EDI TO R I A L
The future of ocean health
H
uman and environmental health are inextricably
linked. Yet ocean ecosystem health is declining
because of anthropogenic pollution, overexploi-
tation, and the effects of global climate change.
These problems affect billions of people depen-
dent on oceans for their lives, livelihoods, and cul-
tural practices. The importance of ocean health is
recognized by scientists, managers, policy-makers, non-
governmental organizations, and stakeholders including
fishers, recreationalists, and cultural practitioners. So
why are the oceans still degrading? Sustainable care of
this vast resource needs a new approach if future genera-
tions are to inherit a legacy of vital marine ecosystems.
Although countries pledge to protect ocean health
through international agreements, including the United
Nations Convention on the Law of the Sea and the Lon-
don Convention and Protocol,
compliance and enforcement are
inadequate, as nations and indus-
tries find loopholes to circumvent
the spirit of essential guiding
“The use of oceans
as a dumping ground
principles. The use of oceans as
a dumping ground for countless
wastes continues as a result. This
is inexcusable. Governance re-
for countless
quires strategies that are effective,
economically feasible, and cultur-
wastes continues…”
ally acceptable for diverse users.
Many ocean pollutants come
from incidental land runoff, unauthorized discharges, and
maritime accidents that release pesticides, heavy metals,
plastics, chemicals, and oil. Others arise from planned
discharges of treated sewage and industrial wastes, such
as last week’s controversial release of radionuclide-con-
taining water from the Fukushima Daiichi nuclear power
plant. The belief that the oceans are limitless in their
capacity to absorb pollutants, while still able to produce
abundant food and support complex ecosystems, is false.
Although pollutant concentrations within “accepted
standards” may seem small when considering the vast
volume of the oceans, many toxicants, even at low con-
centrations, can cause sublethal responses at the cellular,
organismal, population, and ecosystem levels that affect
ecosystem structure, function, and services. Also, con-
taminants act differently in combination, with cumula-
tive effects on ocean life, often making the “soup” more
important than the individual pollutants. Moreover, not
all pollutants share the same fate, as some move with cur-
rents, others bind to seafloor sediments, some adhere to
organic particles, and others are transported and bioaccu-
mulated by marine organisms. Even in a single organism,
SCIENCE science.org
pollutants are bioconcentrated in different organs, bones, Robert Richmond
and tissues, creating a range of potential toxic effects. is a professor and
Effective policies and actions to protect the ocean from director at Kewalo
pollution require sound science. Indeed, the United Na- Marine Laboratory,
tions declared the years 2021 to 2030 as the Decade of University of Hawai‘i
Ocean Science for Sustainable Development, with the at Mānoa, Honolulu,
goal of reversing declines in ocean health. The proclama- HI, USA. richmond@
tion points to pursuing “the science we need for the ocean hawaii.edu
we want.” This is valid, but science alone is not enough.
In addition to closing environmental policy, regula-
Ken Buesseler
tory, and enforcement loopholes, political will is essen-
p
is a senior scientist
tial to bring about healthier oceans. This could be better
at the Woods Hole
achieved by integrating contemporary science with Indig-
enous knowledge about ocean stewardship. This knowl- Oceanographic
edge integration has been leveraged to address issues of Institution, Woods
carrying capacity (the maximum amount of life that an Hole, MA, USA.
ocean ecosystem can sustainably kbuesseler@whoi.edu
g
support), but has not evolved to
tackle mounting, large-scale prob-
lems of transboundary pollution
and global climate change. For
y
example, the 18 sovereign nations
of the Pacific Islands Forum have
demonstrated forward-thinking
cooperative leadership by develop-
ing a “2050 Strategy for the Blue
Pacific Continent” to protect the
marine environment using inte-
grated knowledge. This regional
shared vision provides a potential blueprint for actions
that other nations, in partnership, could pursue.
y g
The Small Island Developing States (also known as
“Large Ocean States”) are part of a group of 57 nations
and political entities that collectively have jurisdiction
over 140 million square kilometers of ocean, or one-quar-
ter of the world’s exclusive economic zones. This gives
p
them control over natural resources in these waters and
as such, they wield a large impact on, and an outsized
stake in, protecting marine ecosystems. Addressing ocean
,
health over multigenerational timescales is an important
characteristic of Indigenous cultures. Native communities
have a strong connection to place, and hence ownership
of problems and solutions. This ethic supported passage
of the UN “high seas” treaty, which requires assessing the
environmental impacts of activities in areas beyond na-
tional jurisdictions, with an emphasis on cleaner oceans.
Integrating contemporary science and traditional
knowledge could strengthen treaties and agreements and
forge new shared visions among nations. It’s not too late
to set the world on a better path to restore ocean health.
–Robert Richmond and Ken Buesseler
10.1126/science.adk5309
1 SEP TEMBER 2023 • VOL 381 ISSUE 6661 927
 1,870,000
NEWS Estimated excess deaths from all causes among those 30 years and
older in China when COVID-19 surged in the 2 months after the end of its
“zero COVID” policy in December 2022. (JAMA Network Open)

Masks, lockdowns vindicated

IN BRIEF | Wearing masks, social
P U B L I C H E A LT H
distancing, and other measures taken
Edited by Jeffrey Brainard
during the COVID-19 pandemic were
effective at slowing viral spread, accord-
ing to a comprehensive report released
last week by the Royal Society. It
describes hundreds of observational stud-
ies and a limited number of randomized
trials of six nonpharmaceutical interven-

p
tions, including contact tracing, indoor
ventilation, travel restrictions, and public
health messaging, used around the world.
All were found to help prevent transmis-
sion, with social distancing, including
lockdowns, coming out on top. The

g
usefulness of travel restrictions varied,
with quarantining policies implemented
early in the pandemic scoring higher
India’s lunar rover, Pragyan, shown descending to the surface, is named for “wisdom” in Sanskrit. than pretravel screening of passengers for

y
high temperatures or other symptoms.
SPACE SCIENCE The report recommends that to prepare
for future pandemics, researchers should
Indian Moon lander begins search for ice collect data about other effects of such
interventions, including impacts on

I
ndia last week put a spacecraft on the Moon, making it the fourth people’s income and mental health.
nation to perform the feat, after the United States, the Soviet
Union, and China. The $75 million, robotic Chandrayaan-3 mis- EPA restricts common pesticide
sion is the first to explore a region near the lunar south pole, E N V I R O N M E N T | The U.S. Environmental
which is thought to hold frozen water that could be used to sup-

y g
Protection Agency (EPA) last week put new
port human explorers. The area also contains large impact craters, limits on malathion, an insecticide com-
monly used to control mosquitoes, aphids,
which could hold clues to the history of the Solar System. Hours after
and other pests, after a federal evaluation
landing, the craft released a solar-powered, suitcase-size rover that determined it could harm 78 endangered

will study soil composition for at least 2 weeks. The mission is a vin-
dication for the Indian Space Research Organisation, which failed to
complete its previous Moon-landing attempt in 2019.
UNC materials scientist slain
WO R K P L AC E | A University of North
Carolina (UNC) at Chapel Hill graduate
student is accused of shooting dead his aca-
demic adviser, applied physicist Zijie Yan, in
a campus research building this week. The
student, Tailei Qi, was apprehended off cam-
pus 95 minutes after the killing. He has been
charged with first-degree murder. The shoot-
ing prompted a 3-hour lockdown on the
flagship campus. The Caudill Labs building,
the site of the shooting, houses chemistry
laboratories and some in applied physical
sciences, where Yan worked. Chinese-born,
he was a graduate of Huazhong University
of Science and Technology. He then came
to the U.S. and earned a Ph.D. in materi-
als engineering at Rensselaer Polytechnic
Institute. He had been at UNC since
2019. Yan and his accused killer, who
complained online about his working life,
co-authored a paper as recently as July.
Qi joined the Yan lab in January 2022 after
earning degrees from Wuhan University and
Louisiana State University.
p
terrestrial species. Manufacturers must
revise product labels to include new
instructions, such as reducing spraying
,
near critical habitat of protected species
such as the sandhill crane and Alabama
cavefish. The Center for Biological
Diversity, which sued EPA more than
a decade ago to change the labeling,
complained about loopholes. Further
restrictions could come by the end of the
year, after the National Marine Fisheries
Service completes an additional evaluation
of malathion’s risk to endangered species
living in coastal waters.
Thefts hit British Museum
| The disappearance of
A R C H A E O L O GY
about 1500 artifacts from the British

928 1 SEP TEMBER 2023 • VOL 381 ISSUE 6661 science.org SCIENCE
 BIODIVERSITY

Gardeners urged to plant endangered species

G
ardens and apartment balconies could make a significant species they identified are currently available to buy from nurseries
contribution to saving endangered plants, according to a study and seed companies. If more gardeners opted to plant those species,

of such species in Germany. Scientists examined the plant spe-
cies that are considered vulnerable, threatened, endangered, or
extinct in the wild and found that nearly half could be grown in
gardens or public green spaces or on green roofs or balconies. They
also found that two-thirds of the 988 native, at-risk, garden-friendly
p
the overall “threat status” for all plants in Germany—the ratio of the
number of at-risk species to all species assessed for risk—could fall
by 25%, the study’s authors report this week in Scientific Reports.
They also found that, on average, the at-risk species tolerate drought
better and need less fertilizer than conventional garden plants.

g
Clammy campion (Viscaria vulgaris Bernh., rosy flowers) and Nottingham catchfly (Silene nutans L., white) are among plant species at risk in parts of Germany.

Museum prompted its director to resign
last week and many to question how a
prestigious institution could allow so
many of its treasures to be lost or stolen.
Among the missing items are jewelry,
gems, and glass dated to the 15th century
B.C.E., The Guardian reported. An internal
investigation by the museum had accused
one of its longtime curators of improp-
erly categorizing antiquities and then
the missing artifacts as police investigate
their whereabouts.
Costly questionnaire challenged
| The Journal of Clinical
P U B L I C H E A LT H
Hypertension has retracted a 2008 article
that provided the empirical foundation for a
questionnaire whose co-creator billed some
researchers several thousand dollars each to
y
in August and reported last week by
Retraction Watch, comes nearly 4 years
after a scientist not involved in the research
told the journal the math in the paper was
implausible. In its retraction notice, the
journal said it conducted an independent
statistical review and concluded that the
article’s results were misleading.
Women pass on elite journals

selling them to private buyers, The Daily
Telegraph reported. The museum’s direc-
tor, Hartwig Fischer, said in a 25 August
statement of resignation that he accepted
responsibility for mismanagement and
y g
use. Public health specialist Donald Morisky
of the University of California, Los Angeles, PUBLISHING | Men are more likely
cited his copyright of the instrument, which than women to submit their scientific
helps predict how likely patients are to manuscripts to Cell, Nature, or Science,
adhere to a drug regimen. Rather than pay, according to a study of 4740 research-

a failure to respond to warnings about
potential thefts in 2021. The museum has
yet to release a comprehensive inventory of
PHOTOS: (TOP TO BOTTOM) CHRISTIAN WIRTH; JASMINA WIEMANN
several scientists chose to retract publica-
tions that used the questionnaire. The
retraction of Morisky’s paper, announced
p
ers. Women may “have a higher internal
standard for what constitutes novel
research,” write the authors of the study,
,
posted as a preprint last week on bioRxiv,
with potential implications for broader

gender disparities in academic science.

IN FOCUS A micrograph of The study team surveyed authors who
a fossilized blood vessel published research from 2008 to 2017,
from a 150-million-year- asking about their manuscript submission
old diplodocid dinosaur experiences and academic rank. Men and
was a runner-up awardee women rated the quality of their research
in this year’s BMC Ecology similarly. But when asked why they didn’t
and Evolution photography submit their work to an elite journal,
competition. The image women were more likely than men to say
was taken by molecular it was not groundbreaking or sufficiently
paleobiologist Jasmina novel. To address the gap, journal editors
Wiemann of the Field could do more outreach to women-led labs
Museum of Natural History. to encourage submission of their work, the
authors write.

SCIENCE science.org 1 SEP TEMBER 2023 • VOL 381 ISSUE 6661 929
NE WS
IN DEP TH
MEDICINE
Hot weight loss drugs tested against addiction
Clinical trials will gauge whether GLP-1 analogs curb drug and alcohol cravings
By Mitch Leslie
W
hen the diabetes treatments known
as GLP-1 analogs reached the mar-
ket in 2005, doctors advised pa-
tients taking the drugs that they
might lose a small amount of
weight. Talk about an understate-
ment. Obese people can drop more than 15%
of their body weight, studies have found, and
two of the medications are now approved by
the U.S. Food and Drug Administration (FDA)
for weight reduction. A surge in demand for
the drugs as slimming treatments has led to
shortages. “This class of drugs is exploding in
popularity,” says clinical psychologist Joseph
Schacht of the University of Colorado School
of Medicine.
But patient reports and animal studies
have yielded tantalizing signs that the drugs
may spur another unexpected and welcome
effect: fighting addiction. Most early trials
were disappointing, but they used less po-
tent versions of the drugs. Now, at least nine
phase 2 clinical trials are underway or being
planned to test whether the more power-
ful compound semaglutide and its chemical
cousins can help patients curb their use of
cigarettes, alcohol, opioids, or cocaine. Hopes
are high. Semaglutide (sold under the trade
names Wegovy, Ozempic, and Rybelsus) “is
truly the most exciting drug for the last few
930 1 SEP TEMBER 2023 • VOL 381 ISSUE 6661
decades,” says neuropharmacologist Leandro
Vendruscolo of the U.S. National Institute on
Drug Abuse.
If the results of the new trials are positive,
addiction science could have its own “Prozac
moment,” says clinical neuroscientist W. Kyle
Simmons of the Oklahoma State University
Center for Health Sciences. In the 1980s, that
drug brought a sea change to psychiatry, be-
coming part of popular culture and leading
to the wider use of antidepressants.
Scientists have long been searching for
new addiction drugs. Although FDA has ap-
proved several, including three for patients
with alcohol use disorder, these medicines
only work for a small percentage of people
who try them. And the pharmaceutical in-
dustry has not delivered new compounds,
in part because companies believe patients
won’t stick with treatments, making their
development a poor investment, says clinical
neuroscientist Lara Ray of the University of
California, Los Angeles. The last “new” drug
treatment for alcohol use disorder received
FDA approval in 2006—and it was an inject-
able version of a drug, naltrexone, that had
been available since the 1980s.
So when patients taking GLP-1 analogs for
diabetes or weight loss reported that their
hankering for substances such as alcohol and
nicotine declined, researchers and doctors in
the addiction field perked up. “You usually
p
g
y
don’t hear people say that a drug makes them
less interested in drinking,” Schacht says.
Researchers are still probing how GLP-1
analogs might pull off that feat. The drugs
replicate the effects of the hormone glucagon-
like peptide-1; by prodding its receptors in
the pancreas, they stimulate release of in-
sulin and trigger other beneficial responses,
y g
which explains how they help people with
diabetes. But several structures in the brain
also produce GLP-1 or carry receptors for the
hormone—including brain areas that are in-
volved in our reward pathways, which drive
p
us to pursue pleasurable activities such as
eating tasty food or hanging out with friends.
Addiction involves “hijacking of the reward
,
pathways in the brain,” says behavioral
neuroscientist Patricia Grigson of the Penn-
sylvania State University College of Medicine.
Researchers think GLP-1 analogs spur weight
loss in part by quelling activity of this system,
and the same mechanism could explain why
people who take the medications report they
are less motivated to drink and smoke.
Studies in rodents and primates have sup-
ported that mechanism and confirmed that
the drugs diminish the desire for substances
such as alcohol, fentanyl, nicotine, and her-
oin. Clinical psychiatrist Anders Fink-Jensen
of the University of Copenhagen and col-
leagues even demonstrated that the drugs
work in an incorrigible group of drinkers:
science.org SCIENCE

the monkeys living on the islands of St. Kitts
and Nevis in the Caribbean. These rowdy pri-
mates are notorious for heavy consumption
of alcoholic drinks, which they often swipe
from tourists.
So far, however, only two clinical trials
have suggested the drugs can curb addiction.
In one, a team led by Luba Yammine, a clini-
cian and researcher at the University of Texas
Health Science Center at Houston, found in
2021 that 46% of patients who wore nico-
tine patches and received weekly injections
of exenatide, a first-generation GLP-1 analog,
stopped smoking, versus 27% of people who
relied only on the patches. “That’s pretty good
in the world of smoking cessation research,”
Yammine says. A second preliminary trial for
binge eating was also positive.
Four other clinical trials came up empty.
Fink-Jensen and colleagues performed one
of these studies on 127 patients with alcohol
use disorder. During the 6-month trial, all
of the participants went through behavioral
therapy to encourage them to drink less,
and 62 patients received weekly injections
of exenatide. Yet both groups cut their alco-
hol consumption and the number of days on
which they drank heavily by about the same
amount, the researchers revealed last year.
“That was a surprise for us,” Fink-Jensen
says. A trial for cocaine consumption was
also negative—as were additional studies on
smoking and binge eating.
But all of the completed trials used older
GLP-1 analogs, Fink-Jensen notes. Semaglu-
tide binds more tightly to the GLP-1 recep-
tor and induces greater weight loss than
previous drugs. So Fink-Jensen and other
researchers have launched new trials, almost
all of which use semaglutide, hoping they will
reveal greater capabilities against addiction,
help identify which patients are most likely
to benefit, and clarify how the drugs work.
To try to nail down whether semaglutide
quells cravings, Schacht and colleagues will
ask participants with alcohol use disorder
to watch as a research technician pours a
brimming glass of their favorite drink. Af-
ter raising the glass and sniffing the aroma,
the patients will then rate their motiva-
tion to imbibe. Participants in two other
studies headed by clinical psychologist
Christian Hendershot of the University of
North Carolina School of Medicine will ac-
tually be given the opportunity to smoke or
drink alcohol in the lab. But they will try to
abstain for as long as possible, earning an
increasing amount of cash the longer they
hold out. Offering participants the chance
to indulge their cravings is advantageous, he
says, because it “can improve sensitivity for
detecting medication effects.”
The current trials are also deploying
techniques that measure brain activity, in-
SCIENCE science.org
cluding electroencephalography and func-
tional MRI (fMRI), to look for clues to the
drugs’ mechanisms. So far, only one pub-
lished study—the clinical trial of exena-
tide by Fink-Jensen and colleagues—has
collected those data for patients receiving
GLP-1 agonists as addiction treatments. The
researchers used fMRI to determine the
activity of three brain structures in the re-
ward system while patients viewed photo-
graphs of alcoholic drinks. The results were
inconclusive. Compared with participants
who took a placebo, patients on exenatide
showed less activity in one structure—the
ventral striatum—but activity in two other
regions didn’t change. Researchers hope the
new work will provide more clarity about
the drugs’ effects on brain circuits.
If GLP-1 analogs prove effective in the cur-
rent phase 2 trials, access to the medications,
which can cost more than $12,000 per year,
could become a concern. Because of high
costs and other barriers, patients with sub-
stance use disorders often can’t access cur-
rent addiction treatments, says psychologist
Katie Witkiewitz of the University of New
Mexico. “I don’t want [semaglutide] to be a
medication that is only available to people
who are wealthy.”
In the United States, insurance might help
if the drugs gain FDA approval as addiction
treatments. That would require pharmaceu-
tical companies to present data from larger,
phase 3 trials. In the meantime, doctors
could still prescribe the drugs off-label. But
because insurance plans usually won’t cover
off-label prescriptions, patients would have
to foot the bill themselves.
Researchers and doctors also fret about
potential side effects—and not just the
gastrointestinal problems such as nausea
and vomiting the drugs typically cause. Ray
worries they might work so well that they
crush patients’ enjoyment of life, a condi-
tion called anhedonia that has been linked
to depression, suicide, and relapse. And if
patients are going to be on GLP-1 analogs
for years, new problems could emerge. The
once-weekly version of semaglutide that
is so popular for weight loss only reached
the market in 2021, notes molecular neuro-
endocrinologist Giles Yeo of the University
of Cambridge. “We need a better idea of
long-term safety,” he says.
Doctors would face a further hurdle if the
drugs gained approval for addiction therapy:
determining how to work them into treat-
ment programs. They won’t be the cure-alls
that some popular articles imply, researchers
warn. They might be more like Prozac and
other antidepressants, which only work for a
fraction of patients. But even that would be a
boon, Ray says. “You can help a lot of people
even if only one in five responds.” j
N E WS
SCIENCE DIPLOMACY
Tensions
endanger U.S.-
China research
agreement
Scientists on both sides
welcome decision to
continue talks on renewing
collaboration
p
By Jeffrey Mervis
R
ising tensions between the United
States and China could derail the re-
newal of a 44-year-old agreement on
scientific cooperation between the
g
two countries.
Last week, U.S. President Joe Biden
invited China to spend the next 6 months dis-
cussing changes to the broad agreement, first
y
signed in 1979, that enables joint research.
The move came after Biden rejected calls
from some Republicans to let the pact lapse
on 27 August, its planned expiration date.
But the 6-month extension leaves little
time to resolve a host of thorny issues, policy
experts say. They include how to protect intel-
lectual property rights to any findings, share
data among collaborators, and ensure that
research outcomes are fully reported. The
y g
Biden administration also faces calls to block
joint work on any technologies that could
have both civilian and military applications.
Chinese officials welcomed the exten-
sion. “As two major R&D countries, China
p
and the United States should maintain con-
tact and exchanges” in science and techno-
logy (S&T), Chinese embassy spokesperson
,
Liu Pengyu said in a statement to Science.
The pact’s history “has fully proved that
China-U.S. exchanges and cooperation are
mutually beneficial,” he said.
Prominent U.S. scientists also applauded
Biden’s move. “This agreement has been
of enormous benefit to the United States,”
wrote Stanford University physicists Steven
Kivelson and Peter Michelson in a letter to
Biden signed by more than 1000 scholars
before the extension was announced. “We
can attest that cutting off ties with China
would directly and negatively impact our
own research, the work of our immediate
colleagues, and the educational mission of
our universities.”
1 SEP TEMBER 2023 • VOL 381 ISSUE 6661 931
NE WS | I N D E P T H
The pact, which has undergone revisions
across several 5-year renewals, essentially
says “that it’s OK with the U.S. government
that [U.S. researchers] collaborate on science
and technology with China, with appropriate
selectivity and management,” says physicist
John Holdren of Harvard University, who
served as science adviser to former President
Barack Obama and helped negotiate an ex-
tension in 2011.
The U.S. has similar bilateral research
agreements with some 60 countries. The one
with China, commonly known as the Science
and Technology Agreement (STA), neither
provides funding for joint projects nor man-
dates research in any particular sector. But
it enables government agencies, universities,
companies, and other entities in each nation
to pursue joint research. Such projects have
included a landmark clinical study showing
how folic acid can prevent birth defects and
a network of clean energy research centers.
“It is particularly important, at a time of
tension in so many dimensions of our rela-
tionship with China, that we show that there
are things worth preserving in that relation-
ship,” Holdren says.
Congressional Republicans and others,
however, say collaborating with China poses
a major threat to U.S. economic and national
security. “The Chinese Communist Party has
abused the openness of the American scien-
tific community to steal American research
and coopt it for its own malign purposes,”
Representative Mike Gallagher (R–WI), chair
of the Select Committee on China in the U.S.
House of Representatives, said after the ex-
tension was announced. Two months ago,
Gallagher and nine Republican colleagues
wrote Secretary of State Antony Blinken urg-
ing him to let the agreement lapse.
No STA project has involved sensitive or
classified research. But even fundamental
research discoveries can help China gain an
advantage over the U.S., Gallagher asserted.
“The evidence available suggests that [China]
will continue to look for opportunities to ex-
ploit partnerships organized under the STA
to advance its military objectives,” the letter
argues. “The United States must stop fueling
its own destruction.”
The STA holds symbolic significance
for China because it was the first agree-
ment the two countries signed after nor-
malizing relations in 1972, says Huiyao
Wang, president of the Center for China
and Globalization, a Beijing-based think
tank. It was renewed during the most tu-
multuous times, such as after the 1989
massacre at Tiananmen Square. Even the
administration of former President Donald
Trump, which was no fan of China, agreed
to several short-term extensions before sign-
ing the current STA, which added language
932 1 SEP TEMBER 2023 • VOL 381 ISSUE 6661
to protect intellectual property, in 2018.
Since the Biden administration took of-
fice in January 2021, it has not launched any
new government-to-government initiatives
under the STA. Although ending the bilateral
agreement wouldn’t prevent government-to-
government scientific cooperation, a U.S. De-
partment of State spokesperson says, “each
agency would have to negotiate arrange-
ments for each individual cooperation.”
Gallagher’s letter to Blinken turned what
had been quiet negotiations between the
two sides into a front-burner political issue.
Ending all collaboration because some future
project might become problematic would
amount to “throwing the baby out with the
bathwater,” says Holdren, who hopes the two
sides can reach an agreement. Even so, sup-
porters of a new deal say changes are needed.
“I’d like to see language relating to
greater transparency and data sharing,” says
Sudip Parikh, CEO of AAAS (which pub-
lishes Science), citing the potential value of
pooling health or environmental data col-
lected by each country. “Reciprocity has not
always been a hallmark of China’s scientific
enterprise.”
The pact should include “some kind of free
flow of information requirement,” says Denis
Simon, an independent specialist on China
policy. He notes that ambiguous wording in
recent Chinese data protection laws has left
foreign scientists unsure that they will have
access to data generated by joint projects.
China may be skeptical of any proposed
changes. “If the U.S. is thinking about im-
proving and strengthening this agreement,”
a deal is within reach, Wang says. But if the
U.S. wants to limit the pact or add inequitable
clauses, he warns, “that could be a problem.”
The issue’s politicization is already hinder-
ing cooperation, says Mu-ming Poo, head of
the Chinese Academy of Sciences’s Institute
of Neuroscience. Renewing the STA “is help-
ful only if there is basic trust and true will-
ingness to collaborate,” says Poo, who spent
decades in the U.S. before returning to China.
U.S. officials acknowledge that striking a
new deal won’t be easy. “We are clear-eyed
about the challenges associated with [China’s]
national strategies related to S&T, as well as
its domestic legal framework,” the State De-
partment spokesperson said. “Strengthened
protections in the agreement will be essential
for any longer-term extension.”
If a deal can be reached, Holdren predicts
the Biden administration will take advantage
of it to launch several joint projects. Areas
ripe for collaboration, he says, include ef-
forts to prevent future pandemics, improve
nuclear reactor safety, and better monitor
earthquake activity. j
With reporting by Dennis Normile.
CONSERVATION
An end to
horseshoe
p
crab bleeds?
Proposal could allow
synthetic proteins
g
to replace harvested
enzyme in drug testing
y
By Freda Kreier
A
proposal released last week by an
obscure U.S. pharmaceutical organi-
zation could help end a decadeslong
practice of bleeding horseshoe crabs
caught along the U.S. East Coast for a
y g
protein used to test drug safety.
On 22 August, the U.S. Pharmacopoeia
(USP)—a not-for-profit group that helps set
industry testing standards—published draft
guidelines on using synthetic alternatives
p
to horseshoe crab blood. The move could
help reduce the need to harvest blood from
hundreds of thousands of crabs each year, a
,
practice that kills up to 30% of the captured
animals. That has put at risk U.S. populations
of the 445-million-year-old arthropods, as
well as the migratory shorebirds that feed on
horseshoe crab eggs, conservationists say.
The draft guidelines are a “first and im-
portant step” toward changing how com-
panies test for bacterial contamination in
products, says Jaap Venema, chief science
officer at USP.
Conservationists welcomed the proposal.
The lack of an industry standard for replac-
ing the horseshoe blood protein has been
“the biggest impediment to [pharmaceuti-
cal companies] transitioning,” says ecologist
David Mizrahi, vice president of research
science.org SCIENCE

and monitoring at New Jersey Audubon
and co-founder of the Horseshoe Crab Re-
covery Coalition.
The discovery in the 1970s that a horse-
shoe crab’s blue-tinged blood coagulates in
the presence of endotoxins—a dangerous
bacterial byproduct—was a breakthrough
moment for pharmaceuticals. The coagu-
lating enzyme, limulus amoebocyte lysate
(LAL), soon became the global standard for
endotoxin testing, used to check everything
from pacemakers to chemotherapy drugs and
COVID-19 vaccines.
But this lifesaving test comes at a high
cost. Because horseshoe crabs cannot be
raised in captivity, companies in North
America harvest blood from Atlantic coast
animals captured as they migrate inshore
each spring to breed. In 2016, an estimated
400,000 horseshoe crabs were captured for
their blood. That number has since shot up
to about 600,000 crabs a year.
Bleeding the crabs doesn’t necessarily kill
them—70% to 95% survive and are put back
in the ocean, according to conservation and
industry estimates. But some released ani-
mals later skip breeding or die from blood
loss and stress, Mizrahi says. Although Atlan-
tic horseshoe crab numbers are recovering
from a population crash in the 1990s, they
remain below historic levels.
SCIENCE science.org
Atlantic horseshoe crabs are bled every year for an
enzyme used to test drug safety.
Not just horseshoe crabs are at risk,
Mizrahi says. Their breeding coincides with
the spring migration of shorebirds, such as
the endangered red knot, that feed on the
crabs’ eggs before continuing north. This
refueling is so critical that some states have
banned harvesting female horseshoe crabs.
Conservationists have pushed for drug
companies to switch to synthetic proteins
such as recombinant factor C (rFC), which
has existed since the 1990s. The European
Pharmacopeia endorsed using rFC in 2019.
But much of the industry has been slow to
switch because rFC is not endorsed by USP,
which sets standards for drug manufacturing
in 150 countries. Companies currently have to
prove to regulators that synthetic alternatives
are as effective as LAL for each product—a
process that takes time and money.
USP’s microbiology committee has pre-
viously rejected proposals to replace LAL.
But this time a large body of data supports
using synthetic alternatives, Venema says.
The standards will be open for public com-
ment in November and December and
could be finalized in 2024. j
Freda Kreier is a freelance journalist in Washington, D.C.
ARTIFICIAL INTELLIGENCE
AI excels
at name
that smell
Neural network
predicts odors from
chemical structures
By Elizabeth Pennisi
W
hen Jonathan Deutsch agreed to
sniff 400 vials of unlabeled liq-
p
uid for science, he didn’t know he
would be competing with a com-
puter. A research chef who helps
with food product development at
Drexel University, he simply welcomed the
chance to hone his sense of smell. But odor
g
profiles generated by Deutsch and 13 other
volunteers served as a test for a computer
program that had been trained to produce
these same types of descriptions—such as
y
fruity, cooling, fishy, piny—using chemical
structure alone.
The results, reported on p. 999, show that
the program, a so-called graph neural net-
work, is excellent at imitating human sniff-
ers, at least when it comes to simple odors.
It reliably predicted what the volunteers
smelled, a feat sensory biologists have been
working toward for decades. It also predicted
the smells of 500,000 other molecules, with
y g
no need to make or sniff them.
The result is a boon for the study of ol-
faction, a field that has “floundered around
for years looking for this information,” says
Stuart Firestein, a neuroscientist at Co-
p
lumbia University who was not involved in
the work. “The approach offers great po-
tential” to speed up the search for better
,
smelling consumer products, adds Andreas
Grasskamp, a neurobiologist who studies
perception at the Fraunhofer Institute for
Process Engineering and Packaging.
The findings may also help establish ol-
factory research as a field on par with sight
or vision. Smell, which in humans involves
a smaller proportion of the brain and fewer
types of receptor cells than in other mam-
mals, was long considered “a primitive sense
that wasn’t worth studying by neurobio-
logists,” says Pablo Meyer Rojas, a physicist at
IBM Research. It has also defied systematic
study. Whereas what we see and hear reflects
quantifiable properties such as wavelength
and frequency, smell doesn’t neatly corre-
1 SEP TEMBER 2023 • VOL 381 ISSUE 6661 933
NE WS | I N D E P T H

spond with the shape of a molecule. Similarly ACADEMIC RESEARCH

structured molecules can smell different,
whereas dissimilar molecules can produce
the same odor.
Ten years ago, a crowdsourced competition
Seizure of Nicaraguan university
challenged researchers to use artificial intel-
ligence (AI) to predict an odor’s smell from its
structure. Asked to “smell” 69 chemicals, the
deals blow to nation’s scientists
winning algorithms could pick out eight of Move marks President Daniel Ortega’s latest effort to crack
19 possible odors that people had attrib- down on perceived political opponents
uted to these samples. But it couldn’t group
the samples according to how similar they
smelled, says Joel Mainland, a neuroscientist By Kata Karáth ducted research and sponsored courses for
at the Monell Chemical Senses Center who teachers and journalists, as well as debates on

R
helped organize the challenge.
Rick Gerkin, a neuroscientist at the AI
company Osmo, was one of the winners.
Because the number of well-characterized
odor molecules for training AI “sniffers” has
greatly expanded since the competition, he
esearchers in Nicaragua say the gov-
ernment’s takeover of a prominent
private university has dealt another
serious blow to academic freedom and
scientific autonomy in the country.
On 15 August, President Daniel
issues including the proposed construction of
an interoceanic canal through Nicaragua.
The university has faced increasing scru-
tiny from Ortega’s government. In 2018,
after UCA became a hub of student-led pro-
democracy demonstrations, Nicaragua’s Na-

thought he could do better. With Alexander
Wiltschko, first at Google Research and
now at Osmo, Gerkin and colleagues in-
put the structures and odor descriptions of
5000 molecules into a more sophisticated AI.
It learned to recognize patterns in the train-
Ortega’s administration confiscated the assets
of the Central American University (UCA) in
Managua and closed its campuses, alleging
the Jesuit-run institution had become a “hub
of terrorism” and its leaders were “traitors to
the country.” It then reopened the campus
p
tional Council of Universities did not renew
its operating license. Since then, “we have
been operating, you could say, illegally,” says a
university official who requested anonymity.
The government also tightly regulated fac-
ulty research, especially in the social sciences.

ing data, correlating a molecule’s smell with
features of its constituent atoms—their iden-
tities, sizes, and connecting bonds.
Then came the testing, led by Mainland.
with new leadership and a new name: the
Casimiro Sotelo Montenegro National Uni-
versity, in honor of a student activist assas-
sinated in 1967.
g
“If I was going to do an investigation [off
campus] … I had to inform the local repre-
sentative of the ruling party, and they had to
accompany [me] to each household,” a UCA

y
Deutsch and his fellow volunteers signed The move was the latest effort by Ortega, professor says. Researchers were barred from
onto Zoom sessions with Monell sensory who has led Nicaragua since 2007, to crack accessing public records such as national sta-
biologist Emily Mayhew to sniff the 400 mys- down on perceived political tistics and faced legal action
tery vials and report which of 55 odors they opponents, including aca- if they failed to report for-
detected in each. Mayhew, now at Michigan demics. In recent years he “The trend is eign funding for their work.
State University, notes that perceived smells
vary greatly between people. So the team cal-
has closed some two dozen
other, mostly smaller, pri-
becoming really As a result, many research-
ers stopped publishing and
culated average human ratings to compare vate universities and the totalitarian. I feel like UCA stopped hosting inter-
with the network’s predictions. In more than nation’s National Academy national meetings.
half of cases, the neural network got even of Sciences. we’re living in a The university once re-
dark medieval village.”

y g
closer to this average than any individual in UCA officials have re- ceived nearly half of its
the group did. “This is a difficult accomplish- jected the government’s $16 million annual bud-
ment,” Grasskamp says. allegations. But faculty Faculty member, get from the government,
Next the team generated 500,000 hypo- members fear the move Central American University but that funding ended in
thetical chemical structures. The network eliminates one of the na- 2018. To cope, the university

quickly inferred how they should smell, pro-
viding a database that should help in the
search for odors for new foods, perfumes,
cleaners, and other products.
Although the neural network shows it’s
possible to map a chemical’s structure to an
odor, it sheds little light on smell’s underlying
biology. From a basic research perspective,
“it’s not clear whether this is an important
development rather than an incremental
one,” says Linda Buck, a neuroscientist at the
Fred Hutchinson Cancer Center. She adds
that the neural network still hasn’t proved
it can evaluate mixtures of molecules—the
kinds of complex odors we encounter in the
real world.
“The frontier is now about mixtures,”
Gerkin agrees. That’s what he hopes to teach
his algorithms to do next. j
tion’s last relatively independent centers of
academic research and will result in more
researchers and students leaving Nicaragua.
“The [university] was the last bastion of
quality higher education,” says one member
of its faculty, who requested anonymity be-
cause of fears of retaliation. “I believe that
independent scientific investigations are over
in Nicaragua.”
Considered one of Central America’s top
private universities, the 63-year-old UCA has
about 500 faculty members and 6000 stu-
dents, mostly undergraduates. (Ortega was
briefly a student there in the 1960s.) It houses
a range of research efforts, including a molec-
ular biology center, a natural sciences insti-
tute, and an herbarium holding some 80,000
plant samples. UCA also provided space and
support to the science academy, which con-
p
downsized staff and increased fees. “I think
they wanted to make it so difficult that we
would close down on our own, but we didn’t,”
,
the university official says. But, “In the end,
[the government] … had to kick us out.”
It’s not clear how many faculty will now
return. Many researchers have been exiled or
left the country under mounting threats from
the government. Some will stay “only because
they don’t have any other options,” the official
says, adding: They will be “forced to say that
everything is wonderful.”
“The trend is becoming really totalitarian,”
a faculty member says. “I feel like we’re living
in a dark medieval village … where there is
no rule of law, no social contract, where no
one can say anything.” j
Kata Karáth is a journalist in Ecuador.

