marked with a white dashed box
in (B). Green lines indicate the Moiré fringes with spacing equal to 20.1 nm
originating from the top and bottom MoS2. Orange lines indicate the Moiré fringes
originate from Au and the bottom MoS2. (D) Higher magnification zoom-in image
of region “a” marked in the white dashed box in (C), showing the smaller-spacing
Moiré fringes. (E) SAED pattern of MoS2-Au-MoS2 in (B). Dga and Dg100 note
the vectors that generate Moiré fringes from the Au-bottom MoS2 and from the
{100} planes of the bilayer MoS2, respectively. Yellow lines show the ideal alignment

Cui et al., Science 383, 212–219 (2024) 12 January 2024 4 of 8

RES EARCH | R E S E A R C H A R T I C L E

of Au with the bottom MoS2. Scale bars: 2 nm−1. (F to T) MoS2-Au-MoS2 samples
with qm = 6.2°, 19.1°, and 31.8°, respectively. When qm = 6.2°, Au reorients to
halfway between the bilayer MoS2. When qm = 19.1°, Au aligns closely with the
bottom MoS2 layer, with misalignment qa determined to be 2.1°. When qm = 31.8°, Au
epitaxially aligns with the bottom MoS2 layer. Scale bar in SAED patterns: 2 nm−1.
Dg'a and Dg110 note the vectors that generate Moiré fringes from the Au-top
MoS2 and the {110} planes of the bilayer MoS2, respectively. (U) The relationship
between the Moiré spacing from Au and the lower MoS2 (S) and the rotation
angle of the top and bottom MoS2 (qm). (V) The relationship between the twist
angle of Au relative to the bottom MoS2 (qa) and the rotation angle of the top
and bottom MoS2 (qm) as determined from the diffraction patterns, showing two
separate regions. Note that the Moiré spacings are ambiguous with respect to the
direction of the twist, although their orientations (perpendicular to Dg−1)
are not. The error bars in our analysis represent the standard deviation of the
small difference of diffraction spot positions in the selected area diffraction
patterns (fig. S30).

small deviations from the ideal epitaxy were
observed, with qa determined to be 2.0° in the
same direction as qm in this case (Fig. 2L, in-
set). The Au in region I was not twisted (Fig.
2J), which indicated that the top layer was
necessary for change in Au orientation. We
note that Au in region III was not twisted after
annealing at 400°C (Fig. 2H), indicating that
Au only reoriented during the nanodisk evo-
lution at 500°C. The {220} lattice spacing of
Au nanodisks was determined to be 0.144 nm
reciprocal lattice orientation of the top MoS2,
Au, and the bottom MoS2, where the MoS2
twist angle qm is 0.8° and the Au orientation
qa is 0°. The bright-field TEM image is shown
in Fig. 3B. A typical Au nanodisk (marked with
a white dashed square in Fig. 3B and shown in
higher magnification in Fig. 3C) had an aver-
age Moiré fringe spacing of Au and the bottom
MoS2 (S) of 1.66 nm (Fig. 3D). This value is
near the theoretical spacing from the parallel
lattice planes, indicating that the Au was not
diffusion (fig. S15). Increasing the annealing
temperature and annealing time extended the
width of the edge region by about three to four
times (fig. S16). A larger-scale synthesis of Au
nanodisks could be achieved by patterning
(fig. S17).
The experimentally observed S, qa, and qm
were highly correlated, as expected. In Fig. 3U,
we plot the S measured from TEM images
versus qm over a 60° range. Figure 3V shows
the relationship between qa, measured from the

from the SAED pattern (fig. S19), consistent
with the natural lattice parameter of Au.
The thickness of Au nanodisks was mea-
sured separately using atomic force micros-
copy (AFM). The TEM and AFM images of
the same Au nanodisks encapsulated in the
twisted in this case (fig. S18). This finding was
further confirmed by the diffraction pattern
(Fig. 3E, insets), as Au {220} spots were epi-
taxially aligned with MoS2 {110} spots.
Figure 3, F to T, shows the samples with
bilayer MoS2 twist angles qm = 6.2°, 19.1°, and
p
diffraction pattern, and qm. The results show
two separate parts. At first, up to about qm =
7°, qa increased linearly with qm with qa = 1/2
qm. Above qm = 7°, qa varied approximately
sinusoidally with qm and roughly followed a
variation qa = 2.2 sin(6qm). The plot of S versus

bilayer MoS2 are shown in Fig. 2, M and N,
respectively (the AFM lateral resolution was
not as high as that of the TEM images in this
case). Figure 2O shows the superposition of the
31.8°. Au nanodisks with a diameter of ~50 nm
were found in each case. In the case where qm =
6.2°, qa was determined to be 3.1°, indicating that
the Au reoriented midway between the top and
g
qa is shown in fig. S18. The data follow the
theoretical relationship S = dmda / [(dm − da)2 +
dmdaqa2]1/2 = |Dga|−1 (45), indicating that the
Au twist solely contributed to the Moiré spac-

TEM and AFM images, in which the brighter
regions (higher protrusions) in the AFM im-
age matched well with the large dark Au NPs
in the TEM image. The height of the large Au
NP in region I (labeled “1”) was determined to
be 15.5 nm, which was similar to its diame-
ter (15.2 nm) measured from the TEM image
(fig. S9). These results show that the NP re-
mained approximately spherical, consistent
with tilting results (fig. S8). The thickness of
bottom MoS2 orientations. As shown in Fig. 3H,
three sets of Moiré fringes coexist in this sys-
tem, which are perpendicular to their corre-
sponding reciprocal lattice vectors noted in the
diffraction pattern. In Fig. 3J, the Au {220} spots
overlap with double diffraction spots from the
bilayer MoS2. The double diffraction effect were
reduced to make the Au diffraction spots clearer
by slightly tilting the sample (fig. S14A). When
qm = 19.1°, Au {220} planes matched closely with
y
ings and orientation changes. To further con-
firm that there was no associated lattice spacing
change, we measured the diffraction spot po-
sitions of MoS2 {110} and Au {220} in the SAED
patterns (fig. S19). Both displayed lattice strains
smaller than 1%.
The intermediate strength of S–Au atomic
interaction is thought to be possibly impor-
tant for the Au anisotropic coupling and nano-
disk evolution. For a brief examination of this

the Au nanodisks in the sandwich region III
(labeled “2,” “3,” and “4”) was measured and
found to be between 2 and 3 nm (Fig. 2P).
Note that the nominal thickness of Au here
was set to 0.3 nm in the electron-beam evap-
oration. Figure S10 shows the samples where
the lower MoS2 {110} planes with a smaller mis-
alignment qa determined to be 2.1° (Fig. 3O).
When qm = 31.8° (near 30°), the Au nanodisks
epitaxially aligned with the bottom MoS2 again
(Fig. 3T). The Au diffraction spots were round
circles when qm was 0° and 19.1° but were
y g
point, we changed our template from MoS2 to
graphene, for which there is a much weaker
C–Au interaction. After annealing at 500°C for
2 hours in Ar, Au nanodisks were found in
MoS2-Au-graphene and graphene-Au-MoS2,
where graphene replaced the top and bottom

y
the nominal thickness of Au was 1.0 and 0.1 nm. elongated in the tangential direction when qm layer MoS2 in MoS2-Au-MoS2, respectively (fig.

Au nanodisks were not formed in the former
sample, whereas smaller (~6.6 nm in diameter)
and thinner (~1.0 nm in thickness) Au nanodisks
are observed in the latter sample compared
with the case when the Au nominal thickness
was 0.3 nm.
Orientation of twisted epitaxial Au nanodisks
To understand further how the top MoS2 af-
fected the orientation of Au nanodisks annealed
at 500°C and the Moiré fringes originating
from the overlapping crystals in the TEM im-
ages, we correlated the Moiré pattens and the
SAED patterns from a series of samples with
different MoS2 bilayer twist angles (0° < qm <
60°) (Fig. 3). Figure 3A shows the schematic
was 6.2° and 31.8°.
Figure 3, I, N, and S, shows the magnified
images of the Moiré fringes from the Au
nanodisks and bottom MoS2 layer, from which
S was measured to be 1.41, 1.47, and 1.63 nm,
respectively, offering direct evidence that the
Moiré spacings changed in response to the
Au twist. The bright-field TEM images and
SAED patterns of all the other samples with
different qm are shown in figs. S11 to S13. Note
that in most instances, Au nanodisks only
evolved at the edge within about 600 nm of
the sandwich region. This spatial effect might
be the result of the confinement of Au at the
peripheral area of the bilayer being lower than
that at the central area, which would favor Au
,
S20). In both cases, Au only epitaxially aligned
with MoS2, regardless of the bilayer twist angle.
In addition, no Au nanodisks were observed
in bilayer graphene (figs. S21 to S23). There-
fore, a strong interaction from one interface
with Au is necessary for nanodisk evolution,
and the interactions from both interfaces
are important for the reorientation of the Au
nanodisks.
Density functional theory calculations
The orientation preference of twisted epitaxial
Au in the twisted bilayer MoS2 is most likely
associated with a lowering of the sum of the
two interfacial energies of MoS2-Au-MoS2. We
define the interfacial energy of Au with the top

Cui et al., Science 383, 212–219 (2024) 12 January 2024 5 of 8

RES EARCH | R E S E A R C H A R T I C L E

p
g
y
y g


Fig. 4. DFT calculations. (A) Schematic configuration of four 111 planes of Au (degrees) at different bilayer MoS2 twist angles (qm, degrees) as obtained from
on a monolayer MoS2 with 0° rotation. Orange, purple, and yellow balls represent fitting the DFT data in (B). (C) When qm = 5°, Etot displays a minimum at qa = 2.5°.
Au, Mo, and S atoms, respectively. (B) Calculated energy of the MoS2-Au model (D) Symmetry breaking of Etot when qm ~ 10°. (E) Etot displays two minima at qa ~ 2°
as a function of the relative twist angle of Au and MoS2 (qa) showing a global and qa ~ 10° when qm = 12°. (F) Etot displays two minima at qa = 0° and qa = 20°

y
minimum at qa = 0° and a local minimum at qa ~ 20°. The inset shows the MoS2-Au when qm = 20°. (G) Etot displays two minima at qa ~ 1° and qa ~ 24° when qm =

,
structure with relative twist angle at 20°. Approximate coincidence sites are 25°. The inset is the magnified image of the plot clearly showing the minimum at
marked with blue circles. Mo atoms and three Au layers are omitted. (C to H) Total qa ~ 1°. (H) Etot displays two minima at qa = 0° and qa = 30° when qm = 30°,
interfacial energy (Etot, electron volts) of MoS2-Au-MoS2 as a function of qa consistent with the MoS2-Au crystallography.

MoS2 as Etop and the interfacial energy of Au
with the bottom MoS2 as Ebottom, and we ap-
proximate the total interfacial energy Etot as
the sum of Etop and Ebottom
Etot = Etop + Ebottom
For a certain qm, the orientation mismatch
of Au with the bottom MoS2 is defined as qa,
and the orientation mismatch of Au with the
top MoS2 is qm − qa. Assuming that Etop and
Ebottom can be described by the same function
Einterfacial(qa), as they are both Au-MoS2 con-
tacts, Etot becomes a function of qa
Etot (qa) = Einterfacial (qa) + Einterfacial (qm − qa)
We performed density functional theory
(DFT) calculations on a MoS2-Au model sys-
tem to find Einterfacial as a function of qa (see
methods in the supplementary materials).
Figure 4A shows the plan view and side view
of this structure with Au epitaxially aligned with
MoS2. Other configurations were obtained by
rotating the Au layers in steps of 1.5° from 0°
to 30° relative to the epitaxial configuration
with the MoS2 layer. The calculated energy of
Au and MoS2 with different twist angles is
plotted in Fig. 4B. Note that the output energy
in DFT is the total energy for the whole struc-
ture, which is not the interfacial energy of the
Au-MoS2 interface. However, as the relative
twist angle is the only variable in this Au-MoS2
structure, we choose the total energy of zero
twist angle as the reference for Einterfacial (qa).
The interfacial energy is lowest at the epitaxial