934 1 SEP TEMBER 2023 • VOL 381 ISSUE 6661 science.org SCIENCE

BIOLOGY
Why do cats love tuna so much?
Study identifies taste receptors that give our feline friends a
craving for meat—and one fish in particular
By David Grimm
A
part from Garfield’s legendary love
of lasagna, perhaps no food is more
associated with cats than tuna. The
dish is a staple of everything from
The New Yorker cartoons to Meow
Mix jingles—and more than 6% of
all wild-caught fish goes into cat food. Yet
tuna is an odd favorite for an animal that
evolved in the desert. Now, researchers say
they have found a biological explanation for
this curious craving.
In a study published last month in Chemi-
cal Senses, scientists report that cat taste
buds contain the receptors needed to detect
umami—the savory, deep flavor of various
meats, and one of the five basic tastes in ad-
dition to sweet, sour, salty, and bitter. Indeed,
umami appears to be the primary flavor cats
seek out. That’s no surprise for an obligate
carnivore. But these cat receptors also seem
to be uniquely tuned to molecules found at
high concentrations in tuna, perhaps reveal-
PHOTO: SVETLANA SULTANAEVA/ISTOCK
ing why our feline friends prefer this deli-
cacy over all others.
“This is an important study that will help
us better understand the preferences of our
familiar pets,” says Yasuka Toda, a molecu-
lar biologist at Meiji University and a leader
in studying the evolution of umami taste in
mammals and birds. The work could help
pet food companies develop healthier diets
SCIENCE science.org
and more palatable medications for cats,
says Toda, who was not involved with the
industry-funded research.
Cats have a unique palate. They can’t taste
sugar because they lack a key protein for
sensing it. That’s probably because there’s
no sugar in meat, says Scott McGrane, a fla-
vor scientist and research manager for the
sensory science team at the Waltham Pet-
care Science Institute, which is owned by pet
food–maker Mars Petcare UK. Cats also have
fewer bitter taste receptors than humans
do—a common trait in uber-carnivores.
But cats must taste something, McGrane
reasoned, and that something is likely
the savory flavor of meat. In humans and
many other animals, two genes—Tas1r1
and Tas1r3—encode proteins that join to-
gether in taste buds to form a receptor that
detects umami. Previous work had shown
that cats express the Tas1r3 gene in their
taste buds. But do they have the other criti-
cal puzzle piece?
McGrane and colleagues biopsied the
tongue of a 6-year-old male cat, eutha-
nized for health reasons. His taste buds
expressed both the Tas1r1 and Tas1r3
genes—the first time scientists showed
that cats have all the molecular machinery
needed to detect umami.
When the researchers compared the pro-
tein sequences encoded by these genes with
those of humans, however, they found a
striking difference: Two critical sites that al-
low the human receptor to bind to glutamic
and aspartic acid—the main amino acids
that activate umami taste in people—were
mutated in cats. “So I began thinking, maybe
cats can’t taste umami,” McGrane says.
To double check, he and his team engi-
neered cells to produce the cat umami re-
ceptor on their surface. They then exposed
the cells to a variety of amino acids and nu-
cleotides. The cells did respond to umami—
but with a twist. In people, the amino acids
bind first and the nucleotides amplify the
response. But in cats, the opposite was true.
In the last part of the experiment,
McGrane and colleagues gave 25 cats a taste
test. In a series of trials, they presented the
felines with two bowls of water, each with
various combinations of amino acids and
nucleotides, or just water alone. The ani-
p
mals showed a strong preference for bowls
that contained molecules found in umami-
rich foods, suggesting this flavor—above all
others—is the primary motivator for cats.
“I think umami is as important for cats as
sweet is for humans,” Toda says. Dogs, she
g
notes, taste both sweet and umami, which
may explain why they’re not such fussy eaters.
But it wasn’t just umami in general the
cats craved. They were especially keen for
y
bowls containing histidine and inosine
monophosphate—compounds found at
particularly high levels in tuna. “It was
one of the most preferred combinations,”
McGrane says. “It really seems to hit that
umami sweet spot.”
That jibes with Toda’s past experience. As
a veterinary student, she got cats with no
appetite to eat by sprinkling their food with
dried flakes of bonito—a common umami
y g
ingredient in Japan and a close relative of
tuna. “It worked very well!” she says.
Indeed, one application of the work could
be developing foods that are more palat-
able to cats, McGrane says. He also thinks
p
a spoonful of umami (figuratively speak-
ing) could help feline medications go down
easier—welcome news for anyone who’s al-
,
most lost a finger trying to pill a cat.
Why cats have a hankering for tuna in
the first place remains a mystery. It may
have developed over time. They’re depicted
eating fish in the art of Ancient Egypt, and
by the Middle Ages, felines in some Middle
Eastern ports were consuming large quan-
tities of fish—including tuna—likely left by
fishers. Cats that evolved the taste may have
had an advantage over their comrades, says
Fiona Marshall, a zooarchaeologist at Wash-
ington University in St. Louis.
“We’re at a starting point—it’s not a fin-
ished story,” McGrane admits. “But all of
this work is building up to our basic under-
standing of what it means to be a cat.” j
1 SEP TEMBER 2023 • VOL 381 ISSUE 6661 935
NE WS

SMOKE
FEATURES

ALARM
As states relax their laws on
cannabis, neuroscientist
Yasmin Hurd is warning about

p
the drug’s dangers
for the developing brain

g
y
y g
p
,

O
ne morning in June, barely
5 months after the first dispensary
for recreational cannabis opened
in New York state, neuroscientist
Yasmin Hurd spoke via Zoom to an
audience of educators and special-
ists who work with or run programs
for children. The session’s organiz-
ers, alarmed by how many children
By Ingrid Wickelgren
in their South Bronx community were now
getting their hands on cannabis, had sought
Hurd’s expertise on the drug’s effects.
Hurd put up a slide of the human brain,
its bumps and grooves tinged blue, green,
yellow, and red to indicate the distribution of
the receptors to which tetrahydrocannabinol
(THC), the psychoactive ingredient in can-
nabis, binds. She showed how they exist
throughout the brain—in the folds of the ce-
rebral cortex, where much of cognition lies;
the cauliflower-shaped cerebellum, the seat
of motor coordination; the hippocampus,
Grand Central for memory; and the amyg-
dala, a crucial hub for emotional regulation.
The receptors, said Hurd, who heads an

936 1 SEP TEMBER 2023 • VOL 381 ISSUE 6661 science.org SCIENCE

addiction research lab at the Icahn School
of Medicine at Mount Sinai, are “really criti-
cal for so many processes in the brain.” And
when a person uses cannabis—in any of its
edible, dabbable, smokable forms—the drug
overwhelms them and disrupts their ability
to calibrate neuronal activity.
That, in turn, can be profoundly prob-
lematic for the developing brain, Hurd’s re-
search suggests. She sees growing evidence
in the field that cannabis use puts children
and adolescents at risk for a variety of psy-
chiatric problems, from dependence on that
drug and others to schizophrenia. In utero
exposure, she believes, can ignite mental
health problems in childhood and beyond.
In studies with rats, human fetal
tissue, and children, her lab has
begun to uncover changes in gene
expression, as well as alterations in
the brain’s chemical communica-
tion systems and wiring, that may
underlie some of these effects.
Hurd’s work is especially com-
pelling because she has been able
CREDITS: (GRAPHIC) D. AN-PHAM/SCIENCE; JULIA GREENWOOD/SCIENCE; (DATA) POTENCY MONITORING PROGRAM QUARTERLY REPORT #153, NATIONAL INSTITUTE ON DRUG ABUSE
to link results across species, col-
leagues say. “It’s so hard to be able
to go back and forth between ani-
mal models and effects in humans,”
says Susan Tapert, an addiction re-
searcher at the University of Califor-
nia, San Diego. “She’s really one of
the leaders in the field in being able
to pull those very different kinds of
studies together.” Tapert agrees the
evidence for harmful effects on the
developing brain are concerning,
although she says the harm likely
varies widely from one individual
to another. The risks for adults are
lower, she says, as the drug’s influ-
ences on memory, mood, sleep, and
motivation tend to wane within
about a month of discontinued use.
Research on cannabis’ devel-
opmental effects has grown in re-
cent years, but Hurd “definitely
pioneered this field,” says Miriam
Melis, a neuroscientist at the Uni-
versity of Cagliari. Hurd also had to
prove herself as a Black woman in
a discipline then dominated by white men,
Melis says. “For me she’s an inspiration of a
woman in science.”
Her work has become increasingly rel-
evant, as U.S. states—23 so far, plus Wash-
ington, D.C.—legalize cannabis for adult
recreational use. Hurd’s findings raise “big
red caution flags,” says Eric Nestler, an addic-
tion researcher and director of the Friedman
Brain Institute, which Hurd’s lab is a part of.
Legalization is “not a free decision. It is a de-
cision that, according Yasmin’s data—and I
would agree with her data—will bring costs.”
SCIENCE science.org
Some adults might be able to use canna-
bis quite safely, experts say, yet the legaliza-
tion trend has made the drug increasingly
accessible to pregnant women and also
children, who may ask adults to buy it,
take it from parents, or use fake IDs to get
it. In the Bronx, kids as young as age 11 or
12 know which shops will sell to minors, ac-
cording to Davon Russell, president of the
Women’s Housing and Economic Develop-
ment Corporation, a Bronx community de-
velopment organization, who invited Hurd
to speak. At the same time, the potency of
the products on offer has risen sharply (see
graphic, below) and is only loosely regulated
by states. Customers “have no clue about the
Higher times
Cannabis plants, resins, and oils have gotten far more potent over
the past quarter-century, according to an analysis of illegal samples
seized by the U.S. Drug Enforcement Administration. Concentrations
of tetrahydrocannabinol (THC), the main psychoactive ingredient, have
risen as states have legalized cannabis use.
20%
15.34
15
Average THC concentration
1996
California
becomes the first
10
state to legalize
medical cannabis.
2012
Washington and
Colorado become
the first states to
5 legalize cannabis
for recreational use.
3.96
0 meant to blend her two main inter-
1995 2000 2005 2010 2015
substances they are consuming,” Hurd says.
She has seen the consequences firsthand:
Besides running her lab, she directs Mount
Sinai’s Addiction Institute, overseeing in-
patient and outpatient treatment centers, as
well as programs for kids.
Although Hurd opposes the criminaliza-
tion of cannabis use and possession, she
believes legalization has come with under-
appreciated downsides. She’s concerned it
has fanned a permissive culture and a per-
ception that the drug is generally safe. “I am
worried about how cavalier we’re becoming
N E WS
and that there is a cannabis smoke shop now
practically, in some places, on every other
block,” she says. “I feel frustrated that people
are willing to sacrifice kids and young people
for their quote-unquote right to get high.”
Her science, she hopes, will foster a greater
awareness of the potential harms.
HURD REMEMBERS HER FIRST science experi-
ment. At about age 6, she set up tin pans
with rice outside her home in Jamaica,
varying the amounts of water and shade to
see how using sunlight to cook rice under
differing conditions affected its quality.
She was an inquisitive child who liked to
question the rules. Why, she recalls asking
her parents, did she have to drink
milk? Did they know that humans
are the only species that drinks
other species’ milk?
p
In the early 1970s, when Hurd
was about 10, her parents divorced
and she moved to New York City
with her mother and siblings. She
loved it from day one. “I’m defi-
nitely not a stereotypical Jamaican,
g
in that some Jamaicans are just
like, ‘No problem man, tomorrow,
tomorrow.’ I came to New York,
and everyone was moving, moving,
y
moving,” she says, and thought:
“This is my place!”
At South Shore High School in
Brooklyn, she was the only Black
person in her honors classes. “You
are a Black girl and they are always
challenging you that you really know
anything,” she says. Yet she excelled
in her science classes and studied
German so she could read classic ex-
y g
periments in their original language.
Hurd says her family valued educa-
tion, and their high expectations
helped propel her to college.
At Binghamton University, she
p
convinced administrators to cre-
ate a new degree for her: a B.A. in
biochemistry and behavior. It was
,
2020
ests, chemistry and behavior—but
she now jokes that it was essentially
a neuroscience major before that became a
thing. In graduate school at the Karolinska
Institute in Stockholm, her colleagues gave
her another lesson in expectations—this
one not based on her skin color. “They said,
‘You’re American; therefore, you must be the
best,’” Hurd says. Their high expectations
motivated her to measure up.
In grad school, Hurd helped develop tech-
niques for measuring neurotransmitters in
the rat brain. In some of her experiments,
she used amphetamine or cocaine to artifi-
cially raise dopamine levels. She was fasci-
1 SEP TEMBER 2023 • VOL 381 ISSUE 6661 937
 p
Neuroscientist Yasmin Hurd (bottom right) examines a slide of donated human brain tissue with students in her lab at the Icahn School of Medicine at Mount Sinai.

g
nated to see a mild-mannered rat suddenly consisting primarily of users of amphetamine cocaine and heroin, inspiring the so-called
become hyperactive, and at higher doses, and heroin. “She was really a pioneer in us- gateway hypothesis. Many scientists and

aggressive and ready to pounce. “If you’ve
never seen a paranoid rat,” she says, “it was
just ferocious.”
As a postdoc at the National Institute
of Mental Health in the early 1990s, she
learned some of the then-new molecular
biology tools to study how cocaine affected
cells and receptors in rodent brains. But
she wasn’t satisfied. “I needed it to have a
human relevance,” Hurd recalls.
ing human brain tissue to understand the
neurobiology of drug addiction,” Nestler
says. Her brain collection, which she ex-
panded at Mount Sinai, “provided assur-
ance that mechanisms scientists study in the
lab, say in rodent models, really focused on
things that had human relevance.”
It was rodents alone, however, that en-
abled Hurd to make her first mark on the
broader debate about cannabis. Epidemio-
y
laypeople believed the effect was strictly
environmental—that is, using cannabis is
likely to introduce people to a drug-using
crowd and to a dealer who also hawks
harder drugs. But Hurd thought there
might also be a biological connection.
What if, she thought, cannabis changes the
developing brain in a way that made some
people vulnerable to addictive substances
more generally?

After finding a National Institutes of
Health pathologist who had started a brain
bank that included cocaine users, she set
out to measure messenger
RNA (mRNA) transcripts
logical studies had suggested people who
use cannabis early in life are more likely
to later become addicted to drugs such as
y g
To investigate, Hurd’s team exposed ado-
lescent rats to THC and found the rodents
later self-administered heroin at increasing
rates, reaching dosages far
higher than controls. Early

lingering in the tissue af-
ter death, hoping to gauge
gene expression in the users’
brains. Other scientists told
her she was on a fool’s errand
because, they said, mRNA be-
comes unstable after death.
But she proved them wrong,
and was able to identify mo-
lecular changes in humans
that matched findings in rats
exposed to cocaine, as well
as some key species differ-
ences in reward regions of
the brain.
When Hurd returned to
Karolinska as an assistant
professor in the early 1990s,
she set up her own brain bank,
p
THC exposure also altered
gene expression in a reward
center of the brain, they found,
,
suggesting the drug can al-
ter the brain’s endogenous
opioid system, which is in-
volved in the perception of
reward, stress, and pain. Re-
viewers were skeptical, Hurd
says, but the manuscript fi-
nally came out in 2007 in
Neuropsychopharmacology,
the year after she became a
professor at Mount Sinai.
It was some of the first
strong biological support
for the gateway hypothesis,
The cannabinoid type 1 receptor, a target for the psychoactive ingredient in cannabis, Nestler says, helping con-
is found throughout the adult human brain (warm colors represent higher concentrations). vince researchers, educators,

938 1 SEP TEMBER 2023 • VOL 381 ISSUE 6661 science.org SCIENCE

and policymakers that biology was part of
the picture.
SPEAKING TO THE AUDIENCE of educators in
June, Hurd painted cannabis dependence as
a biological condition. “Many people think,
‘Oh you can’t become addicted to cannabis,’”
she says, “but when you look at the numbers
out there, cannabis use disorder is actually
quite common.” Estimates vary widely, but
up to 30% of users become unable to stop
using the drug despite negative effects on
their health and well-being, according to the
National Institute on Drug Abuse.
Adolescents are especially vulnerable,
Hurd told the Zoom audience, because the
endocannabinoid system—a network of
natural signaling molecules structurally
similar to THC, along with their receptors—
plays a central role in brain development. It
fine-tunes the maturation of the prefrontal
cortex, a brain area involved in self-
control and decision-making. In 2019,
Hurd and her colleagues reported that re-
peated THC exposure during adolescence
in rats changed the shape and function of
neurons in the animals’ prefrontal cortex.
In her presentation, Hurd showed a neon
green–and–yellow neuron with sparse,
stunted branches next to its much bushier
normal counterpart. The simpler struc-
ture, Hurd explained, means fewer con-
tacts with other neurons.
Her study also revealed a pattern of
gene expression in the rats’ THC-exposed
neurons that overlapped significantly with
gene expression profiles seen in people with
schizophrenia. It was a hint that early can-
nabis use might sometimes pave the way to
this psychiatric disorder, as epidemiological
studies have suggested. Human studies
also support concerns that early use might
have lasting effects. A longitudinal study
of 799 European adolescents published in
2021 linked cannabis use with a thinning
of the prefrontal cortex in regions where
cannabinoid receptors are expressed, and
with higher levels of impulsiveness.
Yet much is not known about adolescent
risk, Tapert says. She adds that a longitu-
dinal look at nearly 12,000 U.S. children
called the Adolescent Brain Cognitive De-
velopment Study, launched in 2016, should
provide critical data on how cannabis use
interacts with characteristics such as a
person’s genetics, history of trauma, stress,
PHOTO: MYSTERY SHOT/GETTY IMAGES
and family mental health history.
However such factors shift the balance,
the increasing potency of cannabis likely
adds to the risks. In a 2022 paper published
in Molecular Psychiatry, Hurd along with
Jacqueline-Marie Ferland, a neuroscientist
in her lab, reported that exposure to
high-dose THC (equivalent to a strong
SCIENCE science.org
recreational human dose of about 20 milli-
grams, or four typical gummies), but not
low-dose (equivalent to one 5-milligram
edible), once every 3 days made rats un-
usually sensitive to environmental stress-
ors such as isolation. After such stress, the
rats tended to avoid other animals—a sign
of social anxiety—and to consume more
sugar than controls, indicating increased
sensitivity to reward. Earlier this year,
Ferland, Hurd, and their colleagues re-
ported in JAMA Psychiatry that high-dose
THC also caused rats to make risky deci-
sions in a “rat gambling task,” in which
a rat must choose between risky and safe
strategies for winning sugar pellets, be-
havior similar to that seen in human study
participants with cannabis use disorder
gambling for money.
High- and low-dose THC also have dis-
tinct effects on the rat brain, they showed.
High doses altered the shape of neuronal
support cells called astrocytes and caused
changes in gene expression that suggest
disruptions to signaling by the inhibitory
neurotransmitter GABA. Low-dose con-
sumption, on the other hand, primarily
distorted the shape and gene-expression
patterns of neurons and spurred changes
in the opioid system. It’s not yet known
whether such changes also happen in the
human brain, but Hurd thinks the rat
studies hint at worrisome biological con-
nections between cannabis use and neuro-
transmitter systems involved in a wide
range of behaviors.
IN A 2019 STUDY, 7% of girls and women ages
12 to 44 reported using cannabis while preg-
nant. That proportion doubled between 2002
and 2017, the researchers found, and it may
be even higher today. The fetus, inevitably, is
exposed: THC readily passes through the pla-
centa to the fetal brain.
To look for possible effects, Hurd began
a collaboration in the early 2000s with
Diana Dow-Edwards, a neuropharmaco-
logist at the SUNY Downstate Health
Sciences University. At the time, Dow-
Edwards had access to fetal tissue from
women with a history of drug use who
had chosen to have an abortion. In women
who had smoked cannabis, Hurd and
Dow-Edwards found alterations to the
fetal brain’s dopamine system, including
reduced expression of dopamine receptors
in the amygdala and nucleus accumbens,
a reward center. The finding hints that in
utero cannabis exposure could interfere
with emotional regulation and boost vul-
nerability to addiction.
It’s just part of the havoc the two re-
searchers uncovered in the fetal brain. THC
reshuffled gene expression in its natural
N E WS | F E AT U R E S
How a component
of cannabis
could fight addiction
p
Y
asmin Hurd has spent much
of her career documenting the
harms caused by the psycho-
active compound in cannabis,
tetrahydrocannabinol (THC).
Ironically, she believes another can-
nabis ingredient, cannabidiol (CBD),
g
could help break cannabis depen-
dence. Her initial focus, though, is on
testing it to help heroin users.
In a seminal study published in
y
2009, she showed CBD could reduce
drug-seeking behavior in rats previ-
ously exposed to heroin, perhaps by
reducing craving triggered by cues
they had associated with the drug.
“CBD could actually do the opposite
of THC,” says Hurd, who heads an
addiction research lab at the Icahn
School of Medicine at Mount Sinai.
In 2019, she and her clinical team re-
y g
ported that compared with a placebo,
a CBD capsule taken once a day for
3 days reduced drug cravings and
anxiety in 45 human heroin users.
p
Hurd’s ballooning clinical crew is
gearing up for larger trials. First up is
a trial of CBD in 200 people with opi-
,
oid use disorder, followed by larger
scale trials with long-term follow-up
that will not only look at substance
abuse and anxiety, but also general
life outcomes such as employment,
parenting, and avoiding trouble with
the law. Mount Sinai neuroscientist
and former social worker Keren
Bachi, who helps direct this effort, is
impressed by Hurd’s ability to think
like a clinician. “She has the vision of
designing the studies in a way of see-
ing whether this intervention would
make a meaningful difference in the
lives of people,” Bachi says. —I.W.
1 SEP TEMBER 2023 • VOL 381 ISSUE 6661 939
NE WS | F E AT U R E S
The opening of the first recreational cannabis dispensary in New York City in December 2022 drew long lines.
Inside, a worker organized jars of cannabis flower.
opioid system. It also tampered with the
cytoskeleton, or internal scaffold, of develop-
ing neurons, reshaping their long extensions
and thereby altering the neuronal wiring in
parts of the fetal cerebral cortex.
When Hurd’s group tried to re-create
the effects of maternal cannabis use in ro-
dents, they saw behavioral consequences.
Male rats exposed to THC in the womb
more readily self-administered heroin as
adults than controls. And in work pub-
lished just last year in Biological Psychia-
try, rats exposed to THC in utero showed
low motivation, depressionlike traits, and
increased sensitivity to stress as adults.
Soon after she arrived at Mount Si-
nai, Hurd got the opportunity to find out
whether something similar happens in chil-
dren exposed to cannabis in utero. A young
assistant professor there, Yoko Nomura,
940 1 SEP TEMBER 2023 • VOL 381 ISSUE 6661
had started an ambitious longitudinal study
of pregnant women to examine how various
aspects of the prenatal environment such as
stress, toxins, and maternal obesity affect
children. Nomura kept running into Hurd
at meetings and was immediately drawn
to her affable personality. “She is very ap-
proachable,” says Nomura, now a professor of
psychology at Queens College at the City Uni-
versity of New York. The two joined forces.
For 15 years, Nomura, Hurd, and their col-
leagues followed the women and their 724
children, evaluating them annually. They
also sequenced mRNA in their placentas to
monitor the activity of thousands of genes. In
a paper published in 2021 in the Proceedings
of the National Academy of Sciences, Hurd,
Nomura, and their colleagues reported that
mothers’ cannabis use was associated with
hyperactivity and heightened anxiety and
aggression among their children at ages 3
to 6, along with increased cortisol, a stress
hormone, in hair samples. It also reduced
expression of placental genes involved in im-
mune function, which the endocannabinoid
system helps regulate. These changes corre-
lated with the children’s future anxiety and
hyperactivity levels.
Superstorm Sandy, which hit the New York
metropolitan area in 2012, enabled Hurd and
Nomura to show that stress exacerbates these
prenatal effects. Women who were pregnant
during the storm used cannabis at high rates,
likely as a coping mechanism, and in the
years since, their children have shown signs
of trouble. At ages 2 to 5, Hurd and Nomura
reported in May, these children were 31 times
more likely to meet the criteria for disrup-
tive behavior disorders and seven times
more likely to have an anxiety disorder than
p
kids exposed to neither cannabis nor Sandy
in utero. “It’s a drastic synergistic increase,”
Nomura says.
That work suggests the effects of can-
nabis on children may also be amplified
in communities and families with “much
g
greater psychosocial challenges,” Hurd
says. These may include neighborhoods of
color beset by poverty, violence, and a dis-
proportionate number of arrests, she says.
y
ALARMED AS SHE IS about these many risks,
Hurd does not support rolling back legal-
ization. Criminalizing cannabis possession,
she notes, has exacted disproportionate
costs to communities of color. It also hurts
people grappling with drug addiction. “It’s
absurd in a civilized society that we think
that locking people up for substance use
will cure the problem. It actually wors-
y g
ens the problem,” she says. Instead, she
favors regulations that limit potency and
using tax revenues from the sale of canna-
bis to educate people about the risks, and
for treatment and research to help those
p
harmed by its use.
Hurd spends much of her time fighting for
space for Mount Sinai’s addiction treatment
,
centers. It’s a constant, depressing battle
against the stigma of substance use disor-
ders, she says, despite an overdose epidemic
that has gripped the nation. A big part of the
problem is money. “Addictions are not a clin-
ically profitable specialty,” she says.
But she’s determined to keep fight-
ing. “There are a lot of people who have
a substance use disorder who would give
everything to get back a normal life. Ev-
erything,” Hurd says. “That’s why I’m so
committed to this career. It’s to help give
people their lives back.” j
Ingrid Wickelgren is a freelance science journalist
based in New Jersey.
science.org SCIENCE

and policymakers that biology was part of
the picture.
SPEAKING TO THE AUDIENCE of educators in
June, Hurd painted cannabis dependence as
a biological condition. “Many people think,
‘Oh you can’t become addicted to cannabis,’”
she says, “but when you look at the numbers
out there, cannabis use disorder is actually
quite common.” Estimates vary widely, but
up to 30% of users become unable to stop
using the drug despite negative effects on
their health and well-being, according to the
National Institute on Drug Abuse.
Adolescents are especially vulnerable,
Hurd told the Zoom audience, because the
endocannabinoid system—a network of
natural signaling molecules structurally
similar to THC, along with their receptors—
plays a central role in brain development. It
fine-tunes the maturation of the prefrontal
cortex, a brain area involved in self-
control and decision-making. In 2019,
Hurd and her colleagues reported that re-
peated THC exposure during adolescence
in rats changed the shape and function of
neurons in the animals’ prefrontal cortex.
In her presentation, Hurd showed a neon
green–and–yellow neuron with sparse,
stunted branches next to its much bushier
normal counterpart. The simpler struc-
ture, Hurd explained, means fewer con-
tacts with other neurons.
Her study also revealed a pattern of
gene expression in the rats’ THC-exposed
neurons that overlapped significantly with
gene expression profiles seen in people with
schizophrenia. It was a hint that early can-
nabis use might sometimes pave the way to
this psychiatric disorder, as epidemiological
studies have suggested. Human studies
also support concerns that early use might
have lasting effects. A longitudinal study
of 799 European adolescents published in
2021 linked cannabis use with a thinning
of the prefrontal cortex in regions where
cannabinoid receptors are expressed, and
with higher levels of impulsiveness.
Yet much is not known about adolescent
risk, Tapert says. She adds that a longitu-
dinal look at nearly 12,000 U.S. children
called the Adolescent Brain Cognitive De-
velopment Study, launched in 2016, should
provide critical data on how cannabis use
interacts with characteristics such as a
person’s genetics, history of trauma, stress,
PHOTO: MYSTERY SHOT/GETTY IMAGES
and family mental health history.
However such factors shift the balance,
the increasing potency of cannabis likely
adds to the risks. In a 2022 paper published
in Molecular Psychiatry, Hurd along with
Jacqueline-Marie Ferland, a neuroscientist
in her lab, reported that exposure to
high-dose THC (equivalent to a strong
SCIENCE science.org
recreational human dose of about 20 milli-
grams, or four typical gummies), but not
low-dose (equivalent to one 5-milligram
edible), once every 3 days made rats un-
usually sensitive to environmental stress-
ors such as isolation. After such stress, the
rats tended to avoid other animals—a sign
of social anxiety—and to consume more
sugar than controls, indicating increased
sensitivity to reward. Earlier this year,
Ferland, Hurd, and their colleagues re-
ported in JAMA Psychiatry that high-dose
THC also caused rats to make risky deci-
sions in a “rat gambling task,” in which
a rat must choose between risky and safe
strategies for winning sugar pellets, be-
havior similar to that seen in human study
participants with cannabis use disorder
gambling for money.
High- and low-dose THC also have dis-
tinct effects on the rat brain, they showed.
High doses altered the shape of neuronal
support cells called astrocytes and caused
changes in gene expression that suggest
disruptions to signaling by the inhibitory
neurotransmitter GABA. Low-dose con-
sumption, on the other hand, primarily
distorted the shape and gene-expression
patterns of neurons and spurred changes
in the opioid system. It’s not yet known
whether such changes also happen in the
human brain, but Hurd thinks the rat
studies hint at worrisome biological con-
nections between cannabis use and neuro-
transmitter systems involved in a wide
range of behaviors.
IN A 2019 STUDY, 7% of girls and women ages
12 to 44 reported using cannabis while preg-
nant. That proportion doubled between 2002
and 2017, the researchers found, and it may
be even higher today. The fetus, inevitably, is
exposed: THC readily passes through the pla-
centa to the fetal brain.
To look for possible effects, Hurd began
a collaboration in the early 2000s with
Diana Dow-Edwards, a neuropharmaco-
logist at the SUNY Downstate Health
Sciences University. At the time, Dow-
Edwards had access to fetal tissue from
women with a history of drug use who
had chosen to have an abortion. In women
who had smoked cannabis, Hurd and
Dow-Edwards found alterations to the
fetal brain’s dopamine system, including
reduced expression of dopamine receptors
in the amygdala and nucleus accumbens,
a reward center. The finding hints that in
utero cannabis exposure could interfere
with emotional regulation and boost vul-
nerability to addiction.
It’s just part of the havoc the two re-
searchers uncovered in the fetal brain. THC
reshuffled gene expression in its natural
N E WS | F E AT U R E S
How a component
of cannabis
could fight addiction
p
Y
asmin Hurd has spent much
of her career documenting the
harms caused by the psycho-
active compound in cannabis,
tetrahydrocannabinol (THC).
Ironically, she believes another can-
nabis ingredient, cannabidiol (CBD),
g
could help break cannabis depen-
dence. Her initial focus, though, is on
testing it to help heroin users.
In a seminal study published in
y
2009, she showed CBD could reduce
drug-seeking behavior in rats previ-
ously exposed to heroin, perhaps by
reducing craving triggered by cues
they had associated with the drug.
“CBD could actually do the opposite
of THC,” says Hurd, who heads an
addiction research lab at the Icahn
School of Medicine at Mount Sinai.
In 2019, she and her clinical team re-
y g
ported that compared with a placebo,
a CBD capsule taken once a day for
3 days reduced drug cravings and
anxiety in 45 human heroin users.
p
Hurd’s ballooning clinical crew is
gearing up for larger trials. First up is
a trial of CBD in 200 people with opi-
,
oid use disorder, followed by larger
scale trials with long-term follow-up
that will not only look at substance
abuse and anxiety, but also general
life outcomes such as employment,
parenting, and avoiding trouble with
the law. Mount Sinai neuroscientist
and former social worker Keren
Bachi, who helps direct this effort, is
impressed by Hurd’s ability to think
like a clinician. “She has the vision of
designing the studies in a way of see-
ing whether this intervention would
make a meaningful difference in the
lives of people,” Bachi says. —I.W.
1 SEP TEMBER 2023 • VOL 381 ISSUE 6661 939
 INSIGHTS
PERSPECTIVES
EVOLUTIONARY BIOLOGY
The raw material of evolution
Estimates of whale mutation rates contribute
to understanding evolutionary processes
By A. Rus Hoelzel1 and Michael Lynch2
E
volution happens when the code of
life, DNA, is changed by the process
of mutation. Mutations include dele-
tions and insertions, rearrangements,
and transpositions (moving DNA).
However, it is the rate of point muta-
tions, which affect a single site in the chain
of nucleotides that make up DNA, that is
most often considered. Knowing the rate
and pattern of mutation is essential to
understanding the process of evolution.
This knowledge increases understanding
of natural selection and has applications
such as tracking demography and dating
phylogenies, but calculating the muta-
tion rate is not straightforward. On page
1
Department of Biosciences, Durham University, Durham,
UK. 2Biodesign Center for Mechanisms of Evolution,
Arizona State University, Tempe, AZ, USA.
Email: a.r.hoelzel@durham.ac.uk
942 1 SEP TEMBER 2023 • VOL 381 ISSUE 6661
990 of this issue, Suárez-Menéndez et al.
(1) use parent-offspring trios and genome
sequences to estimate rates of point-muta-
tional change from one generation to the
next in four species of baleen whales in
the North Atlantic: humpback (Megaptera
novaeangliae), blue (Balaenoptera muscu-
lus), fin (Balaenoptera physalus), and bow-
head (Balaena mysticetus).
Generating sufficient data from such in-
accessible animals is a challenge. However,
Suárez-Menéndez et al. sampled exten-
sively enough to identify parent-offspring
trios—five trios for the humpback whale
and one for each of the other three spe-
cies—using a relatively inexpensive and
rapid genetic method for kinship screen-
ing. Their results matter because a muta-
tion rate cannot simply be calculated for
one species, for which this can be easily
done, and then applied more generally to
other taxa. There is no universal mutation
rate. There is not even a single mutation
p
g
y
y g
rate for a given taxon—it varies across indi-
viduals and among nuclear, mitochondrial,
and chloroplast genomes. The mutation
rate is subject to natural selection like all
p
traits (2, 3), and mutation rates in nuclear
genomes vary 10,000-fold across the tree
of life and 40-fold among vertebrates (3).
,
Estimating mutation rates in very large
mammals is of interest in part to help ad-
dress how species with long generation
times (average time between two consecu-
tive generations) avoid high incidence of
somatic diseases such as cancer. Suárez-
Menéndez et al. estimated the nuclear point-
mutation rate for the four species of baleen
whales combined to be 1.11 × 10−8, which is
high compared with earlier estimates that
were based on phylogenies. Therefore, in
cetaceans a lower than expected incidence
of cancer may be due to selection, as sug-
gested by Tejada-Martinez et al. (4) and oth-
ers, rather than a slower mutation rate.
After a mutation occurs, the individual
science.org SCIENCE