Cui et al., Science 383, 212–219 (2024) 12 January 2024 6 of 8

RES EARCH | R E S E A R C H A R T I C L E

A C E

AA AA
BA
AB
AA AA

SP

B D F
-1.54 eV
-1.56 eV
AA AA
BA
AB
AA AA
-1.72 eV

p
10 nm
AA AB SP BA AA

Fig. 5. 4D-STEM strain analysis. (A) Schematic of 4D-STEM showing scanning are determined by comparing the TEM image with the atomic model in (E), which
convergent beam (purple) and corresponding diffraction patterns at each is also confirmed by multislice simulation (fig. S31). (D) Uniaxial strain exx map
point. Green, red, and blue lines represent top MoS2, Au, and bottom MoS2, of Au, showing small periodic strain (<|±0.5%|) in Au with a 5.5-nm period

respectively. (B) Diffraction pattern of MoS2-Au-MoS2 sample showing qm = 3.5°
and qa = 1.7°. Scale bar: 2 nm−1. (C) Virtual bright-field image reconstructed using
MoS2 {100} diffraction spots and the transmitted beam spot showing Au (dark
area) encapsulated in bilayer MoS2 (whole area) and bilayer Moiré pattern
g
and a compressive strain of Au at AA stacking. The lattice parameter of Au
measured from SAED pattern was chosen as a reference for zero strain, as it
indicates the average lattice parameter of Au. (E) Atomic structure of the Moiré
supercell of twisted bilayer MoS2 with a twist angle of 6°, showing AA, AB, BA,

with a 5.5-nm period. Pixel size: 1 nm. Scare bar: 10 nm. The magnified image
shows the bilayer Moiré pattern (yellow dashed hexagon) and Moiré supercell
(orange diamond). As there is one AA stacking, two AB/BA stackings, and three
SP stackings in one Moiré supercell, the AA, AB/BA, and SP stacking regions
y
and SP stacking configurations. (F) Calculated interaction energy of Au with the
twisted bilayer at different stacking configurations using DFT, showing that the
interaction energy between a single Au atom and the twisted bilayer MoS2 at
AA stacking is ~170 meV lower than that at AB/BA and SP stacking.

position and increases for all the other values.
At small relative twist angles, the interfacial
energy appears as a convex function. There is a
local minimum of Einterfacial (qa) at ~20° that
we attribute to the formation of a relatively
energy minimum position of Etot (qa) to be
around qa = 0°, 1°, and 0°, respectively (Fig. 4,
F to H). The simulation results are consistent
with the experimental observations overall,
except when qm is around 20°. In this case,
semi-angle of 0.4 mrad was rastered across
an Au nanodisk encapsulated in twisted bi-
layer MoS2 with a step size of 1 nm, and the
diffracted electron signal was collected at each
probe position. The resulting diffraction pat-

high areal density of approximate coincidence
sites of Au and S atoms (Fig. 4B, inset).
Figure 4, C to H, shows the plots of Etot (qa)
with different qm. When qm is small (qm = 5°),
given the convex nature of interfacial energy
at small relative twist angle, Etot (qa) displays
the calculations predict no Au reorientation
while a 2° rotation exists in the experiment,
which might be due to the difference between
the simulation model and the real sample (fig.
S24). For example, a set of screwlike interface
dislocations (analogous to those in low-angle
y g
tern is shown in Fig. 5B, from which qm and
qa are determined to be 3.5° and 1.7°, respec-
tively, illustrating the twisted epitaxial align-
ment of Au.
Virtual bright-field and virtual annular dark-
field images of the sample reconstructed from

y
only one minimum at qa = 1/2 qm, which is twist grain boundaries) may need to be taken the diffraction pattern revealed the Moiré pat-

consistent with the reorientation results (Fig.
4C). Figure 4D shows that when qm increases
from 9.5° to 10.5°, Etot (qa) first displays a
single energy minimum at qa ~ 5° and then
quickly bifurcates to two local minima at qa ~
2.5° and qa ~ 7.5°, which explains the rapid
change of qa between the linear and sinu-
soidal region in Fig. 3V. Au always couples
closely with the bottom MoS2 but not the top
one because, in the initial state (before an-
nealing), Au always epitaxially aligns with the
bottom layer. The increase of the energy barrier
at 1/2 qm impedes the Au coupling with the
top MoS2 (Fig. 4E).
When qm increases to 20°, 25°, and 30°, the
energy barrier becomes higher and pushes the
into account for the interface energies in the
low-twist regime, as they may be present in
the real system. We envision that such twisted
epitaxial relationship could happen in other
2D systems when their interfacial energy is a
convex function at a small twist angle regime
but may not happen if the interfacial energy is
a concave function.
4D-STEM strain analysis
The strain variations in the Au nanodisks as-
sociated with the twisted epitaxy were studied
using 4D scanning transmission electron mi-
croscopy (4D-STEM) (46). The schematic 4D-
STEM experiment is shown in Fig. 5A, where
a focused electron beam with a convergence
,
tern of bilayer MoS2 with a period of 5.5 nm,
corresponding to the bilayer twist angle (Fig.
5C and fig. S25). Lattice spacing maps of Au
{440} planes and f224  g planes revealed a pe-
riodic lattice spacing change with the same
period as the bilayer Moiré pattern (fig. S26).
Further strain analysis (uniaxial strain exx and
eyy, shear strain exy, and rotation q) confirmed
the existence of a small periodic strain varia-
tion (<|±0.5%|) in Au with the same period as
the bilayer Moiré pattern (Fig. 5D and fig. S27).
This periodic strain of Au was most likely asso-
ciated with the different chemical environment
of Au at different stacking configurations [AA,
AB, BA, and saddle point (SP)] of twisted bi-
layer MoS2 (Fig. 5E).

Cui et al., Science 383, 212–219 (2024) 12 January 2024 7 of 8

RES EARCH | R E S E A R C H A R T I C L E
The uniaxial strain (exx) map of Au (Fig. 5D)
revealed compressive strain of Au at the AA
stacking region of bilayer MoS2, as well as a
small tensile strain of Au at the AB stacking
(fig. S27A), which we qualitatively corroborated
using DFT calculations (fig. S28). The interac-
tion energy between a single Au atom and the
twisted bilayer MoS2 at the AA stacking region
was calculated to be ~170 meV lower than that
at the AB or SP stacking region (Fig. 5F). This
result indicated that the twisted bilayer MoS2
exerted a periodic potential on Au with a po-
tential well around the AA region ~170 meV
deeper than that at AB or SP, resulting in the
confinement of Au at the AA region and thus
a compressed strain. Further 4D-STEM analysis
showed that, when the twist angle of the bilayer
MoS2 reached a relatively large value, such as
12°, which corresponded to a bilayer Moiré pat-
tern of 1.33 nm, the strain of Au nanodisks encap-
sulated in the bilayer was negligible (fig. S29).
Discussion
The discovery of twisted epitaxial structure of
Au and MoS2 presents possibilities for structure-
function investigation of 2D materials with ad-
vanced electron microscopy. The TEM Moiré
pattern images are representations of the
atomic matching and mismatching of the
participating crystals. The periodicity is directly
associated with the atomic matching of the two
materials, changes of which (as demonstrated
here) are likely to influence the composite
physical properties, such as the band structure
(47), electronic and optical properties of Au
and MoS2 [plasmon-exciton plexciton (48)], etc.
Moreover, twisted epitaxy implies the possi-
bility of using the van der Waals spacing of the
stacked bilayer 2D material as a nanoreactor
to confine material growth and tune the crys-
tal morphology, orientation, and lattice strain,
indicating that the bilayer orientation can act as
Cui et al., Science 383, 212–219 (2024) 12 January 2024
an additional parameter to control the structure
and properties of the encapsulated atoms.
RE FERENCES AND NOTES
1. A. Y. Cho, J. R. Arthur, Prog. Solid State Chem. 10, 157–191 (1975).
2. M. B. Panish, Science 208, 916–922 (1980).
3. D. C. Tsui, H. L. Stormer, A. C. Gossard, Phys. Rev. Lett. 48,
1559–1562 (1982).
4. S. Nakamura, Jpn. J. Appl. Phys. 30, L1705–L1707 (1991).
5. X. Peng, M. C. Schlamp, A. V. Kadavanich, A. P. Alivisatos,
J. Am. Chem. Soc. 119, 7019–7029 (1997).
6. B. O. Dabbousi et al., J. Phys. Chem. B 101, 9463–9475 (1997).
7. S. E. Habas, H. Lee, V. Radmilovic, G. A. Somorjai, P. Yang,
Nat. Mater. 6, 692–697 (2007).
8. O. Chen et al., Nat. Mater. 12, 445–451 (2013).
9. L. J. Lauhon, M. S. Gudiksen, D. Wang, C. M. Lieber, Nature
420, 57–61 (2002).
10. Y. Wu, J. Xiang, C. Yang, W. Lu, C. M. Lieber, Nature 430,
61–65 (2004).
11. T. Mårtensson et al., Nano Lett. 4, 1987–1990 (2004).
12. W. W. Hooper, W. I. Lehrer, Proc. IEEE 55, 1237–1238 (1967).
13. C. Thelander et al., Mater. Today 9, 28–35 (2006).
14. J. Xiang et al., Nature 441, 489–493 (2006).
15. J. Goldberger, A. I. Hochbaum, R. Fan, P. Yang, Nano Lett. 6,
973–977 (2006).
16. S. Nakamura, T. Mukai, M. Senoh, Appl. Phys. Lett. 64,
1687–1689 (1994).
17. A. Tsukazaki et al., Nat. Mater. 4, 42–46 (2005).
18. R. Yan, D. Gargas, P. Yang, Nat. Photonics 3, 569–576
(2009).
19. S. Nakamura et al., Jpn. J. Appl. Phys. 35, L74–L76 (1996).
20. W. Tsang, IEEE J. Quantum Electron. 20, 1119–1132 (1984).
21. X. Duan, Y. Huang, R. Agarwal, C. M. Lieber, Nature 421,
241–245 (2003).
22. S. Gandrothula, T. Kamikawa, J. S. Speck, S. Nakamura,
S. P. DenBaars, Appl. Phys. Express 14, 031002 (2021).
23. M. T. Björk et al., Appl. Phys. Lett. 81, 4458–4460 (2002).
24. M. Law, J. Goldberger, P. Yang, Annu. Rev. Mater. Res. 34,
83–122 (2004).
25. M. Gurioli, Z. Wang, A. Rastelli, T. Kuroda, S. Sanguinetti,
Nat. Mater. 18, 799–810 (2019).
26. T. Kanne et al., Nat. Nanotechnol. 16, 776–781 (2021).
27. Y. Gong et al., Nat. Mater. 13, 1135–1142 (2014).
28. Y. Shi et al., Nano Lett. 12, 2784–2791 (2012).
29. Z. Chen et al., Sci. Adv. 7, eabk0115 (2021).
30. K. Reidy et al., Nat. Commun. 12, 1290 (2021).
31. Y. Kim et al., Nature 544, 340–343 (2017).
32. H. Kim et al., ACS Nano 2, 40 (2022).
33. A. E. White, K. T. Short, R. C. Dynes, J. P. Garno, J. M. Gibson,
Appl. Phys. Lett. 50, 95–97 (1987).
34. T. George, R. W. Fathauer, Appl. Phys. Lett. 59, 3249–3251 (1991).
35. N. Briggs et al., Nat. Mater. 19, 637–643 (2020).
36. G. Algara-Siller et al., Nature 519, 443–445 (2015).
37. Q. Yang et al., Nature 588, 250–253 (2020).
38. K. Mustonen et al., Adv. Mater. 34, e2106922 (2022).
39. Z. Y. Al Balushi et al., Nat. Mater. 15, 1166–1171 (2016).
40. B. Pécz et al., Adv. Mater. 33, e2006660 (2021).
41. A. Allain, J. Kang, K. Banerjee, A. Kis, Nat. Mater. 14, 1195–1205
(2015).
42. I. Popov, G. Seifert, D. Tománek, Phys. Rev. Lett. 108, 156802
(2012).
43. K. Reidy et al., Nano Lett. 22, 5849–5858 (2022).
44. M. H. Jacobs, M. J. Stowell, Philos. Mag. 11, 591–603 (1965).
45. D. B. Williams, C. B. Carter, Transmission Electron Microscopy:
A Textbook for Materials Science (Springer, 2009).
46. C. Ophus, Microsc. Microanal. 25, 563–582 (2019).
47. S. Forti et al., Nat. Commun. 11, 2236 (2020).
48. A. B. Yankovich et al., Nano Lett. 19, 8171–8181 (2019).
49. Y. Cui et al., Twisted epitaxy of gold nanodisks grown between
twisted substrate layers of molybdenum disulfide [Dataset],
Dryad (2023); https://doi.org/10.5061/dryad.wpzgmsbv3.
AC KNOWLED GME NTS
We acknowledge the use and support of the Stanford Nano Shared
Facilities (SNSF). Part of this work was performed at SNSF,
partially supported by the National Science Foundation under
award ECCS-1542152. We also acknowledge helpful discussions
of 4D-STEM data analysis with C. Ophus at the National Center
for Electron Microscopy at the Molecular Foundry, Lawrence
Berkeley National Laboratory. Funding: This work was supported
by the Office of Basic Energy Sciences, US Department of Energy,
p
Division of Materials Science and Engineering, DE-AC02-
76SF00515. Author contributions: Conceptualization: Y.C. (first
author) and Y.C. (corresponding author). Sample preparation:
Y.C. (first author), Y.W., and Y.L. TEM characterization: Y.C. (first
author) and R.S. AFM characterization: Y.C. (first author). DFT
calculations: Y.C. (first author), J.W., and E.B. 4D-STEM: Y.C.
(first author) and R.S. Discussion: W.Z., Z.Z., and J.Z. Writing –
original draft: Y.C. (first author). Writing – review & editing: H.Y.H.,
g
R.S., and Y.C. (corresponding author). Competing interests:
The authors declare that they have no competing interests. Data
and materials availability: All data needed to evaluate the
conclusions in this paper are present in the main text or the
supplementary materials. Codes that support the plots within
y
this paper are available at Dryad (49). License information:
Copyright © 2024 the authors, some rights reserved; exclusive
licensee American Association for the Advancement of Science.
No claim to original US government works. https://www.science.
org/about/science-licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.adk5947
Materials and Methods
Supplementary Notes
Figs. S1 to S31
References (50–57)
Submitted 30 August 2023; accepted 27 November 2023
y g
10.1126/science.adk5947
y
,
8 of 8
RES EARCH