with the mutation may fail to survive or
reproduce because of impairment due to
the mutation (purifying selection) or sim-
ply by failing to reproduce by chance (ge-
netic drift). Therefore, a comparison of
populations or species after 100 or 1000
generations will reflect net genetic differ-
entiation that results from de novo varia-
tion introduced by mutation, as estimated
by trio analysis, which is then modified by
subsequent forces of selection and drift.
Eventually, the measured rate will reflect
new variants shared by all members of the
population (the substitution rate), which
can be more than an order of magnitude
lower than the de novo mutation rate,
but there may be intermediate rates for a
period of time (5). A further factor that is
frequently encountered in the relatively
rapidly changing mitochondrial DNA ge-
nomes of mammals is saturation. This oc-
curs when mutation changes one nucleotide
to another, and then back again. This can
SCIENCE science.org
Humpback whales (Megaptera novaeangliae)
are often inaccessible, which makes assessment
of their mutation rate challenging.
be anticipated and modeled (6), but there
are uncertainties involved. Determining
the mutation rate with trio analyses avoids
these issues, because time is too short for
the loss of anything other than a dominant
lethal mutation.
Knowledge of the mutation rate can re-
veal other important aspects of the biol-
ogy of a species, for example, demography.
The genetic effective population size (Ne) is
an idealized approximation of the average
number of individuals that produce suc-
cessful progeny per generation, which gov-
erns the level of noise in evolutionary pro-
cesses. Suárez-Menéndez et al. used linkage
disequilibrium analysis (nonrandom associ-
ation of loci) and calculated the Ne of hump-
back whales before commercial whaling
began to be 5,700, though, as the authors
point out, there are caveats to these estima-
tions. Ne is generally smaller than the total
number of reproductive adults in the popu-
lation, owing to variance in reproductive
success among individuals, demographic
fluctuations in population size, and delete-
rious mutations (which can purge variation
from linked chromosomal regions). The ra-
tio of Ne to the census population size var-
ies widely (7) and, for the humpback whale,
may be roughly 0.1 (8).
All methods for estimating demographic
trajectories depend critically on accurate
measures of both the mutation rate and the
generation time. An alternative method for
calculating this trajectory is based on co-
alescence (the time at which evolutionary
lineages come together) and the knowledge
that the interval between these coalescence
points varies with Ne [(9), compare with
(10)]. For example, the skyline model can
estimate these demographic trajectories
and incorporate ancient DNA to estimate
the relevant mutation rate [but see (11, 12)].
These rates may be somewhat slower than
trio-based rates (6) but can provide remark-
ably close correlation with known histori-
cal events (13) and are typically faster than
substitution rates calculated from fossil-
calibrated phylogenies (1, 5).
A long-term average Ne can be calculated
from the diversity of the population at pu-
tatively neutral sites within the genome
(the behavior of which is not biased by se-
lection) and the mutation rate. A notable
negative correlation between taxon-spe-
cific nuclear mutation-rate estimates and
their long-term Ne is found when organ-
isms are compared across the entire tree
of life (14), and now there are enough data
to see this trend emerging among mam-
mals. Estimates of mammalian mutation
rates (per generation) range approximately
fourfold, from 0.4 × 10−8 (pig) to 1.6 × 10−8
(gray mouse lemur), with cetaceans having
relatively high estimated mutation rates,
in the range of 0.9 × 10-8 to 1.4 × 10-8 (1,
3). The observed negative scaling is consis-
tent with the drift-barrier hypothesis (15),
which postulates that as Ne declines, drift
reduces the ability of selection to purge
mutator alleles. When Ne is large, selection
is more efficient at reducing the mutation
rate. Other factors that may influence ver-
tebrate mutation rates include parental
age, species-level fecundity, a higher muta-
tion rate in males than females [in birds
and mammals (3)], and genome size (14).
Several uncertainties remain with respect
to the new data from Suárez-Menéndez et
al. as well as recent estimates from other
p
species of vertebrates (3). Estimates of Ne
that are derived by dividing standing lev-
els of variation at putatively neutral sites
by mutation rate are subject to biases,
with the bias direction depending on the
relative magnitude of sampling error that
g
is associated with estimates of nucleotide
variation and mutation rates. Additionally,
most mammalian mutation-rate estimates
are based on very small numbers of trios
y
(often only one), although it is known that
mutation rates can vary by a factor of at
least two among individuals within a spe-
cies (14). Thus, new estimates such as the
one presented by Suárez-Menéndez et al.,
which have been notably lacking for long-
lived species other than humans, represent
progress toward a fuller understanding of
the rate of origin of the raw material of evo-
lution. At the same time, caution is needed
y g
when dating past evolutionary events with
present-day mutation rates and drawing
inferences about the biological basis for
taxon-specific mutation-rate differences. j
p
RE FE REN C ES AN D N OT ES
1. M. Suárez-Menéndez et al., Science 381, 990 (2023).
2. W. Wei et al., Nat. Commun. 13, 4752 (2022).
3. L. A. Bergeron et al., Nature 615, 285 (2023).
,
4. D. Tejada-Martinez et al., Proc. R. Soc. Lond. Ser. B 288,
20202592 (2021).
5. S. Y. W. Ho et al., Syst. Biol. 60, 366 (2011).
6. D. Graur, Molecular and Genome Evolution (Sinauer,
2016).
7. S. Hoban et al., Biol. Conserv. 248, 108654 (2020).
8. A. L. Cypriano-Souza, T. F. da Silva, M. H. Engel, S. L.
Bonatto, Genet. Mol. Biol. 41, 253 (2018).
9. S. Y. W. Ho, B. Shapiro, Mol. Ecol. Resour. 11, 423 (2011).
10. K. V. Parag, O. G. Pybus, C.-H. Wu, Syst. Biol. 71, 121
(2021).
11. P. Johri et al., PLOS Biol. 20, e3001669 (2022).
12. S. Möller, L. du Plessis, T. Stadler, Proc. Natl. Acad. Sci.
U.S.A. 115, 4200 (2018).
13. M. de Bruyn, A. R. Hoelzel, G. R. Carvalho, M. Hofreiter,
Trends Ecol. Evol. 26, 405 (2011).
14. M. Lynch et al., Nat. Rev. Genet. 17, 704 (2016).
15. M. Lynch, Genome Biol. Evol. 3, 1107 (2011).
10.1126/science.adk0121
1 SEP TEMBER 2023 • VOL 381 ISSUE 6661 943
I NS I GHTS | P E R S P E C T I V E S
HEALTH
Unanswered questions about
the causes of obesity
Obesity is now a global pandemic, but there is little
consensus about the causes
By John R. Speakman1,2,3,4,
Thorkild I. A. Sørensen5,6,
Kevin D. Hall7, and David B. Allison8
O
besity is a major health issue that
has reached pandemic status with
no clear solutions. Much has been
learned over the past 50 years about
the regulation of body fat. Examples
include the discovery of the hormone
leptin, finding thermogenic brown adipose
tissue in adult humans, elucidating path-
ways in the brain that affect hunger and
feeding behavior, quantifying adipocyte
turnover and the lipids therein, identifying
single genes that produce rare but severe
obesity, and finding thousands of genetic
variants associated with individual differ-
ences in body mass index (BMI). Despite
this progress, there remain several key
questions to be answered to aid the preven-
tion and treatment of obesity.
Confusion about the causes of obesity has
arisen based on the false dichotomy of genes
versus environment (rather than the com-
bined effects of genes and environment).
At any point in time, most of the variance
in levels of obesity among individuals may
be genetic. But, changes across time are
predominantly driven by the environment.
Which individuals deposit the most fat in
response to environmental change is influ-
enced by both.
Adult human body weight is often stable
over long periods within a given environ-
ment. Moreover, individuals often respond
to imposed perturbations of energy balance
by altering components of energy expen-
diture and intake to resist weight change.
Frequently, body weight (and fatness) re-
turns to a similar level to what it was before
the perturbation after it ends. This might
suggest that body weight is regulated around
an individual set point (1). But if that is the
case, why is there an obesity pandemic? Two
alternatives to a strict set point model in-
clude that the so-called regulation is an illu-
1
Shenzhen Key Laboratory of Metabolic Health, Center for Energy Metabolism and Reproduction, Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences, Shenzhen, China.
2
School of Biological Sciences, University of Aberdeen, Aberdeen, Scotland, UK. 3State Key Laboratory of Molecular Developmental Biology, Institute of Genetics and Developmental Biology,
Chinese Academy of Sciences, Beijing, China. 4China Medical University, Shenyang, Liaoning, China. 5Department of Public Health, University of Copenhagen, Copenhagen, Denmark. 6Novo Nordisk
Foundation Center for Basic Metabolic Research, University of Copenhagen, Copenhagen, Denmark. 7National Institute of Diabetes and Digestive and Kidney Diseases, National Institutes of Health,
Bethesda, MD, USA. 8School of Public Health, Indiana University, Bloomington, IN, USA. Email: j.speakman@abdn.ac.uk; tias@sund.ku.dk; kevinh@niddk.nih.gov; allison@iu.edu
944 1 SEP TEMBER 2023 • VOL 381 ISSUE 6661
sion because when individuals gain weight,
they also increase expenditure (2), which
brings them to a new equilibrium (the “dy-
namic equilibrium point”). Or, there are two
“control points” that stop humans from get-
ting too fat or too thin, but between those
limits there is very little control (the “dual
intervention point model”) (1). However,
these alternative models are inadequate de-
scriptors of the whole obesity phenomenon.
To develop more-comprehensive models,
it is necessary to understand the molecu-
lar mechanisms of intake and expenditure
control. For example, during periods of en-
ergy deficit, hunger is increased, and energy
expenditure is suppressed. Leptin appears
to play a major role in these responses but
does not fully explain them. Little is known
about the mechanisms behind responses to
overfeeding. Rodent studies suggest that
activation of thermogenic brown adipose
tissue or futile energy cycling elsewhere are
potentially involved, but their contributions
in humans remain unclear. A system that
directly senses overall weight (a gravitostat)
might be involved (3), but the molecular ba-
sis is ill-defined.
The sizes of individual organs and tis-
sues also seem to be regulated. People with
obesity, almost without exception, also
have increased lean body mass (all non-
adipose parts of the body, including skel-
etal muscles, bones, and viscera) (4). The
mechanisms involved in partitioning en-
ergy imbalances between lean versus adi-
pose tissues have not yet been elucidated.
The mass of adipose tissue depends on the
balance between adipogenesis and apopto-
sis and between lipogenesis and lipolysis
in adipocytes, but the mechanisms con-
trolling this are incompletely understood.
Moreover, the factors that influence the
relative deposition of lipids into adipose
versus nonadipose tissues are particularly
important to understand because ectopic
lipid deposition seems to be implicated in
metabolic disease. Elucidating these con-
trol mechanisms has the potential to reveal
therapeutic targets.
There appears to be a balance between
the nutrient and energy needs of the body
and food intake. However, little is known
about how these needs are sensed and how
the corresponding signals influence appe-
tite and behavior. For example, although
short-term bouts of exercise may not reli-
ably influence food intake (5), prolonged
exercise interventions and high activity
stimulate food consumption (6). Although
leptin acts as a long-term circulating signal
of overall body energy stores and stimu-
lates hunger when they are shrinking, other
currently unknown mechanisms appear to
link expenditure to intake through rest-
ing metabolic rate and fat-free mass (7).
Discovery of these signals will likely be im-
portant targets for modulating fat storage.
p
However, it is important to recognize that
the different routes of energy expenditure
(physical activity, rest, and thermoregula-
tion) may not be related to each other in an
independent additive fashion. Elevations of
one aspect of expenditure may cause com-
g
pensatory decreases in other components
and/or changes in intake. The long- and
short-term consequences of manipulating
different components of energy balance re-
y
main unclear.
Food intake and energy expenditure are
controlled by the brain. Yet, it is unclear
how the brain orchestrates eating behavior
in response to signals from the periphery
and the environment. The brain integrates
signals from the body regarding nutrient
and energy needs to modulate food choice
and the amount consumed. The hindbrain
and hypothalamus regulate food intake in
y g
response to signals from the gut and circu-
lating hormones. The recent success of glu-
cagon-like peptide 1 receptor agonists for
the treatment of obesity is because they en-
gage these brain regions at pharmacological
p
doses. There are probably many other brain
regions involved. For example, there is less
knowledge about the roles of brain regions
,
that support motivation, reward, habit for-
mation, cognitive abilities, and emotional
control and how these integrate mutually
and with environmental cues to generate
the total feeding response.
It is presently unclear what social and
physical factors in the environment are
driving increased obesity prevalence.
Physical changes in the food environment
have likely been important contributors to
increasing obesity prevalence, but the most
science.org SCIENCE

Some aspects of the changing food environment, combined with other environmental factors and genetic predisposition, have created the current obesity pandemic.
However, the details of this interaction remain elusive, leading to many unanswered questions about what might appear to be a simple problem of positive energy balance.
important factors in food that induce obe-
sity remain unknown. Intake is not only for
energy but for specific nutrients as well.
How nutrients are sensed and the signals
that control nutrient-specific appetites are
poorly understood. The protein leverage
model (8) suggests that protein content has
been diluted by excess carbohydrate and
fat, thereby promoting excess energy intake
to meet the body’s protein needs. However,
the determinants of protein requirements,
the signals of protein deficiency and sur-
plus, and the resultant mechanisms con-
trolling total protein intake are unclear.
Alternatively, the carbohydrate-insulin
model (9) suggests that the increase in high
glycemic carbohydrates in food results in
increased insulin levels and thereby drives
adipose tissue lipid accumulation and con-
sequently increased appetite and decreased
energy expenditure. But, the mechanisms
remain to be elucidated. Fructose does not
stimulate insulin but may elicit different
mechanisms that lead to fat storage.
Energy density of food, determined
partly by water and fat content, has a pro-
found effect on energy intake, at least over
the short term (10). Laboratory rodents and
humans consuming high-fat diets eat less
weight of food but overconsume energy.
This leads to some intriguing questions
about why the intake of high–energy den-
sity foods is so difficult to regulate and to
PHOTO: SORBIS/SHUTTERSTOCK
what extent increased availability of high–
energy density foods has contributed to
the pandemic. Conversely, a reduction in
low–energy density alternatives may be a
driver. For instance, dietary fiber has low
energy density, and its intake has declined
over the past 30 years in the USA. Are there
SCIENCE science.org
mechanisms by which fiber exerts effects
on satiety and thereby limits energy intake
other than by lowering food energy density?
The availability and marketing of ultrapro-
cessed foods have substantially increased
over several decades, and diets high in such
foods promote greater energy consumption
(11). It is not yet known what attributes of
ultraprocessed food result in excess energy
intake. Certain nutrient combinations may
be hyperpalatable and influence the activ-
ity of brain regions involved in reward and
motivation, thereby promoting their over-
consumption, but the specific mechanisms
remain unclear. Whether noncaloric sweet-
eners and the thousands of food additives
common in modern foods affect satiety and
energy intake is also unclear.
Social factors may be a feature of the
obesogenic environment that differs among
and within societies, and the distribu-
tion and impact of these factors may have
changed over time. A strong environmen-
tal correlate of obesity is social adversity,
including poverty, particularly in high-
income countries (12). However, in low- and
middle-income countries, poverty tends to
be associated with lower BMI. Perhaps, a
combination of food availability and food
insecurity explains this pattern, so that in-
dividuals who experience food insecurity
despite food availability become obese. But
what triggers the perception of food insecu-
rity, and what the physiological reaction to
it is, including increased fat storage, remain
unknown. Chronic stress may be an impor-
tant factor. Moreover, these relationships
appear more prominent among women
than men and may contribute to obesity
occurring more frequently in women than
p
in men. Such processes may be reinforced
owing to the stigmatization and discrimina-
g
tion that people with obesity suffer, creating
a vicious cycle.
The environment and societal changes
may also have affected physical activity.
y
Occupational physical activity in the USA
has declined since the 1950s (13), but leisure-
time activity has increased. Energy expendi-
ture resulting from physical activity has not
declined since the late 1980s in the USA and
Europe (14). Moreover, low energy expendi-
ture from physical activity does not predis-
pose individuals to obesity. Although people
with obesity are less active, they do not
spend less total energy on activity. Although
y g
it is often asserted that increasing sedentary
behavior is a major cause of the obesity pan-
demic, this is far from clear, and present evi-
dence does not support this conclusion.
Environmental drivers of the obesity
p
pandemic have not resulted in uniform
increases in adiposity within and between
populations, potentially also because of
,
genetic differences between individuals.
Individual variation in responses to many
of the physiological and environmental fac-
tors contributes to variation in obesity sus-
ceptibility. Much is known about the rare,
extreme monogenic forms of obesity, but
there are thousands of single-nucleotide
variants (SNVs) associated with BMI (15).
However, knowledge about the mechanisms
by which these SNVs become associated
with BMI is limited, both because of the
inadequacy of BMI as a measure of obe-
sity and because of uncertainty about the
causal mechanisms linking genetic variants
to phenotypes. Moreover, at the fetal stage
and early in life, individuals are wholly de-
1 SEP TEMBER 2023 • VOL 381 ISSUE 6661 945
I NS I GHTS | P E R S P E C T I V E S

pendent on their mothers for nourishment. CLIMATE CHANGE

Studies of rats and mice show that mater-
nal obesity and diet influence offspring sus-
ceptibility to obesity, possibly by epigenetic
mechanisms (15), but whether this also ap-
Arctic sea ice, ocean,
plies to humans is uncertain.
Given the diversity of environmental and
genetic causes of obesity, there are many
and climate evolution
routes to excess adiposity. There may be Wind variability affects the rate of Arctic sea ice decline
subtypes of obesity, which could guide spe-
cific prevention and treatment. At present,
however, there are many uncertainties, By Sheldon Bacon regional atmospheric circulation can control
such as how the different subtypes might be decadal changes in the rate of sea ice decline.

defined. Individual variation in food pref-
erence is enormous, but the basis of this
variation and its consequences for obesity
development remain unknown. Whether
different food choices lead to adaptations
of the gut and the resident microbiota to
alter energy absorption efficiency is also
A
rctic sea ice decline is caused by global
warming and is accelerated by regional
feedbacks resulting from changes in
absorption and reflection of heat by
ice and snow (albedo), among other
processes. By the end of the century,
Arctic sea ice is confidently expected to dis-
The Arctic Ocean is stratified, or lay-
ered, nearly horizontally in density. Sea
ice floats on the cold, fresh, ~200-m-thick
halocline. The halocline in turn overlies
the saline and relatively warm Atlantic
water layer, which is ~1 km thick. Below
are deep, dense, and very slow-moving

unclear. In some laboratory rodent models,
differences in absorption efficiency can be
sufficient to drive large differences in fat
storage, but whether the same applies to
humans has not been sufficiently studied.
Gut microbiota differences may be associ-
appear in summertime, although this will
probably happen sooner (1). The Arctic sea
ice area decline further increases global
warming rates because the albedo reduction
means that more heat is absorbed by Earth.
The Arctic sea ice changes by themselves are
p
waters. Summertime heating and melting
generate a thin mixed layer in the surface
that largely disappears with winter freez-
ing. Thus, the halocline largely insulates
the underlying ocean from surface fluxes
of heat and momentum. As a result, across

ated with absorption efficiency and may
have effects on other aspects of energy
balance. It remains unclear whether dif-
ferences in the microbiota between people
concerning, but remote atmospheric and oce-
anic linkages also force changes in weather,
climate, and extreme events at lower lati-
tudes (2). On page 972 of this issue, Polyakov
g
most of the Arctic Ocean, circulation is
sluggish, turbulence and mixing are largely
quiescent, and sea ice loss is mainly caused
by atmospheric forcing—wind and heat

with and without obesity are consistent and
reproducible and whether these differences
are consequences or causes of obesity.
By building on the considerable advances
made in the past 50 years, the study of the
causes of obesity promises to be a rich area
for discovery that may be transformative for
the lives of millions of people. Improving
our understanding of environmental driv-
ers and how these interact with genetic
y
et al. (3) show that mid-term variability in the acting on the sea surface.
The Arctic Ocean in the past and future
In the past, sea ice almost entirely covered the Arctic Ocean, and the ocean displayed slow
currents (left). In the future, sea ice is predicted to disappear, with currents' speed increasing
and causing turbulence to intensify, particularly over the continental shelf edge (right).
This could cause upward heat flux to the surface that might further reduce sea ice, even in winter.
PAST

y g
composition is vital to making future in- Wind Sea ice
roads to this serious medical condition. j
RE F ER E NC ES AND NOTES
1. J. R. Speakman et al., Dis. Model. Mech. 4, 733 (2011).

p
2. K. D. Hall et al., Lancet 378, 826 (2011).
3. J.-O. Jansson et al., Proc. Natl. Acad. Sci. U.S.A. 115, 427
(2018). FUTURE
4. H. Pontzer et al., Science 373, 808 (2021).
5. M. Hopkins et al., Int. J. Obes. 43, 1466 (2019).
,
6. P. Caudwell et al., Am. J. Clin. Nutr. 97, 7 (2013).
7. S. J. Simpson, D. Raubenheimer, Obes. Rev. 6, 133
(2005).
8. D. S. Ludwig et al., Am. J. Clin. Nutr. 114, 1873 (2021).
9. J. H. Ledikwe et al., Am. J. Clin. Nutr. 83, 1362 (2006).
10. B. J. Rolls, Physiol. Behav. 97, 609 (2009).
11. K. D. Hall et al., Cell Metab. 30, 67 (2019).
12. E. Hemmingsson, P. Nowicka, S. Ulijaszek, T. I. A.
Sørensen, Obes. Rev. 24, e13514 (2023).
13. T. S. Church et al., PLOS ONE 6, e19657 (2011).
14. J. R. Speakman et al., Nat. Metab. 5, 579 (2023).
15. R. J. F. Loos, Curr. Opin. Genet. Dev. 50, 86 (2018).

AC KN OWL EDG M ENTS

The authors thank the Royal Society for supporting the Halocline
scientific meeting in October 2022 about Causes of obesity:
Theories, conjectures and evidence, which was the inspiration Atlantic water Currents
for this Perspective.
Deep dense water
Continental
10.1126/science.adg2718 Turbulence shelf

946 1 SEP TEMBER 2023 • VOL 381 ISSUE 6661 science.org SCIENCE
 Differences to this relatively simple pic-
ture of the Arctic Ocean used to exist in the
western Eurasian Basin (north and east of
Svalbard, north of Norway) where Atlantic
water entering the Arctic through the Fram
Strait (between Svalbard and northeast
Greenland), at or near the surface, remained
in contact with the atmosphere. In a previ-
ous study (4), the Atlantic character was seen
to be extending eastward, toward eastern
Siberia, between the early 2000s and the
mid-2010s, leading to the notion of Arctic
Ocean “Atlantification.” It was shown that sea
ice reductions, weakening of the halocline,
and reduction in the depth of the Atlantic
water layer were together making the eastern
Eurasian Basin, north of Siberia, resemble the
western basin in terms of its Atlantification.
The consequent increased ocean heat flux
was reducing the winter sea-ice formation
rate, thus explaining the then-recent reduc-
tion in Arctic sea ice cover in the eastern ba-
sin. The hypothesis was that Atlantification
was moving part of the Arctic Ocean toward
a new climate state.
Polyakov et al. build on these observed
changes in atmospheric circulation and in ice
and ocean behavior. After decades of Arctic
sea ice decline, they show that ice area and
thickness have been stable, although variable
since 2007. The decadal-scale variability of
the atmospheric circulation explains this sea
ice behavior.
Atmospheric circulation variability is de-
scribed by using normal modes: patterns of
variability in which all parts of the system
change together. In the Arctic, the leading
mode, the Arctic Oscillation, represents the
variability of the polar high-pressure system.
The second mode, the Arctic Dipole, is as-
sociated with anticyclonic winds over North
America and cyclonic winds over Eurasia.
The Arctic Dipole was roughly neutral before
2007 and increasingly positive thereafter. In
its positive phase, the Arctic Dipole weakens
inflows to the Arctic Ocean across the Fram
Strait while strengthening inflows through
the Barents Sea, northeast of northern
Norway. The resulting stronger Arctic Ocean
circulation has been responsible for increased
freshwater accumulation in the Amerasian
Basin, which in turn has increased stratifica-
tion and reduced upward oceanic heat fluxes,
slowing Arctic sea ice loss. This atmospheric
forcing is consistent with observed and mod-
eled variability of Arctic Ocean freshwater
accumulation and export (5, 6). An eventual
return to dominance of the Arctic Oscillation
mode would reinstate the faster sea ice re-
duction. Regardless of the mode in place,
both the Arctic atmosphere and the Arctic
National Oceanography Centre, Southampton SO14 3ZH,
UK. Email: s.bacon@noc.ac.uk.
SCIENCE science.org
Ocean are predicted to continue to warm
faster than the rest of the planet through the
21st century (1,7).
There remain unexplored and plausible
ocean mechanisms that may further acceler-
ate Arctic warming, with consequences for
the Arctic ice and ocean system, and with
possible atmospheric feedbacks. When the
Arctic Ocean is completely covered by sea
ice, stresses exerted by winds are largely ab-
sorbed by the ice and ultimately transmit-
ted to land, greatly reducing the mechanical
forcing of the ocean, hence the slow circu-
lation. When the sea ice declines, more of
the ocean will be directly exposed to wind,
increasing the efficiency of atmosphere-to-
ocean momentum transfer and accelerating
the ocean circulation, called ocean “spin-up.”
Consequently, parts of the ocean, particularly
where it flows over rough topography around
the edge of the continental shelf, would be-
come more turbulent, increasing the upward
heat flux out of the Atlantic water layer (8)
and likely spinning off more eddies (circu-
lar currents), carrying that heat toward the
North Pole (see the figure). The strength of
the Coriolis force—the force on a fluid that
results from the rotation of Earth—in high
latitudes changes the balance of physical pro-
cesses. Exotic turbulence mechanisms such
as unsteady lee (standing) waves can be ex-
pected to dominate (9, 10), and they would be
less season-dependent, further threatening
Arctic sea ice survival outside summertime.
Such mechanisms are not represented in any
current forced ocean or coupled climate mod-
els and are the subject of ongoing research.
In the Arctic Ocean, Atlantic-layer heat is
presently trapped below the halocline, except
where Atlantification has been taking place.
However, ocean spin-up will increase turbu-
lence and mixing, which might release this
heat more widely and in turn may accelerate
the year-round sea ice decline. This would
bring further and unknown consequences
for both Arctic and mid-latitude weather, cli-
mate, and extreme events. j
R EFER ENC ES A ND N OT ES
1. Intergovernmental Panel on Climate Change, Climate
Change 2021: The Physical Science Basis (Cambridge
Univ. Press, 2021).
2. J. A. Screen et al., Nat. Geosci. 11, 155 (2018).
3. I. V. Polyakov et al., Science 381, 972 (2023).
4. I. V. Polyakov et al., Science 356, 285 (2017).
5. A. Proshutinsky, D. Dukhovskoy, M.-L. Timmermans, R.
Krishfield, J. L. Bamber, Philos. Trans.- Royal Soc., Math.
Phys. Eng. Sci. 373, 20140160 (2015).
6. C. Florindo-López et al., J. Clim. 33, 8849 (2020).
7. Q. Shu et al., Sci. Adv. 8, eabn9755 (2022).
8. T. P. Rippeth et al., Nat. Geosci. 8, 191 (2015).
9. T. P. Rippeth et al., Geophys. Res. Lett. 44, 12,349 (2017).
10. L. E. Baker, A. Mashayek, J. Geophys. Res. 127,
e2022JC018995 (2022).
ACK NOWL EDG M E N TS
S.B. receives funding from the UK Natural Environment Research
Council CANARI (Climate change in the Arctic–North Atlantic
region and impacts on the UK) project (NE/W004984/1).
10.1126/science.adj8469
ANTHROPOLOGY
Did our
ancestors
nearly die out?
Genetic analyses suggest an
ancient human population
crash 900,000 years ago
By Nick Ashton1 and Chris Stringer2
E
arth’s climate system began to change
during the Middle Pleistocene transi-
p
tion, which is associated with a severe
cooling phase about 900,000 years ago.
How this change might have affected
human populations is difficult to deter-
mine, because the human fossil and archaeo-
logical records are relatively sparse for this
g
period and lie beyond the reach of ancient
DNA recovery. On page 979 of this issue, Hu
et al. (1) use a new method of analysis called
FitCoal to project current human genetic
y
variation backward in time, to estimate the
size of populations at specific points in the
past. The results suggest that our ancestors
suffered a severe population bottleneck that
started around 930,000 years ago and lasted
for almost 120,000 years. This is estimated
to have reduced the number of breeding in-
dividuals to ~1300, bringing our ancestors
close to extinction.
Hu et al. argue that the proposed bottle-
y g
neck correlates with a chronological gap in
the African and Eurasian fossil records and
may have led to the evolution of a new hu-
man species, ancestral to Homo sapiens. They
favor a widely defined H. heidelbergensis as
p
this ancestral species, probably emerging in
Africa by 800,000 years ago. Two lineages of
large-brained humans have long been recog-
,
nized in the later Pleistocene: H. sapiens and
Neanderthals (H. neanderthalensis). A third
group—Denisovans—was identified more re-
cently from ancient DNA in fossils and sedi-
ments at Denisova Cave in Siberian Russia
(2). Considering how the inferred bottleneck
might have affected human evolution in-
evitably leads to debate about the nature of
the last common ancestor of H. sapiens, Ne-
anderthals, and Denisovans, and when and
where this ancestor lived.
It is not yet clear whether the last common
1
Department of Britain, Europe and Prehistory, British
Museum, London, UK. 2Centre for Human Evolution
Research, Natural History Museum, London, UK.
Email: nashton@britishmuseum.org; c.stringer@nhm.ac.uk
1 SEP TEMBER 2023 • VOL 381 ISSUE 6661 947
I NS I GHTS | P E R S P E C T I V E S

ancestor lived in Europe, Asia, or Africa. Hu
et al. attributed the East Asian fossil record
from that time to the more ancient human
species H. erectus. However, there is evidence
from sites such as Yunyang, China, of a dis-
tinct species that is morphologically closer to
later humans from Eurasia and Africa such
as H. sapiens and Neanderthals (3), suggest-
ing their ancestral lineage(s) might already
have diverged from H. erectus by the time
of the Yunyang fossils, which are dated at
800,000 to 1,100,000 years ago. Moreover,
some genetic models for the deep ancestry
of H. sapiens and Neanderthals suggest that
concepts of a single last common ancestor in
time and space might be illusory (2, 4).
Estimates using genomic data from H.
sapiens, Neanderthals, and Denisovans
calibrate the last common ances-
tor to between about 500,000
cupation in Africa, Asia, and Europe that
have been attributed to this period of re-
versed polarity and, with varying degrees of
certainty, to the inferred time of the bottle-
neck. These include sites in Kenya such as
Kilombe (GqJh1 and Gqjh2), Kariandusi
and Isinya (8), and locations in Bed IV of
Oldupai in Tanzania (9). Of particular im-
portance are the human fossils from Gom-
bore II in Ethiopia that have an estimated
age of 850,000 years before present (10),
which lies within the bottleneck window.
Sites in Europe, such as Gran Dolina and
Boella in Spain, Monte Poggiolo in Italy, and
Happisburgh 3 in the UK, have also been
attributed to this period (11). Of further im-
portance are the increasing number of sites
in China, particularly in the Qinling Moun-
tain Range, which cover the time span of
the proposed bottleneck (12). This includes
the human fossils from Yunyang (3).
These fossil records dating to the inferred
bottleneck period 813,000 to 930,000 years
ago suggest that humans were widespread
inside and outside of Africa. Therefore, what-
ever caused the proposed bottleneck may
have been limited in its effects on human
populations outside the H. sapiens lineage,
or its effects were short-lived. This also im-
plies that the cause of the bottleneck was
unlikely to have been a major environmental
event, such as severe global cooling, because
this should have had a wide-ranging impact.
Nevertheless, the provocative study of Hu et
al. brings the vulnerability of early human
populations into focus, with the implication
that our evolutionary lineage was
nearly eradicated. Recent work

and 700,000 years ago (2), which
would relate the inferred bottle-
neck to this ancestral population,
wherever it lived, with compa-
rable signals expected from Den-
isovan and Neanderthal genomes
p
Archaeological evidence during the bottleneck suggests that Europe was prob-
Evidence of human occupation, including fossilized human bones (sites in ably completely depopulated after
bold), dated to the proposed population bottleneck [813,000 to 930,000 a previously unrecognized cold
years before present (kyr B.P.)], is recorded in Asia, Europe, and Africa. phase about 1,100,000 years ago,
Gray bars represent the range of age estimates for when site occupation is before the proposed bottleneck (11).
thought to have occurred, and arrows indicate that occupation is thought If methods such as FitCoal can be

as well as from those of H. sapi-
ens. However, some recent stud-
ies of dental and cranial variation
in fossilized skulls place the last
g
to have extended beyond the date range shown [based on (3, 8–14)]. This further applied to growing data
suggests that the effects of the bottleneck were limited geographically and from the genomes of H. sapiens,
chronologically. The stone tools from Trinil are disputed. Neanderthals, Denisovans, and
hopefully others, this should clarify

y
common ancestor earlier, between 1000 950 900 850 800 750 kyr B.P. ancient bottlenecks, and which re-
about 800,000 and 1,200,000 years gions under habitation might have
Matuyama Chron
ago (5, 6), which might mean that been the most or least affected.
already separated basal lineages Jaramillo Subchron Proposed Brunhes However, the genomic data must
of Neanderthals and Denisovans population Chron also be tested against the fossil and
Human fossils found
avoided the bottleneck, or suffered at bolded sites bottleneck archaeological records of early hu-
it to a different extent. Morphologi- man populations, records that need
cal analysis of fossils suggests that Yunyang further enhancement from many
the last common ancestor might be Guanmenyan regions of the ancient world. j
H. heidelbergensis (as favored by Hu Yuelianghu
R E F E R E N C ES A N D N OT ES
Meipu Eastern

y g
et al.), H. rhodesiensis, H. antecessor, 1. W. Hu et al., Science 381, 979 (2023).
Huixinggou-Shuigou Asia
or H. bodoensis (7). 2. A. Bergström et al., Nature 590, 229 (2021).
Wutaicun 3. D. Lewis, Nature 612, 200 (2022).
The proposed bottleneck needs 4. A. P. Ragsdale et al., Nature 617, 755 (2023).
Longgangsi
to be tested against the human fos- 5. A. Gómez-Robles, Sci. Adv. 5, 1268 (2019).
Trinil 6. X. Ni et al., Innovation (Camb.) 2, 100130
sil and archaeological evidence by

p
(2021).
assessing how much of it might lie Evron Southwest 7. E. Delson, C. Stringer, Evol. Anthropol. 31,
within the designated time span Durnsunlu Asia 233 (2022).
of 930,000 to 813,000 years be- 8. S. Hoare et al., J. Archaeol. Sci. 125, 105273
(2021).
,
Gombore II
fore present. Most sites are dated 9. A. Deino et al., Palaeogeogr. Palaeoclimatol.
Garba XIIIB Palaeoecol. 571, 109990 (2021).
through magneto-stratigraphy,
Konso 10. R. Gallotti, M. Mussi, J. Anthropol. Sci. 95, 1
whereby changes in the polarity Bouri (2017).
(normal or reversed) of Earth Eastern 11. V. Margari et al., Science 381, 693 (2023).
Olorgesailie M7
can be identified, in combina- Africa 12. D. Liu et al., Palaeogeogr. Palaeoclimatol.
Kariandusi Palaeoecol. 605, 119229 (2022).
tion with isotopic dating methods Isinya 13. J. Galway-Witham, J. Cole, C. Stringer, J.
and in some cases changes in the Quat.Sci. 34, 355 (2019).
Kilombe 14. J. S. Brink et al., in African Paleoecology and
species of mammals present. The Oldupai IV Human Evolution, S. Reynolds, R. Bobe, Eds.
putative bottleneck occurred dur- (Cambridge Univ. Press, 2022), pp. 120–134.
Cornelia-Uitzoek Southern
ing a reversed phase (end of the Africa AC K N OW L E D G M E N TS
Matuyama Chron) between the Monte Poggiolo The research of the authors is supported by the
normal polarity of the Jaramillo Calleva Foundation. C.S. also acknowledges the
Cueva Negra support of the Human Origins Research Fund.
Subchron and Brunhes Chron (see la Boella Europe The authors thank J. Gowlett and C. Shipton for
the figure). Despite issues with Gran Dolina discussion of the Kenyan sites.
the resolution of dating, there are Happisburgh 3 or
sites with evidence of human oc- 10.1126.science.adj9484