CONSERVATION ments to the number of IUCN criteria applied

and uncertainties in species information (tables
Comprehensive conservation assessments reveal high S2 to S4). Finally, we predicted the conserva-
tion status of tropical forests worldwide using
extinction risks across Atlantic Forest trees the relationship between species threat and
habitat loss observed in the Atlantic Forest.
Renato A. F. de Lima1,2*, Gilles Dauby3, André L. de Gasper4, Eduardo P. Fernandez5, Alexander C. Vibrans6,
Conservation status of the Atlantic Forest
Alexandre A. de Oliveira7, Paulo I. Prado7, Vinícius C. Souza2, Marinez F. de Siqueira5,8, Hans ter Steege1,9
tree flora
Biodiversity is declining globally, yet many biodiversity hotspots still lack comprehensive species We classified two-thirds of the 4950 tree spe-
conservation assessments. Using multiple International Union for Conservation of Nature (IUCN) Red List cies populations that occur in the Atlantic
criteria to evaluate extinction risks and millions of herbarium and forest inventory records, we present Forest as threatened following the IUCN Red
automated conservation assessments for all tree species of the Atlantic Forest biodiversity hotspot, List Categories and Criteria (Fig. 1A). The per-
including ~1100 heretofore unassessed species. About 65% of all species and 82% of endemic species centage of threatened species increases to 82%
are classified as threatened. We rediscovered five species classified as Extinct on the IUCN Red List when only endemic species are considered
and identified 13 endemics as possibly extinct. Uncertainties in species information had little influence (Fig. 1B), with 2025 endemic Atlantic Forest
on the assessments, but using fewer Red List criteria severely underestimated threat levels. We suggest trees globally threatened with extinction (data
that the conservation status of tropical forests worldwide is worse than previously reported. S2). The Atlantic Forest’s Red List Index (RLI),
which measures the overall conservation status

H
for a list of species and ranges from zero (all
uman pressure on nature has increased They require detailed species information and species are extinct) to one (no threatened spe-

in recent decades, particularly in the
tropics, where most of the planet’s bio-
diversity resides (1, 2). Consequently, we
face a global biodiversity crisis (3). Re-
versing this crisis is a pressing challenge and
begins by classifying species based on extinction
the time, training, and resources to apply the
IUCN Red List Categories and Criteria on a
species-by-species basis (9), all of which are
limited, especially in the tropics (4, 5). There-
fore, automated assessments are increasingly
being proposed as complements or alterna-
p
cies) (23), was 0.542 [95% confidence interval
(CI): 0.534 to 0.550]. These numbers are worse
than the global averages of threatened species
(25%) and RLI values (95% CI: 0.55 to 1) reported
for other groups of organisms (2), indicating
that threat levels in highly modified regions,

risks, which are used to monitor biodiver-
sity and prioritize conservation actions (4, 5).
Also known as red listing, these conservation
assessments are a cornerstone of global con-
tives to manual assessments (4, 10–13) to
provide fast-track conservation assessments
for megadiverse regions (14–16).
Assessments for tropical biodiversity hot-
g
such as the Atlantic Forest, can be much higher
than global averages.
Many emblematic endemic trees of the
Atlantic Forest were classified as threatened in

servation programs, such as the International
Union for Conservation of Nature (IUCN) Red
List, which categorizes species extinction risks
based on one or multiple criteria, including
population size decline (criterion A), geographic
range (criterion B), and very small populations
(criterion D).
Efforts to include species on the IUCN Red
List have grown in recent years, but much
remains to be done (4–7). Even for the well-
spots, where most threatened species occur
(17), remain rare. One of these hotspots is the
Atlantic Forest in eastern South America,
which has more than 15,000 plant species, of
which half are endemic (18). With 35% of the
South American human population living with-
in its borders, about 80% of its original cover
has been lost, and deforestation and degrada-
tion remain high (19, 20). Species conservation
assessments are limited to about 25% of the
y
this work. The iconic Paubrasilia echinata
(brazilwood), the tree that gave its name to
Brazil, was listed as Critically Endangered (CR)
owing to an estimated 84% drop in its pop-
ulation size over the past three generations.
The once common Araucaria angustifolia
(Paraná pine), Euterpe edulis (palm heart),
and Ilex paraguariensis (yerba mate) also ex-
perienced declines in their wild populations
of at least 50% and are thus classified as En-

studied trees of Europe, red listing efforts have
been published only recently (8). One reason
why only a small part of global biodiversity
has up-to-date conservation assessments is the
difficulty in carrying out these assessments.
Atlantic Forest flora and are mostly being con-
ducted using few IUCN Red List criteria (21). A
comprehensive assessment at the Atlantic Forest
scale could provide insights into the conserva-
tion status of other tropical biodiversity hot-
spots, which do not all have the same amount
y g
dangered (EN). Endemic timber species, such as
Cariniana legalis, Dalbergia nigra, Melanoxylon
brauna, Myrocarpus frondosus, Ocotea odorifera,
Ocotea porosa, Parapiptadenia rigida, and
Paratecoma peroba, also experienced declines
ranging from 53 to 89% and are thus classified

y
1
Tropical Botany, Naturalis Biodiversity Center, Darwinweg 2, of information available as the Atlantic Forest. as EN or CR.

2333 CR Leiden, Netherlands. 2Departamento de Ciências
Biológicas, ESALQ, Universidade de São Paulo, Avenida
Pádua Dias, 11, 13418-900 Piracicaba, Brazil. 3Botanique et
Modélisation de l’Architecture des Plantes et des Végétations
(AMAP), Université de Montpellier, IRD, CNRS, INRAE, CIRAD,
Montpellier, France. 4Departamento de Ciências Naturais,
Universidade Regional de Blumenau, Rua Antônio da Veiga, 140,
89030-903 Blumenau, Brazil. 5Centro Nacional de Conservação
da Flora (IUCN SSC Brazil Plant Red List Authority), Instituto
de Pesquisas Jardim Botânico do Rio de Janeiro, Rua Pacheco
Leão, 915, 22460-030 Rio de Janeiro, Brazil. 6Departamento de
Engenharia Florestal, Universidade Regional de Blumenau, Rua
São Paulo, 3250, 89030-000 Blumenau, Brazil. 7Departamento
de Ecologia, Instituto de Biociências, Universidade de São
Paulo, Rua do Matão, trav. 14, 321, 05508-090 São Paulo,
Brazil. 8Departamento de Biologia, Pontifícia Universidade
Católica do Rio de Janeiro, Rua Marquês de São Vicente 225,
22451-900 Rio de Janeiro, Brazil. 9Quantitative Biodiversity
Dynamics, Department of Biology, Utrecht University, 3584 CS
Utrecht, Netherlands.
*Corresponding author. Email: raflima@usp.br
We present the conservation status of the
Atlantic Forest tree flora, which represents a
third of the entire hotspot’s plant diversity and
is crucial to providing people with ecosystem
services (5, 7, 12). We automated the conser-
vation assessments for nearly 5000 species
using more than 800,000 herbarium records,
1.3 million tree counts from forest inventories,
and information on species life histories, com-
mercial uses, and long time-series of habitat
loss (fig. S1 and data S1) (22). We developed a
replicable workflow that strictly adheres to the
IUCN Red List Categories and Criteria (9) and
delivers conservation assessments based on
the IUCN criteria A to D (table S1). This work-
flow also evaluates the sensitivity of the assess-
,
Most species (75%) were classified as threat-
ened under IUCN criterion A, which evaluates
population decline in the past three gener-
ations. The high deforestation of the Atlantic
Forest led to 57% of endemic tree species having
estimated population declines above the IUCN
threshold of 30% (9). By contrast, only 7% of
the endemics showed declines below 30%. An-
other important IUCN criterion to detect threat
in the Atlantic Forest was B2 (28% of the
cases), which is related to small areas of oc-
cupancy (AOO). AOO was below the thresh-
old of 2000 km2 for most species (median of
208 km2), but more than two-thirds of these
species occurred in more than 10 locations
or were not severely fragmented (figs. S2 and

de Lima et al., Science 383, 219–225 (2024) 12 January 2024 1 of 6

RES EARCH | R E S E A R C H A R T I C L E

S3) and hence did not meet the conditions
necessary to be classified as threatened under
criterion B (9). Only 5% of all species were
classified as threatened under the IUCN crite-
ria C and/or D (small and declining popula-
tions or very small populations; fig. S4), with
most species (94%) exceeding the critical size
of 10,000 mature individuals.
Comparison with previous IUCN assessments
We found previous assessments for 59 and
49% of the Atlantic Forest tree flora at global
(24) and national levels (21), respectively. We
thus present the first assessments for 1120
species, 456 of which are Atlantic Forest en-
demics. We rediscovered five species previously
classified as Extinct (EX) on the IUCN Red
List: Campomanesia lundiana, Chrysophyllum
januariense, Myrcia neocambessedeana, Pouteria
stenophylla, and Pradosia glaziovii. These spe-
cies were classified as EX because they were
to keep the IUCN Red List up to date (25–27) and
how approaches like the one we used can facil-
itate a higher frequency of reassessments (13).
Overall, our reassessments rarely resulted in
up-listing species by more than two threat cat-
egories (4%) or down-listing species previously
assessed as threatened (also 4%), mainly when
comparisons considered assessments using only
criterion B (table S5). This confirms that auto-
mated assessments can provide accurate pictures
of species conservation statuses for tropical re-
gions (14, 15). About 3% of the species moved
from the Least Concern (LC) or Near Threatened
(NT) IUCN Red List categories to CR. These spe-
cies were previously classified as LC owing to their
large extent of occurrence (EOO) (>20,000 km2),
but we estimated population reductions greater
than 80%, which is the IUCN threshold to clas-
sify species as CR. Another 3% that were pre-
viously classified as threatened were assessed
in this work as LC or NT. For these species, we
in our reassessments (Fig. 2). This deterioration
may be due to differences in the amount of in-
formation available, the number of IUCN criteria
used, or a genuine decline in the species’ con-
servation statuses (28). To separate these causes,
we compared changes for 1170 endemic species
that we reassessed using only IUCN criterion B.
About 31% of these species had changes in their
assessments, resulting in a significantly better
RLI value (95% CI: 0.82 to 0.85) than the pre-
vious one (95% CI: 0.74 to 0.78). This unexpected
finding is likely due to the inclusion of new
occurrences, which tends to classify species
under lower levels of threat (29). Thus, the
observed RLI deterioration is more related
to the inclusion of more IUCN criteria in our
reassessments than to a genuine decline in
the Atlantic Forest conservation status (see
next section).
The influence of the number of IUCN criteria

known only from their type specimens at the
time of the previous assessments (1998). We
found taxonomically and geographically vali-
dated herbarium records for 2008 or later for
all five species. Only C. lundiana had no recent
taxonomically vetted record, suggesting that it
found no indications of population declines
≥30%, severe fragmentation, or occurrence in
fewer than 10 locations (figs. S2 and S3), which
are the necessary conditions to detect threat un-
der the IUCN criterion A or B (9).
The RLI value for the endemic Atlantic Forest
p
on species assessments
Around 69% of the plants on the IUCN Red
List are assessed using only criterion B (spe-
cies geographic range) (24) owing to limited
data on population size and trends for most
organisms (6, 28, 29). To evaluate the impact

could indeed be assessed as EX or that its tax-
onomic delimitation remains uncertain. All five
species remained classified as threatened in
this work but under different categories, which
tree flora deteriorated from 0.74 (95% CI: 0.73
to 0.76) in the assessments that are available on
the IUCN Red List to 0.50 (95% CI: 0.49 to 0.51)
in our reassessments (Fig. 2). Only 18% of the
g
of using multiple IUCN criteria on conserva-
tion assessments, we compared the results for
a subset of 2698 species that were assessed in
this work using four IUCN criteria (A, B, C,

y
emphasizes the importance of new information species previously assessed as LC remained so and D; table S1). We found that assessments

B2 (2.7%)
A All populations B Endemics
A2 A2 B2
(5.8%) (4.2%) (5.1%)

B1+2
(10.8%)
CR CR
(17.1%) (9.2%) (10.8%)
A2
NA LC

y g
(30.4%) All
(17.1%) (6.1%) (16.9%)
A2
(34.9%)
EN NT (0.9%)
(40.9%) B2
LC (3.4%) EN

y
(17.0%) VU (52.8%)
B1+2 B1+2 (1.8%)
(18.7%)

,
(12.5%)
NT A2, B1+2
(1.4%)
(1.3%)
VU B1+2 A2, B1+2 A2
All (14.4%) (4.2%) (1.5%) (12.1%)
(4.5%) B2
(4.7%)
B2 B1+2
A2 (6.9%) (8.2%) A2, B1+2
(9.7%) (2.8%)
B2 (2.8%)
B1+2 (1.2%)

Fig. 1. The proportion of populations classified under each threat category
and the corresponding IUCN criteria assigned to the categories. (A and
B) Results are presented for the populations of (A) all tree species occurring
in the Atlantic Forest (n = 4953) and (B) only endemic species (n = 2464).
The proportion of populations classified under each threat category and the
corresponding IUCN criteria (indicated by letters) are shown in the central pie
chart and external donut chart, respectively. Populations classified as NA mainly
correspond to the Not Applicable category of IUCN regional assessments,
including vagrant species. For clarity, panels do not include the Data
Deficient category (too few species). Here, we define “population” as a group
of individuals of the same species that inhabit the same geographical
area (the Atlantic Forest, in our case). CR, Critically Endangered (red); EN,
Endangered (orange); LC, Least Concern (dark green); NA, Not Applicable (gray);
NT, Near Threatened (green); VU, Vulnerable (yellow).

de Lima et al., Science 383, 219–225 (2024) 12 January 2024 2 of 6

RES EARCH | R E S E A R C H A R T I C L E

A B
Global EX Regional
Previous assess. CR CR New assess. Previous assess. CR CR New assess.