948 1 SEP TEMBER 2023 • VOL 381 ISSUE 6661 science.org SCIENCE

P OLICY FORUM
CLIMATE AND ECOLOGY
Unlock the Endangered Species
Act to address GHG emissions
For the first time, ESA evaluations can include impacts on
polar bears from greenhouse gas emissions
By Steven C. Amstrup1,2 and Cecilia M. Bitz3
I
n 2008, projections that up to two-thirds
of the world’s polar bears could disap-
pear by mid-century (1) led to polar bears
becoming the first species listed under
the US Endangered Species Act (ESA)
because of threats from anthropogenic
climate warming. Updated analyses (2) cor-
roborated the 2008 projections and showed a
linear but inverse relationship between Arctic
sea ice extent and global mean temperature.
Despite the relationship between warming
and sea ice loss, absence of a quantitative
link between anthropogenic greenhouse gas
(GHG) emissions, sea ice loss, and declin-
ing polar bear vital rates has foiled full ESA
implementation for polar bears. By quantify-
ing the relationship between anthropogenic
GHG emissions and polar bear recruitment,
we show that sensitivities to cumulative
anthropogenic emissions explain observed
population trends, allow estimation of demo-
graphic impacts from new emissions sources,
and enable ESA procedures to assess global
warming impacts of proposed actions—along
with impacts on the ground.
Section 7 of the ESA provides a process by
which federal agencies ensure that actions
they take, including those they fund or au-
thorize through leasing activities or issuance
of permits (e.g., oil and gas production), do
not jeopardize the continued existence of
listed species [see supplementary materials
(SM)]. But in October 2008, then-Solicitor of
the Department of Interior David Bernhardt
issued a memorandum [M-Opinion M-37017
(see SM)] stating that Section 7 consultations
would not be required “unless it is estab-
lished that emissions from a proposed action
would cause an indirect effect to listed spe-
cies or critical habitat.” Addressing his “un-
less” requirement, Solicitor Bernhardt con-
cluded: “Any observed climate change effect
on a member of a particular listed species
1
Polar Bears International, Bozeman, MT, USA. 2Department
of Zoology and Physiology, University of Wyoming, Laramie,
WY, USA. 3Atmospheric Sciences, University of Washington,
Seattle, WA, USA. Email: samstrup@pbears.org; bitz@uw.edu
SCIENCE science.org
or its critical habitat…would be the conse-
quence of the collective GHG accumulation
from natural sources and the worldwide an-
thropogenically produced GHG emissions
since at least the beginning of the industrial
revolution” and therefore “cannot be attrib-
uted to the emissions from any particular
source.” This inability to attribute negative
consequences for polar bears to emissions
from anthropogenic GHG sources has pre-
vented the ESA from considering anthropo-
genic GHG emissions when evaluating im-
pacts on polar bears from proposed actions.
This limitation eviscerated the ESA with re-
gard to the global warming threat that justi-
fied the 2008 listing of polar bears.
POLAR BEARS, SEA ICE, EMISSIONS
Polar bears occur in 19 somewhat distinct
subpopulations within four major eco-
regions (see SM). Throughout their range,
they rely on sea ice over productive conti-
nental shelf waters to catch their prey (1–3).
When sea ice concentration falls to levels
preventing effective foraging, polar bears
are forced onto land or onto ice that has
retreated far from shore and over deep un-
productive waters. In both situations, they
are largely food deprived (4, 5). During these
forced fasting durations (FDs), polar bears
survive on accumulated fat reserves and lose
nearly a kilogram of body mass each day (6).
Anthropogenic climate warming shortens
the period during which polar bears can for-
age from sea ice to build up their fat reserves
and lengthens their fasting periods.
The fundamental dependence of polar
bears on sea ice (1–6) and the documented re-
lationship between declining sea ice and cu-
mulative anthropogenic carbon dioxide (CO2)
emissions (7) assures that polar bear distri-
bution and abundance ultimately can only
decline as cumulative emissions increase.
Yet, a quantifiable link between polar bear
demographics and sea ice attributes such
as extent, area, thickness, and volume has
previously evaded discovery. In the absence
of such a quantifiable link, previous polar
bear studies have depended on associations
between modeled future sea ice changes and
present demographic or population viability
estimates, or assumptions based on the gen-
eral relationship between demographic per-
formance and sea ice availability (8, 9).
By contrast, Molnár et al. (10) recognized
that demographic performance is determined
by energy reserves of bears at fast initiation,
their energy expenditures while fasting, and
fast duration. The polar bear’s physiological
constraints, unlike demographic estimates
derived from current population assess-
ments, will remain unchanged, even as sea
ice is changing, and provide a more robust
foundation for assessment of future impacts.
Molnár et al. (10) established two critical
features of FD: (i) fasting impact thresh-
olds—FDs beyond which the percentage of
recruitment and survival failure increase rap-
idly—and (ii) the demographic “sensitivity”
p
when FD exceeds the fasting impact thresh-
olds (determined by the regression intercept
and slope of increasing recruitment and sur-
vival failure—versus FD; see fig. S2a). Molnár
et al. (10) defined FD as 24 days shorter than
the number of ice-free days (IFD). Because
g
IFD determines FD, IFD is the critical sea ice
attribute providing the previously missing
quantifiable link between observed sea ice
decline and the polar bear’s ability to main-
y
tain physiological and reproductive health.
Molnár et al. (10) showed that survival of
cubs (recruitment into the next generation),
determined by the mother’s declining abil-
ity to provide enough milk for a cub’s early
survival, is the first demographic threshold
crossed as FD increases (see SM). Because
populations that are not successfully recruit-
ing young can only decline, and because the
recruitment impact threshold is the most
y g
sensitive to FD, we focused on the sensitivity
of cub recruitment to emissions (see SM).
Armed with new knowledge of FD impacts
on polar bear recruitment (10), here we link
FD to cumulative anthropogenic GHG emis-
p
sions. This reveals a causal connection be-
tween emissions from proposed actions and
impacts on polar bears and helps explain re-
,
cent population trends. Further, by quantify-
ing polar bear impacts from anthropogenic
GHG emissions rather than atmospheric con-
centrations, we disproved claims in M-37017
that consequences from current anthropo-
genic GHG emissions cannot be separated
from natural sources or past accumulation of
anthropogenic emissions.
Because climate warming and its associ-
ated sea ice loss result from all anthropogenic
GHG emissions, not just CO2, we considered
CO2 equivalent (CO2-eq) emissions—the com-
bination of GHG pollutants that contribute to
climate warming—adjusted for their global
warming potential. Considering CO2-eq also
is important because emissions and impacts
1 SEP TEMBER 2023 • VOL 381 ISSUE 6661 949
I NS I GHTS | P O L I C Y F O RU M

of some GHGs can be more quickly and easily
mitigated than CO2.
We set FD to 24 days less than the number
of IFD in the Seasonal Ice Ecoregion and equal
to IFD elsewhere (see SM). As in Molnár et
al. (10), we defined crossing the recruitment
impact threshold as the first time FD exceeds
117 days for at least three of five consecutive
years (10). This decision rule recognizes that
FD is increasing on average with increasing
cumulative emissions, that bears cannot re-
cover from clusters of years above the impact
threshold, and that future FD will increase on
average as anthropogenic emissions increase.
Detailed methods, including uncertainty esti-
mates, are described in the SM.
Sensitivity of FD and recruitment failure to cumulative CO2-eq emissions
Sensitivity to cumulative CO2-eq emissions of FD (g1) and recruitment failure (b1g1) for all polar bear subpopulations that have
sea ice showed the most rapid change with
increasing emissions (steepest FD/CO2-eq
slopes). In the Chukchi Sea, for example, FD
increased from ~12 days when satellite imag-
ery of sea ice first became available in 1979
(see SM) to ~137 days in 2020 (table S2), and
another day of fasting was added for each 14
(10, 18) Gt CO2-eq (hereafter 95% confidence
intervals appear in parentheses after cen-
tral estimates) released into the atmosphere
(see the table). Contrary to what might have
been expected, the Seasonal Ice Ecoregion,
where sea ice always has melted entirely dur-
ing summer, saw the slowest rate of FD in-
crease. The most gradual slope was observed
in southern Hudson Bay, where FD increased
The sensitivity (regression slope) of recruit-
ment to cumulative emissions ranges from
0.11 (0.084, 0.15) % Gt-1 in the Chukchi Sea
to 0.016 (0.0, 035) % Gt-1 in southern Hudson
Bay (see the table). These sensitivities, when
multiplied by the emissions accumulated
since each subpopulation region crossed the
117-day recruitment impact threshold, sug-
gest recruitment impacts that are consistent
with independent estimates of population
status and trend.
For example, in western Hudson Bay
(WHB), the best-known of all polar bear sub-
populations, the 117-day recruitment impact
threshold was reached in 1994 (table S1). The
1193 Gt of CO2-eq emitted from 1994 to 2020
and the recruitment failure
rate of 0.025 (0.0072, 0.046)
% Gt–1 emissions (see the ta-
ble) would mean an ~30 (9,

p
experienced at least 10 years with ice-free seasons from 1979 to 2020. 55) % decline in recruitment
since crossing the thresh-
OCCUPIED AREA FD/CO2-eq (d Gt–1) CO2-eq/FD AREA WEIGHTED FD/CO2-eq RECRUITMENT FAILURE/ old. This percentage decline
SUBPOPULATION (km2) A g1 (Gt d–1) 1/g1 (d km2 Gt–1) Ag1 CO2-eq (% Gt-1) b1g1 would mean that the annual
South Beaufort Sea 135,000 0.043 (0.024, 0.065) 23 (15, 41) 5800 (3300, 8800) 0.065 (0.037, 0.099) cub survival rate in WHB,
which was estimated as
Chukchi Sea 1,280,000 0.073 (0.056, 0.096) 14 (10, 18) 93,000 (71,000, 120,000) 0.11 (0.084, 0.15)

g
~70% during the 1980s (11),
Laptev Sea 1,610,000 0.048 (0.036, 0.065) 21 (16, 28) 77,000 (57,000, 100,000) 0.073 (0.054, 0.098) would now be ~49%. Such
Kara Sea 980,000 0.067 (0.048, 0.091) 15 (11, 21) 65,000 (47,000, 89,000) 0.010 (0.073, 0.14) a decline in annual cub sur-
Barents Sea 786,000 0.064 (0.035, 0.098) 16 (10, 29) 51,000 (27,000, 77,000) 0.098 (0.052, 0.15)
vival is consistent with the

y
low proportion of yearlings
East Greenland 617,000 0.023 (0.009, 0.098) 43 (25, 113) 14,000 (5,400, 24,000) 0.035 (0.013, 0.060) observed in recent years and
Kane Basin 48,200 0.029 (0.016, 0.044) 35 (23, 62) 1,400 (770, 2100) 0.043 (0.024, 0.067) population estimates sug-
Lancaster Sound 229,000 0.12 (0.002, 0.025) 84 (39, 360) 2,700 (550, 5700) 0.018 (0.0035, 0.038)
gesting an ~30% decline in
numbers between 1987 and
Baffin Bay 616,000 0.027 (0.016, 0.040) 38 (25, 63) 16,000 (9,800, 25,000) 0.040 (0.024, 0.061) 2016 (12). The most recent
McClintock Channel 133,000 0.018 (0.006, 0.033) 54 (30, 159) 2,500 (830, 4400) 0.028 (0.0094, 0.051) (2021) estimate, however,
Gulf of Boothia 61,000 0.037 (0.023, 0.052) 27 (19, 43) 2,300 (1,400, 3,200) 0.056 (0.035, 0.079) suggests that the WHB pop-
ulation may have declined
Foxe Basin 502,000 0.023 (0.011, 0.037) 44 (27, 88) 12,000 (5700, 19,000) 0.034 (0.017, 0.056) by nearly half since the late

y g
West Hudson Bay 176,000 0.017 (0.005, 0.031) 60 (32, 199) 2,900 (840, 5400) 0.025 (0.0072, 0.046) 1980s (11). The recruitment
South Hudson Bay 399,000 0.010 (0.000, 0.023) 93 (44, inf) 4,100 (0, 9100) 0.016 (0.00, 0.035) sensitivity to emissions in
WHB, therefore, may be
Davis Strait 554,000 0.024 (0.009, 0.042) 44 (24, 115) 13,000 (4,800, 23,000) 0.036 (0.013, 0.064)
closer to the upper end of
Total 8,120,000 366,000 (319,000, 420,000) our confidence interval than

OUTCOMES
Emissions and fasting duration
Molnár et al. (10) found an essentially linear
relationship between FD and recruitment
impact (10). We found that the relationship
between FD and cumulative CO2-eq emis-
sions, for subdomains where polar bears oc-
cur, is also largely linear—by estimating the
regression relationship between FD and each
gigatonne (Gt) of cumulative emissions (see
the figure, table, figs. S1 and S2, and SM). For
ease of interpreting impact on polar bears, we
took the reciprocals of regression line slopes
to reveal the emitted amount of CO2-eq that
prolongs FD by one additional day (see the ta-
ble). Emission levels that add 1 day to FD var-
ied greatly among polar bear subpopulations.
FD in areas historically covered by perennial
from ~139 days to ~157 days since 1979, and
93 Gt CO2-eq were emitted for each day added
to FD (see the figure, table S2, and SM).
Emissions and demographic impact
The relationship between FD and cumulative
CO2-eq emissions can be combined with the
relationship between declining recruitment
and FD [from Molnár et al. (10)] to calculate
the rate at which polar bear recruitment has
declined with cumulative emissions. The re-
lationship between declining recruitment
(increasing percentage recruitment failure)
and cumulative emissions not only explains
observed recruitment trends but also allows
estimating the recruitment response to emis-
sions from proposed future actions (see SM
methods and fig. S2).
p
to the middle, meaning a
current annual cub survival rate of ~32%.
Alternatively, a larger numerical decline may
,
mean that WHB crossed the recruitment im-
pact threshold earlier. Molnár et al. (10) esti-
mated that the effect of not including the 24-
day adjustment to the onset of fasting would
mean threshold crossings 10 to 30 years
earlier than with the adjustment. Indeed, if
WHB bears begin to fast as soon as sea ice
extent drops below 30%, they would have
exceeded the 117-day recruitment impact
threshold before 1979, suggesting a 44 (13,
81) % decline in recruitment and a current
annual cub survival rate of 40%. Crossing
the recruitment impact threshold earlier is
consistent with reduced body condition in
females, disappearance of early weaning of
cubs, and reduced cub recruitment reported

950 1 SEP TEMBER 2023 • VOL 381 ISSUE 6661 science.org SCIENCE
 Linking CO2-eq to polar bear fasting and recruitment failure
Shown are annual observations (dots) and regressions (solid lines) by subpopulation. Dashed lines are mean 95%
confidence intervals (CIs) for sampling plus observational uncertainty. Shading illustrates sampling uncertainty
only for comparison. See table for regression slopes and 95% CIs, and supplementary materials for details.
S. Beaufort Sea
150
Fast duration (days)
100
50
0 0
0 500 1000 1500
Recruitment failure (%)
60
0 25
-60
-120
0 500 1000 1500
before 1994 (13). With or without the 24-day
offset in FD, the recruitment declines that we
estimate corroborate recent observations of
poor cub survival and major population de-
cline in WHB. See also the table and figure,
fig. S3, and SM.
As when evaluating surface disturbances
or toxic chemical releases, impacts of GHGs
emitted from individual actions may appear
small. This emphasizes the importance of as-
sessing cumulative impacts at the program-
matic rather than individual action level. For
example, emissions from each of the many
new actions anticipated to occur on US pub-
lic lands between 2020 and 2050 may appear
to have minimal impact on polar bears in iso-
lation. Cumulatively, however, their projected
24.1 Gt of CO2-eq emissions (see SM) will
decrease polar bear cub recruitment by 0.6
(0.17, 1.1) % in WHB and by 2.7 (2.0, 3.6) % in
the Chukchi Sea. Similarly, although each of
the hundreds of power plants in the US make
a relatively small contribution to emissions,
together their nearly 2 Gt annual CO2 emis-
sions (see SM) and 60+ Gt emissions over
30+ year life spans reduce recruitment in the
southern Beaufort Sea by ~4% (see SM).
Interpreting demographic impacts of in-
dividual actions can be likewise confounded
by propagation of uncertainties through
GRAPHIC: D. AN-PHAM/SCIENCE
our estimation process—resulting in wide
interval estimates. We therefore cannot
overemphasize the need to refine emissions
estimates at all levels of reporting, as well
as to fill the gap in measurements of body
masses at which polar bears in most regions
initiate summer fasts (10).
SCIENCE science.org
Chukchi Sea W. Hudson Bay S. Hudson Bay
150 150 150
100 100 100
50 50 50
0 0
0 500 1000 1500 0 500 1000 1500 0 500 1000 1500
100 50 50
25
0
0 sphere or on the ground. In addition to broad-
-100 0
-25
-25
0 500 1000 1500 0 500 1000 1500 0 500 1000 1500
Cumulative emissions since 1979 (gigatonnes of CO2-eq)
OTHER SPECIES AND SYSTEMS
We focus on polar bears because the Bern-
hardt memo (M-37017) was prompted by the
ESA listing of polar bears and because we have
the energetics data (10) to make the quantita-
tive link between emissions, sea ice, and po-
lar bear demographics. But ramifications of
our findings go far beyond polar bears and
sea ice. In presenting this polar bear case his-
tory, our outcomes show by example how the
impact of emissions can be determined and
included in Section 7 consultations for other
species and habitats threatened by climate
warming. For example, although we person-
ally lack data, it is logical to believe that links
between GHG emissions, sea level rise, and
nesting habitat for marine turtles and beach
nesting birds could be established. Also, it
seems reasonable that similar regression
links could be established between emissions,
water temperature change, and impacts on
both freshwater and marine species of plants
and animals. We are confident that our polar
bear example will lead other investigators to
uncover parallels in their data, providing nu-
merous other species with previously unavail-
able ESA protections from climate warming.
It is also reasonable that GHG emissions
could be similarly linked to other phenom-
ena—beyond the purview of the ESA.
We have confirmed the direct link be-
tween CO2-eq emissions and polar bear de-
mography. In addition to aiding interpre-
tation of observed population trends, this
allows estimation of impacts from proposed
GHG-emitting actions. Because the ESA re-
quires following the best available science,
the Department of the Interior now has the
scientific justification and duty to rescind
M-Opinion M-37017—allowing the US Fish
and Wildlife Service to include anthropo-
genic GHG emissions in Section 7 consulta-
tions and meet its responsibility under the
ESA to protect polar bears as well as other
species imperiled by climate warming. This
will provide the government an important
and previously unavailable tool for address-
ing anthropogenic climate warming.
Small impacts of most individual actions
emphasize that GHG emissions are the epito-
mic cumulative impacts issue—requiring mit-
igation at programmatic, regional, national,
and international levels. We must remember,
however, that the growth in cumulative GHG
emissions by over 1750 Gt from 1979 to today
happened largely one action at a time with-
out regard for cumulative effects in the atmo-
p
scale efforts, if future impact assessments
require separate evaluations, decisions must
be based on merit, contribution to cumulative
impact, and the best available science.
Rescinding M-37017 goes beyond con-
g
sideration of polar bears and sea ice—ad-
hering to President Biden’s mandate for a
“Government-wide approach that reduces
climate pollution in every sector of the econ-
y
omy” (SM). The role of the Arctic sea ice sys-
tem in moderating global climate means that
protecting polar bears from anthropogenic
climate warming also will benefit the rest of
life on Earth—including humans. j
RE FE REN CES A ND N OT ES
1. S. C. Amstrup, B. G. Marcot, D. C. Douglas, “A Bayesian
network modeling approach to forecasting the 21st
century worldwide status of polar bears,” in Arctic Sea
Ice Decline: Observations, Projections, Mechanisms, and
Implications, E. T. DeWeaver, C. M. Bitz, L.-B. Tremblay,
Eds. American Geophysical Union, Geophys. Monogr.
y g
Ser. (2008), vol. 180, pp. 213–268.
2. S. C. Amstrup et al., Nature 468, 955 (2010).
3. G. M. Durner et al., Ecol. Monogr. 79, 25 (2009).
4. J. P. Whiteman et al., Science 349, 295 (2015).
5. J. P. Whiteman et al., Oecologia 186, 369 (2018).
p
6. N. W. Pilfold et al., Physiol. Biochem. Zool. 89, 377 (2016).
7. D. Notz, J. Stroeve, Science 354, 747 (2016).
8. C. M. Hunter et al., Ecology 91, 2883 (2010).
9. E. V. Regehr et al., Biol. Lett. 12, 20160556 (2016).
10. P. K. Molnár et al., Nat. Clim. Chang. 10, 732 (2020).
,
11. E. V. Regehr et al., J. Wildl. Manage. 71, 2673 (2007).
12. S. N. Atkinson et al., Aerial survey of the Western Hudson
Bay polar bear subpopulation 2021. Final Report.
Government of Nunavut, Department of Environment,
Wildlife Research Section, Status Report 2022, Igloolik,
Nunavut, Canada (2022).
13. I. Stirling, A. E. Derocher, Glob. Change Biol. 18, 2694 (2012).
14. C. Bitz, S. C. Amstrup, cmbitz/PolarBearSurvival_vs_
GHGEmissions, Zenodo (2023); https://doi.org/
10.5281/zenodo.8263597.
AC KN OW LED G M E N TS
S.C.A. received funding from Polar Bears International. C.M.B.
received funding from National Science Foundation grant
OPP-2237964. The authors thank T. McDonald for advice on
statistical procedures, and P. Molnár for providing data for
recruitment failure as a function of fasting duration (10). Data
and code for replication are available at Zenodo (14).
SU PP L EM E N TARY M AT E RIA LS
science.org/doi/10.1126/science.adh2280
10.1126/science.adh2280
1 SEP TEMBER 2023 • VOL 381 ISSUE 6661 951

B O OKS et al .
HISTORY OF MEDICINE
Saving mothers with a simple act
A compelling new play revisits the discovery that
drastically reduced maternal mortality
By Stuart Firestein
“S
mell!” shouts the strange, diminu-
tive man emphatically from the edge
of the stage at the beginning of the
new play Dr Semmelweis. He repeats
it several more times—to himself, to
the audience, to the other characters,
to anyone who will listen—as he moves up-
stage. This mystifying pronouncement is a
deliberate foreshadowing of the events that
are about to unfold, as the audience jour-
neys alongside the man—soon identified
as Ignaz Semmelweis, a 19th-century Hun-
garian physician whose simple discovery
would eventually change medicine forever.
Co-written by Stephen Brown and the
esteemed British actor Mark Rylance, the
play—which also stars Rylance in the title
role—is, among its many other values, an
excellent play about science. It is to Semmel-
weis, after all, that we owe the beginnings
of germ theory and the once-radical idea of
washing our hands.
Semmelweis speaks with just the hint of
a stutter and an Eastern European accent.
Despite his position as assistant to the di-
rector of the Vienna General Hospital, he is
an outsider and suffers daily frustrations,
small and large, as a result.
The reviewer is at the Department of Biological Sciences,
Columbia University, New York, NY 10027, USA; is a member
of the Fractal Faculty, Santa Fe Institute, Santa Fe, NM 87501,
USA; and is a former professional theater director.
Email: sjf24@columbia.edu
952 1 SEP TEMBER 2023 • VOL 381 ISSUE 6661
Semmelweis is obsessed with find-
ing possible causes of “childbed fever,” a
deadly infection that plagues an alarming
number of postpartum women who give
birth in hospital obstetric wards. He dis-
misses variables one by one, but one curi-
ous fact remains: Maternal death rates in
doctor-run obstetric wards are three times
that of patients whose births are adminis-
tered by midwives. The doctors, however,
find excuse after excuse for this anomaly.
A janitor who has been clean-
ing up after autopsies in the
background throughout the en- Dr Semmelweis
tire play offers a hint. “I’ll mop Stephen Brown with
up and kill the stink,” he tells Mark Rylance
Harold Pinter Theatre,
Semmelweis. The smell! “How London, UK,
do you kill the smell?” the physi- through 7 October 2023
cian inquires. The janitor points
to his bucket of chlorine bleach. Semmelweis
plunges his hands into the bucket, the first
act of hygienic handwashing, as the first half
of the play comes to a close. An audible sigh
fills the darkened theater, conveying the au-
dience’s palpable relief for the thousands of
women who will subsequently be saved by
this simple action.
During the intermission, I overheard
a fellow patron remark, “But how could
they not have been washing their hands?
Doesn’t everyone just know that?” How
easily we forget that what seems like com-
mon sense is often the result of long and
painstaking scientific work.
The difference between midwives and doc-
Physician Ignaz Semmelweis (left; played by Mark
Rylance) makes his important discovery.
tors, the audience learns as the second act
opens, is that only the doctors perform au-
topsies, and they typically do so in the morn-
ings before going to the maternity ward to
deliver babies. Washing one’s hands between
doing autopsies and delivering babies proves
transformative: The incidence of childbed fe-
ver is miraculously reduced from as high as
10% of patients to less than 1%.
Compelling as the data are, the medical
establishment is not convinced. Physicians
scoff at the idea that such a simple practice
could make a difference, and they resent
the implication that they are in some way
contaminated.
Madness descends on Semmelweis as
frictions between the doctor and his col-
p
leagues come to the surface. He is overcome
by thoughts of the women who have died
and the women who will soon be dead if
his revelation is not embraced immediately.
He calls eminent surgeons “murderers” and
“assassins.” Hospitals are “slaughterhouses.”
g
Everyone who fails to recognize the impor-
tance of his discovery is “stupid.” Semmel-
weis’s outbursts and impatience make him
easy to dismiss. Remember, the microorgan-
y
isms responsible for the infections, which
he calls “cadaveric particles,” have not yet
been discovered. Is it ego or urgency that
drives Semmelweis? Rylance and the script
blur the distinction.
Semmelweis is ultimately committed to
an insane asylum, driven there by an over-
whelming sense of the tragedy of a woman
giving birth and dying soon after. Ironically,
he dies during his confinement from sepsis
y g
due to an unwashed wound.
This is no simple historical
retelling; all of the techniques of
the theater are put to use. There
is music—four violins and a
p
cello—that drives the action for-
ward. There are dancers, mostly
women, who embody the victims
,
of childbed fever and heighten the show’s
sense of urgency. At two and a half hours, the
play is substantial, but Rylance’s Semmelweis
is riveting throughout—charming and funny
one moment, obsessive and abusive the next.
Semmelweis is a tragic character—a brave
scientist who embraced data above all else
and the mostly forgotten pioneer who first
made the connection between microbes
and disease. Yet his story is ultimately a tri-
umph. His unyielding vision led to a simple
solution that has reduced maternal mortal-
ity worldwide. His life, and this play, are a
scientific story worth celebrating. j
10.1126/science.adj8740
science.org SCIENCE
 I N SI G H T S

ECOLOGY Crossings: How Road

Ecology Is Shaping the

Coexistence at the crossroads Future of Our Planet

Ben Goldfarb
Norton, 2023. 384 pp.