EN
EN
VU

EN

EN
VU
NT

NT
LC

LC
VU

U
V

p
NT NT
DD LC DD LC
DD DD

Fig. 2. A comparison between previous conservation assessments and the of the two assessments. Previous conservation assessments were obtained
new ones presented in this work. (A and B) A comparison between previous at the global level from version 2022-2 of the IUCN Red List (www.iucnredlist.

g
assessments (left arcs) and the assessments in this work (right arcs) is org) and from the national red lists from Argentina, Brazil, and Paraguay.
presented at (A) global and (B) regional levels. The widths of the linking lines The color legend is the same as in Fig. 1. DD, Data Deficient (gray);
correspond to the proportion of species shared between categories of threat EX, Extinct (black).

y
in the population (p) (tables S2 and S3), based
Table 1. A comparison of the assessments based on individual and multiple IUCN criteria for on a combination of species’ growth forms
the populations of all species and only endemic species. A comparison was conducted for a subset of (e.g., large trees) and ecological groups (e.g.,
Atlantic Forest populations that had enough information available to assess criteria A, B, C, and D. For the pioneers; data S3). To assess the sensitivity
RLI, values in brackets represent the 95% CI around mean estimates obtained from 50,000 bootstraps; of our assessments to these generalizations,
different superscript lowercase letters indicate differences of RLI means among categories based on we compared them with assessments that
the 50,000 bootstraps that were run for the same set of populations. We define “population” as a group of were generated using varying values of GL and
individuals of the same species that inhabit the same geographical area (the Atlantic Forest, in our case). p. We found that only the use of GLs smaller
than 25 years could considerably change the

y g
overall proportion of threatened species (fig.
Populations of all species (n = 2698) Only endemic species (n = 1586) S5). This occurred because, for most species,
Criteria
Threatened (%) RLI Threatened (%) RLI the peak of Atlantic Forest loss (1950–1980) is
more recent than the three GLs defined by the
A 91.4 0.471 (0.464‒0.479)a 90.3 0.498 (0.488–0.507)a
.....................................................................................................................................................................................................................
b IUCN to evaluate population size decline. So, 57
B 10.7 0.951 (0.945‒0.956) 16.5 0.924 (0.915‒0.932)b
.....................................................................................................................................................................................................................
and 92% of species threatened under criterion

y
c
C 2.5 0.993 (0.991‒0.995) 3.2 0.992 (0.990‒0.995)c
.....................................................................................................................................................................................................................
c A would have remained so if we had considered
0.993 (0.991‒0.995)c
,
D 2.3 0.994 (0.993‒0.996) 3.2
.....................................................................................................................................................................................................................
a only one or two GLs, respectively. We also found
A, B, C, and D 91.7 0.470 (0.463‒0.478) 90.6 0.489 (0.479‒0.499)a
..................................................................................................................................................................................................................... that using smaller values of p changed few as-
sessments under criteria C and D (figs. S6 and
S7) because populations mostly remained well
would be substantially different if only crite- use only criterion B can severely underestimate above the IUCN critical population size of 10,000.
rion B, C, or D was used (Table 1), with six the conservation status of regional biotas, espe- Our group-specific approach is biologically mean-
times fewer species found to be threatened if cially in highly modified regions such as the ingful because it is based on species life histories
we had used only criterion B. This discrep- Atlantic Forest and other biodiversity hotspots. and evidence from long-term monitoring of
ancy is likely due to the fact that criterion B tropical trees (22).
does not consider population declines with- Uncertainty in species information Only records with species identifications that
in species ranges. Common endemic species and identifications were vetted by taxonomists should be included
with large ranges (>300,000 km2) and pop- Much of the species’ information that is needed in conservation assessments (25). In our data-
ulation sizes (>2,000,000 mature individu- to apply IUCN criteria A, C, and D is missing set, only 38% of the records were taxonomically
als), such as Metrodorea nigra or Picramnia for tropical regions. Therefore, we had to make vetted, which would lead to fewer records avail-
ramiflora, had estimated population declines generalizations of species’ generation lengths able per species and thus less-reliable assessments
greater than 90%. Therefore, assessments that (GLs) and proportions of mature individuals (10, 30). Therefore, we implemented an approach