Ideology meets infrastructure for road ecologists helping

flora and fauna live well in the presence of motorways
By Sarah Boon activity such as birdsong. Roads can lead tioned them from white ones,” he writes.
to habitat fragmentation, which isolates Meanwhile, road noise and pollutants dis-

W
hen we think of “road ecology,” most
of us think of wildlife crossings—
the overpasses, underpasses, and
fences, designed by engineers and
biologists to connect fragmented
habitats, into which animals are
funneled. But road ecology, a relatively new
field of science, is about so much more than
just wildlife crossings.
individual animals from the rest of their
population. Contaminants released from
tires and exhaust also affect the near-road
environment. Meanwhile, remote wilder-
ness roads enable hunting access, and the
increased contact between humans and
animals when new roads are built into un-
tracked wilderness can enhance the spread
of zoonotic diseases.
proportionately affect low-income commun-
ities and racial minorities.
But not everything about roads is bad.
Goldfarb finds that road verges—vegetation
next to roads—can help preserve endan-
gered habitats, such as prairies. Verges can
also provide insect habitats and are less of a
barrier, as they provide habitat that is paral-
lel, rather than perpendicular, to roads. This

p
In Crossings: How Road Ecology Is “For all of human history, roads have been reduces the likelihood of insects being killed
Shaping the Future of Our Planet, Ben instruments of conquest,” writes Goldfarb, by vehicle traffic.
Goldfarb explores the science of road ecol-
ogy and its impacts on our world. “It’s really
interesting to meet someone in college who
says, ‘I’m a road ecologist,’” Washington State

g
biologist Paul Wagner tells Goldfarb. “[I]t’s
like, huh, I remember when we made that
term up [in the 1990s].”
Since that time, the field of road ecology

y
has expanded exponentially, as evidenced
by the sheer number of experts Goldfarb
interviews. The scientists, engineers, policy-
makers, and First Nations representatives
he talks to are working to help all manner of
flora and fauna—from mammals, to reptiles,
to fish, to plants, to insects, and even hu-
mans—better coexist with roads. Goldfarb’s
many field excursions with these experts help
readers experience road ecology firsthand, re-

y g
vealing how it started and where it is headed.
Goldfarb contends that roads are more
than mere infrastructure, they are an ideol-
ogy that promotes sprawl, champions car
traffic, and—when they are built in the

backcountry—encourages human–animal
interactions. “Like abortion or critical race
theory, roads [have] become shibboleths,
entrenched cultural markers that [distin-
guish] warring factions,” he writes. “The
practice of road ecology is not merely a set
of engineering principles but a moral man-
date.” Road ecologists and everyday people
who rehabilitate injured wildlife found
on roads and shepherd migrating animals
across roads answer this mandate.
Roads are problematic in many ways:
The number of cars they can support and
the driving speeds they enable lead to wild-
life deaths, and the noise pollution that
PHOTO: DAVID STEEN
A box turtle undertakes a perilous road crossing in the state of Georgia.
but the Confederated Salish and Kootenai
Tribes believe that “the road is a visitor” that
should “respond to and be respectful of the
land and the Spirit of Place.” Culverts that
prevent salmon from spawning upstream
affect Indigenous people for whom salmon
is part of their creation story. New roads
built in the Amazon connect “undiscovered”
Indigenous groups with the outside world,
often without their consent.
Many roads, notes Goldfarb, were osten-
sibly built to combat “urban blight” in areas
p
,
Roads are growing faster than any other
human infrastructure, and although road
ecologists cannot stop road building, they
can push to be involved in road design. Some
strategies that would help mitigate the dam-
aging effects of roads include designing some
to be driven on more slowly, closing some at
night, and incorporating wildlife overpasses
and underpasses as needed.
Crossings is a deeply researched and com-
pelling read, enlivened with interviews and
on-the-ground fieldwork. The book offers

often accompanies them changes animal that were, in reality, vibrant minority com- readers a look behind the scenes of a rich but
munities. “Freeways were the erasers with underappreciated field of study that has the
The reviewer is a freelance writer and editor from Vancouver which cities rubbed out Black communi- potential to affect our everyday lives. j
Island, Canada. Email: snowhydro1@gmail.com ties and the walls with which they parti- 10.1126/science.adi6285

SCIENCE science.org 1 SEP TEMBER 2023 • VOL 381 ISSUE 6661 953
 LET TERS
When China’s first solar panel installations reach the end of their life spans, the country will need a plan to sustainably process the decommissioned equipment.
Edited by Jennifer Sills
Retraction
On 9 March 2001, Science published the
Report “Binding of DCC by netrin-1 to medi-
ate axon guidance independent of adenosine
A2B receptor activation” (1). In 2015, Science
agreed to the publication of an Erratum to
address issues that had come to our atten-
tion, specifically: in Figures 1 and 3, beauti-
fication of some Western blots outside the
data-containing portions (which was not for-
mally disallowed at the time of publication)
and some image breaks caused by format-
ting (“tiling”) issues, for which we provided
corrected images. However, because of an
error on the part of the journal, the Erratum
was not posted. In 2022, when new concerns
were raised, Science posted an Editorial
Expression of Concern (2) while an insti-
tutional investigation was conducted. The
investigation is complete and has identified
further issues, including manipulation of
data-containing portions of Western blot
images in Figures 1, A and C, and 3, B and
D, undermining confidence in the paper’s
conclusions (3). As a result, we are retract-
ing the paper. We regret the impact of these
issues on the scientific community. Author
Elke Stein disagrees with the decision to
retract the paper.
Yimin Zou1, Mu-Ming Poo2, Marc Tessier-Lavigne3*
1
University of California San Diego, San Diego, CA,
USA. 2Institute of Neuroscience, Chinese Academy
of Sciences, Shanghai, China. 3Stanford University,
Stanford, CA, USA.
*Corresponding author.
Email: tessier3@stanford.edu
954 1 SEP TEMBER 2023 • VOL 381 ISSUE 6661
R EFER ENC ES AN D N OT ES
1. E. Stein, Y. Zou, M.-M. Poo, M. Tessier-Lavigne, Science
291, 1976 (2001).
2. H. H. Thorp, Science 378, 1284 (2022).
3. H. Cline, K. Dzirasa, S. E. Hyman, R. Schekman, S.
M. Tilghman, “Report of the Scientific Panel of the
Special Committee of the Stanford University Board of
Trustees” (2023); https://boardoftrustees.stanford.
edu/wp-content/uploads/sites/5/2023/07/Scientific-
Panel-Final-Report.pdf.
10.1126/science.adk1521
Retraction
On 8 February 2001, Science published the
Research Article “Hierarchical organiza-
tion of guidance receptors: Silencing of
netrin attraction by Slit through a Robo/
DCC receptor complex” (1). In 2015,
Science agreed to the publication of an
Erratum to address issues that had come
to the authors’ attention, specifically: in
Figures 4, 5, and 6, beautification of some
Western blots outside the data-containing
portions (which was not formally disal-
lowed at the time of publication) and some
image breaks caused by formatting (“til-
ing”) issues, for which I provided corrected
images, and panel duplications in Figures
2D (micrographs) and 4B (blank Western
blots), and an incorrect blank panel in
Figure 5E. However, because of an error on
the part of the journal, the Erratum was
not posted. In 2022, when new concerns
were raised, Science posted an Editorial
Expression of Concern (2) while an insti-
tutional investigation was conducted. The
investigation is complete and has identified
further issues, including manipulation of
data-containing portions of Western blot
p
g
images in Figure 4E and an additional panel
duplication in Figure 5C, undermining con-
fidence in the paper’s conclusions (3). As a
result, I am retracting the paper. I regret the
y
impact of these issues on the scientific com-
munity. Author Elke Stein disagrees with
the decision to retract the paper.
Marc Tessier-Lavigne
Stanford University, Stanford, CA, USA.
Email: tessier3@stanford.edu
RE FE REN C ES AN D N OT ES
1. E. Stein, M. Tessier-Lavigne, Science 291, 1928 (2001).
2. H. H. Thorp, Science 378, 1284 (2022).
3. H. Cline, K. Dzirasa, S. E. Hyman, R. Schekman, S.
M. Tilghman, “Report of the Scientific Panel of the
y g
Special Committee of the Stanford University Board of
Trustees” (2023); https://boardoftrustees.stanford.
edu/wp-content/uploads/sites/5/2023/07/Scientific-
Panel-Final-Report.pdf.
10.1126/science.adk1517
p
Green decommissioning
for renewables in China ,
Carbon neutrality depends on the devel-
opment of renewables such as wind and
solar energy (1, 2). China is implementing
ambitious renewable energy plans, with
the goal of exceeding 1.2 billion kilowatts
of total installed capacity of wind and
solar energy by 2030 (3). However, wind
and photovoltaic equipment have limited
life spans. Wind turbines and photovoltaic
modules installed in China around 2000
will soon reach the end of their life of 20
to 25 years. At that time, China will need
to determine an environmentally respon-
sible way to decommission them (4).
science.org SCIENCE

By 2040, the cumulative scale of decom-
missioned wind and solar modules in
China is estimated to reach 280 and 250
GW, respectively (4). The decommissioning
of wind and photovoltaic equipment will
generate a large amount of solid waste,
including blades, glass, and semiconductor
materials (5). By 2028, China’s decommis-
sioned wind turbines will generate approx-
imately 74,000 tons of solid waste (6),
and by 2050, the discarded photovoltaic
modules in China will reach 20 million
tons (7). Disposing of this solid waste in
improper ways, such as burial or incinera-
tion, will waste resources and pollute the
soil, groundwater, and atmosphere (4, 8).
China plans to prevent waste and pol-
lution by recycling decommissioned wind
turbines and photovoltaic modules (3,
9), but more comprehensive action is
required. An integrated plan should clarify
decommissioning programs and actions
for wind and photovoltaic equipment at
both national and local scales. The plan
must involve all relevant parties, including
industry, academic institutions, investment
institutions, and local stakeholders. China
should also establish the circular economy
that will be required to recover and reuse
decommissioned wind turbines and photo-
voltaic modules (10) as well as the techni-
cal specifications, policies, and regulations
that will be required (11). As China pre-
pares to manage decommissioned equip-
ment, the government should lead invest-
ments in the research and development of
new energy materials and extend the life
cycle of wind turbines and photovoltaic
modules through technological innovation
and green retrofitting.
China’s wind and photovoltaic decommis-
sioning will be concentrated in ecologically
fragile areas (11). In addition to clarifying
the logistics of retiring equipment, the coun-
try should prioritize in-depth assessments
of the impacts of decommissioned wind
and photovoltaic equipment on ecosystems.
If the environment suffers adverse effects,
China should adopt management practices
such as ecological restoration to minimize
them. China can maximize the benefits of
wind and solar energy by preparing green
and sustainable decommissioning plans.
Siyou Xia1,2, Yu Yang1,2*, Yi Liu1,2, Jessie Poon3
1
Institute of Geographic Science and Natural
Resources Research, Chinese Academy of
Sciences, Beijing, China. 2College of Resources
and Environment, University of Chinese Academy
of Sciences, Beijing, China. 3Department of
Geography, University at Buffalo–State University
of New York, Buffalo, NY, USA.
*Corresponding author.
Email: yangyu@igsnrr.ac.cn
RE F ER E NC ES AND NOTES
1. H. Duan et al., Science 372, 378 (2021).
2. G. Lin, Y. Zhao, J. Fu, D. Jiang, Science 380, 6646 (2022).
SCIENCE science.org
3. The State Council of the People’s Republic of
China, “Action plan for carbon dioxide peaking
before 2030” (2021); http://english.www.gov.
cn/policies/latestreleases/202110/27/content_
WS6178a47ec6d0df57f98e3dfb.html.
4. PowerLab, “Zero-waste future of renewable energy: A
study on the development of recycling industry for wind
and solar energy” (2022); https://powerlab-recycling.
greenpeace-carbon-tracker.com/.
5. N. Majewski, J. Florin, R. Jit, R. Stewart, Renew. Sustain.
Energ. Rev. 164, 112538 (2022).
6. H. Qin, Wind Energ. 3, 1 (2022) [in Chinese].
7. International Renewable Energy Agency, International
Energy Agency–Photovoltaic Power Systems Program,
“End-of-life management: Solar photovoltaic panels”
(2016); https://www.irena.org/-/media/Files/IRENA/
Agency/Publication/2016/IRENA_IEAPVPS_End-of-
Life_Solar_PV_Panels_2016.pdf?rev=49a75178e38c46
288a18753346fb0b09.
8. J. Li, J. Shao, X. Yao, J. Li, Resour. Conserv. Recycling 196,
107027 (2023).
9. Ministry of Industry and Information Technology of the
People’s Republic of China, “Implementation plan on
accelerating the comprehensive utilization of industrial
resources” (2022); https://ythxxfb.miit.gov.cn/
ythzxfwpt/hlwmh/tzgg/sbyw/dzxxjsbzhfwpt/
art/2022/art_ece63eff27d3477085ea06a933523bd9.
html [in Chinese].
10. İ. M. Eligüzel, E. Özceylan, J. Clean. Prod. 380, 135004
(2022).
11. S. Li, Guizhou Today 16, 78 (2023) [in Chinese].
10.1126/science.adj5769
Strengthen management
of offshore aquaculture
China is accelerating toward deeper
offshore aquaculture (1) in an effort to
diversify the food supply system to ensure
domestic food security in the face of
unstable international markets. However,
offshore aquaculture presents risks that
differ from those in nearshore fisheries (2,
3). The government and the aquaculture
industry in China need to identify and
address the unique challenges of offshore
farms before expanding into deep water.
As of 2022, China had 20,744 km2 of
marine aquaculture cover (4). Nearshore
aquaculture, a mature and well-regulated
industry (5), takes place in enclosed or
semi-enclosed waters that are less than 20
m deep and uses cages, rafts, and bottom-
seeding to produce the desired products.
In 2022, nearshore aquaculture produced
more than 22 million metric tons of fish,
shellfish, and aquatic plants (4). However,
nearshore sites are fast approaching the
farming capacity limit (6) as they com-
pete for space with recreational fishing
and tourism and suffer from pollution
(5). Offshore farms have been proposed
because they have better water quality and
larger farming capacity (7).
Offshore aquaculture takes place in
open water deeper than 20 m and is domi-
nated by cage culture. In 2022, offshore
aquaculture production totaled just 0.4
million tons (4). Still in its infancy, the
I N SI G H T S
offshore industry lacks efficient produc-
tion chains and market mechanisms (7).
Offshore farms require sophisticated
infrastructure and equipment (2) and are
more vulnerable than nearshore sites to
natural disasters such as typhoons (3).
Many nearshore sites produce introduced
nonnative and migratory native species
(8, 9), but farming such species in
open water could pose risks that are
not well understood.
In June, China developed its first
guiding opinion on the development
of deeper offshore aquaculture (10),
but current policies focus on nearshore
activities and may be insufficient for
offshore farms. Before expanding off-
shore aquaculture, China should refine
regulations for managing ecological and
environmental protection, financial credit,
p
policy-based insurance, and the selec-
tion of aquaculture species. Additional
research should be funded to establish
environmentally safe offshore infrastruc-
ture (3) and production procedures spe-
cific to the offshore industry (10). National
g
and regional communication and coopera-
tion should promote the integrated devel-
opment of the offshore production chain
by sharing scientific, technological, and
y
human resources.
Daomin Peng1,2, Yugui Zhu3,4, Jiansong Chu1,4*
1
College of Marine Life Sciences, Ocean University
of China, Qingdao 266003, China. 2Institute for
the Oceans and Fisheries, University of British
Columbia, Vancouver, BC V6T 1Z4, Canada. 3Key
Laboratory of Mariculture (Ministry of Education),
College of Fisheries, Ocean University of China,
Qingdao 266003, China. 4Southern Marine
Science and Engineering Guangdong Laboratory,
Guangzhou 511458, China.
*Corresponding author. Email: oucjs@ouc.edu.cn
y g
RE FE REN C ES AN D N OT ES
1. K. S. Mai, “Exploring the road to high-quality develop-
ment of deep-sea farming in China” (Ocean University
of China News, 2021); http://ouceducnxiaobao.ihwrm.
com/index/article/articleinfo.html?doc_id=3659710
p
[in Chinese].
2. S. L. Dong et al., J. Fish. Chin. 47, 1 (2022) [in Chinese].
3. M. Lin, J. Manage. World 38, 39 (2022) [in Chinese].
4. Ministry of Agriculture and Rural Affairs, “China fishery
statistical yearbook” (China Agriculture Press, 2023) [in
,
Chinese].
5. K. S. Mai et al., Strateg. Stud. CAE 18, 90 (2016) [in
Chinese].
6. D. H. Li, L. M. Han, Soc. Sci. J. 245, 109 (2019) [in
Chinese].
7. L. N. Long et al., Rev. Aquacult. 15 (S1), 1 (2023).
8. R. L. Naylor, S. L. Williams, D. R. Strong, Science 294,
1655 (2001).
9. B. Kang et al., Rev. Aquacult. 15, 676 (2023).
10. Ministry of Agriculture and Rural Affairs, Ministry
of Industry and Information Technology, National
Development and Reform Commission, Ministry of
Science and Technology, Ministry of Natural Resources,
Ministry of Ecology and Environment, Ministry of
Transport, China Coast Guard, “Opinions on acceler-
ating the development of deeper offshore aquacul-
ture” (2023); http://www.moa.gov.cn/govpublic/
YYJ/202306/t20230612_6430042.htm [in Chinese].
10.1126/science.adj4352
1 SEP TEMBER 2023 • VOL 381 ISSUE 6661 955
 RESEARCH
and base for coupling of particu-
lar reactant pairs. Ten products
were isolated in more than 85%
yield under the individualized
conditions predicted by the
model. —JSY
Science, adg2114, this issue p. 965

IN S CIENCE JOU R NA L S
NATURAL HAZARDS
Edited by Michael Funk Two powerful quakes
The Kahramanmaraş earth-
quake sequence in Turkey on
6 February 2023 caused a
tremendous amount of damage
and loss of life. The sequence
occurred across several faults,
including and associated with
the East Anatolian Fault, a
strike-slip fault that has had

p
many major earthquakes in the
past. Jia et al. used an array of
geophysical observations to
produce models of how the rup-
tures occurred. The earthquake
sequence ruptured at least six

g
faults, including a large portion
of the East Anatolian Fault. The
rupture sequence was complex
and contained surprises in

y
the details of how the rupture
occurred. These observations
and models are important
for understanding strike-slip
faults and forecasting seismic
Artist’s depiction of a black hazards. —BG
hole, with an associated Science, adi0685, this issue p. 985
star and accretion disk,
outputting light at
x-ray, optical, and radio CALCULUS INSTRUCTION

y g
wavelengths Improving student
success with engagement
Across US universities, calculus
BLACK HOLES
is a gateway course for STEM

p
Magnetic field halts accretion degrees. Of all students who
initially pursue STEM degrees,

A
s material falls toward a black hole, it forms an accretion disk that emits x-rays and optical
more than half graduate
light and sometimes also a jet that is visible at radio wavelengths. Theory predicts that if the
,
without one, often after strug-
disk contains a sufficiently strong magnetic field, then it can resist the gravitational pull of the
gling through coursework.
black hole and temporarily halt the accretion process. You et al. compared observations of a
Instructors defaulting to tradi-
transient black hole accretion event at x-ray, optical, and radio wavelengths. They found time
tional lecture-based instruction
delays between brightening at the different wavelengths and used models to show that this behav-
exacerbates disparities in
ior is produced by a magnetically arrested disk. —KTS
failure rates; this disproportion-
Science, abo4504, this issue p. 961
ately affects women, Hispanic,
and Black students, depriving
the workforce of talent and
insights from diverse groups.
ORGANIC CHEMISTRY is one of the most widely used often requires a trial-and-error Kramer et al. conducted a large
reactions in pharmaceutical process to identify pertinent trial that randomized students
A carbon–nitrogen research and manufacturing. optimal conditions. Rinehart into calculus classrooms where
coupling guide Nonetheless, it depends sensi- et al. trained and validated a instructors actively engaged
The palladium-catalyzed cou- tively on the structure of the two machine learning model to pre- students collaboratively (treat-
pling of amines with aryl halides coupling partners and therefore dict appropriate ligand, solvent, ment) or relied on traditional

956 1 SEP TEMBER 2023 • VOL 381 ISSUE 6661 science.org SCIENCE
 lecture styles that treated them
as passive learners (control).
Across demographic groups,
the treatment was more effec-
tive, as engagement fostered
a deeper understanding of
calculus, improved grades,
and promoted the inclusion of
underrepresented students.
—EEU
Science, ade9803, this issue p. 995
NEUROSCIENCE
A perceptual space
map for smells
For vision and hearing, there
to the amyloplast settling sites
IN OTHER JOURNALS Edited by Caroline Ash
when the gravity direction is
and Jesse Smith
altered, indicating that these
proteins transmit informa-
tion about gravity direction
to downstream signaling
components. The experiments
support the idea that gravi-
sensing is carried out through a
positional sensing mechanism
rather than by sensing force.
—MRS
Science, adh9978, this issue p. 1006
SILICONE CHEMISTRY NEUROSCIENCE
Alcohols keep the Segmented and reconstructed
chains long

S
are well-developed maps that egmentation allows a visual system to distinguish
relate physical properties such Mass production of silicones which regions of a visual scene belong to the back-

as frequency and wavelength
to perceptual properties such
as pitch and color. The sense of
olfaction does not yet have such
a map. Using a graph neural
network, Lee et al. developed
proceeds by the steady growth
of silicon–oxygen chains.
However, there is a small
fraction of these chains that
inevitably bite back on them-
selves late in the reaction to
p
ground, which regions belong to different objects,
and where the borders that constrain the regions of
these objects are located. The brain uses a diversity of
cues, including luminance, texture, disparity, and motion, to
segregate objects from the background. Working with rhesus
macaques, Hesse and Tsao used electrophysiology guided

a principal odor map (POM)
that faithfully represents known
perceptual hierarchies and dis-
tances. This map outperforms
produce cyclic impurities. Shi
et al. report that benzyl alcohol
can complex with the chain end
through hydrogen bonding and
g
by functional magnetic resonance imaging to investigate
whether the visual cortex has a functional organization for
segmentation. They found that the visual cortex contains
discrete modules spanning visual areas V2, V3, V3A, V4,
and V4A for segmenting visual scenes. These segmentation

previously published models
to the point that replacing a
trained human’s responses
with the model output would
improve overall panel descrip-
tion. The POM coordinates were
able to predict odor intensity
and perceptual similarity,
even though these perceptual
features were not explicitly part
inhibit the back-biting process.
Moreover, they describe a
phosphonium counterion that
is also stabilized by the alcohol
but decomposes in its absence
to deactivate chain growth,
likewise preventing the by-
product formation. —JSY
Science, adi1342, this issue p. 1011
y
modules are functionally distinct from both color-selective
regions and other neighboring regions. —PRS
Proc. Natl. Acad. Sci. U.S.A. (2023) 10.1073/pnas.2221122120
Functional magnetic resonance imaging scan showing
activation of the human visual cortex

of the model training. These
results were used to build a vari-
ety of olfactory predictions that
outperformed previous feature
sets even without fine-tuning.
y g
REGENERATION cross-talk mechanism, in which
CANCER
ß-adrenergic receptor signaling
My oh myeloid Adrenergic nerves and promoted interleukin-22
Myeloid cell populations intestinal regeneration production from type 3 innate
control immunosuppression The intestine is a highly regen- lymphoid cells. —PNK

—PRS
Science, ade4401, this issue p. 999
PLANT SCIENCE
Lazily sensing gravity
The growth of plants in
response to gravity relies on
PHOTO: LIVING ART ENTERPRISES/SCIENCE SOURCE
in the tumor microen-
vironment (TME), and
targeting the TME is a promis-
ing therapeutic strategy for
immunosuppressed tumors.
Juric et al. developed an
afucosylated humanized
monoclonal antibody against
p
erative organ. The epithelial Cell Stem Cell (2023)
lining serves as a barrier against 10.1016/j.stem.2023.07.013
pathogens, and its cells have a
,
high turnover rate, consistently
responding to wear and tear to RACISM
regenerate. Although peripheral
Vicarious shocks to
nerves innervate the intestine,
how neurons contribute to epithe- mental health

the sedimentation of starch
granules called amyloplasts.
Transduction of the sedimen-
tation signal relies on both
LAZY1-LIKE (LZY) proteins and
redirection of auxin transport,
but the links between each of
these processes have not been
fully established. Nishimura et
al. found that LZY proteins at
the plasma membrane relocate
a proinflammatory receptor
expressed on many myeloid
cells. Treatment of a synge-
neic mouse model with this
antibody promoted anti-
tumor efficacy, suggesting a
promising strategy for future
immunotherapy that requires
further study. —DLH
Sci. Transl. Med. (2023)
10.1126/scitranslmed.add9990
lial cell repair is currently unclear.
Wang et al. studied adrenergic
nerves and their influence on
the renewal and regeneration
of intestinal epithelium. Using
a mouse model of irradiation-
induced injury, the authors
found that the density of gut
adrenergic nerves increased after
irradiation. Intestinal regenera-
tion occurred by a neuroimmune
Can learning about or watch-
ing someone else in your social
network experience racism
undermine your own mental
health? A growing literature sug-
gests that vicarious racism could
be as impactful as direct racism,
but we need to understand its
prevalence and social patterning.
Quinn et al. parsed the specific
contributions of social contexts

SCIENCE science.org 1 SEP TEMBER 2023 • VOL 381 ISSUE 6661 957

ALSO IN SCIENCE JOURNALS
HEALTH
What causes obesity?
The incidence of obesity is
increasing substantially. Obesity
is recognized as a medical
condition and as a risk factor for
various other diseases. There
are clearly genetic influences on
the development of obesity, as
well as environmental factors
such as the amount and type
of food intake. In a Perspective,
Speakman et al. argue that we
should approach obesity as
arising from a combination of
genes and environment rather
than genes versus environ-
ment. Considering these factors
together highlights the mul-
tifactorial influences on the
development of obesity and
raises many questions about
its causes, which need to be
addressed to better prevent and
treat this disease. —GKA
Science, adg2718, this issue p. 944
ATHEROSCLEROSIS
A protein that helps
prevent heart disease
Apolipoprotein B (apoB) is
a protein that carries lipids
and cholesterol. Lipoproteins
composed of apoB and its lipid
cargo play a key role in the
pathogenesis of atherosclerosis.
Tissue plasminogen activator
(tPA), an anticoagulant protein
that can be used to break down a
blood clot causing a heart attack
or stroke, inversely correlates
with the amount of cholesterol-
loaded apoB lipoproteins, but it
was not clear why this inverse
correlation exists. Dai et al. dis-
covered that in both mouse and
human cells, tPA produced by
the liver binds directly to apoB,
preventing it from being loaded
with lipids and assembled into
lipoproteins. Conversely, an
inhibitor of tPA has the opposite
effect, binding up free tPA and
preventing its interaction with
apoB. —YN
Science, adh5207, this issue p. 959
SCIENCE science.org
Edited by Michael Funk
PLANT MITOCHONDRIA
Initiating protein
translation
Mitochondrial protein transla-
tion machinery has diverged
within groups of eukaryotes,
with differing ribosome struc-
tures and initiation sites. Tran
et al. found that two proteins,
mTRAN1 and mTRAN2, are
required for translation initiation
in Arabidopsis mitochondria. The
mTRAN proteins form part of the
ribosome small subunit and bind
to motifs in the 59 untranslated
regions of mitochondrial genes,
allowing protein translation to
begin. This work offers insight
into translation initiation in
plants and highlights how dif-
ferent mitochondrial translation
mechanisms have evolved from
a common bacterial ancestor.
—MRS
Science, adg0995, this issue p. 960
ARCTIC CLIMATE
Going with the flow
One of the reasons that Arctic
sea ice has been disappearing
over the past decades is that
warm water from the Atlantic
is being advected into the
high-latitude ocean in increas-
ing amounts, a process called
“atlantification.” But what
drives this process? Polyakov
et al. show that the large-scale
weather pattern called the Arctic
Dipole causes atmospheric wind
patterns that modulate North
Atlantic inflows across the Fram
Strait and within the Barents
Sea, resulting in variations in
Arctic Ocean circulation, fresh-
water fluxes into the Amerasian
Basin, ocean stratification, and
heat fluxes (see the Perspective
by Bacon). —HJS
Science, adh5158, this issue p. 972;
see also adj8469, p. 946
HUMAN EVOLUTION
A close call
Today, there are more than 8 bil-
lion human beings on the planet.
We dominate Earth’s landscapes,
and our activities are driving
large numbers of other species
to extinction. Had a researcher
looked at the world sometime
between 800,000 and 900,000
years ago, however, the picture
would have been quite different.
Hu et al. used a newly developed
coalescent model to predict
past human population sizes
from more than 3000 present-
day human genomes (see the
Perspective by Ashton and
Stringer). The model detected
a reduction in the population
size of our ancestors from about
100,000 to about 1000 individu-
als, which persisted for about
100,000 years. The decline
appears to have coincided with
both major climate change and
subsequent speciation events.
—SNV
Science, abq7487 this issue p. 979;
see also adj9484, p. 947
EVOLUTIONARY BIOLOGY
Not so different after all
The rate of genetic mutation
that occurs across generations
is regularly used to estimate a
wide array of measures, includ-
ing past population sizes. It also
varies depending on a suite of
considerations, including gener-
ation time and body size, and is
notoriously difficult to estimate
in wild animal species. Suárez-
Menéndez et al. used a pedigree
approach in four wild baleen
whale species, producing a
mutation rate different from that
estimated using a phylogenetic
approach (see the Perspective
by Hoelzel and Lynch). The
new rate is faster than previ-
ously estimated for these large
animals, being more consistent
with primates and other smaller
species with similar generation
times. —SNV
Science, adf2160, this issue p. 990;
see also adk0121, p. 942
1 SEP TEMBER 2023 • VOL 381 ISSUE 6661
R E S E A RC H
CANCER STEM CELLS
Breaking the cycle in
leukemic stem cells
Leukemic stem cells are
resistant to various therapies
and promote relapse in many
chronic and acute leukemias.
Naef et al. identified a pathway
that was necessary for stemness
in leukemic stem cells but not
in normal hematopoietic stem
cells. Fusion proteins that drive
leukemias promoted leukemic
stem cell proliferation through a
cell-autonomous loop involv-
p
ing Wnt signaling and ST2, a
receptor for the cytokine inter-
leukin-33. ST2 was not detected
on normal hematopoietic stem
cells, indicating a potentially
cell-targeted approach for more
g
durable treatment outcomes in
patients. —LKF
Sci. Signal. (2021)
10.1126/scisignal.add7705
y
T CELLS
Replenishment of self-
renewing T cells
Stem-like CD8 T cells develop
during periods of persistent
antigen exposure. During chronic
infection and cancer, therapies
y g
can stimulate the differentiation
of antigen-specific stem-like
CD8 T cells into cells with effec-
tor function. However, factors
controlling the integrity of the
p
stem-like T cell pool have been
unknown. Studying chronic viral
infection in mice, Humblin et al.
,
found that the costimulatory
molecule CD28 regulated the
metabolism of stem-like CD8 T
cells expressing the inhibitory
receptor PD-1 to promote self-
renewal and differentiation. In
companion work, Gill et al. found
that anti–PD-1 blockade, while
promoting effector cell differen-
tiation, likewise stimulated the
self-renewal of these stem-like
T cells, maintaining the size and
functionality of the progenitor
pool. Thus, immune check-
point therapies may rejuvenate
immune responses without
958-B
R ES E ARCH | I N S C I E N C E J O U R NA L S

diminishing the durability of the

stem-like T cells that respond to
treatment. —SHR
Sci. Immunol. (2023)
10.1126/sciimmunol.adg0878,
10.1126/sciimmunol.adg0539

PHYSICAL CHEMISTRY
Attochemistry of
nuclear motion
Photoionization by absorption
of an extreme ultraviolet photon
can result in ultrafast nuclear
motion in the resulting cationic
molecule. In joint experimental
and theoretical work, Ertel et
al. investigated how nuclear
dynamics affect the outcome of

p
isotopically different chemical
reactions by steering the elec-
tronic dynamics triggered by an
attosecond pulse. The authors
used methane containing hydro-
gen (CH4) or deuterium (CD4),

g
which present different nuclear
dynamics after the absorption
of an extreme ultraviolet photon.
They suggest that the physical

y
origin of the differences can
be traced back to the nuclear
motion or the nuclear autocor-
relation function of a molecule
after the absorption. —SMK
Sci. Adv. (2023)
10.1126/sciadv.adh7747

y g
p
,

958-C 1 SEP TEMBER 2023 • VOL 381 ISSUE 6661 science.org SCIENCE
 lecture styles that treated them
as passive learners (control).
Across demographic groups,
the treatment was more effec-
tive, as engagement fostered
a deeper understanding of
calculus, improved grades,
and promoted the inclusion of
underrepresented students.
—EEU
Science, ade9803, this issue p. 995
NEUROSCIENCE
A perceptual space
map for smells
For vision and hearing, there
to the amyloplast settling sites
IN OTHER JOURNALS Edited by Caroline Ash
when the gravity direction is
and Jesse Smith
altered, indicating that these
proteins transmit informa-
tion about gravity direction
to downstream signaling
components. The experiments
support the idea that gravi-
sensing is carried out through a
positional sensing mechanism
rather than by sensing force.
—MRS
Science, adh9978, this issue p. 1006
SILICONE CHEMISTRY NEUROSCIENCE
Alcohols keep the Segmented and reconstructed
chains long

S
are well-developed maps that egmentation allows a visual system to distinguish
relate physical properties such Mass production of silicones which regions of a visual scene belong to the back-

as frequency and wavelength
to perceptual properties such
as pitch and color. The sense of
olfaction does not yet have such
a map. Using a graph neural
network, Lee et al. developed
proceeds by the steady growth
of silicon–oxygen chains.
However, there is a small
fraction of these chains that
inevitably bite back on them-
selves late in the reaction to
p
ground, which regions belong to different objects,
and where the borders that constrain the regions of
these objects are located. The brain uses a diversity of
cues, including luminance, texture, disparity, and motion, to
segregate objects from the background. Working with rhesus
macaques, Hesse and Tsao used electrophysiology guided

a principal odor map (POM)
that faithfully represents known
perceptual hierarchies and dis-
tances. This map outperforms
produce cyclic impurities. Shi
et al. report that benzyl alcohol
can complex with the chain end
through hydrogen bonding and
g
by functional magnetic resonance imaging to investigate
whether the visual cortex has a functional organization for
segmentation. They found that the visual cortex contains
discrete modules spanning visual areas V2, V3, V3A, V4,
and V4A for segmenting visual scenes. These segmentation

previously published models
to the point that replacing a
trained human’s responses
with the model output would
improve overall panel descrip-
tion. The POM coordinates were
able to predict odor intensity
and perceptual similarity,
even though these perceptual
features were not explicitly part
inhibit the back-biting process.
Moreover, they describe a
phosphonium counterion that
is also stabilized by the alcohol
but decomposes in its absence
to deactivate chain growth,
likewise preventing the by-
product formation. —JSY
Science, adi1342, this issue p. 1011
y
modules are functionally distinct from both color-selective
regions and other neighboring regions. —PRS
Proc. Natl. Acad. Sci. U.S.A. (2023) 10.1073/pnas.2221122120
Functional magnetic resonance imaging scan showing
activation of the human visual cortex

of the model training. These
results were used to build a vari-
ety of olfactory predictions that
outperformed previous feature
sets even without fine-tuning.
y g
REGENERATION cross-talk mechanism, in which
CANCER
ß-adrenergic receptor signaling
My oh myeloid Adrenergic nerves and promoted interleukin-22
Myeloid cell populations intestinal regeneration production from type 3 innate
control immunosuppression The intestine is a highly regen- lymphoid cells. —PNK

—PRS
Science, ade4401, this issue p. 999
PLANT SCIENCE
Lazily sensing gravity
The growth of plants in
response to gravity relies on
PHOTO: LIVING ART ENTERPRISES/SCIENCE SOURCE
in the tumor microen-
vironment (TME), and
targeting the TME is a promis-
ing therapeutic strategy for
immunosuppressed tumors.
Juric et al. developed an
afucosylated humanized
monoclonal antibody against
p
erative organ. The epithelial Cell Stem Cell (2023)
lining serves as a barrier against 10.1016/j.stem.2023.07.013
pathogens, and its cells have a
,
high turnover rate, consistently
responding to wear and tear to RACISM
regenerate. Although peripheral
Vicarious shocks to
nerves innervate the intestine,
how neurons contribute to epithe- mental health

the sedimentation of starch
granules called amyloplasts.
Transduction of the sedimen-
tation signal relies on both
LAZY1-LIKE (LZY) proteins and
redirection of auxin transport,
but the links between each of
these processes have not been
fully established. Nishimura et
al. found that LZY proteins at
the plasma membrane relocate
a proinflammatory receptor
expressed on many myeloid
cells. Treatment of a synge-
neic mouse model with this
antibody promoted anti-
tumor efficacy, suggesting a
promising strategy for future
immunotherapy that requires
further study. —DLH
Sci. Transl. Med. (2023)
10.1126/scitranslmed.add9990
lial cell repair is currently unclear.
Wang et al. studied adrenergic
nerves and their influence on
the renewal and regeneration
of intestinal epithelium. Using
a mouse model of irradiation-
induced injury, the authors
found that the density of gut
adrenergic nerves increased after
irradiation. Intestinal regenera-
tion occurred by a neuroimmune
Can learning about or watch-
ing someone else in your social
network experience racism
undermine your own mental
health? A growing literature sug-
gests that vicarious racism could
be as impactful as direct racism,
but we need to understand its
prevalence and social patterning.
Quinn et al. parsed the specific
contributions of social contexts

SCIENCE science.org 1 SEP TEMBER 2023 • VOL 381 ISSUE 6661 957
R ES EARC H | I N O T H E R J O U R NA L S

EXTREME RAINFALL

Rising up all over

M
esoscale convective systems, large, organized, long-lasting, and prodigiously precipitating events, occur
worldwide and are responsible for major fractions of the rainfall in the areas where they take place. Li et al.
report that these systems have become more frequent and intense in the East Asian rainband, accounting for
most of the increase in rainfall that the region has experienced in the past 20 years. The authors attribute
this pattern to anthropogenic temperature increases and expect it to continue as the climate continues to
warm, with the consequent impacts on events such as extreme flooding. —HJS
Geophys. Res. Lett. (2023) 10.1029/2023GL103595

p
g
A mesoscale convective system over

y
the South China Sea as seen from the
International Space Station

and demographic characteristics complexes that differ widely in PHYSICS GLYCOSYLATION

among African American adults.
Findings differed by gender, and
vicarious discrimination was more
stability and persistence. The E3
ubiquitin ligase UBR5 is a mem-
Looking for bound Sugar-coated methylation
magnons O-linked glycosylation and
ber of a family of enzymes that
methylation are dynamic post-

prevalent than direct discrimina-
tion, particularly among educated
groups. Compared with women,
men reported higher levels of
both vicarious and direct discrimi-
target specific proteins for pro-
teolysis and plays an important
role in regulating gene expression
in human stem cells. Mark et al.
found that UBR5 monitors the
y g
Floquet engineering, which relies
translational modifications of
on applying a judiciously chosen
proteins and DNA, respectively,
periodic series of pulses, has
that enable cells to encode an
been used to generate various
additional level of information
spin interactions in cold atom

nation, particularly in the context
of police encounters. Women
reported more direct discrimina-
dissociation of multiprotein tran-
scription factor complexes that
serve as hubs or networks con-
systems. Kranzl et al. used this
method to create a long-range
anisotropic Heisenberg model in
p
in response to external stimuli.
Shin et al. explored an intriguing
link between these processes.

tion in their workplaces and high
vicarious discrimination related
to police. Therefore, research
may be neglecting an important
contributor to poor health for this
stigmatized group. —EEU
Soc. Sci. Med. (2023)
10.1016/j.socscimed.2023.116104
SIGNALING
Protein degradation and
transcription
Metazoan development requires
the coordination of protein
trolling cell functions. In a network
centered on the c-Myc oncopro-
tein, dissociation of c-Myc and
other subunits exposed motifs
that were recognized by UBR5,
leading to subunit degradation
at different rates. Such targeted
degradation by UBR5 forces
cycles of complex reformation
that support dynamic regula-
tion of transcription and ensure
that stem cells are responsive
to changes in their environment.
—LBR
Cell (2023)
10.1016/j.cell.2023.06.015
an array of 20 trapped calcium
ions. The researchers then
searched for bound magnon
excitations in this system, start-
ing from a state in which two
adjacent spins were “flipped”
with respect to the rest. They
found that bound magnons
propagate with a finite velocity
that depends on the interac-
tion strength. For even stronger
interactions, it may be possible
to create multiple bound states
using this technique. —JS
Phys. Rev. X (2023)
10.1103/PhysRevX.13.031017
,
In human cells and mouse liver
samples, the DNA methyltrans-
ferase DNMT-1 is glycosylated
in response to elevated glucose.
This modification reduces DNA
methyltransferase activity and
results in a global loss of DNA
methylation, particularly at sites
known as partially methylated
domains. Using a mutant that
is not subject to this regulatory
mechanism, the authors showed
that the loss of methylation under
high-glucose conditions is detri-
mental to DNA stability. —MAF
eLife (2023) 10.7554/eLife.85595

958 1 SEP TEMBER 2023 • VOL 381 ISSUE 6661 science.org SCIENCE
RES EARCH

◥ and/or mTRAN2 were impaired in Arabidopsis,

RESEARCH ARTICLE SUMMARY the plants showed growth reductions ranging
from moderate growth inhibition to embryo
PLANT MITOCHONDRIA lethality, indicating that they are essential
for normal plant development. In the mtran
An mTRAN-mRNA interaction mediates mitochondrial double mutants, mitochondrial biogenesis and
function were impaired, with reduced abun-
translation initiation in plants dance and activity of all oxidative phosphoryl-
ation complexes that contain mitochondrially
Huy Cuong Tran, Vivian Schmitt, Sbatie Lama, Chuande Wang, Alexandra Launay-Avon, Katja Bernfur, encoded subunits (Complexes I, III, IV, and V).
Kristin Sultan, Kasim Khan, Véronique Brunaud, Arnaud Liehrmann, Benoît Castandet, Fredrik Levander, Using coimmunoprecipitation assays, we
Allan G. Rasmusson, Hakim Mireau, Etienne Delannoy, Olivier Van Aken* found that mTRAN1 and -2 are part of the
plant mitoribosomal “small” subunit (mtSSU),
in line with previous studies. Transcriptome
INTRODUCTION: Mitochondria are eukaryotic RATIONALE: We characterized two “proteins of analysis of mtran double mutants showed a
organelles required for energy conversion and unknown function” that were suggested to be rearrangement of nuclear, mitochondrial, and
metabolism. As mitochondria are likely rem- part of plant mitoribosomes, using Arabidopsis chloroplast gene expression that further sug-
nants of bacteria that were once incorporated thaliana (Arabidopsis) as a model system. To gested mitochondrial translation defects. In
through endosymbiosis, they have retained assess their functions, we produced plants with agreement, we showed that mtran double
many features similar to their bacterial an- reduced expression of these proteins and mutants have reduced mitochondrial protein

cestors. Mitochondria in most eukaryotes have
retained their own genome, which encodes a
set of proteins required for mitochondrial ope-
ration. During evolution many mitochondrial
components have, however, diverged substan-
tially between eukaryotic kingdoms, of which
assessed how their loss of function affected
mitochondrial biogenesis and proteostasis. In
addition, we searched the 5′ regions of mito-
chondrial mRNAs for conserved motifs that
could act as potential mitoribosomal bind-
ing sites.
p
synthesis rates. A large proportion of mitochon-
drial mRNAs were not bound by mitoribo-
somes in mtran double mutants, suggesting
an important role for mTRANs in translation
initiation. Structural modeling indicated that
mTRAN proteins are alpha-solenoid proteins

the mitochondrial ribosome (mitoribosome) is
a prime example. How mitochondrial mRNAs
are recognized by mitoribosomes to initiate
protein translation in plants has been an
enigma, especially because mitochondrial
RESULTS: We found that the identified genes,
which we named “mitochondrial translation
factors” mTRAN1 and mTRAN2, are conserved
in land plants but absent in green algae, ani-
g
related to pentatricopeptide repeat (PPR) pro-
teins and are thus potentially RNA-binding
proteins. Because we could not find putative
RNA motifs that could be bound by mTRANs
using classical “PPR-code” detection tools, we

mRNAs lack bacterial-type Shine-Dalgarno
ribosome-binding sequences.
mals, and fungi. mTRANs were exclusively
found inside mitochondria, and when mTRAN1
Mitochondrion
y
searched the 5′ sequences of thousands of
plant mitochondrial genes for conserved motifs.
We identified A/U-rich motifs (CUUUxU and
AAGAAx/AxAAAG) and showed that mTRAN1
can directly bind these motifs in vitro and
in vivo. Finally, using ribosome footprinting
(Ribo-seq), we showed that mTRANs are re-
quired for binding and translation of likely all
plant mitochondrial mRNAs.

y g
CONCLUSION: This study establishes mTRAN
proteins as factors required for mitochon-
drial translation in plants. Our results in-
dicate that mTRANs can bind conserved

p
A/U-rich motifs present in the 5′ regions
mtLSU of plant mitochondrial mRNAs, which may
act as ribosome binding sites. mTRANs may
fMet
mtLSU thus act as “universal” homing factors to
A/U-rich motif UAC guide the mitoribosome to mitochondrial ,
AUG mRNA AUG mRNA mRNAs and initiate translation. Plant mito-
5′ 3′ 5′ 3′
mTRAN
chondrial translation initiation therefore
mTRAN
appears to use a protein-mRNA interaction
mtSSU mtSSU that is divergent from bacteria or mam-

Mitoribosome
malian mitochondria.
▪
mTRAN proteins mediate mitochondrial translation in plants. Plant mitochondria are required for
energy metabolism, growth, and survival. Several proteins of the mitochondrial respiration chain and Adenosine The list of author affiliations is available in the full article.
*Corresponding author. Email: olivier.van_aken@biol.lu.se
triphosphate (ATP)-synthase are encoded by the mitochondrial genome and translated by mitoribosomes.
Cite this article as H. C. Tran et al., Science 381, eadg0995
Plant-specific mTRAN proteins are part of the plant mitoribosome “small” subunit (mtSSU). Our results indicate (2023). DOI: 10.1126/science.adg0995
that mTRANs likely recognize A/U-rich sequence motifs in the 5′ regions of mitochondrial mRNAs to mediate
translation initiation. After mRNA-binding, the “large” mitoribosomal subunit (mtLSU) joins to start translation. fMet, READ THE FULL ARTICLE AT
N-formylmethionine; A, adenosine; U, uridine; G, guanine; C, cytosine. https://doi.org/10.1126/science.adg0995

Tran et al., Science 381, 960 (2023) 1 September 2023 1 of 1

RES EARCH

◥ mTRAN1 and mTRAN2 are plant-specific proteins

RESEARCH ARTICLE Using a BLAST search, we found that mTRAN1
(AT4G15640) and mTRAN2 (AT3G21465) are
PLANT MITOCHONDRIA present only in Embryophyta (land plants). Pro-
tein sequence alignment showed that mTRAN1
An mTRAN-mRNA interaction mediates mitochondrial and -2 are highly similar (84% identical),
suggesting that they are paralogs in Arabi-
translation initiation in plants dopsis thaliana. Some angiosperms, includ-
ing Brassicaceae and Glycine max, contain
Huy Cuong Tran1, Vivian Schmitt1, Sbatie Lama1†, Chuande Wang2, Alexandra Launay-Avon3,4, two mTRAN proteins (Fig. 1A and fig. S1), in-
Katja Bernfur5, Kristin Sultan6, Kasim Khan1, Véronique Brunaud3,4, Arnaud Liehrmann3,4,7, dicating that the duplication event of mTRAN
Benoît Castandet3,4, Fredrik Levander6,8, Allan G. Rasmusson1, Hakim Mireau2, protein-encoding genes occurred relatively late
Etienne Delannoy3,4, Olivier Van Aken1* in evolution, with the divergence of Fabaceae
dating back around 70 million years (22).
Plant mitochondria represent the largest group of respiring organelles on the planet. Plant Physcomitrium patens, a nonvascular moss,
mitochondrial messenger RNAs (mRNAs) lack Shine-Dalgarno-like ribosome-binding sites, so it is also has two mTRAN proteins, suggesting that
unknown how plant mitoribosomes recognize mRNA. We show that “mitochondrial translation independent duplication events have occurred
factors” mTRAN1 and mTRAN2 are land plant–specific proteins, required for normal mitochondrial (23). The remaining land plant species exam-
respiration chain biogenesis. Our studies suggest that mTRANs are noncanonical pentatricopeptide ined here contain only one mTRAN protein.
repeat (PPR)–like RNA binding proteins of the mitoribosomal “small” subunit. We identified

p
conserved Adenosine (A)/Uridine (U)-rich motifs in the 5′ regions of plant mitochondrial mRNAs. mTRAN1 and mTRAN2 are targeted
mTRAN1 binds this motif, suggesting that it is a mitoribosome homing factor to identify mRNAs. We to mitochondria
demonstrate that mTRANs are likely required for translation of all plant mitochondrial mRNAs. Prediction analyses suggested that mTRAN1 is
Plant mitochondrial translation initiation thus appears to use a protein-mRNA interaction that is mitochondrially targeted (24), and this was sup-
divergent from bacteria or mammalian mitochondria. ported by mitochondrial complexome profiling
(25). To confirm mTRAN mitochondrial local-

lthough most mitochondrial proteins are
nuclear encoded, mitochondria have par-
tially retained their genome and transla-
tional machinery (1, 2). In plants, the
there is poor understanding of how plant mito-
ribosomes interact with 5′ UTRs of mitochon-
drial mRNAs and how mitoribosomal small
subunits (mtSSU) identify correct translation
g
ization, we stably expressed mTRAN1-GFP and
mTRAN2-GFP fusions in Arabidopsis Col-0
plants. mTRAN1-GFP and mTRAN2-GFP co-
localized with the MitoTracker fluorescent dye
that labels mitochondria (Fig. 1B). Addition-

number of unique mitochondrial proteins
is estimated around 1000 to more than 2000
(2, 3), but the mitochondrial genome only
encodes 20 to 40 proteins (4, 5). Despite a bac-
terial origin, mitochondrial translation is sub-
stantially different from its bacterial counterpart
(6–8). In yeast, mitochondrial translation in-
itiation is regulated by general translational
factors associated with the mitoribosome and
mRNA-specific factors that bind 5′-untranslated
start codons. In contrast to other species, the
plant mtSSU is larger than the mitoribosomal
large subunit (mtLSU) (14, 15). The plant mito-
ribosome structure and core components have
recently been studied (14–16), showing that
plant-specific ribosomal pentatricopeptide re-
peat (rPPR) proteins have become bona fide
ribosomal constituents. PPR proteins contain
repeats of 35 amino-acid tandem motifs, each
forming two antiparallel α-helices, which in-
y
ally, we performed immunoblot analysis on
purified cytosolic, mitochondrial, and chloroplas-
tic fractions isolated from Arabidopsis seedlings
expressing mTRAN1-GFP or mTRAN2-GFP
(Fig. 1C). mTRAN1-GFP or mTRAN2-GFP could
only be detected in the mitochondrial fractions,
further confirming their mitochondrial location.
The mtran mutants have a growth
reduction phenotype

regions (5′ UTRs) to position translation ini-
tiation sites (9, 10). Mammalian mitochon-
drial mRNAs have absent or extremely short
5′ UTRs, and translation can start directly on
teract with each other to generate a helix-turn-
helix motif (17, 18). The helix-turn-helix PPR
domains form a superhelix with a central
groove, allowing PPR proteins to interact with
y g
mTRAN1 and mTRAN2 genes show relatively
similar expression patterns during plant de-
velopment (fig. S2). Of note, mTRAN proteins
are expressed very early in seeds during ger-

p
non-AUG (adenine-uracil-guanine) start codons RNA (17, 19). mination (one day after imbibition) (fig. S2),
(11, 12). We characterized Arabidopsis thaliana genes suggesting that they might be critical for plant
Bacterial-type Shine–Dalgarno ribosome- AT4G15640 and AT3G21465—currently annotated development. Three independent knockout lines

,
binding sequences are not found in flowering as adenylyl cyclases (ACs), but also suggested to were obtained for each mTRAN gene (fig. S3A).
plant mitochondrial 5′ UTRs (13). Therefore, be mitochondrial ribosome components rPPR10/ We obtained double homozygous knockout
mS83 (ribosomal pentatricopeptide repeat protein plants for mtran1-1 × mtran2-1 (mtran1-1/2-1)
1
Department of Biology, Lund University, Lund, Sweden. 10) (14, 15). ACs catalyze conversion of ATP to and mtran1-2 × mtran2-2 (mtran1-2/2-2) and
2
Université Paris-Saclay, INRAE, AgroParisTech, Institut
Jean-Pierre Bourgin (IJPB), 78000 Versailles, France.
3′-5′-cyclic adenosine monophosphate (cAMP)— verified loss of mTRAN1 and mTRAN2 tran-
3
Université Paris-Saclay, CNRS, INRAE, Université d’Évry, a second messenger that affects different scripts by quantitative reverse–transcription
Institute of Plant Sciences Paris-Saclay (IPS2), 91405 Orsay, physiological and biochemical processes. ACs polymerase chain reaction (qRT-PCR) (fig. S3D).
France. 4Université Paris Cité, CNRS, INRAE, Institute of Plant
Sciences Paris-Saclay (IPS2), 91405, Orsay, France.
have been extensively studied in animals, but Only plants homozygous for either mtran1-3
5
Department of Chemistry, Lund University, Lund, Sweden. little is known about them in plants (20, 21). or mtran2-3 and hemizygous for the other
6
Department of Immunotechnology, Lund University, Lund, Our results indicate that AT4G15640 and mutations could be obtained, indicating that
Sweden. 7Université Paris-Saclay, CNRS, Université d’Évry,
AT3G21465 are not classical PPR proteins but mtran1-3/2-3 was nonviable (fig. S3, B and C).
Laboratoire de Mathématiques et Modélisation d’Évry, 91037
Évry-Courcouronnes, France. 8National Bioinformatics are required for translation initiation by plant The mtran1 single mutants were smaller than
Infrastructure Sweden, Science for Life Laboratory, Lund mitoribosomes, and as such we propose a new the Col-0 wild type (WT), whereas the mtran2
University, Lund, Sweden. annotation for AT4G15640 and AT3G21465: mutants appeared more similar in size to the
*Corresponding author. Email: olivier.van_aken@biol.lu.se
†Present address: Department of Plant Breeding, Swedish MITOCHONDRIAL TRANSLATION FACTOR WT (Fig. 1D). mtran1-1/2-1 was also smaller
University of Agricultural Sciences, Alnarp, Sweden. 1 (mTRAN1) and mTRAN2, respectively. than the WT, whereas mtran1-2/2-2 showed

Tran et al., Science 381, eadg0995 (2023) 1 September 2023 1 of 12

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

Fig. 1. mTRAN1 and mTRAN2 A P. trichocarpa (Potri.008G205300)

995 G. max (Glyma.17G023300)
are land plant–specific mito- 1000
G. max (Glyma.07G251000)
chondrial proteins required 0.1 997 S. parvula (Tp3g19500)
for normal development. 1000 A. thaliana mTRAN2 (AT3G21465)
997
(A) Phylogenetic tree of mTRAN1 1000 A. lyrata (AL3G35540)
(AT4G15640) and mTRAN2 1000
S. parvula (Tp7g13890)
1000 A. thaliana mTRAN1 (AT4G15640)
(AT3G21465) from Arabidopsis 1000
A. lyrata (AL7G39790)
thaliana with mTRAN proteins
O. sativa (LOC_Os07g47201)
from other Embryophyta 1000
Z. mays (Zm00001d022498)
570
(land plants). Scale bar indicates A. trichopoda (ATR0706G238)
10% sequence divergence and 936 P. abies (PAB00014411)
node numbers indicate bootstrap P. patens (Pp3c14_15200)
694
values. (B) Full-length coding 872 P. patens (Pp3c14_7470)
M. polymorpha (Mapoly0002s0143)
sequences of mTRAN1 and
S. moellendorffii (SMO202G0041)
mTRAN2 were fused with GFP at
B GFP MitoTracker Merge D
the C-termini to assess GFP Col-0 mtran1-1 mtran1-2
targeting in stably transformed
mTRAN1

Arabidopsis plants. Imaging was

done by confocal microscopy
on root cells. Mitochondria were

p
labeled with MitoTracker Red.
mtran1-3 mtran2-1 mtran2-2
mTRAN2

Scale bars: 5 mm. (C) Immunoblot

analysis of cytosolic, mitochon-
drial, and chloroplastic fractions
isolated from homozygous
Arabidopsis seedlings expressing C mTRAN1-GFP mTRAN2-GFP

g
mTRAN1-GFP and mTRAN2-GFP.
to
to

mtran2-3 mtran1-1/2-1 mtran1-2/2-2

t
t

Cp
Cp

Cy
Cy

Mi
Mi

Blots were probed with antibodies kDa

against chloroplast-targeted PGRL1 29
PGRL1, mitochondrially targeted TOM40 40

y
TOM40, and GFP. The molecular
weight (kDa) is shown on the right GFP 70
side of each blot. Cyt, cytosol;
E 0.1 0.50.71.0 1.02 1.04 F
Mito, mitochondria; Cp, chloro- 25
plast. (D) A representative picture Col-0 Col-0
*** *** *** mtran1-1
of 25-day-old soil-grown plants. mtran1-1 mtran1-2
Rosette leaf area (cm2)

20
Scale bar: 2 cm. (E) Plate-based ** *** *** *** mtran1-3
***
phenotypic analysis (means ± SD, mtran1-2 mtran2-1
mtran2-2 ***
n > 60). Arrows indicate growth *** *** *** *** 15 mtran2-3
mtran1-3 mtran1-1/2-1
stages as described previously
*** *** *** *** mtran1-2/2-2

y g
(26): 0.1, Seed imbibition; 0.5, mtran2-1 10
***
radicle emergence; 0.7, hypocotyl *** *** ***
and cotyledon emergence; 1.0, mtran2-2 ***
** *** *** *** 5
cotyledons fully open; 1.02, two ***
mtran2-3 ***

p
rosette leaves > 1 mm; 1.04, ***
*** *** *** *** *** *** *** ***
four rosette leaves > 1 mm. The mtran1-1/2-1 0
15 18 21 24 27 30
boxes indicate the time between *** *** *** ***
the growth stages. (F) Rosette mtran1-2/2-2 Days after sowing

,
leaf area of soil-grown plants
0 5 10 15 20
(means ± SD, n = 9). Statistical
Days after sowing
significance was based on
Student’s t test with Bonferroni correction (*P < 0.05, **P < 0.01, ***P < 0.001).

severe growth retardation. mtran single and
double mutants were significantly slower to
germinate (stage 0.5) and develop true leaves
(stage 1.02/1.04) (Fig. 1E) (26). Primary roots of
the mtran single and double mutants were
significantly shorter than those of the WT
(fig. S3E). Again, the double mutants showed a
>80% reduction in primary root growth. Ad-
ditionally, mtran single and double mutants
were significantly smaller and slower to flower
than the WT (fig. S3F). mtran1-2/2-2 had a
stronger phenotype compared with mtran1-1/
2-1 (Fig. 1, D to F). The milder phenotype of
mtran1-1/2-1 is likely because of the presence
of the T-DNA in the 10th intron of mTRAN1,
potentially resulting in a partially active pro-
tein. For mTRAN1, the T-DNA is inserted in
the 5th and 7th exon, likely resulting in severe
growth deficiency and lethality of mtran1-2/2-
2 and mtran1-3/2-3, respectively.
mTRAN1 and mTRAN2 are unlikely to be
adenylyl cyclases
mTRAN1 and mTRAN2 are annotated as ACs,
but we could not find experimental evidence
nor a defined source for this annotation. To
evaluate mTRAN1/2 AC activity, we performed

Tran et al., Science 381, eadg0995 (2023) 1 September 2023 2 of 12

RES EARCH | R E S E A R C H A R T I C L E
a bacterial cAMP synthase complementation
assay in the BTH101 (cya) E. coli strain that
lacks endogenous AC activity (27). The cells
transformed with the positive control formed
colored colonies on MacConkey/maltose and
LB/X-gal media and could also grow on M63
medium (fig. S4A). mTRAN1 and mTRAN2
were cloned into the pUT18 vector and trans-
formed into E. coli cya. On all tested media,
pUT18-mTRAN1 and pUT18-mTRAN2 colo-
nies had the same phenotype as the negative
pUT18 control (fig. S4A). Thus, mTRAN1 and
mTRAN2 do not appear to have AC activity in
the bacterial system, though we cannot rule
out a false negative result as E. coli may not
provide the correct environment for plant ACs.
Furthermore, there was no significant difference
in cAMP concentration between the WT and
mtran double-knockout mutants (fig. S4B) de-
spite their clear growth retardation phenotype
(Fig. 1, D to F), supporting the idea that mTRAN1
and mTRAN2 are unlikely to be ACs.
Mitochondrial oxidative phosphorylation
complexes show reduced activity in mtran
double mutants
To find the cause for slow growth of mtran
double mutants, mitochondria isolated from
mtran1-1/2-1 and mtran1-2/2-2 were analyzed
by blue native polyacrylamide gel electropho-
resis (BN-PAGE). The mtran double mutants
showed a decrease in abundance of supercom-
plex I/III and complexes I, III, and V compared
with the WT (Fig. 2, A and B). Furthermore,
the activity of complexes I, III, IV, and V was
reduced in the mtran double mutants, espe-
cially mtran1-2/2-2 (Fig. 2C). Freshly isolated
mitochondria from mtran1-1/2-1 and mtran1-
2/2-2 were used for oxygen consumption mea-
surements (Fig. 2D). The state III respiration
rates (succinate+ADP+NADH) of mtran dou-
ble mutants were significantly lower com-
pared with the WT. The mtran double mutants,
however, had a significantly higher alterna-
tive oxidase (AOX) pathway capacity (KCN+
dithiothreitol+pyruvate). Additionally, the
ratio of AOX capacity to state III was signif-
icantly higher in the mtran double mutants,
suggesting that they rely heavily on alterna-
tive respiration.
Loss of mTRAN1 and mTRAN2 causes changes in
abundance of mitochondrial proteins
In the mtran double mutants, especially
mtran1-2/2-2, proteins from complexes I, III,
IV, and V were less abundant (Fig. 2E), in
agreement with the decrease in abundance
and activity of these complexes (Fig. 2, A to C).
Mitochondrially encoded mitoribosomal sub-
units RIBOSOMAL PROTEIN S4 (RPS4) and
RIBOSOMAL PROTEIN L16 (RPL16), and nu-
clear encoded mitochondrial translation fac-
tor LEUCINE ZIPPER-EF-HAND-CONTAINING
TRANSMEMBRANE PROTEIN 1 (LETM1) were
Tran et al., Science 381, eadg0995 (2023) 1 September 2023
more abundant in the mtran double mutants.
Nuclear-encoded mtSSU subunit RIBOSOMAL
PROTEIN S10 (RPS10) did not show clear
differences in abundance in the mutants.
Proteins encoded by mitochondrial retro-
grade signaling marker genes, including
ALTERNATIVE OXIDASE (AOX), UP REGU-
LATED BY OXIDATIVE STRESS (UPOX), and
TRANSLOCASE OF THE INNER MEMBRANE
17 (TIM17) (28, 29), were more abundant in
both double mutants. This indicates that loss
of mTRAN1 and mTRAN2 induced mitochon-
drial retrograde/unfolded protein response
(UPRmt) signaling (Fig. 2E). Up-regulation of
AOX protein is consistent with the reliance of
mtran double mutants on alternative respira-
tion (Fig. 2C).
To assess global changes in mitochondrial
protein abundance, we analyzed isolated mito-
chondria with tandem mass spectrometry
(MS/MS) (data S1). We detected mTRAN1 in
the WT and mtran1-1/2-1 mitochondria, with
comparable abundance. However, mTRAN1
was not detected in mtran1-2/2-2. This is con-
sistent with our hypothesis regarding mTRAN1
activity and the location of the T-DNA insertion
in mTRAN1. mTRAN2 could not be detected in
any of the genotypes, suggesting that mTRAN1
is much more abundant than mTRAN2, which
is in agreement with a recent study (3). This
may explain why we found 684 and 24 pro-
teins to be significantly different in abundance
in mtran1-2/2-2 and mtran1-1/2-1 (Padj < 0.2),
respectively, compared with the WT. Using a
less-stringent unadjusted P < 0.05, 247 pro-
teins were found to be differentially abundant
in the mtran1-1/2-1 mitochondria, represent-
ing subunits of complexes I, III, IV, and V, in
agreement with immunoblot results (Fig. 2E).
Of the 684 proteins affected in mtran1-2/2-2,
mitochondrially encoded oxidative phospho-
rylation (OXPHOS) proteins, including NADH
DEHYDROGENASE 1, 7, and 9 (Nad1, Nad7, and
Nad9), APOCYTOCHROME B (Cob), CYTO-
CHROME OXIDASE 2 (Cox2), and ATP SYN-
THASE SUBUNIT 1, 4, and 8 (Atp1, Atp4, and
Atp8), were less abundant. However, sub-
units of complex II, all of which are nuclear-
encoded, were not affected in the mutants.
Mitochondrially encoded RIBOSOMAL PRO-
TEIN L5 (Rpl5) and Rpl16, and nuclear-encoded
AOX1A, AOX1D, and LETM1 were more abun-
dant. This is consistent with the results of our
immunoblot analysis. Furthermore, heat shock
proteins, adenosine diphosphate/adenosine tri-
phosphate (ADP/ATP) carrier proteins, ALTER-
NATIVE NAD(P)H DEHYDROGENASES NDA1/
NDB2, proteases, mitochondrially targeted ribo-
somal proteins, PPR/MORF proteins, ovule
abortion proteins, and amino acid-tRNA lig-
ases were more abundant. In conclusion, loss
of both mTRAN proteins results in severe per-
turbation of the mitochondrial proteome and
reduction of many OXPHOS components.
mTRAN1 and mTRAN2 are part of the
mitoribosome small subunit
To identify mTRAN1/2-interacting proteins,
plants expressing mTRAN1-GFP or mTRAN2-
GFP were used for coimmunoprecipitation
(co-IP) followed by MS/MS. A mitochondria-
targeted GFP line (mito-GFP) (30) was used as
a negative control. The phenotypes of comple-
mentation lines mtran1-2/2-2 35S:mTRAN1-GFP
and mtran1-2/2-2 35S:mTRAN2-GFP were
similar to the WT (fig. S5), indicating that
mTRAN1-GFP and mTRAN2-GFP proteins are
fully functional. A summary of proteins co-
immunoprecipitated with mTRAN1-GFP and
mTRAN2-GFP is presented in data S2A. We
found that mTRAN1 and mTRAN2 interacted
with 27 components of the mtSSU, as pro-
posed by previous work (14, 15). By contrast,
no mitoribosome core subunits were pulled
down using the mito-GFP control, indicating
p
the specificity of the interactions. This indi-
cates that mTRAN1 and mTRAN2 are part of
the mtSSU. All ribosomal proteins interacting
with mTRAN proteins belonged to the mtSSU,
except mitochondrial ribosomal protein L11,
which is part of the mtLSU. mtLSU L11 was,
g
however, only identified in the first replicate.
Most rPPR proteins found in Waltz et al. (15)
and Rugen et al. (14) interacted with mTRAN
proteins, though rPPR3a (AT1G55890) only
coimmunoprecipitated with mTRAN1. An in-
y
dividual mitochondrion contains on average
around 520 copies of mtSSU proteins (3),
which is very close to the estimated 534 copies
found for mTRAN1, supporting the idea that
mTRANs are integral parts of the plant mito-
ribosome. We also found proteins interacting
with mTRAN proteins that were only detected
in our analysis (data S2A, gray color), includ-
ing three additional mitochondrially targeted
ribosomal proteins, a PPR protein (AT3G61520)
y g
and MORF8. mTRAN1 and mTRAN2 also inter-
acted with mitochondrially encoded Nad7 and
Nad9, which are less abundant in the mtran
double mutants (Fig. 2 and data S1). Overall,
p
our co-IP results demonstrate that mTRAN1/2
are part of the mtSSU, suggesting that they
might be involved in mitochondrial translation.
,
Transcriptome analysis of mtran1-2/2-2
indicates mitochondrial translation defects
RNA-seq analysis was carried out on the WT
and mtran1-2/2-2 (Fig. 1, D to F). 2143 nuclear-
encoded differentially expressed genes (DEGs;
Padj<0.05, fold change >2 or <0.5) were found
in mtran1-2/2-2 versus the WT, of which 1117
were up-regulated and 1026 were down-
regulated (data S3). Gene ontology analysis
showed that stress-responsive genes were up-
regulated whereas oxygen level-, hypoxia-, and
hormone (auxin)–responsive genes were down-
regulated (data S4). Mitochondrial retrograde
signaling marker genes such as AOX1a were
significantly up-regulated (data S3), indicating
3 of 12
RES EARCH | R E S E A R C H A R T I C L E

Fig. 2. mtran double mutants A C

a)
1

1
1

1
2
2

2
/2-

/2-

/2-
/2-

/2-

/2-
/2-
/2-

/2-

/2-
(k D
show defects in mitochondrial

1-1

1-2

1-1
1-1

1-1

1-1
1-2
1-2

1-2

1-2
er
ran

ran

ran
biogenesis, protein content,

ran

ran

ran
ran
ran

ran

ran
dd

l- 0
l-0

l- 0

l- 0

l- 0
La
mt

mt

Co
mt
Co

Co

Co

Co
mt

mt

mt
mt
mt

mt

mt
and function. (A) Analysis of
abundance of mitochondrial
complexes by Coomassie-Colloidal I+III2
I+III2
stained–BN-PAGE. Arrows indi- I
cate respiratory complexes and I
supercomplexes. The molecular III2 V
V 669
weight (kDa) is shown on the right III2
side of the gel. (B) Western blot
F1
using anti-RISP antibody for native F1 440
complex III on mitochondrial pro- IV
tein extracts of wild-type and IV
mtran double mutant separated 232
by BN-PAGE. (C) Activity mea-
surement of the respiratory 158
complexes I, III, IV, and V in
BN-PAGE. Arrows indicate respi-
ratory complexes and super- CI CIII CIV CV

p
complexes. I+III2, supercomplex B / 2-
1
/ 2-
2
I+III2; CIII, complex III; CIV, 1-1 1-2
complex IV; CV, complex V; F1, l-0 tr an tr an
Co m m E 20 µg 10 µg
F1-subcomplex of complex V.
RISP

(D) Oxygen consumption rates

III2
of isolated mitochondria using a

g
Clark-type oxygen electrode. The
ratio of maximal AOX respiration/ D State III kDa
(succinate + ADP + NADH) Nad6 24
nmol O2 min-1 mg-1

state III respiration, and maxi- 40

mized KCN-resistant respiration Complex I Nad7 42 Mito-encoded
30

y
(KCN+DTT+pyruvate)/state III Nad9 23
* *
respiration (succinate+ADP 20
Complex III RISP 30 Nuclear-encoded
+NADH). Statistical significance
10 Complex IV Cox2 30 Mito-encoded
was based on Student’s t test
(means ± SE, n = 3) (*P < 0.05, 0 ATPβ 55 Nuclear-encoded
**P < 0.01). (E) 20 mg and 10 mg AOX ATP1 55
of mitochondrial proteins were (KCN + DTT + pyruvate) Complex V
nmol O2 min-1 mg-1

25 ATP4 22 Mito-encoded
loaded onto SDS-PAGE for Western **
20 * ATP8 18
Blot analysis. The antibodies and
the molecular weight (kDa) are 15 RPS4 40 Mito-encoded
Ribosomal

y g
shown on the left and right sides of 10 RPS10 27 Nuclear-encoded
subunit
each blot, respectively. The arrows 5 RPL16 20 Mito-encoded
next to the molecular weight of 0 Translation LETM1 86
proteins indicate whether proteins factor
1.5 AOX/State III * UPOX 14

p
are more abundant (up arrowheads) Retrograde
* AOX 34
or less abundant (down arrow- signalling
1.0 marker OM66 66 Nuclear-encoded
Ratio

heads). Organellar origin of the

protein (nuclear encoded or mito- TIM17 30

,
chondrially encoded) is shown on 0.5
OMM TOM40 40
the right side of the blots. OMM, protein VDAC 30
0.0
outer mitochondrial membrane.
Col-0 mtran1-1/2-1 mtran1-2/2-2

that loss of mTRAN1 and mTRAN2 activates
UPRmt signaling likely through ANAC017 (31).
We confirmed up-regulation of UPRmt marker
genes in mtran1-2/2-2 and mtran1-1/2-1 by qRT-
PCR (fig. S6A).
Next, the RNA-seq dataset of mtran1-2/2-2
was compared with the transcript profiles of
22 additional Arabidopsis mutants and chem-
ical treatments affecting mitochondrial func-
tion (Fig. 3A and data S5A). The similarities
among all datasets were evaluated by com-
parison of common pairwise DEGs calculated
through the Sørensen-Dice similarity coeffi-
cient (DSC) (data S5B), followed by hierarchi-
cal clustering of the complete DSC matrix (Fig.
3A). mtran1-2/2-2 clustered most closely with
the rps10 P2 and P3 mutants, in which transcript
levels of mtSSU-subunit RPS10 are reduced (32).
This further supports that mTRANs are in-
volved in mitochondrial translation. mtran1-2/
2-2 also clustered relatively closely to anti-
mycin A (AA) (inhibitor of complex III), oligo-
mycin (inhibitor of complex V), and rotenone
(inhibitor of complex I) datasets, in line with
the observed loss of activity of these complexes
in mtran double mutants (Fig. 2). We then
visualized the similarities of nuclear transcript