de Lima et al., Science 383, 219–225 (2024) 12 January 2024 3 of 6

RES EARCH | R E S E A R C H A R T I C L E
Fig. 3. The spatial distribution
A B
of tree species threat in the
All species
Atlantic Forest biodiversity hot-
spot. (A and B) Maps present the
spatial interpolation of the RLI
across the Atlantic Forest bio-
diversity hotspot, considering the
populations of (A) all tree species
and (B) only endemic species. The
RLI interpolation was obtained
based on the list of species
recorded (and their threat
categories) across the cells of a
grid covering the entire region.
The RLI ranges from zero (all
species are classified as extinct)
to one (all species are not threat-
ened). The color scale ranges
from lower, or worse, RLI values
(dark red) to higher, or better,
values (yellow).
to add records while losing as little taxonomic
confidence as possible (fig. S8 and table S4).
We compared this approach to assessments
using only taxonomically vetted records (fig.
S8A) and found that our approach resulted in
fewer threatened species (17%) than the latter
approach (27%). Consequently, the RLI was
significantly higher here (0.906; 95% CI: 0.900
to 0.912) than when using only taxonomically
vetted records (0.843; 95% CI: 0.841 to 0.855).
Differences between approaches were larger
for assessments that used less than 60% of
taxonomically vetted records (fig. S8B). These
differences emerge from the different number
of records available: Assessments that use more
records often yield higher EOO and AOO esti-
mates (30, 31). Alternatively, adding misiden-
tified records outside species’ natural ranges
overestimates their EOO and AOO (25). Together,
these explanations suggest that we likely under-
estimated the threat status of species that have
fewer taxonomically vetted records.
Threatened species in time and space
Of the 815,000 valid herbarium records, the
first dates from the 17th century, but 79% were
made after the 1980s. We found no valid records
over the past 50 years for 41 endemic species.
Thirteen species are only known from their
type specimen (fig. S9), which is one of the
criteria for tagging threatened species as pos-
sibly extinct (9). These are priority species for
new studies (data S2) to assess whether they
have limited sampling and/or taxonomic treat-
ment or are probably extinct in nature (9).
de Lima et al., Science 383, 219–225 (2024) 12 January 2024
Endemics
0.7
0.6
0.5
0.4
Red List Index
0.3
The spatial distribution of threatened spe-
cies in the Atlantic Forest was similar when
considering all species or only endemic ones.
We found that the western, central, and north-
ern regions had the worst RLI values (Fig. 3)
and the highest proportions of threatened spe-
cies (figs. S10 to S13). Previous studies have
shown that the western and northern regions
have fewer species and endemism than the
central region (18, 32), but they all share the
highest fragmentation levels of the Atlantic
Forest (19), except for the Misiones region in
Argentina. In addition, we found that the Serra
do Mar and Araucaria regions, which are home
to the largest Atlantic Forest remnants and
systems of strictly protected areas (19, 20), had
smaller proportions of threatened species. This
highlights that threatened species are concen-
trated where habitat loss and fragmentation
are greater. Therefore, in situ conservation ac-
tions in the Atlantic Forest (e.g., reverting forest
degradation, fostering landscape restoration,
and the creation and strengthening of pro-
tected areas) should target not only areas with
high species richness and endemism (18, 33)
but also highly modified areas where threat-
ened species are less likely to sustain their pop-
ulations owing to low habitat quantity, quality,
and connectivity (34, 35).
We found that most threatened endemic
trees (82%) had at least one confirmed occur-
rence inside strictly protected areas. However,
75% of them had less than a quarter of their
records and one-tenth of their EOO inside
protected areas (fig. S14). Species classified as
0.7
0.6
p
0.5
0.4
Red List Index
g
0.3
CR had fewer occurrences in protected areas
than those in other threat categories, which is
y
partially explained by the smaller proportion
of protected areas within their EOO (fig. S14).
Also, the terrestrial area of habitat (AOH) that
remained in 2018 was significantly lower for
species classified as CR (median of 19%) than
for those in other threat categories (fig. S15).
These results indicate that many threatened
Atlantic Forest tree species occur mostly in
unprotected areas and have limited habitat
left. A comprehensive study on the type, design,
y g
and extent of conservation actions that are able
to increase habitat availability, quality, and
protection of threatened Atlantic Forest trees
is needed. Maps of regions with higher con-
centrations of threatened species (figs. S11 and
S13), especially CR endemics (fig. S16), are cru-
y
cial for prioritizing those conservation actions.
,
Implications for tree species conservation
The conservation status of the Atlantic Forest
tree flora is alarming but probably worse in
reality. Our assessments focused more on the
decline of habitat quantity (i.e., deforestation)
rather than quality (i.e., forest fragmentation
and degradation). Estimated population de-
clines would have been greater if only the
intact Atlantic Forest was considered (3.5 to
7% instead of the 12 to 28% of all forest cover)
(19, 36). Additionally, we used conservative
values of GL, p, and exploitation levels for
valuable species and incorporated as much
data as possible for the assessments of species
with fewer occurrences. These choices likely
4 of 6
RES EARCH | R E S E A R C H A R T I C L E
2 6
8 15
13
16
12
1
0.3 0.4
Fig. 4. Predicted RLI values for the 18 main tropical forest areas of the
world. Predictions are based on the relationship between population size
reductions and habitat loss, which could be obtained for all forests from forest
cover maps and reports of their number of endemic species. Thus, these
predictions are based solely on the population size reductions (i.e., IUCN
led to more conservative estimates of species
threat (i.e., smaller species ranges, population
sizes, and/or declines), particularly for late-
successional species (37). This is less the case for
early-successional species, whose population de-
clines due to habitat loss are often mitigated by
changes in habitat quality (e.g., increase of forest
edges). These changes were accounted for in
our assessments (22), but there are still un-
certainties related to their population declines.
Furthermore, most Atlantic Forest loss occurred
in the past 50 to 70 years, which, for many tree
species, falls within two to three GLs into the
past. So, despite the smaller Atlantic Forest
deforestation today, the effects of past hab-
itat loss, fragmentation, and selective logging
on these long-lived species may not have had
enough time to fully express themselves (38),
which suggests an extinction debt yet to be
paid in the coming decades (39, 40).
The status of the Atlantic Forest tree flora
has direct implications for the Global Tree
Assessment initiative (5, 7) and the IUCN post-
2020 global biodiversity framework. We pro-
vide all the IUCN-required information to ease
the incorporation of our assessments into their
Red List (data S4), seeking to bridge the gap
between research and its integration into con-
servation practice (13). We also highlight hab-
itat loss as the main threat to tropical tree
diversity, which raises the question about its
impact on other tropical forests (14, 16, 17, 41–43).
Thus, based on the present forest cover of 18
main tropical forests and the relationship be-
tween species threat and Atlantic Forest loss
(fig. S17) (22), we roughly estimate that 20,504
to 24,910 tropical tree species are likely threat-
de Lima et al., Science 383, 219–225 (2024) 12 January 2024
5 4
17
3
7
0.5 0.6 0.7
criterion A2) of endemic species. The RLI ranges from zero (all species are
classified as extinct) to one (no species are classified as threatened). The color
scale ranges from lower, or worse, RLI values (dark red) to higher, or better,
values (green). See supplementary materials and tables S6 and S7 for details
and the full tropical forest names that correspond to the numbers in the map.
ened because of habitat loss alone (Fig. 4 and
tables S6 and S7). This represents 35 to 43% of
tree species worldwide (7, 14) and confirms
that tropical forests shelter most of the global-
ly threatened species (17). Despite its assump-
tions and limits, this prediction includes only
tree species that are endemic to these 18 trop-
ical forests (22). Therefore, it is higher than the
present estimate of 30% of threatened tree
species (7) and closer to the 43% that is
estimated using artificial intelligence (12).
If we account for temperate species and trop-
ical species that are shared among tropical
forests or woodlands (i.e., nonendemics),
threats of habitat loss to global tree diversity
will be even greater than previously recog-
nized, making trees one of the most threat-
ened groups of organisms on the planet (2).
Considering the ecological and sociocultural
importance of tree species (5, 7) and the con-
tinued pressure on tropical forests (1–3),
fighting tropical deforestation and effectively
implementing both in situ and ex situ conser-
vation must be prioritized if we are to prevent
the extinction of thousands of tree species in
the next decades.
RE FERENCES AND NOTES
1. S. L. Lewis, D. P. Edwards, D. Galbraith, Science 349, 827–832 (2015).
2. S. Díaz et al., Science 366, eaax3100 (2019).
3. G. Ceballos et al., Sci. Adv. 1, e1400253 (2015).
4. S. P. Bachman et al., Biol. Conserv. 234, 45–55 (2019).
5. A. Newton et al., Oryx 49, 410–415 (2015).
6. N. A. Brummitt et al., PLOS ONE 10, e0135152 (2015).
7. Botanic Gardens Conservation Initiative (BGCI), “State of the
world’s trees” (BGCI, 2021).
8. M. Rivers et al., “European Red List of Trees” (International
Union for Conservation of Nature, 2019).
9. IUCN Standards and Petitions Committee, “Guidelines for using the
IUCN Red List Categories and Criteria, Version 14” (IUCN, 2019).
10
18
11
14
9
Red List Index
0.8 0.9 1
p
g
10. E. Nic Lughadha et al., Philos. Trans. R. Soc. London Ser. B 374,
20170402 (2018).
11. S. E. Darrah, L. M. Bland, S. P. Bachman, C. P. Clubbe,
A. Trias-Blasi, Divers. Distrib. 23, 435–447 (2017).
12. S. V. Silva, T. Andermann, A. Zizka, G. Kozlowski, D. Silvestro,
y
Front. Plant Sci. 13, 839792 (2022).
13. V. Cazalis et al., Trends Ecol. Evol. 37, 359–370 (2022).
14. H. Ter Steege et al., Sci. Adv. 1, e1500936 (2015).
15. T. Stévart et al., Sci. Adv. 5, eaax9444 (2019).
16. V. H. F. Gomes, I. C. G. Vieira, R. P. Salomão, H. ter Steege,
Nat. Clim. Chang. 9, 547–553 (2019).
17. T. M. Brooks et al., Conserv. Biol. 16, 909–923 (2002).
18. R. A. F. de Lima, V. C. Souza, M. F. de Siqueira, H. ter Steege,
Biol. Conserv. 252, 108825 (2020).
19. M. R. Rosa et al., Sci. Adv. 7, eabc4547 (2021).
20. R. A. F. de Lima et al., Nat. Commun. 11, 6347 (2020).
21. E. Martins, R. Loyola, G. Martinelli, Ann. Mo. Bot. Gard. 102,
347–356 (2017).
22. Materials and methods are available as supplementary materials.
y g
23. S. H. M. Butchart et al., PLOS ONE 2, e140 (2007).
24. IUCN, The IUCN Red List of Threatened Species, Version 2022-2
(2023); https://www.iucnredlist.org [accessed 26 January 2023].
25. E. M. Nic Lughadha et al., Conserv. Biol. 33, 511–522 (2019).
26. A. Newton, S. Oldfield, Endanger. Species Res. 6, 137–147 (2008).
27. A. M. Humphreys, R. Govaerts, S. Z. Ficinski, E. Nic Lughadha,
y
M. S. Vorontsova, Nat. Ecol. Evol. 3, 1043–1047 (2019).
28. L. Santini et al., Conserv. Biol. 33, 1084–1093 (2019).
,
29. Ł. Tracewski et al., Conserv. Biol. 30, 1070–1079 (2016).
30. M. C. Rivers et al., Biol. Conserv. 144, 2541–2547 (2011).
31. S. Bachman, B. E. Walker, S. Barrios, A. Copeland, J. Moat,
Biodivers. Data J. 8, e47018 (2020).
32. V. P. Zwiener, R. A. F. de Lima, A. Sánchez-Tapia, D. S. B. Rocha,
M. C. M. Marques, The Atlantic Forest, M. C. M. Marques,
C. E. V. Grelle, Eds. (Springer, 2021), pp. 115–131.
33. A. Bonn, A. S. L. Rodrigues, K. J. Gaston, Ecol. Lett. 5,
733–741 (2002).
34. K. V. Root, Ecol. Appl. 8, 854–865 (1998).
35. J. A. Hodgson, A. Moilanen, B. A. Wintle, C. D. Thomas,
J. Appl. Ecol. 48, 148–152 (2011).
36. S. Sloan, C. N. Jenkins, L. N. Joppa, D. L. A. Gaveau,
W. F. Laurance, Biol. Conserv. 177, 12–24 (2014).
37. D. J. Bender, T. A. Contreras, L. Fahrig, Ecology 79, 517–533 (1998).
38. M. B. Lauterjung et al., For. Ecol. Manage. 435, 144–150 (2019).
39. M. Vellend et al., Ecology 87, 542–548 (2006).
40. Q. Cronk, Science 353, 446–447 (2016).
41. N. S. Sodhi et al., Biodivers. Conserv. 19, 317–328 (2010).
42. E. W. Seabloom, A. P. Dobson, D. M. Stoms, Proc. Natl. Acad.
Sci. U.S.A. 99, 11229–11234 (2002).
5 of 6
RES EARCH | R E S E A R C H A R T I C L E
43. M. Barstow, T. Jimbo, K. Davies, . Plants People Planet 5,
508–519 (2023).
44. D. A. L. Canhos et al., Biota Neotrop. 22, e20221394 (2022).
45. L. A. E. da Silva et al., Rodriguésia 68, 391–410 (2017).
46. GBIF.org, GBIF Occurrence Download (2019); https://doi.org/
10.15468/dl.mzmat2.
47. R. A. F. de Lima et al., Biodivers. Conserv. 24, 2135–2148 (2015).
48. R. A. F. de Lima, G. Dauby, Scripts, functions and data to
reproduce the results of “Comprehensive conservation
assessments reveal high extinction risks across Atlantic Forest
trees,” Version 0.0.1, Zenodo (2023); https://doi.org/10.5281/
zenodo.8253431.
49. ESA, “Land Cover CCI Product User Guide Version 2” (2017);
http://maps.elie.ucl.ac.be/CCI/viewer/download/ESACCI-LC-
Ph2-PUGv2_2.0.pdf.
50. M. Hoffman, K. Koenig, G. Bunting, J. Costanza, K. J. Williams,
Biodiversity Hotspots, Version 2016.1, Zenodo (2016);
https://doi.org/10.5281/zenodo.3261807.
51. E. Dinerstein et al., Bioscience 67, 534–545 (2017).
52. UNEP-WCMC, IUCN, Protected planet: The World Database on
Protected Areas (WDPA) (2020); www.protectedplanet.net.
ACKN OW LEDG MEN TS
We thank H. R. Akçakaya and CNCFlora staff (P. da Rosa,
G. Martinelli, E. Martins, R. Loyola, P. H. A. de Melo, and
T. L. B. da Cunha) for advice on the IUCN Red List Categories
and Criteria. Special thanks go to M. Rivers and E. Beech [Botanic
Gardens Conservation International (BGCI)] for reviewing the
methods used to generate the conservation assessments. We
acknowledge J. S. Wright and H. Muller-Landau (Smithsonian
Tropical Research Institute) for their input on the estimation of age
de Lima et al., Science 383, 219–225 (2024) 12 January 2024
of first reproduction and C. Canteiro (Global Center for
Species Survival, Indianapolis Zoo) for the clarifications of the
GL used in previous IUCN Red List assessments. We also thank
M. Barstow and E. Beech (BGCI) for providing their list of
commercial timber species. Funding: This research was funded by
the European Union’s Horizon 2020 research and innovation
program under the Marie Skłodowska-Curie grant agreement
no. 795114 (R.A.F.d.L. and H.t.S.). Forest inventory data was funded
by grants 2013/08722-5 [São Paulo Research Foundation
(FAPESP)], 312075/2013-8 [Brazilian National Council for
Scientific and Technological Development (CNPq)], and
2017TR1922 [Santa Catarina Research Foundation (FAPESC)] to
R.A.F.d.L., P.I.P., A.C.V., and A.L.d.G. A.L.d.G., A.C.V., M.F.d.S., and
P.I.P. were supported by CNPq grants 311303/2020-0, 305199/
2022-6, 310206/2019-7, and 313055/2020-3, respectively.
Author contributions: Conceptualization: R.A.F.d.L., H.t.S.; Data
curation: R.A.F.d.L., A.L.d.G., A.C.V., V.C.S.; Formal analysis:
R.A.F.d.L., G.D.; Funding acquisition: R.A.F.d.L., A.L.d.G., A.C.V.,
P.I.P., H.t.S.; Methodology: R.A.F.d.L., G.D., E.P.F., M.F.d.S., A.L.d.G.,
H.t.S.; Software: R.A.F.d.L., G.D.; Visualization: R.A.F.d.L.; Writing –
original draft: R.A.F.d.L.; Writing – review and editing: G.D.,
A.L.d.G., E.P.F., A.C.V., A.A.d.O., P.I.P., V.C.S., M.F.d.S., H.t.S.
Competing interests: The authors declare that they have no
competing interests. Data and materials availability: Herbarium
occurrence data were compiled from speciesLink, Jabot, and GBIF
(44–46). The forest inventory and species trait data were extracted
from the Neotropical Tree Communities database (version 4.1;
http://labtrop.ib.usp.br/doku.php?id = projetos:treeco:start) (20, 47)
and are available to researchers by request from the database
managers. The full sources of the publicly available herbarium,
inventory, and species trait data used in this study are provided as
supplementary materials (data S1). All scripts, functions, and data
that support the findings of this study are openly available at https://
github.com/LimaRAF/THREAT and deposited at Zenodo (48). The
maps used to assess habitat cover change, assess species occurrences
in protected areas, and delimit the main tropical forests are available
online (49–52). A summary of species information, population metrics,
and conservation assessment results are provided as supplementary
materials (data S2). Reference values of population parameters
from the literature and an exploration of parameters used for the
group-specific generalizations performed in this study are also provided
as supplementary materials (data S3). All files necessary to enter
the assessments in the IUCN Species Information Service (SIS)
system (SIS Connect) are provided as supplementary materials
(data S4). Data S1 to S4 are also deposited at Zenodo (48).
License information: Copyright © 2024 the authors, some
rights reserved; exclusive licensee American Association
for the Advancement of Science. No claim to original US government
works. https://www.science.org/about/science-licenses-journal-
article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.abq5099
Materials and Methods
Figs. S1 to S17
Tables S1 to S7
References (53–163)
Data S1 to S4
p
Submitted 12 April 2022; resubmitted 6 May 2023
Accepted 5 December 2023
10.1126/science.abq5099
g
y
y g
y
,
6 of 6
RES EARCH

CONSERVATION finning rather than to curtail fishing or reten-

tion of sharks outright. Currently, 94 juris-
Global shark fishing mortality still rising despite dictions and one RFMO have finning regulations
that require fishers to land whole sharks with
widespread regulatory change their fins naturally attached. A further four
jurisdictions and two RFMOs have regulations
Boris Worm1*, Sara Orofino2,3,4, Echelle S. Burns2,3,4, Nidhi G. D’Costa1, Leonardo Manir Feitosa4, requiring that sharks and their fins are landed
Maria L. D. Palomares5, Laurenne Schiller6, Darcy Bradley2,3,4,7 in a prescribed fin-to-carcass ratio, and 27 juris-
dictions and one RFMO have mixed finning
Over the past two decades, sharks have been increasingly recognized among the world’s most threatened regulations that differ by shark species. There
wildlife and hence have received heightened scientific and regulatory scrutiny. Yet, the effect of are 16 jurisdictions with unspecified finning
protective regulations on shark fishing mortality has not been evaluated at a global scale. Here we regulations and 75 jurisdictions with no relevant
estimate that total fishing mortality increased from at least 76 to 80 million sharks between 2012 and measures on record (Fig. 1D and tables S4 to
2019, ~25 million of which were threatened species. Mortality increased by 4% in coastal waters but S5). Regulatory attempts to eliminate shark
decreased by 7% in pelagic fisheries, especially across the Atlantic and Western Pacific. By linking finning and associated fishing mortality cur-
fishing mortality data to the global regulatory landscape, we show that widespread legislation designed rently occur in nearly 70% of maritime juris-
to prevent shark finning did not reduce mortality but that regional shark fishing or retention bans dictions globally; few such regulations existed
had some success. These analyses, combined with expert interviews, highlight evidence-based solutions to 20 years ago (Fig. 1). Concomitantly, we observed
reverse the continued overexploitation of sharks. a rapid increase in trade-restricted shark spe-
cies listed under the Convention on International

S
Trade in Endangered Species (CITES) (Fig. 1,

harks and their relatives (class Chon-
drichthyes) have persisted as powerful
ocean predators for over 400 million years,
yet many shark species are recently threat-
ened by overfishing, raising serious con-
cerns about species extinction (1) and associated
from 2012 to 2019 and contrasted these pat-
terns with relevant regulations adopted during
this time. We took a synthetic approach, col-
lating and analyzing all available shark-catch
data and regulatory data reported by individual
fisheries, countries, and RFMOs (fig. S1 and
p
A and B, red lines; and table S1). Likewise,
there has been a rapid increase in the number
of pelagic tuna fishing companies seeking
Marine Stewardship Council (MSC) ecocertifi-
cation and engaging in related fishery improve-
ment projects, two prominent market-based

consequences for ocean ecosystems (2, 3). Large
numbers of sharks have historically been caught
incidentally by the world’s pelagic tuna fisheries,
and elevated mortality has been linked to in-
tables S1 to S6). Fishery- and country-level data
were derived from detailed, spatially explicit
catch reconstructions based on United Nations
Food and Agriculture Organization (FAO)–
g
measures that prohibit shark finning on board
certified vessels (Fig. 1C and table S6).
Mapping spatial patterns of shark fishing
mortality, we found that current hotspots of

creasing demand for their fins, a valuable com-
modity in Asian markets (4, 5). In response,
protective regulations were introduced by na-
tional governments and regional fisheries
management organizations (RFMOs). Most
regulations aimed to eliminate the wasteful
practice of shark finning, in which valuable
fins are retained and shark carcasses are dis-
carded at sea (6). These efforts were further
supported by changing market forces (7), inter-
reported catches combined with regional data
and expert sources informing estimates of
unreported catches and discards that are not
included in the FAO statistics (11). Publicly
available RFMO shark-catch data recorded by
scientific observers or self-reported by fishers
were collated, evaluated, and spatially allocated
by using a recently developed Random Forest
machine learning approach (12) (tables S7
to S10). All catch data were converted to fishing
y
mortality are concentrated in coastal environ-
ments such as the Atlantic coast of North and
South America; West Africa; the northern Indian
Ocean; and the Coral Triangle, a particularly
biodiverse region spanning the national waters
of Indonesia, Malaysia, Papua New Guinea, and
the Philippines (Fig. 2A). This distribution also
holds broadly for species designated as threat-
ened with extinction by the International Union
for the Conservation of Nature (IUCN) (Fig. 2B).

national agreements to limit the trade of threat-
ened species (8), sustained nongovernmental
organization (NGO) advocacy to address pelagic
fisheries bycatch (9), and public awareness
campaigns intended to curb demand for shark
fins (10). No studies to date have investigated
mortality estimates by using species-, gear-,
and, where available, location-specific shark-
catch fate and postrelease mortality information
(figs. S1 to S7). We further conducted in-depth
interviews with 22 experts, whose deep knowl-
edge helped contextualize current trends in
y g
Individual taxa have fishing mortality hotspots
primarily in pelagic (e.g., silky shark, Fig. 2C)
or coastal environments (e.g., hammerhead
sharks, Fig. 2D), respectively, depending on
their preferred habitat and intersection with
fisheries (see fig. S8 for other taxa).

y
whether such shark finning and fishing regu- shark finning and mortality and the mecha- From 2017 to 2019, national waters accounted

lations have successfully reduced shark fishing
mortality globally.
In this study, we calculated global patterns
of shark fishing mortality at 1° by 1° resolution
1
Biology Department, Dalhousie University, Halifax, NS
B3H 4R2, Canada. 2Environmental Markets Lab, University
of California, Santa Barbara, CA 93106, USA. 3Marine
Science Institute, University of California, Santa Barbara,
CA 93106, USA. 4Bren School of Environmental Science and
Management, University of California, Santa Barbara,
CA, 93106, USA. 5Sea Around Us, Institute for the Oceans
and Fisheries, University of British Columbia, Vancouver,
BC, Canada V6T 1Z4. 6School of Public Policy and
Administration, Carleton University, Ottawa, ON K1S 5B6,
Canada. 7The Nature Conservancy, California Oceans
Program, Santa Barbara, CA 93117, USA.
*Corresponding author. Email: bworm@dal.ca
nistic drivers of these trends and served to vali-
date our quantitative assessment (full details
on data sources and processing are found in
materials and methods).
Tracking changes in the regulatory landscape
(Fig. 1, A to C, and tables S3 to S5), we docu-
mented a >10-fold increase in international (Fig.
1A) and national (Fig. 1B) management mea-
sures addressing shark fishing and finning since
2000. As of 2022, 29 countries and overseas
territories (hereafter “jurisdictions”) have declared
shark sanctuaries, no-take protected areas, or
other protective measures that prohibit shark
fishing within their national waters. Most
jurisdictions, however, have focused shark con-
servation efforts on measures to eliminate shark
,
for 95% of shark fishing mortality by number of
individuals and 71% of catch by tonnage (tables
S14 and S15). Global fishing mortality increased
from 76 million sharks in 2012 to more than
80 million in 2017, averaging 79 million from
2017 to 2019 (table S15). The number of threat-
ened shark species caught during this timeframe
fluctuated between 22 million and 28 million
each year (Fig. 2B). From 2012 to 2019, fishing
mortality increases were observed in 35% of
nonzero shark catch cells and decreases in 65%,
with an overall increase in mortality of 3.6% in
national waters and a 7.4% decrease in pelagic
fisheries managed by RFMOs (Fig. 2E). This
decrease coincides with the introduction of new
RFMO regulations prohibiting the retention of

Worm et al., Science 383, 225–230 (2024) 12 January 2024 1 of 6

RES EARCH | R E S E A R C H A R T I C L E

A Fins naturally attached B Shark fishing prohibition C Longline

Fin−to−carcass 100 Fins naturally attached 100 Purse seine
40 14 100
Species specific Fin−to−carcass Other
Other Fins artificially attached

CITES−listed pelagic species n

CITES−listed coastal species n

Finning regulation unspecified

MSC and FIP tuna fisheries (n)

Active domestic measures (n)
CITES 12
Active RFMO measures (n)

80 CITES
75
30 75
10

60
8
20 50 50
6 40

4
10 25 25
20
2

0 0 0 0 0
1992 1998 2004 2010 2016 2022 1992 1998 2004 2010 2016 2022 1992 1998 2004 2010 2016 2022
Year Year Year

p
Shark fishing
WCPFC
prohibition
Fins naturally
ICCAT attached

g
Fin−to−carcass
ratio
Mixed finning
IATTC
regulations
Finning regulation

y
IOTC unspecified
No finning
regulation
WCPFC

Fig. 1. Global regulatory landscape. (A and B) Increasing trends in number of (C) Trends in tuna fisheries involved with MSC ecocertification and related
shark fishing and finning regulations adopted (A) internationally through tuna fishery improvement projects (FIPs). (D) Spatial pattern of active shark
regional fisheries management organizations (RFMOs) and (B) nationally through regulations in 2022. WCPFC, Western & Central Pacific Fisheries Commission;

y g
domestic laws and policies. Number of CITES-listed threatened shark species IATTC, Inter-American Tropical Tuna Commission; ICCAT, International
additionally regulated through international trade restrictions are superimposed Commission for the Conservation of Atlantic Tunas; IOTC, Indian Ocean
as red lines. New listings introduced in 2022 are not yet fully implemented. Tuna Commission.

specific threatened species (Fig. 1A), particu- Bank’s Voice and Accountability index (15), regulations and fishing intensity—and negative

y
larly those listed under CITES. Indeed, avail- were associated with reduced mortality of outliers, where outcomes for sharks are worse

,
able species-specific data indicate a decrease in sharks. The main drivers predicted to increase than expected (e.g., the shark sanctuaries of
retention and an increase in observed live re- mortality were total catch (combined tonnage Sint-Maarten and the Dominican Republic, re-
lease for hammerhead, thresher, and oceanic of all species landed) and overall fishing effort spectively; Fig. 3B).
whitetip sharks across various RFMOs (fig. S9). [total kilowatt vessel hours, a measure of fish- Interviews with a globally diverse group of
In addition, although longline vessels fishing ing intensity (13)]. Established finning regu- shark science, conservation, fishery, and indus-
for tuna, billfish, and sharks have the widest lations had little effect on mortality (Fig. 3A) try professionals independently corroborate and
spatial footprint of all shark-related fishing gears and may have even increased it, possibly by contextualize the above trends (Fig. 4A and
(13), most mortality hotspots coincide with incentivizing full use of sharks and creating tables S11 to S13). Almost all interviewees per-
coastal gears such as gillnets and trawls (Fig. additional markets for shark meat and carti- ceived that shark finning had declined over
2F), both of which are known to incur sub- lage, among other products (16–18). Compar- the past two decades, whereas trends in fishing
stantial shark mortality (14). ing shark mortality standardized by fishing mortality were perceived differently among
When we analyzed the relationship between effort across all countries reveals the complex- regions (Fig. 4B), and 45% noted a concurrent
shark fishing mortality rates and prevalent ity of assessing regulatory effectiveness, with increase in the demand for shark meat. One
regulations by country (Fig. 3A), we found that the same regulation in some cases resulting in NGO representative suggested, “Shark finning
only shark fishing prohibitions and accounta- both positive outliers—that is, locations with legislation particularly didn’t have an impact
ble governance, as measured by the World lower-than-expected shark mortality given their on reducing overall shark mortality [because]

Worm et al., Science 383, 225–230 (2024) 12 January 2024 2 of 6

RES EARCH | R E S E A R C H A R T I C L E

A B

Annual global Annual global mortality

mortality (n) 7 403 22026 threatened species (n) 7 403 22026

C D

p
Annual global mortality Annual global mortality
silky shark (n) 2 54 1096 hammerhead sharks (n) 2 54 1096

g
E F

y
Longline Purse seine
Relative change −100−0 No change Dominant
Trawlers Small−scale gears
in annual mortality 0−100 100−500 500+ fishing gear

y g
Other coastal gears

Fig. 2. Global patterns of shark fishing mortality. Shown are average (Sphyrna spp.). (E) Relative trend in annual mortality (percent increase or
annual mortalities per 1° by 1° grid cell from 2016 to 2018 for (A) all sharks, decrease) for all sharks from 2012 to 2019. (F) Fishing gear type that dominates
(B) threatened species (IUCN critically endangered, endangered, or vulnerable), shark mortality for each cell (further details about fishing gear categories are

y
(C) silky sharks (Carcharhinus falciformis), and (D) hammerhead sharks shown in table S19).

after the prohibitions [countries that were
finning sharks] just landed the sharks whole,
[which] resulted in new markets for shark meat,
oil, and other products in countries that didn’t
consume shark meat previously” (Interviewee
NGO-3; table S11). Many interviewees further
perceived that fisheries are now catching smaller
sharks, including juveniles (19), because of de-
clines in the fin trade, regional declines in the
abundance of large sharks, and increasing de-
mand for shark meat (table S11). Almost two-
thirds of experts (64%) highlighted regions of
the Indian Ocean or Indo-Pacific as primary
areas of shark bycatch concern, with West
Africa identified as an additional hotspot by
23% of experts. In terms of fishing gears, 91%
highlighted gillnets as key contributors of shark
mortality because of their unselective nature and
their unregulated use in many coastal fisheries.
Regarding the reduction of shark finning,
domestic and RFMO finning regulations were
perceived as being most effective (54 and 45%,
respectively), followed by public awareness
campaigns and pressure from large seafood
retailers for sustainable seafood products (Fig.
4C and tables S12 and S13). One industry repre-
sentative noted, “We are seeing a big push from
[the] market side, which is having a bigger
impact at this stage relative to the regulatory
side [because] if you lose your [eco-]certifica-
tion, it would be diabolical for business and a ,
huge company risk” (Interviewee IND-2; table
S11). All three industry representatives and 18%
of all interviewees perceived that RFMO reten-
tion prohibitions had effectively reduced the
catch of CITES-listed species (e.g., silky, hammer-
head, oceanic whitetip; fig. S9). Conversely, do-
mestic fisheries management, catch monitoring,
and enforcement were deemed to be most in
need of improvement (Fig. 4C and tables S12
and S13). Multiple interviewees also suggested
that improved engagement by government agen-
cies with coastal fishers was essential, espe-
cially where shark meat is contributing to local
food security.