Tran et al., Science 381, eadg0995 (2023) 1 September 2023 4 of 12

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

Fig. 3. Loss of mTRAN proteins A B

impairs mitochondrial
translation and mitoribosome 10 min 30 min 60 min Control
loading onto mitochondrial
mRNAs. (A) Heat map representing
common pairwise differentially
expressed genes (DEGs) among kDa

the transcriptomic datasets

70
analyzed by the Sørensen-Dice Atp1
55
similarity coefficient (DSC).
Cob
The complete matrix of DSC 40

values for mtran1-2/2-2 and 22

35 Cox2
additional datasets was hierarchi-
25
cally clustered using euclidean
distance. The subbranch including
mtran1-2/2-2 is highlighted for 15 Atp9
clarity. Color bar indicates linear 10
fold change. (B) Autoradiogram
of in organello protein synthesis for
10, 30, and 60 min using purified
C

p
mitochondria from WT, mtran1-1/2- 50 nad6 50 nad7 60 nad9
1, and mtran1-2/2-2. Sodium ace- 40 40 50
tate was used as a substrate 40
30 30
to assess bacterial contamination
30
in the controls (60 min). The 20 20
20
molecular weight (kDa) and names 10 10

g
10
of identified mitochondrially
0 0 0
encoded proteins are indicated. 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10
(C) Polysome profiling for 80 50 50
Percentage of total RNA present in all fractions

cox2 cob atp1

mitochondrially encoded tran- 70
40 40

y
scripts: OXPHOS subunits, 60
50 30 30
mitoribosomal subunit RPS4
40
and mitochondrial rRNAs rrn18/ 30 20 20
rrn26, and nuclear-encoded 20
10 10
transcript UBC21. Percentage 10
0 0 0
of total mRNA in all fractions was
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10
measured by qRT-PCR. 50 atp8 50 atp9 60 rps4
40 40 50
40
30 30
30

y g
20 20
20
10 10 10

0 0 0
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10

p
60 rrn18 80 rrn26 40 UBC21
70
50
60 30 Col-0
40 50 mtran1-1/2-1
mtran1-2/2-2
,
30 40 20
30
20
20 10
10 10
0 0 0
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10

profiles among mtran1-2/2-2, rps10 mutants, genotypes, including many UPRmt markers compensation response to reduced translation.
and prohibitin atphb3 mutant (29), which did such as AOX1a (fig. S6C). All six genes induced more than four times en-
not cluster closely to mtran1-2/2-2 (fig. S6B). We also analyzed mtran1-2/2-2 mitochon- coded tRNAs, further suggesting an attempt to
mtran1-2/2-2 shared 326 and 651 DEGs with drial and chloroplast transcript profiles (data S6). compensate for reduced translation. The mito-
rps10 P2 and P3, respectively, whereas 106 A number of mitochondrially encoded genes chondrial transcriptome was further com-
common DEGs were found between mtran1- (57) were significantly up-regulated (FC>1.4, pared with mitochondrial/chloroplast RNA
2/2-2 and atphb3. 288 DEGS were commonly Padj<0.05), with fold change up to ~4.39 for polymerase rpotmp and atphb3 mutants (29, 33).
found between mtran1-2/2-2 and both rps10 trnY (data S6A). Mitochondrial transcription In rpotmp and atphb3 mutants a mix of up-
mutants. 31 DEGs were common among all thus seems globally increased, which may be a and down-regulated genes could be observed,

Tran et al., Science 381, eadg0995 (2023) 1 September 2023 5 of 12

RES EARCH | R E S E A R C H A R T I C L E
whereas in mtran1-2/2-2 nearly all genes were
up-regulated (fig. S6D and data S6C). By con-
trast, all 117 annotated chloroplast-encoded
genes were slightly down-regulated (Padj<0.05)
(data S6B). In conclusion, loss of mTRAN1/2
resulted in increased expression of nearly all
mitochondrially encoded genes, suggesting
that the cells attempted to compensate for re-
duced mitochondrial translation.
Next, we analyzed the mitochondrial tran-
scriptome for alterations in splicing and edit-
ing (data S7). The rate of C-to-U editing was
significantly different between mtran1-2/2-2
and the WT for 262 cytosines (FDR < 0.05)
(data S7A). Most (208) of these editing sites
were relatively less edited in the mutant than
in the WT (data S7A). However, considering
that mitochondrial transcripts were more abun-
dant in mtran1-2/2-2, for 244 differentially
edited sites there were still more edited tran-
scripts in absolute terms in mtran1-2/2-2 com-
pared with that in the WT. We therefore suggest
that there are more than enough correctly
edited mitochondrial transcripts in mtran1-2/
2-2 mitochondria to allow production of the
correct proteins. For six mitochondrial tran-
script splice sites, a significant difference was
observed in the mutants versus the WT (FDR
< 0.05) (data S7B). Also here, for all transcripts
there was a higher absolute number of spliced
transcripts in the mutant, indicating that the
lower abundance of many mitochondrially en-
coded proteins (Fig. 2) is not caused by a lack
of correctly spliced/edited mitochondrial tran-
scripts. The observed similarities in nuclear
transcriptome between mtran1-2/2-2 and rps10
mutants, along with excessive expression of
mitochondrial transcripts, thus point to a de-
fect in mitochondrial translation.
mTRAN1 and mTRAN2 are required for
translation in mitochondria
To validate whether mitochondrial translation
was indeed affected in the mtran double mu-
tants, we performed in organello protein syn-
thesis assays using freshly isolated mitochondria
from Arabidopsis seedlings. Synthesized radio-
labeled mitochondrial translation products were
evaluated after 10, 30, and 60 min of translation
(Fig. 3B and fig. S7). Sodium acetate, a sub-
strate only for bacterial translation, was used to
evaluate bacterial contamination during mito-
chondrial purification. We could identify sev-
eral distinct protein bands, including Atp1,
Cob, Cox2, and Atp9, based on their molecular
mass (34, 35). Atp1, Cob, Cox2, and Atp9 had a
slower rate of translation in the mtran double-
knockout mutants than in the WT. This was
consistent with our immunoblot and MS/MS
analysis (Fig. 2E and data S1), in which we
observed a reduction in the abundance of Atp1
and Cox2 in the mtran double-knockout mu-
tants. Thus, mTRAN1 and mTRAN2 are required
for efficient translation in mitochondria.
Tran et al., Science 381, eadg0995 (2023) 1 September 2023
The efficiency of translation depends on
both translation initiation and elongation. If
translation initiation is impaired, one would
expect fewer ribosomes attached to a given
mRNA. If translation elongation is defective,
one would expect more ribosomes attached
to the mRNA, due to difficulties in completing
translation. To allow us to distinguish be-
tween these two possibilities in the mtran
double mutants, we assessed ribosome load-
ing along mitochondrially encoded transcripts
in mutants and the WT (Fig. 3C). Polysomes
are composed of mRNAs bound to two or
more ribosomes, whereas monosomes consist
of mRNAs bound to a single ribosome and/or
“vacant couples,” which are stable associations
of small and large ribosomal subunits without
binding mRNAs (36). In this analysis, poly-
somes were separated from monosomes and
free mRNAs by sucrose gradient ultracentrifu-
gation. After centrifugation, we collected 10
fractions along the gradient (fractions 1 to 10
from light to heavy) and performed qRT-PCR
to determine the percentage of mitochondrial
mRNA present in each fraction (Fig. 3C). Over-
all, all analyzed mitochondrially encoded
mRNAs of both mtran1-1/2-1 and mtran1-2/
2-2 were distributed mainly in fractions 2 and
3, whereas mRNAs in the WT were mostly
found in fractions 3 and 4 (Fig. 3C). These
shifts toward the lighter fractions indicate that
the number of ribosomes associated with the
mRNAs was lower in the mtran double mu-
tants, suggesting inefficient translation initiation/
mitoribosomal binding to mRNAs. To deter-
mine the distribution of the mitoribosomes,
we analyzed 18S rRNA (rrn18) of mtSSU and
26S rRNA (rrn26) of mtLSU. The rrn18 of
the WT was mainly found in fractions 3 and 4
whereas rrn26 was mainly in fraction 4, colo-
calizing with protein-encoding mRNAs (Fig.
3C). This indicates that most WT mRNAs are
bound by mitoribosomes and are actively trans-
lated. In mtran1-2/2-2, rrn18 was still present
mainly in fractions 3 and 4, as observed in the
WT. rrn26 was, however, spread out over
lighter fractions 2 to 4, suggesting that mtLSU
is not efficiently translating in mtran1-2/2-2.
In mtran1-1/2-1, the mitoribosomes were main-
ly found in the heavy fractions 8 and 9, and to
a lesser extent in fractions 3 and 5, as opposed
to in the lighter fractions 2 to 4 as observed in
mtran1-2/2-2.
In all cases, the mRNA levels in fraction 2
were much higher in the mutants than in the
WT. Because fraction 2 contains almost no
mitoribosomes according to our analysis (Fig.
3C and fig. S8), our results indicate that a large
proportion of mitochondrial mRNAs are likely
devoid of ribosomes in the mtran double mu-
tants. In mtran1-1/2-1, a small percentage of
several mRNAs was found in fractions 8 and 9,
indicating that they are heavily translated by
polysomes (see discussion). However, no obvi-
ous effect of loss of mTRANs could be detected
on cytosolic translation, as assessed by ribo-
some loading of nuclear-encoded UBIQUITIN-
CONJUGATING ENZYME 21 (UBC21). Together,
our findings provide evidence that mTRAN
proteins are key components of plant mito-
ribosomes required for translation initiation.
Structural modeling suggests that mTRAN
proteins have PPR protein–like structures
Because mTRAN proteins are part of the
mtSSU (data S2) and required for translation
(Fig. 3), we assessed predicted protein struc-
tures to anticipate whether the proteins could
bind mitochondrial mRNAs to help mtSSU ini-
tiate translation. Structures of mTRAN proteins
without the predicted mitochondrial target-
ing sequences (MTS) (37) were modeled using
AlphaFold (38), RoseTTAFold (39), and iTASSER
(40) (fig. S9). The modeled partial mTRAN1
p
structure extracted from the cryo-EM structure
of the mitoribosome from Brassica oleracea
var. botrytis (cauliflower) (16) is also presented
for comparison (fig. S9). Overall, the protein
structures shown by prediction tools and
Waltz et al. (16) suggest that mTRAN1 and
g
mTRAN2 have tetratricopeptide-repeat (TPR)/
PPR protein–like structures. PPR proteins are
known as RNA binding proteins involved in
RNA processing, splicing, stability, editing, and
translation (19, 41). The three programs pre-
y
dicted that mTRAN1 and mTRAN2 contain 8
to 10 PPR-like repeats. The cryo-EM structure
of the plant mitoribosome by Waltz et al. (16)
proposed that mTRAN1 is formed by 6 PPR-
like repeats, based on the number of PPR re-
peats predicted by “TPRpred” software (42).
We then attempted to find potential mTRAN-
RNA target sites using aPPRove (43) and the
PPRCODE prediction server (44) but were un-
able to find potential RNA binding sites for
y g
mTRAN1 and mTRAN2. This suggests that
mTRANs belong to a PPR-like protein class
that does not obey the standard PPR code and
may bind RNA differently.
p
mTRAN1 binds to putative mitoribosome
binding sites in mitochondrial 5' UTRs
,
As potential RNA binding sites could not be
predicted for mTRANs using the PPR code, we
searched for potential conserved mitoribosome
binding sites in the 5′ UTRs of mitochondrial
mRNAs. Using the Multiple Expectation-
maximization for Motif Enrichment (MEME)
motif search tool, we identified a potential
CUUUxU-like mitoribosome binding site in the
5′ UTRs of 26 mitochondrial mRNAs (Fig. 4A,
fig. S10, and data S8, B and C). Waltz et al. (16)
suggested that the plant mitoribosome might
recognize an AxAAA-related motif, which is
located 19 bases upstream of the start codons
in 17 mitochondrial mRNA 5′ UTRs. The MEME
motif search tool was unable to identify the
AxAAA-related motif, but we found a similar
6 of 12
RES EARCH | R E S E A R C H A R T I C L E

Fig. 4. mTRAN1 binds potential A

mitoribosome binding sites
in vitro. (A) The motif analysis
of the 5′ UTRs of mitochondrial
mRNAs by the MEME online B
motif search tool identified
Gene Motif WT sequence (5'-3') Mutated sequence (5'-3')
potential mitoribosome binding
AGAAAG
site CUUUxU and AAGAAx. The nad9 CGGAACUACAAGAAAGCUUUCUUUAU CGGAACUACACCCCCCCUCCCUUUAU
CUUUCU
other potential mitoribosome nad7 AAAAAG UGUGUAAAAAAAAAAGCUGGUGGGGC UGUGUAAAAAGGGGGGCUGGUGGGGC
binding site AxAAAG was identified
cox3_1 AAAGAA CAAUGCAAUUAAAGAACCAUCCCUUC CAAUGCAAUUGCGCGCCCAUCCCUUC
by a manual search. The motifs
cox3_2 AGAAAG CCUGUCAGAUAGAAAGAAAAUUAGAC CCUGUCAGAUGCGCGCAAGCUUAGAC
were illustrated with WebLogo
(68). (B) The 5′ UTR regions of cox2 CUUUUU AUUCCCAUCUCUUUUUUCUUGGUUGG AUUCCCAUCUCCCCCCUCUUGGUUGG
selected mitochondrial mRNAs cob/rpl5 AGAAAG GAGGAUGGGGAGAAAGCGGGCCGAGA GAGGAUGGGGGGGGGGCGGGCCGAGA
containing potential mitoribosome
WT Mut
binding sites (underlined) identi- C WT Mut

fied by the motif analysis. Binding mTRAN1 [µM] mTRAN1 [µM] mTRAN1 [µM] mTRAN1 [µM]
of recombinantly purified mTRAN1 t 25 25 .5 25 25 .5 t 25 25 .5 25 25 .5
to the fluorescently labeled WT Mu 0.6 1. 2 5 0.6 1. 2 5 WT Mu 0.6 1. 2 5 0.6 1. 2 5
RNA probes was tested by RNA

p
electrophoretic mobility shift
assays (REMSAs). The binding B B
sites were specifically changed in
the mutated probes (underlined).
(C) REMSAs confirmed the binding U U
of mTRAN1 to the 5′ UTRs of

g
mitochondrial mRNAs. Increasing nad9 nad7
amounts of mTRAN1 were incu- WT Mut WT Mut
bated with the RNA probes to mTRAN1 [µM] mTRAN1 [µM] mTRAN1 [µM]
mTRAN1 [µM]
allow estimation of the Kd binding
25 25 .5 25 25 .5

y
ut t
coefficient, according to the indi- WT M 1.25 2.5 5 10 1.2
5
2.5 5 10 WT Mu 0.6 1. 2 5 0.6 1. 2 5
cated concentrations. mTRAN1
had strong specific bindings with
nad9, nad7, cox3, cob/rpl5, and
B B
cox2 that could not be observed
when the binding sites were
mutated.
U U

cox3_1 cox3_2
WT Mut

y g
WT Mut
mTRAN1 [µM] mTRAN1 [µM]
mTRAN1 [µM] mTRAN1 [µM]
ut t 25 25 .5 25 25 .5
WT M 2.5 5 10 20 2.5 5 10 20 WT Mu 0.6 1. 2 5 0.6 1. 2 5

p
B
B

,
U U

cox2 cob/rpl5

AAGAAx-like motif in 30 5′ UTRs when in-
cluding the intercistronic UTRs (Fig. 4A, fig.
S10, and data S8C). Furthermore, we manually
found an overlapping AxAAAG-like motif in
the 5′ UTRs of 17 mitochondrial mRNAs (Fig.
4A, fig. S10, and data S8, B and C). In sum-
mary, we identified at least one CUUUxU or
AAGAAx/AxAAAG motif in the 5′ regions of all
30 mitochondrial mRNAs (fig. S10).
Next, we searched the 5′ upstream genomic
regions for hexamers enriched at any given
position in 49706 mitochondrially encoded
genes of seed plants, 4256 of mosses, 2659 of
liverworts, and 7867 of green algae, using
“PLMdetect” (45, 46) (fig. S11). In seed plants,
there is a preference for enriched motifs around
position −25 upstream of the start codon, but
the location of the motifs is more variable, ex-
tending beyond 100 bases upstream and into
the coding sequence. Similar diffuse prefer-
ences were observed in mosses and liverworts,
with peaks situated around −25/−28 or −27, re-
spectively. Analysis of 7867 green algae mito-
chondrial gene 5′ genomic regions revealed an
extremely tight peak at −25 and nearly no signal
elsewhere (fig. S11). The sequences of the motifs
enriched around −25 in seed plants were A-rich

Tran et al., Science 381, eadg0995 (2023) 1 September 2023 7 of 12

RES EARCH | R E S E A R C H A R T I C L E
and related to AAGAAA/AGAAAA, similar to
those we defined in Arabidopsis (fig. S11 and
data S9). Over half of the tested seed plant 5′
regions had such a motif around the −25 po-
sition. Also in mosses and liverworts, AAAAAA/
AAAAAG-like motifs were present in the major-
ity of 5′ UTRs around −25/−28 or −27, respectively.
In green algae almost complete conservation
of strictly A/U containing hexamers was found
around −25 of the start codon (fig. S11 and data
S9). Being based on genomic data, the pres-
ence of these motifs in mature mRNAs re-
quires further validation. The CUUUxU motif
was found in all lineages in 3 to 4% of the
genes with a different preferential position:
−28 in seed plants and −69 in green algae. Over-
all, A/U-rich motifs appear to be a common
feature in most 5′ sequences of plant mito-
chondrial mRNAs, suggesting that they may
serve as ribosome binding sites.
As mTRAN1 is the main isoform in Arabidopsis,
we performed RNA electrophoretic mobility
shift assays (REMSAs) to test whether mTRAN1
can bind to the CUUUxU or AAGAAx/AxAAAG
motifs identified in 5′ UTRs of mitochondrial
mRNAs (Fig. 4 and fig. S12). Synthesized Cy5-
labeled RNA probes, containing one or more of
the predicted mitoribosome binding motifs in
the 5′ UTRs of nad9, nad7, cox2, cox3, and rpl5/
cob were obtained (Fig. 4, A and B). Mutated
probes were also synthesized to verify the spe-
cificity of mTRAN1 binding. The REMSAs
showed that mTRAN1 could bind the WT nad9,
nad7, cox2, rpl5/cob and cox3 probes specifi-
cally, whereas the interactions were lost when
the predicted CUUUxU/AAGAAx/AxAAAG mo-
tifs were specifically mutated (Fig. 4C). Al-
though a weak shift was observed in the nad7
mutated probe, it was clear that mTRAN1
bound more strongly (lower Kd) to the WT
than mutated probe (Fig. 4C and fig. S12, C and
D). The binding of mTRAN1 appeared stronger
(lower Kd) to AAGAAx/AxAAAG motifs than to
CUUUxU motifs (fig. S12, C and D).
To establish whether mTRAN1 binds the 5′
UTRs of mitochondrial mRNAs in vivo, we
performed RNA immunoprecipitation sequenc-
ing (RIP-seq) on isolated mitochondria from
the mTRAN1-GFP lines, using mito-GFP as a
control. To discriminate the mTRAN1-specific
immunoprecipitated RNA fragments from the
GFP-control background, we used “DiffSegR”
(47, 48), which delineates boundaries of dif-
ferential regions without using preexisting
annotations, allowing precise identification of
enriched regions. The RIP-seq analysis showed
that mTRAN1 binds 5′ UTR regions, generally
preferring the UTRs to CDS (Fig. 5A). Further-
more, we observed clear evidence that the
mTRAN1-bound 5′ UTR regions contained the
same CUUUxU/AAGAAx/AxAAAG motifs for
nad9, cox2, and cox3 as used for the REMSAs
(Fig. 4 and 5B). The rps4 mRNA has no 5′ UTR,
but we found an mTRAN1-interacting segment
Tran et al., Science 381, eadg0995 (2023) 1 September 2023
at the 5′ of the mRNA, with a clear peak cov-
ering an AAAAAA motif positioned 13 bases
after the ATG (Fig. 5B). Significant mTRAN1-
bound regions were found in the 5′ UTR or
just downstream of the AUG for 17 protein-
encoding genes. A MEME motif search of these
regions confirmed enrichment of CUUUxU
and AAGAAx motifs (data S10). Together, our
findings indicate that mTRAN1 binds to CUUUxU
and AAGAAx/AxAAAG-like motifs, which are
potential mitoribosome binding sites in the 5′
regions of plant mitochondrial mRNAs, to in-
itiate translation.
mTRAN proteins are likely universal plant
mtSSU translation initiation factors
To further assess how translation of mitochon-
drial mRNAs is affected in the mtran mutants,
we performed ribosome profiling analysis
(Ribo-seq), in which short RNAse-protected
mRNA fragments covered by the translating
ribosomes (called ribosome footprints) were
isolated and sequenced. To compare the rela-
tive density of ribosome footprints on each
mitochondrial mRNA between mtran mu-
tants and WT, the calculated densities were
normalized to both mRNA length and abun-
dance determined by qRT-PCR (data S11).
As observed in our polysome fractionation
analysis, all mitochondrial mRNAs in the
mtran mutants showed much lower ribosome
loading levels compared with the WT (Fig.
6A). A single, possibly artefactual, strong foot-
print observed in ccmFN1 masked the reduc-
tion in footprints across the rest of the ccmFN1
mRNA in both mtran double mutants. In gen-
eral, mtran1-2/2-2 have lower ribosome den-
sities compared with the mtran1-1/2-1 mutant,
especially on mRNAs encoding OXPHOS pro-
teins (~0.5 times on average). Visualization of
mitoribosome footprints of representative mito-
chondrial mRNAs clearly showed that transla-
tion activity was very low in mtran1-2/2-2 and
moderately low in mtran1-1/2-1 (Fig. 6B). This
shows that not only a subset—but virtually
all—mitochondrial mRNAs depend on mTRAN
proteins for translation, indicating that mTRAN
proteins likely are “universal” (for all plant
mRNAs) mtSSU subunits involved in transla-
tion initiation.
Discussion
Phylogenetic analysis revealed that mTRANs
are likely land plant–specific proteins existing
as single-copy genes in most plant species.
Based on the more-pronounced growth defects
in mtran1 mutants than mtran2 mutants, in-
consistent detection of mTRAN2 in proteomic
studies, and higher mTRAN1 protein number
per mitochondrion (3), mTRAN1 is likely the
major isoform in Arabidopsis. When mTRAN1
is knocked out, however, mTRAN2 can appar-
ently take over most functions. When knocking
out both isoforms, a range of phenotypes was
observed, including embryo lethality (mtran1-
3/2-3) and severe (mtran1-2/2-2) to moderate
(mtran1-1/2-1) growth defects. However, it is
not as clear why mtran1-2/2-2 is (just) viable
and mtran1-3/2-3 is not. We thus propose that
mTRANs are essential for normal develop-
ment of plants.
Knocking out mTRANs reduced complex I,
III, IV, and V abundance and activity while
increasing AOX capacity. Because all complexes
containing mitochondrially encoded subunits
were strongly reduced in abundance and ac-
tivity, mTRANs must have a very fundamental
function in plant OXPHOS biogenesis. In com-
parison, mutants in mitochondrial/chloroplast
RNA polymerase rpotmp are only affected in
complex I and IV abundance (32). Together,
impaired OXPHOS-dependent ATP production
likely causes the growth retardation phenotype
in mtran mutants. This may be conveyed sys-
p
temically by negatively affecting, for example,
auxin signaling, as observed in the RNA-seq
data (data S4). An antagonistic relation between
UPRmt and auxin was suggested previously (49)
and is underlined by the up-regulation of UPRmt
markers. Down-regulation of chloroplast-
g
encoded transcripts by impairing mitochon-
drial respiration was also previously reported
(32, 50), suggesting that an unknown retrograde
pathway may reduce chloroplast transcription.
mTRAN1/2 interacted with 17 mtSSU ribo-
y
somal proteins. Two proteomic studies also
identified mTRAN1/2 as mtSSU components
using independent methods (14, 15). Our co-IP
additionally identified five plant-specific mtSSU
rPPR proteins (14, 15). We discovered proteins
interacting with mTRANs identified only in
our analysis, including nad7/nad9, which were
less abundant in mtran double mutants. We
therefore suspect that we captured nascent
Nad7/Nad9 peptides that were actively being
y g
translated by the mitoribosome. The in organ-
ello protein synthesis assays, showing reduced
translation rates in the mtran double mutants,
further support that mTRANs are key mtSSU
p
components required for efficient translation
in mitochondria.
Polysome fractionation showed that a much
,
larger proportion of mitochondrial mRNAs is
not bound by mitoribosomes in mtran double
mutants compared with the WT, suggesting
inefficient translation initiation. The reduced
ribosome loading was also observed for nad6
and rps4 mRNAs, although Nad6 showed no
changes and rps4 was more abundant at the
protein level. This suggests that Nad6 protein
may be quite stable, possibly even if not in-
corporated into the fully assembled complex I.
The ribosome fractionation showed that a
small fraction of mRNAs (generally <10%) has
high ribosome loading in mtran1-1/2-1 (Fig. 3C),
observed in fractions 8 and 9. We propose that
this small fraction of mRNAs could still be bound
by the mtran1-1/2-1 mutant mitoribosomes,
8 of 12
RES EARCH | R E S E A R C H A R T I C L E

Fig. 5. mTRAN1 binds A/U-rich A

motifs in the 5' regions of DiffSegR
mitochondrial mRNAs in vivo.
(A) RIP-seq analysis on isolated 5’UTR 3’UTR
mitochondria from mTRAN1- Gene nad9
GFP plants versus mito-GFP as
control. Enriched mTRAN1-binding
fragments of mitochondrially DiffSegR
encoded nad9, cox2, cox3, and
rps4 mRNAs were visualized using Gene 5’UTR 3’UTR
cox2
integrative genomics viewer
(IGV). Overrepresentation of
mTRAN1-specific immunoprecipi- DiffSegR
tated RNA fragments versus
the GFP-control background
Gene 3’UTR 5’UTR
was calculated by DiffSegR cox3
(Padj<0.05). mTRAN1-specifically
bound regions are marked in
green (highlighted by black DiffSegR
arrows), underrepresented

p
regions in purple, and nonen- rps4
Gene
riched regions in gray. The 3’UTR
“gene” track shows the gene
structure of the specific genes B Data range: -0.591-3.921
with untranslated regions (UTR)
log2FC
and coding sequences (CDS).

g
(B) Closeup of mTRAN1-binding DiffSegR
-20 -14
sites in the 5′ UTRs of nad9, Sequence
cox2, and cox3 and in the 5′ end nad9_5’UTR
Gene nad9
of the rps4 CDS. Red bar charts

y
indicate log2FC>0 (Padj<0.05) Data range: -0.009-1.507
obtained by DiffSegR, which
represents the mTRAN1-specific
log2FC
immunoprecipitated RNA frag-
ments statistically different from DiffSegR
-119
the GFP-control background. Sequence
“Data range” indicates minimum Gene cox2_5’UTR
and maximum values of the
respective log2FC tracks. The Data range: 0-1.714
A/U-rich motifs are highlighted log2FC

y g
for clarity. The positions of the
DiffSegR
motifs from the start codon
-83 -121
AUG are indicated. The black
Sequence
arrows indicate the direction of
Gene cox3_5’UTR

p
the genes.
Data range: -1.216-1.659

log2FC
DiffSegR
+13 ,
Sequence
Gene rps4

perhaps due to different secondary/tertiary
structure of the 5′ UTR allowing better mito-
ribosome entry. As a result, these rare “bind-
able” mRNAs may then be heavily translated
by all available mitoribosomes, partially com-
pensating for the lack of initiation on the vast
majority of “unbindable” mRNAs. This com-
pensation in the mtran1-1/2-1 mutant (in which
truncated mTRAN1 protein is detected) likely
explains the higher vigor of mtran1-1/2-1 com-
pared with mtran1-2/2-2 plants.
The question arises as to how mTRAN pro-
teins play such a fundamental role in mitochon-
drial translation. The polysome fractionation
suggests that the problem lies in the very first
steps, where the mitoribosome detects and
binds the mRNA. Plant mitochondrial mRNAs
lack a Shine-Dalgarno (SD) sequence, and plant
mitochondrial rRNAs do not have anti-SD se-
quences (13). Our data suggest that mTRANs
are recognition factors, which would mean that
in plant mitochondria a protein-mRNA inter-
action mediates ribosome binding, rather than
an rRNA-mRNA interaction. A cryo-EM study
of the structure of plant mitoribosomes (16)
proposed that mTRAN1 (named as ms83 and
rPPR10 by the authors) is a PPR protein, which

Tran et al., Science 381, eadg0995 (2023) 1 September 2023 9 of 12

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

Fig. 6. mTRAN proteins are A

RPKM normalized by mRNA abundance

likely universal mtSSU
translation initiation 1.5
factors. (A) Loss of mTRANs
decreases mitochondrial
1.0
ribosome footprints. Reads
per kilobase million (RPKM)
normalized by mRNA abundance
0.5
of the mtran double mutants
were normalized to WT.
Means ± SE (n = 2). (B) Density 0.0
of mitochondrial ribosome

ttb
p1
p4
p6
p8

cc 9
cc C

x1
x2

m 3

d1
d2

na 3
R

L
d4
d5
d6
d7
d9

6
l2

rp l5
2
s3
s4
s7
m C
m 1
2
cc FN
FN
cc F

x
p

d
co

d4

l1

s1
m

at

rp
rp
co
co
co

rp
rp
rp
na
na
na

na
na
na
na
na
at
at
at
at
at

m
m
footprints on mitochondrially

rp
encoded cox2, atp1, atp9, Col-0 mtran1-1/2-1 mtran1-2/2-2
nad9 and nad7 was obtained
by integrative genomics B
viewer (IGV) software using 5’UTR cox2 3’UTR
Col-0
autoscale.
mtran1-1/2-1
mtran1-2/2-2

p
5’UTR atp1 3’UTR
Col-0
mtran1-1/2-1
mtran1-2/2-2

3’UTR atp9 5’UTR

Col-0

g
mtran1-1/2-1
mtran1-2/2-2

5’UTR nad9 3’UTR

y
Col-0
mtran1-1/2-1
mtran1-2/2-2

3’UTR nad7 5’UTR

Col-0
mtran1-1/2-1
mtran1-2/2-2

y g
are known RNA binding proteins. Indeed, an- codon) according to PastDB (56). However, region that might recognize an AxAAA-related
other study identified mTRAN1 as an mRNA- protein structure prediction tools AlphaFold, motif of mitochondrial mRNA 5′ UTRs, thus
binding protein (51). However, mTRAN proteins RoseTTAFold, and iTASSER showed that acting as a SD–anti-SD–like recognition system.
are far from “classical” PPR proteins. First, the mTRAN proteins are likely to have α-solenoid/ Additionally, their motif analysis suggested that

p
largest PPR database based on 44,562 PPR PPR protein–like structures (57). Unfortunate- the 5′ UTRs of only 17 mitochondrial mRNAs
protein sequences from >1000 transcriptomes ly, potential RNA binding sites could not be contained a loosely conserved AxAAA consen-
(52, 53) does not identify mTRANs as PPR predicted by aPPRove (43) and PPRCODE (44). sus. However, our study showed that mTRANs
proteins. Furthermore, “TPRpred” predicts other
,
This further suggests that mTRAN proteins are are required for translation of all mitochon-
rPPRs as PPR proteins with 100% confidence, α-solenoid proteins that bind RNA differently drially encoded genes, of which many do not
yet only gives a confidence score of 4 to 14% to classical PPR proteins. Based on available contain such an AxAAA 5′ UTR motif but rather
for mTRAN1/2. Additionally, the gene struc- protein structure predictions, 6 to 10 repeated a CUUUxU or AAGAAx motif.
ture of mTRAN with 10 introns is highly aber- motifs were found in the mTRAN protein struc- Indeed, mitoribosome loading was lower for
rant compared with classical PPR genes (fig. ture, which suggests that they could interact all mRNAs in mtran mutants. Thus, mTRAN
S3A), which usually have 0 to 1 introns (54). with a mRNA motif of around 6 to 10 bases, proteins may be “universal” recognition fac-
Introns are associated with higher mRNA sta- matching our identified CUUUxU/AAGAAx/ tors for all mitochondrial mRNAs, rather than
bility, and indeed mTRANs have longer average AxAAAG motifs. mTRANs were suggested to mRNA-specific initiation cofactors (9, 10). The
transcript half-lives of 4.4 hours, as compared sit in a cleft where the incoming mRNA may universally reduced mitoribosome loading in
with 1.7 hours for the average PPR gene with 0 be positioned in the cryo-EM mitoribosome the mtran double mutants is also different
to 1 introns (55), resulting in 1.8 to 2.4 times structure (16). It should be noted that the reso- from the ribosome loading in rps10 mutants
higher relative expression (data S12). Never- lution of several regions in the currently available (58), a mtSSU subunit thought to be involved
theless, only one mTRAN1 splice form seems plant mitoribosome structure is, however, too in elongation (59). In the rps10 mutants, varia-
dominant while two are expressed for mTRAN2 low to accurately position mTRANs. Waltz et al. ble increased or decreased ribosome loading
(of which one is low abundant with an early stop (16) hypothesized that mTRAN1 is part of the was observed on mRNAs encoding OXPHOS