Worm et al., Science 383, 225–230 (2024) 12 January 2024 3 of 6

RES EARCH | R E S E A R C H A R T I C L E
A B
Shark fishing
prohibition (n = 19)
Fins naturally
attached (n = 79)
Fin−to−carcass
ratio (n = 16)
Unspecified finning
regulations (n = 15)
Total effort
(kwh)
Total catch
(tonnes)
World Bank
Index
2 4 6 8
Coefficient estimate
Our analyses represent a first global synthe-
sis of spatial and temporal trends in shark
fishing mortality in the context of widespread
regulatory change. Although finning regula-
tions have been successful in reducing waste
and animal cruelty, there is little evidence that
they have reduced shark mortality overall (Figs.
2E, 3A, and 4B). Indeed, domestic measures
adopted to eliminate the fin trade were insuffi-
cient to halt overexploitation, and interviewees
suggested that they even contribute to incentiv-
izing retention of whole sharks and, by exten-
sion, markets for their meat. This is consistent
with results from our regulatory analyses (Fig. 3)
and mapping exercise (Fig. 2 and tables S14 and
S15), which suggest that shark mortality is
increasingly concentrated in coastal hotspots.
We found that territorial waters of just six
coastal nations incurred 50% of global shark
mortality from 2017 to 2019 (table S15), four
of which (Indonesia, Brazil, Mauritania, and
Mexico) were also highlighted by interviewees
as places where high shark fishing mortality
coincides with insufficient regulatory capacity.
It is noteworthy that these countries are also
major international suppliers or domestic con-
sumers of shark meat, reflecting growing mar-
kets for nonfin shark products (17).
These findings suggest a shifting global land-
scape of shark fishing mortality that is moving
away from finning of larger pelagic species (4)
toward full use of smaller coastal species, pre-
senting new regulatory and conservation chal-
lenges. On the positive side, shark mortality
under the oversight of the tuna RFMOs appears
in decline overall, most notably in the Atlantic
and Western Pacific, where species-specific
retention bans and comprehensive observer
coverage on purse seine fishing vessels combine
Worm et al., Science 383, 225–230 (2024)
Dominican Republic
Shark fishing Sint−Maarten
prohibition
Fins naturally
attached
Regulation category
Fin−to−
carcass ratio
Unspecified finning
regulations
−15 −10
ln (Shark mortality per unit effort)
with strong incentives to provide ecocertified
tuna to global markets. It is yet unclear whether
this is enough to reduce pelagic shark threat
status similarly to that of tuna and billfish, the
traditional target of improved management
(20). In this regard, it is notable that only since
2019—50 years after its establishment—have
member states of the International Commis-
sion for the Conservation of Atlantic Tunas
(ICCAT) been legally mandated to manage
sharks in the Atlantic in the same way that they
manage target tuna species. Across national
waters, some small island nations lead the
charge with respect to reducing shark fishing
mortality. Shark sanctuaries in the Bahamas
and Maldives, for example, have managed to
maintain relatively healthy shark populations
(21), which fuel successful dive tourism indus-
tries (22, 23). Other measures, such as large no-
take protected areas in the Pacific Remote Islands
Marine National Monument south of Hawaii,
also appear successful in maintaining relatively
low shark fishing mortality, although outcomes
are variable (Fig. 3A). Nations that are more
democratic are also consistently associated
with better outcomes for sharks (Fig. 3A),
echoing similar results from a global analysis
of reef shark abundance (21). Yet, our analysis
suggests that when viewed through a global
lens, current risks for coastal sharks still appear
to be escalating on average, a conclusion that
is supported by recent IUCN assessments (1, 24).
We caution that data quality and transparency
were key limitations for our analyses. Catch data
are rarely reported with uncertainty estimates,
although there are many known uncertainties
and data gaps (25). The detailed catch recon-
structions used here (11) acknowledge this un-
certainty by cross-referencing multiple data
12 January 2024
Fig. 3. Linking shark fishing mortality to reg-
ulation. Shown are (A) slope coefficients [± 95%
confidence interval (CI)] from a generalized
linear model predicting shark fishing mortality
for all countries’ national waters as a function of
active shark conservation regulations. The verti-
cal hatched line denotes a neutral effect, with
points (unfilled circles) to the right signifying
increased and points to the left decreased shark
fishing mortality given the regulation; points are
the exponentiated coefficient estimates, and
horizontal lines are the 95% CIs. Total catch refers
Indonesia
to the combined tonnage of all species landed and
total effort to the number of kilowatt vessel hours
observed fishing. World Bank Index refers to the
Voice and Accountability composite index, with higher
scores indicating improved democratic governance.
(B) The distribution of (natural log-transformed)
shark fishing mortality per kilowatt fishing hour and
regulatory regime for each country. Outliers of lower-
or-higher-than-expected mortality were 2 SDs below
p
−5 0
or above the mean and are labeled (black circles).
inputs and assigning a data-source specific,
qualitative uncertainty score to all catch esti-
g
mates (tables S14 and S15). Mapping the catch-
weighted uncertainty across coastal and high-seas
areas highlights hotspots of high shark catch
and high uncertainty where more comprehen-
y
sive and reliable data are critical for improving
global estimates of shark mortality (figs. S10 to
S12). Likewise, although our RFMO catch models
are specifically designed to overcome spatio-
temporal observation biases arising from in-
complete reporting, catches are still largely
self-reported by fishing countries and almost
certainly underestimate total fishing mortality,
especially for rare and endangered species (26).
We provide upper and lower 95% confidence
y g
limits around RFMO shark catch estimates to
explicitly account for this uncertainty and to
highlight areas with particularly poor data
reporting (tables S16 to S19 and figs. S11 and
S12). We further present expert interviews as an
independent source of data to further mitigate
y
these uncertainties (Fig. 4 and tables S11 to S13).
,
We offer a detailed discussion of data uncer-
tainty in the materials and methods and make
all data and models publicly available for fu-
ture work that could make use of new and im-
proved reporting (data S1 to S5).
We further caution that country-level shark
catch is also likely underestimated because
24% of the annual catch from national waters
is being reported at the subclass (Elasmobranchii)
level (fig. S3), preventing meaningful analysis of
these data at the species level. Including these
taxonomically unresolved elasmobranch catches
(which include rays and skates in addition to
sharks) and assuming a similar proportion of
true sharks as in the taxonomically resolved data
(71%) would increase our global shark mortality
4 of 6
RES EARCH | R E S E A R C H A R T I C L E

p
g
y
Fig. 4. Expert perceptions on shark fishing and regulation. (A) Geographical location of 22 regional experts interviewed for this study. (B) Interviewee perceptions
on current trends in shark finning (left) and shark mortality (right). (C) Interviewee perceptions on effective (in boldface) and insufficient regulatory and market
measures affecting trends in shark finning and mortality (see tables S11 to S13 for detailed interview data summaries).

y g
estimate to 101 million sharks (in 2019). This estimate of total shark fishing mortality is given the recent expansion of CITES listings to
figure represents a 4% increase over previous conservative. include 54 species of requiem and hammerhead
global fishing mortality estimates for the year Our analysis shows that shark fishing con- sharks (Fig. 1, A and B, and table S1). Our find-

y
2010 (5). tinues to present a substantial threat to shark ings signal that such measures can be effective

These results highlight the importance of
improved data reporting. Market and regulatory
pressures to enhance reporting requirements
and data dissemination are expected over time,
and trends of decreasing shark fishing mortal-
ity in industrial tuna fisheries should be scruti-
nized further when more data are available.
Additional datasets that more comprehensively
document discarding practices are also required
to improve mortality estimates, especially for
poorly observed longline fisheries (fig. S9). Last-
ly, illegal, unreported, and unregulated (IUU)
shark fishing is only in part captured by our
analyses but is very evident both from expert
interviews (table S11) and from the literature
(19, 27). Combined, these data indicate that our
populations over much of the world, despite the
widespread adoption of antifinning legislation
and related measures. This regulatory shortfall
needs to be addressed through a combination
of area-based conservation (28) and improved
shark-specific fisheries management measures
(29) that address overcapacity and disincentivize
retention of overfished and threatened species.
Science-based harvest control rules have been
recently adopted for most tuna stocks under
RFMO oversight (7, 30) and should be used to
address overexploitation of pelagic sharks, most
of which lack catch limits. Effective bycatch
mitigation is a pressing issue in this regard,
both for management bodies and for fishing
companies seeking ecocertification, especially
,
when paired with other regulations in interna-
tional fisheries (fig. S9) and also highlight the
importance of similarly expanding improved
regulation and oversight in national fisheries.
Increased transparency and accountability of
fishing companies, fleets, and management
bodies are needed to support successful imple-
mentation of these measures. We observe that
some of the most effective regional solutions have
been spearheaded by low-income countries with
a high dependence on a healthy marine environ-
ment for food and livelihood security. Such
localized efforts are part of an emerging trend
in ocean sustainability (31) and demonstrate
that positive change can be achieved where
the long-term needs of both nature and people

Worm et al., Science 383, 225–230 (2024) 12 January 2024 5 of 6

RES EARCH | R E S E A R C H A R T I C L E
are adequately valued. As these solutions are
more widely adopted, our analyses can pro-
vide a spatially explicit baseline against which
future progress in recovering threatened shark
populations can be assessed, supporting time-
ly efforts to rebuild resilient ocean ecosystems
and sustainable fisheries.
RE FE RENCES AND N OT ES
1. N. K. Dulvy et al., Curr. Biol. 31, 4773–4787.e8 (2021).
2. F. Ferretti, B. Worm, G. L. Britten, M. R. Heithaus, H. K. Lotze,
Ecol. Lett. 13, 1055–1071 (2010).
3. J. A. Estes, M. Heithaus, D. J. McCauley, D. B. Rasher, B. Worm,
Annu. Rev. Environ. Resour. 41, 83–116 (2016).
4. S. C. Clarke et al., Ecol. Lett. 9, 1115–1126 (2006).
5. B. Worm et al., Mar. Policy 40, 194–204 (2013).
6. Convention on Migratory Species, Memorandum of
Understanding on the Conservation of Migratory Sharks,
3rd Meeting on International Cooperation on Migratory Sharks
under the Convention on Migratory Species (CMS, 2010);
https://www.cms.int/sharks/en/meeting/3rd-meeting-
international-cooperation-migratory-sharks-under-convention-
migratory-species.
7. L. Schiller, M. Bailey, Fish Fish. 22, 592–604 (2021).
8. A. C. Vincent, Y. J. Sadovy de Mitcheson, S. L. Fowler,
S. Lieberman, Fish Fish. 15, 563–592 (2014).
9. L. Schiller, G. Auld, H. Sinan, M. Bailey, Conserv. Lett. 14,
e12827 (2021).
10. M. Fabinyi, N. Liu, Conserv. Soc. 12, 218–228 (2014).
11. D. Pauly, D. Zeller, Nat. Commun. 7, 10244 (2016).
12. E. S. Burns, D. Bradley, L. R. Thomas, Front. Mar. Sci. 9, 1062447
(2023).
13. D. A. Kroodsma et al., Science 359, 904–908 (2018).
14. S. Oliver, M. Braccini, S. J. Newman, E. S. Harvey, Mar. Policy
54, 86–97 (2015).
Worm et al., Science 383, 225–230 (2024) 12 January 2024
15. D. Kaufmann, A. Kraay, M. Mastruzzi, “The worldwide governance
indicators: Methodology and analytical issues” (World Bank
Policy Research Working Paper No. 5430, SSRN, 2010); https://
papers.ssrn.com/sol3/papers.cfm?abstract_id=1682130.
16. L. N. Davidson, M. A. Krawchuk, N. K. Dulvy, Fish Fish. 17,
438–458 (2016).
17. F. Dent, S. Clarke, “State of the global market for shark products”
(FAO Fisheries and Aquaculture Technical Paper 590, 2015), p. 187.
18. World Wildlife Fund, “The shark and ray meat network: A deep
dive into a global affair” (WWF, 2021); https://sharks.panda.
org/images/downloads/392/WWF_MMI_Global_shark__ray_
meat_trade_report_2021_lowres.pdf.
19. E. Bonaccorso et al., Sci. Rep. 11, 14959 (2021).
20. M. J. Juan-Jordá et al., Science 378, eabj0211 (2022).
21. M. A. MacNeil et al., Nature 583, 801–806 (2020).
22. A. J. Gallagher, N. Hammerschlag, Curr. Issues Tour. 14, 797–812
(2011).
23. J. S. Zimmerhackel, M. E. Kragt, A. A. Rogers, K. Ali,
M. G. Meekan, Mar. Policy 100, 21–26 (2019).
24. C. S. Sherman et al., Nat. Commun. 14, 15 (2023).
25. G. D. Blasco, D. M. Ferraro, R. S. Cottrell, B. S. Halpern,
H. E. Froehlich, Front. Mar. Sci. 7, 612831 (2020).
26. G. Mucientes, M. Vedor, D. W. Sims, N. Queiroz, Biol. Conserv.
269, 109534 (2022).
27. D. J. Madigan, N. S. Arnoldi, N. E. Hussey, A. B. Carlisle,
Conserv. Lett. 14, e12796 (2021).
28. C. A. Hyde et al., Front. Mar. Sci. 9, 968853 (2022).
29. N. K. Dulvy et al., Curr. Biol. 27, R565–R572 (2017).
30. R. M. Hillary et al., Fish Fish. 17, 469–482 (2016).
31. N. J. Bennett, P. Le Billon, D. Belhabib, P. Satizábal, npj Ocean
Sust. 1, 3 (2022).
32. B. Worm et al., Global shark fishing mortality still rising despite
widespread regulatory change, Dryad (2023); https://doi.org/10.
25349/D9JK6N.
33. B. Worm et al., Global shark fishing mortality still rising despite
widespread regulatory change, version 4.0.0, Zenodo (2023);
https://doi.org/10.5281/zenodo.8422551.
AC KNOWLED GME NTS
The authors express deep gratitude to all data providers and
interviewees sharing their expertise and insight. Special thanks
to C. Field and E. Lawler for statistical advice and support.
Funding: We acknowledge funding from the Natural Sciences
and Engineering Council of Canada, the Sustainable Fisheries
and Communities Trust, and the Blue Marine Foundation. E.S.B.
acknowledges funding from the Waitt Foundation; L.M.F. received
funding from CAPES and Fulbright Brazil; and L.S. is grateful
for support from the Liber Ero Fellowship Program. Author
contributions: This study was originally conceptualized by
B.W., L.S., and N.G.D. and supervised by B.W. and D.B. Fisheries
data were sourced and curated by M.L.D.P., D.B., S.O., E.S.B.,
L.S. and L.M.F. Interview data were collected and analyzed
by N.G.D., L.S., and B.W. Model development was led by D.B.
and E.S.B. Further data analyses, interpretation, and writing were
shared among all authors. Competing interests: The authors
declare that they have no competing interests. Data and materials
availability: All data are available in the supplementary materials
(tables S1 to S19 and data S1 to S5) or available for download
on Dryad (32). The complete code repository is available at Zenodo
(33). License information: Copyright © 2024 the authors, some
rights reserved; exclusive licensee American Association for the
Advancement of Science. No claim to original US government works.
https://www.science.org/about/science-licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
p
science.org/doi/10.1126/science.adf8984
Materials and Methods
Figs. S1 to S12
Tables S1 to S19
References (34–78)
MDAR Reproducibility Checklist
Data S1 to S5
g
Submitted 18 November 2022; resubmitted 13 May 2023
Accepted 2 November 2023
10.1126/science.adf8984
y
y g
y
,
6 of 6
 WORKING LIFE
By Joe Higham

Missing the signs

“I
’m worried about you,” a friend said. During the 2 weeks prior, my mind had been racing with more
ideas than I could keep up with and I had rarely slept more than 4 hours per night. But I wasn’t tired;
I felt incredible, invincible. It wasn’t the first time that this friend—another researcher I worked with
during my postdoc—had voiced their concern. But this time, the words resounded in my mind. As we
continued to talk, they began recounting my other unusual behavior, and I began to notice just how
quickly, loudly, and forcefully I had been talking. That’s when the reality began to sink in: I was man-
ic. It was a realization that not only changed my approach to mental health, but also my career path.

p
It still came as a shock when I was with chasing productivity regard-
later diagnosed with bipolar disor- less of the cost to my well-being
der, a condition that leads to extreme returned, triggering bouts of ma-
mood swings—from manic highs to nia and depression, including one
depressive lows. As I learned more of the worst depressive episodes of
about the condition, though, and my life.

g
looked back on my past experiences, My family, seeing that I could
it all made sense. I had been fighting barely get out a few words on the
the symptoms for most of my adult phone before bursting into tears,
life. But within the pressure cooker convinced me to speak with a thera-

y
that is academia, I had ignored them. pist. After a few weeks of sessions,
My first experience with mania things began to improve. But a few
was as an undergraduate student, in months later, triggered by the anxi-
the run-up to exams. The pressure ety that had gripped me from day
and stress—which I now understand one of my postdoc, I plunged into
are triggers for bipolar episodes— the 2-week manic episode that had
would sometimes give way to a wave
of boundless energy. That energy
“I know I’m not alone in alarmed my friend.
I had already realized my post-
was useful for powering through
my studying. But once exam season
my struggle with doc wasn’t a healthy environment
for me and had decided to leave.
mental health in academia.”

ended, I crashed and could barely
function beyond eating and sleeping.
My symptoms worsened after I started my Ph.D., when
the stress wasn’t restricted to certain times of year but was
nearly constant. I put a tremendous amount of pressure on
y g
My later diagnosis confirmed I
had made a good decision. I would
have loved to make a career in research. But I’ve learned
that for me, it’s risky to take on the stress and unending
sacrifice such a life demands. Since then, I’ve begun to

myself to excel, equating my professional success with my
personal self-worth. Each time an experiment failed, the
pressure intensified.
I mostly brushed off my worsening anxiety and depres-
sion as a normal part of the graduate school experience, as
everyone around me seemed to be stressed and anxious,
too. I didn’t seek professional help. Instead, I employed
stress management techniques such as meditation, mind-
fulness, and exercise. Those helped, but they didn’t solve
the problem entirely.
Then, soon after starting my first postdoctoral position,
I became unable to cope. I was working in an unfamiliar
field and was struggling to adapt to the dynamic of my new
research group. My self-confidence dwindled and my stress
management strategies no longer worked. My obsession
y
explore my passions beyond chemistry and work as a free-
lance writer, a career path that I find far less stressful and
,
that fulfills my passion for writing, which I discovered
while writing my Ph.D. thesis.
I know I’m not alone in my struggle with mental health
in academia. I might have been able to cope more effectively
had I sought out professional help sooner. But better men-
tal health support, a normalization of work-life balance,
and improved mentorship to help deal with the stressful
realities of research would also have helped me persevere.
Because if the culture around you sees stress and anxiety
as the norm, it’s hard to see your own mental struggles as a
cause for concern. j
Joe Higham is a freelance writer working primarily in nonfiction.

234 12 JANUARY 2024 • VOL 383 ISSUE 6679 science.org SCIENCE

Produced by the Science/AAAS Custom Publishing Office LIFE SCIENCE TECHNOLOGIES
new p rodu cts: ge neral lab equipment
        ( 
CHO Cell Growth Media Platform   "  % &  '!")  
   / !#$#% '  % %   ' #' *'  '
!#$( $)#+#"", '#    $ ") ' #' #+
&+# (- (+0 ! !#'' ( 00     )   '
!- #++1'/'*0 !-'# )#+#"",  %     $ ' 
&(%$#')#+#"",    ! '%%   * % &)  #   %
)#+#"",  !0#'' $#'$*'  #+ %)) ,     '
-#-*0 !#+ 1#''%  1(      ")     '% 
"!$,#-$  #1.",  +++$# #++$ #+$($# !#' ) '   * ")  %' #+
-#%(# %*(-$-#++$!##$$ +#*# $!$    #

%') %%)
0#'' $ ' !+# (-0 !$ / -'#'$*'  ")*#+)%%)
$%   +++$#'(%$ %+ %+# (-$( !    &      
-- #+# (- +#+# '$ -#  $     %     *+ 
$-+#$#' !-+ %+# !#%  $$*'      &      
'+!*!  $%'!*! ++'$ 1 ' $)#+#"",   &')  *%%)  %#+ 
 -'%-$$ / #++1'+'( -#+*0 !#'    )   *
'%  1(",   +++$#' !-'%-$% $ ! TESTA Analytical Solutions
# !#'('
# ! %+ %$ $ -# +0 !#0'#*(    
%+ %+# (-$#' #$( #* $!-'%-$#-#+ !##$%&&(#*%#+++,'#-+,(-.
- (#'- $'%$ 1$ #'#'$(+%-  

p
'%  1 ++
$ / (%$# $#'$$++%-$ Compact Solvent Recycler
'#1  -$ -+ #'#'1 %$(## 1((%$ /#*-!,0-%!+%+-*0#!**,*+%*#%*).+#)+1*
%+ %+# (-$ +%'*(%$-- #'(%$
##+ %,1*#*-'-, # 0*% *0#*-1*$# "#!*
# $ %,1*#%*0#!*. ,*$!+%* '*0*-# #!*$*%,1*#
   
 
##!*%,1*#*%*10 %-+#-(!*).+#)+1*
     .$1*%,+ %%#++ ,#'+0%0*% *0#-,'$+'#%*,

g
  
 '*+ , *%*+-!*%#*-'-,*+,.%#+,,#!*%,1*#%*0
%-+#-%'%#*.%().+#)+1*%$**00*-#,'.'
High-Performance Photoreactor -!.+# +$!'0+#+%'%#*. +0$*+0+$#*%**0(
"!"+!"! ! !"$!%"&'"(' "+!"
#-.*%#!%#+*#- *#!*$++.*#*%+0$*.

! ! % " '" "'' " "!$"%! "'"(! !)"
"!""- '-"$!"$" -"! ! % "+!"
' "'-".-" ""$ "' ! /)"!"0' " -' "
,!." !" ' ' "  1" "#"%"0("'""' "
!$",-"0'  1"1"1"1"'-""'%*""!-"
(""! %("!"."  ". !%  "-  " "
0' "!$" "!"#"%!-."'-"-. " "!. . "
   " !-' )"
" % ."'!"'-"$ " . !. "
" -" !"!  " "#"$!%"!0 ' )"
'"' " ". "
'. ""!0--"$!"! % .".)"!" "." "!"
+!"! ! !1"" !" !""' -"'-"" !-" !"
y
++0##+,#!*%,1*#*-1**0(.$+# ,*#!+,.%#
+' ).+#)+1*-#%,'.#%#!*#$#% +,
#!*-!.+# +$!-%#.*#0*#*-#( ++, #0 #+,
-1*#*$0-*%0 #+,0+#+*1+,+# '+#* +,%.+#
$-*%%( #!**+'1+,1*#!*).+#)+1*%%#-!*0#+%#*
*-'-,*$%#+--0 #$#% +,,*1*,(!*%#-! 
1+,1*%-*-#*0##!*0*#*-#. ,*$!+%*#$#(
Biotech Fluidics
%#&"'%&&$!$($)&
*++,-../0"%+12*4"6"2-/2%7

y g
 !"%!-."'-" %-"'" ".' " ", '"-. )" Colored 96-Well Skirted PCR Plates
"!"#"%".' "" '"' -"' !" " !" !"'-" 1'+8&5"1&2"1'21-&*8-&#118-16&"+-& #871+8#&!1&-"#+16& &
!'' -" !" "! %%("!".)"" !" !" ,8+1&"'&8&"61&#8' 1&%&2%%#- &61-" '16&+%&18-1&+*1&%# 8'"8+"%'&
% ."" !' !-"(" !'' ' " " !-" !" !" "" !- 8'6&"-8&"61'+"28+"%'&%&6"1#1'+&-1+-&%& &,8+1-&"'&7+"-1#&
 "."$!" !'" !" !"!!%" % .*1"!"$(" 80-/&*1&'1&,8+1&2%%#- &*"2*&"'261&,#,1 &01 &218# & #11' &
!""  !'" %! . !'" %) #16 &082 &8'6&*"+1 &8%&80-&+%&"7,#%1&%#4%&12"1'2%#&

y
Uniqsis Ltd. 1387,1&0&8--" '"' &182*&-1#&+*1"#&%'&2%%# &%#&-"' &8&,8#+"28#&

,
    2%%#&%#&182*&6"1#1'+&8--8&+,1&%#&-%#21/& #871+8#&,8+1-&28'&
 !!$%  $'(! 8-%&*1,& &-87,1&-2*16"' &0&#1412+"' &8'6&28#""' &+*1&
6""-"%'&%&80%#&"+*&2%%#16&,8+1- &"+&"-&18-&+%&-11&*%-1&-87,1-&
HPLC Flow Rate Monitor 8#1&#''"' &8'6&*%&-+"&'116-&+"71&%'&+*1&+*1#7%221#/&*1&
")&%*(+,+#-())& (""+-#'+-  #871+8#&,8+1-&#" "6&,%28#0%'8+1&#871&"7,#%1-&712*8'"28&
+)+- ()(,*)')+()+ ) &+*& # )( -+80""+&0&1"7"'8+"' &8#,"' &8'6&6"-+%#+"%'&6#"' & /& #11&
()(+')$"+*+--#)+-+,)%+-#*++ -87,1-&8#1&88"801&,%'&#1
1-+&#%7&1'+8/&&
(& # )( -) )*'+--#')'
,+ )()+ )* Azenta Life Sciences
++())  +-#*) + + ) # )(  %#&"'%& &$!$)&&
)(,++()+* +# --)+ )+--%+-#$ *++,../81'+8/2%7&&&

Electronically submit your new product description or product literature information! Go to www.science.org/about/new-products-section for more information.

Newly offered instrumentation, apparatus, and laboratory materials of interest to researchers in all disciplines in academic, industrial, and governmental organizations are featured in this
space. Emphasis is given to purpose, chief characteristics, and availability of products and materials. Endorsement by Science or AAAS of any products or materials mentioned is not
implied. Additional information may be obtained from the manufacturer or supplier.

SCIENCE science.org/custom-publishing 12 JANUARY 2024 • VOL 383 ISSUE 6679 231