Tran et al., Science 381, eadg0995 (2023) 1 September 2023 10 of 12

RES EARCH | R E S E A R C H A R T I C L E
components, but higher ribosome loading
was found for translation- and cytochrome
c maturation–related mRNAs. This difference
further suggests that mTRANs are involved in
translation initiation and not elongation.
Most Arabidopsis mitochondrial mRNAs
have long 5′ UTRs, which is different from
mammals but relatively similar to yeast (9).
Our results indicate that the putative mito-
ribosome binding sites CUUUxU and AAGAAx/
AxAAAG in plant mitochondrial mRNAs are
not at a particular distance from the start
codon (figs. S10 and S11), though with some
preference for −25 bases upstream of AUG.
Therefore, we hypothesize that either the 2/3D
structure of the 5′ UTRs may bring the ribo-
some binding sites physically closer to the
start codon, or that plant mtSSU can effi-
ciently scan long 5′ UTRs to find the correct
AUG, perhaps with the help of other mRNA-
specific cofactors (60). In green algae, an A/U-
motif is conserved around −25 bases from the
start codon (fig. S11); yet mTRAN genes are
absent. So perhaps mTRANs evolved in land
plants to compensate for increasing variability
in mitochondrial ribosome binding site posi-
tion, or conversely, the appearance of mTRANs
allowed this variability to occur. These findings
indicate that translational initiation by mito-
chondrial ribosomes occurs in a different way in
plants as compared with in fungi and animals
(12, 61).
Materials and methods summary
Plants were grown under long day condition
(16h light/8h dark, approximately 120 mmol
photons m−2s−1). A. thaliana Col-0 was used
as the WT. The T-DNA insertion lines SALK_
044671, GABI_915G12, WiscDsLox485-488E21,
SALK_054298, SALK_096907, and SALK_099373
were obtained from the European Arabidopsis
Stock Centre. The double-knockout mutants
mtran1-1/2-1 and mtran1-2/2-2 were obtained
by crossing. Stable transgenic and complemen-
tation lines were generated by transforming
pB7FWG2-mTRAN1/2 into the WT or mtran1-2/
2-2 by floral dipping, respectively. Plant mito-
chondria were purified as described in Tran et al.
(62) for further analysis, e.g., BN-PAGE, West-
ern blot, MS/MS analysis, RIP-seq, and in or-
ganello translation assays. For whole genome
RNA-seq analysis, 11-day-old WT and 18-day-
old mtran1-2/2-2 seedlings grown on MS agar
plates were harvested at developmental growth
stage 1.04 (26). Total RNA was isolated to gen-
erate RNA-seq libraries for sequencing. Bioin-
formatics and statistical analysis were carried
out according to Baudry et al. (63). Polysomal
RNA purified as described in (64) was used for
qRT-PCR. mTRAN1 protein was expressed and
purified from E. coli. 5′Cy5 labeled RNA probes
were obtained from Sigma-Aldrich. REMSAs
were performed as described previously (65, 66).
Mitoribosome footprints were prepared from
Tran et al., Science 381, eadg0995 (2023) 1 September 2023
Arabidopsis flower buds as previously described
(8). Isolated mitochondria from mito-GFP and
mTRAN1-GFP lines were subjected to immuno-
precipitations for RIP-seq analysis as previously
described (67). The RIP-seq data were processed,
mapped against Col-0 mitogenome, and ana-
lyzed using DiffSegR to identify mTRAN1-
binding sequences. PLMdetect (45) was used
to analyze preferred hexameric motif locations
in the −300/+200 genomic upstream regions
from the start codon of mitochondrial protein-
coding genes from all species available in the
NCBI database. A complete “Materials and
Methods” section is provided in the Supple-
mentary Information.
RE FERENCES AND NOTES
1. O. Schmidt, N. Pfanner, C. Meisinger, Mitochondrial protein
import: From proteomics to functional mechanisms. Nat. Rev.
Mol. Cell Biol. 11, 655–667 (2010). doi: 10.1038/nrm2959;
pmid: 20729931
2. R. S. Rao et al., The proteome of higher plant mitochondria.
Mitochondrion 33, 22–37 (2017). doi: 10.1016/
j.mito.2016.07.002; pmid: 27405097
3. P. Fuchs et al., Single organelle function and organization as
estimated from Arabidopsis mitochondrial proteomics. Plant J.
101, 420–441 (2020). doi: 10.1111/tpj.14534; pmid: 31520498
4. J. P. Mower, Variation in protein gene and intron content
among land plant mitogenomes. Mitochondrion 53, 203–213
(2020). doi: 10.1016/j.mito.2020.06.002; pmid: 32535166
5. I. M. Møller, A. G. Rasmusson, O. Van Aken, Plant mitochondria -
past, present and future. Plant J. 108, 912–959 (2021).
doi: 10.1111/tpj.15495; pmid: 34528296
6. N. Mai, Z. M. Chrzanowska-Lightowlers, R. N. Lightowlers, The
process of mammalian mitochondrial protein synthesis.
Cell Tissue Res. 367, 5–20 (2017). doi: 10.1007/s00441-016-
2456-0; pmid: 27411691
7. B. J. Greber, N. Ban, Structure and Function of the
Mitochondrial Ribosome. Annu. Rev. Biochem. 85, 103–132
(2016). doi: 10.1146/annurev-biochem-060815-014343
pmid: 27023846
8. N. Planchard et al., The translational landscape of Arabidopsis
mitochondria. Nucleic Acids Res. 46, 6218–6228 (2018).
doi: 10.1093/nar/gky489; pmid: 29873797
9. K. S. Derbikova, S. A. Levitsky, I. V. Chicherin,
E. N. Vinogradova, P. A. Kamenski, Activation of Yeast
Mitochondrial Translation: Who Is in Charge? Biochemistry
(Mosc.) 83, 87–97 (2018). doi: 10.1134/S0006297918020013;
pmid: 29618295
10. N. S. Green-Willms, T. D. Fox, M. C. Costanzo, Functional
interactions between yeast mitochondrial ribosomes and
mRNA 5′ untranslated leaders. Mol. Cell. Biol. 18, 1826–1834
(1998). doi: 10.1128/MCB.18.4.1826; pmid: 9528754
11. J. Montoya, D. Ojala, G. Attardi, Distinctive features of the 5′-
terminal sequences of the human mitochondrial mRNAs.
Nature 290, 465–470 (1981). doi: 10.1038/290465a0;
pmid: 7219535
12. E. Kummer et al., Unique features of mammalian mitochondrial
translation initiation revealed by cryo-EM. Nature 560,
263–267 (2018). doi: 10.1038/s41586-018-0373-y;
pmid: 30089917
13. T. Hazle, L. Bonen, Comparative analysis of sequences
preceding protein-coding mitochondrial genes in flowering
plants. Mol. Biol. Evol. 24, 1101–1112 (2007). doi: 10.1093/
molbev/msm030; pmid: 17301062
14. N. Rugen, H. Straube, L. E. Franken, H. P. Braun, H. Eubel,
Complexome Profiling Reveals Association of PPR Proteins
with Ribosomes in the Mitochondria of Plants. Mol. Cell.
Proteomics 18, 1345–1362 (2019). doi: 10.1074/mcp.
RA119.001396; pmid: 31023727
15. F. Waltz et al., Small is big in Arabidopsis mitochondrial
ribosome. Nat. Plants 5, 106–117 (2019). doi: 10.1038/s41477-
018-0339-y; pmid: 30626926
16. F. Waltz, H. Soufari, A. Bochler, P. Giegé, Y. Hashem, Cryo-EM
structure of the RNA-rich plant mitochondrial ribosome.
Nat. Plants 6, 377–383 (2020). doi: 10.1038/s41477-020-
0631-5; pmid: 32251374
17. I. D. Small, N. Peeters, The PPR motif - a TPR-related motif
prevalent in plant organellar proteins. Trends Biochem. Sci. 25,
45–47 (2000). doi: 10.1016/S0968-0004(99)01520-0;
pmid: 10664580
18. S. Manna, An overview of pentatricopeptide repeat proteins
and their applications. Biochimie 113, 93–99 (2015).
doi: 10.1016/j.biochi.2015.04.004; pmid: 25882680
19. C. Schmitz-Linneweber, I. Small, Pentatricopeptide repeat
proteins: A socket set for organelle gene expression.
Trends Plant Sci. 13, 663–670 (2008). doi: 10.1016/
j.tplants.2008.10.001; pmid: 19004664
20. A. Moutinho, P. J. Hussey, A. J. Trewavas, R. Malhó, cAMP acts
as a second messenger in pollen tube growth and
reorientation. Proc. Natl. Acad. Sci. U.S.A. 98, 10481–10486
(2001). doi: 10.1073/pnas.171104598; pmid: 11517303
21. I. Al-Younis, A. Wong, C. Gehring, The Arabidopsis thaliana
K(+)-uptake permease 7 (AtKUP7) contains a functional
cytosolic adenylate cyclase catalytic centre. FEBS Lett. 589,
3848–3852 (2015). doi: 10.1016/j.febslet.2015.11.038;
pmid: 26638082
22. N. K. Centeno-González, H. I. Martínez-Cabrera,
H. Porras-Múzquiz, E. Estrada-Ruiz, Late Campanian fossil of a
legume fruit supports Mexico as a center of Fabaceae
radiation. Commun. Biol. 4, 41 (2021). doi: 10.1038/s42003-
020-01533-9; pmid: 33446929
23. S. A. Rensing et al., An ancient genome duplication contributed
to the abundance of metabolic genes in the moss
Physcomitrella patens. BMC Evol. Biol. 7, 130 (2007).
p
doi: 10.1186/1471-2148-7-130; pmid: 17683536
24. C. M. Hooper, I. R. Castleden, S. K. Tanz, N. Aryamanesh,
A. H. Millar, SUBA4: The interactive data analysis centre for
Arabidopsis subcellular protein locations. Nucleic Acids Res.
45, D1064–D1074 (2017). doi: 10.1093/nar/gkw1041;
pmid: 27899614
25. J. Senkler et al., The mitochondrial complexome of Arabidopsis
thaliana. Plant J. 89, 1079–1092 (2017). doi: 10.1111/tpj.13448;
g
pmid: 27943495
26. D. C. Boyes et al., Growth stage-based phenotypic analysis of
Arabidopsis: A model for high throughput functional genomics
in plants. Plant Cell 13, 1499–1510 (2001). doi: 10.1105/
TPC.010011; pmid: 11449047
27. G. Karimova, J. Pidoux, A. Ullmann, D. Ladant, A bacterial two-
y
hybrid system based on a reconstituted signal transduction
pathway. Proc. Natl. Acad. Sci. U.S.A. 95, 5752–5756 (1998).
doi: 10.1073/pnas.95.10.5752; pmid: 9576956
28. O. Van Aken et al., Mitochondrial type-I prohibitins of
Arabidopsis thaliana are required for supporting proficient
meristem development. Plant J. 52, 850–864 (2007).
doi: 10.1111/j.1365-313X.2007.03276.x; pmid: 17883375
29. O. Van Aken, E. Ford, R. Lister, S. Huang, A. H. Millar,
Retrograde signalling caused by heritable mitochondrial
dysfunction is partially mediated by ANAC017 and improves
plant performance. Plant J. 88, 542–558 (2016). doi: 10.1111/
tpj.13276; pmid: 27425258
30. D. C. Logan, C. J. Leaver, Mitochondria-targeted GFP highlights
y g
the heterogeneity of mitochondrial shape, size and movement
within living plant cells. J. Exp. Bot. 51, 865–871 (2000).
doi: 10.1093/jexbot/51.346.865; pmid: 10948212
31. S. M. Kacprzak, A. Dahlqvist, O. Van Aken, The transcription
factor ANAC017 is a key regulator of mitochondrial proteotoxic
p
stress responses in plants. Philos. Trans. R. Soc. B 375,
20190411 (2020). doi: 10.1098/rstb.2019.0411;
pmid: 32362262
32. A. Adamowicz-Skrzypkowska, M. Kwasniak-Owczarek,
O. Van Aken, U. Kazmierczak, H. Janska, Joint inhibition of
,
mitochondrial complex IV and alternative oxidase by genetic or
chemical means represses chloroplast transcription in
Arabidopsis. Philos. Trans. R. Soc. B 375, 20190409 (2020).
doi: 10.1098/rstb.2019.0409; pmid: 32362248
33. K. Kühn et al., Decreasing electron flux through the
cytochrome and/or alternative respiratory pathways triggers
common and distinct cellular responses dependent on growth
conditions. Plant Physiol. 167, 228–250 (2015). doi: 10.1104/
pp.114.249946; pmid: 25378695
34. M. Kwasniak-Owczarek, A. Tomal, H. Janska, Assessment of
Protein Synthesis in Mitochondria Isolated from Rosette
Leaves and Liquid Culture Seedlings of Arabidopsis.
Methods Mol. Biol. 2363, 183–197 (2022). doi: 10.1007/
978-1-0716-1653-6_14; pmid: 34545494
35. R. Kolli, J. Soll, C. Carrie, OXA2b is Crucial for Proper
Membrane Insertion of COX2 during Biogenesis of Complex IV
in Plant Mitochondria. Plant Physiol. 179, 601–615 (2019).
doi: 10.1104/pp.18.01286; pmid: 30487140
36. M. Noll, B. Hapke, M. H. Schreier, H. Noll, Structural dynamics
of bacterial ribosomes. I. Characterization of vacant couples
11 of 12
RES EARCH | R E S E A R C H A R T I C L E
and their relation to complexed ribosomes. J. Mol. Biol. 75,
281–294 (1973). doi: 10.1016/0022-2836(73)90021-1;
pmid: 4580677
37. Y. Fukasawa et al., MitoFates: Improved prediction of
mitochondrial targeting sequences and their cleavage sites.
Mol. Cell. Proteomics 14, 1113–1126 (2015). doi: 10.1074/mcp.
M114.043083; pmid: 25670805
38. J. Jumper et al., Highly accurate protein structure prediction
with AlphaFold. Nature 596, 583–589 (2021). doi: 10.1038/
s41586-021-03819-2; pmid: 34265844
39. M. Baek et al., Accurate prediction of protein structures and
interactions using a three-track neural network. Science 373,
871–876 (2021). doi: 10.1126/science.abj8754;
pmid: 34282049
40. J. Yang, Y. Zhang, I-TASSER server: New development for protein
structure and function predictions. Nucleic Acids Res. 43, W174–
W181 (2015). doi: 10.1093/nar/gkv342; pmid: 25883148
41. E. Delannoy, W. A. Stanley, C. S. Bond, I. D. Small,
Pentatricopeptide repeat (PPR) proteins as sequence-
specificity factors in post-transcriptional processes in
organelles. Biochem. Soc. Trans. 35, 1643–1647 (2007).
doi: 10.1042/BST0351643; pmid: 18031283
42. M. R. Karpenahalli, A. N. Lupas, J. Söding, TPRpred: A tool for
prediction of TPR-, PPR- and SEL1-like repeats from protein
sequences. BMC Bioinformatics 8, 2 (2007). doi: 10.1186/1471-
2105-8-2; pmid: 17199898
43. T. Harrison, J. Ruiz, D. B. Sloan, A. Ben-Hur, C. Boucher,
aPPRove: An HMM-Based Method for Accurate Prediction of
RNA-Pentatricopeptide Repeat Protein Binding Events.
PLOS ONE 11, e0160645 (2016). doi: 10.1371/
journal.pone.0160645; pmid: 27560805
44. J. Yan et al., Delineation of pentatricopeptide repeat codes for
target RNA prediction. Nucleic Acids Res. 47, 3728–3738
(2019). doi: 10.1093/nar/gkz075; pmid: 30753696
45. J. Rozière et al., Plant-PLMview: a database for identifying
cis-regulatory sequences with preferential positions in
gene-proximal regions of plants. bioRxiv, 2022.2012.2020.521192
(2022). doi: 10.1101/2022.12.20.521192
46. GitLab, PLMdetect_tool (Project ID 5699, 2022); https://
forgemia.inra.fr/GNet/plmdetect/plmdetect_tool.
[accessed 24 May 2023].
47. A. Liehrmann, E. Delannoy, B. Castandet, G. Rigaill, DiffSegR:
An RNA-Seq data driven method for differential expression analysis
using changepoint detection. bioRxiv, 2023.2006.2005.543691
(2023). doi: 10.1101/2023.06.05.543691
48. GitHub, DiffSegR (2023); https://aliehrmann.github.io/
DiffSegR/index.html. [accessed 24 May 2023].
49. A. Ivanova et al., A Functional Antagonistic Relationship
between Auxin and Mitochondrial Retrograde Signaling
Regulates Alternative Oxidase1a Expression in Arabidopsis.
Plant Physiol. 165, 1233–1254 (2014). doi: 10.1104/
pp.114.237495; pmid: 24820025
50. Y. O. Zubo, T. V. Potapova, V. I. Tarasenko, T. Börner,
Y. M. Konstantinov, The rate of transcription in Arabidopsis
chloroplasts depends on activity of alternative electron
transfer pathway in mitochondria. Dokl. Biochem. Biophys.
455, 76–79 (2014). doi: 10.1134/S1607672914020094;
pmid: 24795105
51. M. Bach-Pages et al., Discovering the RNA-Binding Proteome
of Plant Leaves with an Improved RNA Interactome Capture
Method. Biomolecules 10, 661 (2020). doi: 10.3390/
biom10040661; pmid: 32344669
Tran et al., Science 381, eadg0995 (2023) 1 September 2023
52. B. Gutmann et al., The Expansion and Diversification of
Pentatricopeptide Repeat RNA-Editing Factors in Plants.
Mol. Plant 13, 215–230 (2020). doi: 10.1016/
j.molp.2019.11.002; pmid: 31760160
53. The University of Western Australia, ARC Centre of Excellence in
Plant Energy Biology, PPR: 446262 Pentatricopeptide repeat
proteins from 1121 species (2022); https://ppr.plantenergy.
uwa.edu.au/. [accessed 10 November 2021].
54. C. Lurin et al., Genome-wide analysis of Arabidopsis
pentatricopeptide repeat proteins reveals their essential role in
organelle biogenesis. Plant Cell 16, 2089–2103 (2004).
doi: 10.1105/tpc.104.022236; pmid: 15269332
55. R. Narsai et al., Genome-wide analysis of mRNA decay rates
and their determinants in Arabidopsis thaliana. Plant Cell 19,
3418–3436 (2007). doi: 10.1105/tpc.107.055046;
pmid: 18024567
56. G. Martín, Y. Márquez, F. Mantica, P. Duque, M. Irimia,
Alternative splicing landscapes in Arabidopsis thaliana across
tissues and stress conditions highlight major functional
differences with animals. Genome Biol. 22, 35 (2021).
doi: 10.1186/s13059-020-02258-y; pmid: 33446251
57. D. Fournier et al., Functional and genomic analyses of alpha-
solenoid proteins. PLOS ONE 8, e79894 (2013). doi: 10.1371/
journal.pone.0079894; pmid: 24278209
58. M. Kwasniak-Owczarek, U. Kazmierczak, A. Tomal,
P. Mackiewicz, H. Janska, Deficiency of mitoribosomal S10
protein affects translation and splicing in Arabidopsis
mitochondria. Nucleic Acids Res. 47, 11790–11806 (2019).
doi: 10.1093/nar/gkz1069; pmid: 31732734
59. S. Hermann-Le Denmat, M. Sipiczki, P. Thuriaux, Suppression
of yeast RNA polymerase III mutations by the URP2 gene
encoding a protein homologous to the mammalian ribosomal
protein S20. J. Mol. Biol. 240, 1–7 (1994). doi: 10.1006/
jmbi.1994.1412; pmid: 8021936
60. N. Haïli et al., The MTL1 Pentatricopeptide Repeat Protein Is
Required for Both Translation and Splicing of the Mitochondrial
NADH DEHYDROGENASE SUBUNIT7 mRNA in Arabidopsis.
Plant Physiol. 170, 354–366 (2016). doi: 10.1104/pp.15.01591;
pmid: 26537562
61. V. Singh et al., Structural basis of LRPPRC-SLIRP-dependent
translation by the mitoribosome. bioRxiv 2022.2006.2020.496763
[Preprint] (2022); doi: 10.1101/2022.06.20.496763
62. H. C. Tran, O. Van Aken, Purification of Leaf Mitochondria
from Arabidopsis thaliana Using Percoll Density Gradients.
Methods Mol. Biol. 2363, 1–12 (2022). doi: 10.1007/978-1-
0716-1653-6_1; pmid: 34545481
63. K. Baudry, E. Delannoy, C. Colas des Francs-Small, Analysis of
the Plant Mitochondrial Transcriptome. Methods Mol. Biol.
2363, 235–262 (2022). doi: 10.1007/978-1-0716-1653-6_17;
pmid: 34545497
64. A. Barkan, Nuclear Mutants of Maize with Defects in Chloroplast
Polysome Assembly Have Altered Chloroplast RNA Metabolism.
Plant Cell 5, 389–402 (1993). doi: 10.2307/3869720; pmid: 12271069
65. P. Kindgren, A. Yap, C. S. Bond, I. Small, Predictable alteration
of sequence recognition by RNA editing factors from
Arabidopsis. Plant Cell 27, 403–416 (2015). doi: 10.1105/
tpc.114.134189; pmid: 25649437
66. M. Schallenberg-Rüdinger, P. Kindgren, A. Zehrmann, I. Small,
V. Knoop, A DYW-protein knockout in Physcomitrella affects
two closely spaced mitochondrial editing sites and causes a
severe developmental phenotype. Plant J. 76, 420–432 (2013).
doi: 10.1111/tpj.12304; pmid: 23909746
67. T. T. Nguyen et al., A Case of Gene Fragmentation in Plant
Mitochondria Fixed by the Selection of a Compensatory Restorer of
Fertility-Like PPR Gene. Mol. Biol. Evol. 38, 3445–3458 (2021).
doi: 10.1093/molbev/msab115; pmid: 33878189
68. G. E. Crooks, G. Hon, J. M. Chandonia, S. E. Brenner, WebLogo:
A sequence logo generator. Genome Res. 14, 1188–1190
(2004). doi: 10.1101/gr.849004; pmid: 5173120
69. Y. Perez-Riverol et al., The PRIDE database resources in 2022:
A hub for mass spectrometry-based proteomics evidences.
Nucleic Acids Res. 50, D543–D552 (2022). doi: 10.1093/nar/
gkab1038; pmid: 34723319
70. ModelArchive, doi: 10.5452/ma-9aynd
AC KNOWLED GME NTS
We thank O. Gustafsson for technical assistance for confocal
microscopy analysis, K Flärdh for his advice on the cAMP
complementation assay, M. Kwasniak-Owczarek and H. Janska for
their advice on in organello protein synthesis and polysome
profiling assays, POPS platform for the RNA-seq analysis,
E. Krupinska and W. Knecht for their advice on recombinant
protein expression, P. Kindgren for his advice on REMSAs, and
I. Max Møller for his critical comments on the manuscript.
Funding: This project was supported by the Swedish Research
Council (Vetenskapsrådet 2017-03854; 2021-04358), Crafoord
Foundation (20170862; 20190868), Carl Trygger Foundation
(CTS 17 487, 22 1981), Kungliga Fysiografiska Sällskapet i Lund
p
and Jörgen Lindström's Foundation, and by grants from the
French National Research Agency ANR-20-CE11-0021 to H.M.
and ANR-20-CE20-0004 to B.C. The POPS platform, the IPS2, and
IJPB benefit from the support of Saclay Plant Sciences-SPS
(ANR-17-EUR- 0007). Author contributions: O.V.A., H.C.T., E.D.,
H.M., and A.G.R. conceived and planned the project. H.C.T., V.S.,
S.L., K.B., K.K., A.L-A., B.C, K.S., and C.W. performed experiments.
H.C.T., H.M., E.D., C.W., F.L., V.B., A.L., and O.V.A. analyzed data,
g
and H.C.T. and O.V.A wrote the manuscript with input from all
coauthors. Competing interests: The authors have no competing
interests to declare. Data and materials availability: All data
are available in the manuscript, supplementary materials, or the
following repositories. The RNA-Seq project: Gene Expression
Omnibus (31) GSE186586. Ribo-seq: ArrayExpress E-MTAB-13059;
y
RIP-Seq: ArrayExpress E-MTAB-13060. The mass spectrometry
proteomics data have been deposited to the ProteomeXchange
Consortium via the PRIDE (69) partner repository with the
dataset identifier PXD043729. The predicted mTRAN structural
models are available in ModelArchive (70). License information:
Copyright © 2023 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.adg0995
y g
Materials and Methods
Figs. S1 to S12
References (71–99)
MDAR Reproducibility Checklist
Data S1 to S13
p
Submitted 2 December 2022; resubmitted 8 June 2023
Accepted 2 August 2023
10.1126/science.adg0995
12 of 12
RES EARCH

◥ increased intrahepatic apoB lipidation. apoB is

RESEARCH ARTICLE SUMMARY lipidated in the endoplasmic reticulum (ER) by
microsomal triglyceride transfer protein (MTP),
ATHEROSCLEROSIS which incorporates neutral lipids onto nascent
apoB. Transfecting human primary hepato-
Intracellular tPA–PAI-1 interaction determines cytes with a plasmid encoding tPA led to lower
secretion of newly synthesized [3H]-labeled
VLDL assembly in hepatocytes apoB. Proximity ligation, confocal imaging,
and immunoprecipitation assays revealed that
Wen Dai*, Heng Zhang, Hayley Lund, Ziyu Zhang, Mark Castleberry, Maya Rodriguez, tPA interacts with apoB in the hepatocyte ER.
George Kuriakose, Sweta Gupta, Magdalena Lewandowska, Hayley R. Powers, Swati Valmiki, In addition, recombinant tPA interacts with
Jieqing Zhu, Amy D. Shapiro, M. Mahmood Hussain, José A. López, Mary G. Sorci-Thomas, solid-phase immobilized LDL, inhibits MTP-
Roy L. Silverstein, Henry N. Ginsberg, Daisy Sahoo, Ira Tabas*, Ze Zheng* apoB interaction, and reduces neutral lipid
transfer to apoB. Moreover, the serine pro-
tease inactive tPA (S513A) also interacts with
INTRODUCTION: Tissue plasminogen activator sion was silenced or deleted in the hepatocytes solid-phase immobilized LDL, reducing both
(tPA) is a serine protease that initiates fibrinol- of mice. These manipulations resulted in higher apoB secretion by human primary hepatocytes
ysis to remove excessive blood clots and restore plasma apoB and cholesterol levels, indepen- and MTP-mediated lipid transfer activity to
blood flow. Intravenous infusion of recombinant dent of any changes in hepatic low-density the same degree as wild-type tPA, which in-
tPA is approved as a thrombolytic therapy in lipoprotein receptor (LDLR) or apolipoprotein dicates that this action of tPA is independent

thrombotic cardiovascular diseases, including
ischemic stroke and myocardial infarction. Low
plasma tPA activity is associated with a higher
risk of atherosclerotic cardiovascular disease and
atherogenic apolipoprotein B (apoB)–lipoprotein
cholesterol levels in humans. However, whether
E (apoE) expression or Apob mRNA level. The
higher plasma cholesterol in these mice was
distributed in the very-low-density lipopro-
tein (VLDL) and low-density lipoprotein (LDL)
fractions. In human primary hepatocytes,
silencing tPA increased the secretion of new-
ly synthesized [3H]-labeled apoB despite no
p
of its protease activity. The tPA-LDL inter-
action is inhibited by antibodies against
the Kringle 2 (K2) domain of tPA, the MTP-
interacting regions at the N terminus of apoB,
and the lysine analog tranexamic acid. Fur-
ther, deleting the K2 domain or mutating the

low tPA elevates apoB-lipoprotein cholesterol
is unknown.
RATIONALE: Given the central role of hepato-
cytes in apoB-lipoprotein production and recent
change in Apob mRNA. Thus, hepatocyte-
tPA deficiency increases the secretion of
apoB-lipoproteins.
Lipidation is a key factor determining the
g
lysine-binding site in the K2 domain of tPA
abrogates the effects of tPA on limiting apoB
secretion. These data indicate that tPA, par-
tially through the lysine-binding site on its
K2 domain, binds to the N terminus of apoB,

studies showing that hepatocytes synthesize
tPA, we sought to identify possible links be-
tween tPA and apoB-lipoprotein assembly and
secretion from hepatocytes.
RESULTS: To investigate the role of hepatocyte
tPA in apoB-lipoprotein metabolism, tPA expres-
fate of intrahepatic apoB. Poorly lipidated apoB
undergoes intracellular degradation, whereas
fully lipidated apoB is efficiently secreted as
VLDL particles with larger size and lower den-
sity. Silencing hepatocyte tPA in Ldlr−/− mice led
to larger VLDL particles in the plasma, with
more triglyceride per VLDL particle, indicating
y
blocking the interaction between apoB and
MTP in hepatocytes. This process reduces
VLDL assembly and plasma apoB-lipoprotein
cholesterol levels.
Plasminogen activator inhibitor 1 (PAI-1) is a
rapidly acting serine protease inhibitor (serpin)
of tPA. Upon lipid loading of hepatocytes, PAI-1
forms a complex with tPA and sequesters tPA
away from apoB, which allows apoB to be
lipidated and facilitates VLDL assembly and

y g
secretion. Consistent with these findings, hu-
mans with PAI-1 deficiency have smaller VLDL
particles and lower plasma levels of apoB-
lipoprotein cholesterol.

p
CONCLUSION: The findings in this study sug-
gest a mechanism that fine-tunes VLDL assem-

,
bly through intracellular interactions among
tPA, PAI-1, and apoB in hepatocytes, thereby
affecting the plasma levels of atherogenic
apoB-lipoproteins. Knowledge of this mech-
anism of hepatic lipoprotein regulation may
suggest therapeutic strategies for lowering
atherogenic apoB-lipoproteins and cardiovas-
cular risk.
▪
The list of author affiliations is available in the full article online.
*Corresponding author. Email: zzheng@mcw.edu (Z.Z.);
iat1@columbia.edu (I.T.); wdai@versiti.org (W.D.)
Intracellular tPA–PAI-1 interaction in the ER of hepatocytes determines apoB lipidation, VLDL
Cite this article as W. Dai et al., Science 381, eadh5207 (2023).
assembly, and secretion. In the basal state, tPA interacts with apoB and inhibits MTP-apoB interaction in DOI: 10.1126/science.adh5207
the ER of hepatocytes, limiting MTP-mediated apoB lipidation, VLDL assembly, and secretion. When
hepatocytes are loaded with lipid, PAI-1 sequesters tPA away from apoB by forming a complex with tPA, READ THE FULL ARTICLE AT
allowing the lipidation of apoB and the assembly and secretion of VLDL. [Figure created with Biorender] https://doi.org/10.1126/science.adh5207

Dai et al., Science 381, 959 (2023) 1 September 2023 1 of 1

RES EARCH

◥ anisms involved remain unknown. Given the

RESEARCH ARTICLE central role of hepatocytes in apoB-lipoprotein
production and our recent studies in mice
ATHEROSCLEROSIS showing that hepatocytes synthesize tPA and
contribute ~40 to 50% to both the plasma basal
Intracellular tPA–PAI-1 interaction determines tPA concentration and to fibrinolysis (19, 20),
we sought to identify possible links between
VLDL assembly in hepatocytes hepatocyte tPA and apoB-lipoprotein assem-
bly and secretion. Our investigation revealed
Wen Dai1*, Heng Zhang1, Hayley Lund2, Ziyu Zhang1, Mark Castleberry1, Maya Rodriguez1,3, that endogenous hepatocyte tPA limits VLDL
George Kuriakose4, Sweta Gupta5, Magdalena Lewandowska5, Hayley R. Powers6, Swati Valmiki7,8, assembly by directly interacting with apoB,
Jieqing Zhu1,6, Amy D. Shapiro5, M. Mahmood Hussain7,8, José A. López9,10, Mary G. Sorci-Thomas2,11,12, interrupting the interaction between apoB and
Roy L. Silverstein1,2, Henry N. Ginsberg13, Daisy Sahoo2,6,12, Ira Tabas4,14,15*, Ze Zheng1,2,12,16* MTP and thereby impairing MTP-dependent
neutral lipid transfer and apoB lipidation. We
Apolipoprotein B (apoB)–lipoproteins initiate and promote atherosclerotic cardiovascular disease. also show that plasminogen activator inhibitor
Plasma tissue plasminogen activator (tPA) activity is negatively associated with atherogenic apoB- 1 (PAI-1) binds to tPA within hepatocytes and
lipoprotein cholesterol levels in humans, but the mechanisms are unknown. We found that tPA, abolishes the effect of tPA on limiting VLDL as-
partially through the lysine-binding site on its Kringle 2 domain, binds to the N terminus of apoB, sembly. Our in vivo data suggest that the tPA–
blocking the interaction between apoB and microsomal triglyceride transfer protein (MTP) in hepatocytes, PAI-1 interaction enables physiologic postprandial
thereby reducing very-low-density lipoprotein (VLDL) assembly and plasma apoB-lipoprotein cholesterol VLDL assembly. Thus, our findings that tPA

p
levels. Plasminogen activator inhibitor 1 (PAI-1) sequesters tPA away from apoB and increases VLDL can alter the production of atherogenic apoB-
assembly. Humans with PAI-1 deficiency have smaller VLDL particles and lower plasma levels of apoB- lipoproteins have potential therapeutic impli-
lipoprotein cholesterol. These results suggest a mechanism that fine-tunes VLDL assembly by cations for lowering atherosclerotic CVD risk.
intracellular interactions among tPA, PAI-1, and apoB in hepatocytes.
Lowering hepatocyte tPA raises plasma apoB

A
To investigate the role of hepatocyte tPA in de-

polipoprotein B (apoB)–lipoproteins ini-
tiate and promote atherosclerotic cardio-
vascular disease (CVD) (1, 2). The major
source of blood apoB-lipoproteins is he-
patocytes, where very-low-density lipo-
However, these treatments have only a modest
effect on other atherogenic apoB-containing
lipoproteins, such as VLDL and IDL (4, 5),
which contribute to the residual CVD risk in
populations with well-controlled LDL cho-
g
termining plasma apoB-lipoprotein levels in a
setting where hepatic clearance by the LDL
receptor is not a factor, tPA expression was
silenced in the hepatocytes of Ldlr−/− mice fed
a high-fat, high-cholesterol Western diet (WD).

proteins (VLDLs) are assembled and secreted.
Circulating VLDL is then progressively hydro-
lyzed in the blood to form intermediate-density
lipoprotein (IDL) and low-density lipopro-
tein (LDL) (2). Currently recommended lipid-
intervention therapies to prevent CVD primarily
use statins or protein convertase subtilisin/
kexin type 9 (PCSK9) inhibitors, both of which
lower LDL by enhancing hepatic LDL re-
ceptor (LDLR)–mediated LDL clearance (3).
lesterol (6, 7). Thus, therapies that inhibit
hepatic VLDL production might be useful in
decreasing CVD risk because they would
lower all atherogenic apoB-lipoproteins. apoB,
a large complex protein, is the structural scaf-
fold for forming VLDL (8). VLDL is assembled
in hepatocytes by incorporating triglyceride,
cholesteryl ester, and phospholipid onto apoB
to form spherical particles (9). This process,
known as apoB lipidation, depends on both
y
This silencing was accomplished by admin-
istering an adenovirus-associated virus 8 (AAV8)
expressing a hairpin RNA against Plat mRNA,
which encodes tPA protein, driven by the H1
promoter, AAV8-H1-shPlat (sh-tPA) (19, 20).
The hepatocyte tPA–silenced mice showed a
47% higher plasma total cholesterol level (P <
0.01) and 28% higher plasma apoB-100 (P <
0.05) compared with mice receiving AAV8-H1-
scrambled RNA (scr) (Fig. 1A and fig. S1A).

y g
lipid availability and a neutral lipid transporter, Plasma lipoprotein fraction profiling by fast
microsomal triglyceride transfer protein (MTP) protein liquid chromatography (FPLC) re-
1
Versiti Blood Research Institute, Milwaukee, WI 53226, USA. (10, 11). MTP binds to apoB in the hepatocyte vealed that the hepatocyte tPA–silenced mice
2
Department of Medicine, Medical College of Wisconsin, endoplasmic reticulum (ER) and transfers lip- had more cholesterol and apoB-100 in the

Milwaukee, WI 53226, USA. 3College of Arts and Sciences,
Marquette University, Milwaukee, WI 53233, USA.
4
Department of Medicine, Columbia University Irving Medical
Center, New York, NY 10032, USA. 5Indiana Hemophilia
p
ids to apoB (10). When MTP and/or lipids are VLDL and LDL fractions and more triglyceride
unavailable, VLDL cannot be assembled, and in VLDL (Fig. 1A). Using a different hypercho-
newly translated apoB is then targeted for deg- lesterolemic model, we showed that silenc-

,
and Thrombosis Center, Indianapolis, IN 46260, USA. radation (12). Although the essential role of MTP ing hepatocyte tPA in WD-fed Apoe knockout
6
Department of Biochemistry, Medical College of Wisconsin,
Milwaukee, WI 53226, USA. 7Department of Cell Biology, in apoB lipidation has been well established, (Apoe−/−) mice led to 56% higher plasma total
SUNY Downstate Medical Center, Brooklyn, NY 11203, USA. much less is known about the regulation of cholesterol (P < 0.05) and 28% higher apoB-
8
Department of Foundations of Medicine, NYU Long Island apoB-MTP interaction and MTP-mediated lipid 100 (P < 0.05) accompanied by increases in
School of Medicine, Mineola, NY 11501, USA. 9Bloodworks
Research Institute, Seattle, WA 98102, USA. 10Department of
transfer to apoB. VLDL and LDL cholesterol, triglyceride, and
Medicine, University of Washington, Seattle, WA 98195, USA. Previous studies have shown that low plas- apoB-100 (Fig. 1B and fig. S1B). Similar results
11
Department of Pharmacology and Toxicology, Medical College ma tissue plasminogen activator (tPA) activity were observed in WD-fed wild-type (WT) mice
of Wisconsin, Milwaukee, WI 53226, USA. 12Cardiovascular
Center, Medical College of Wisconsin, Milwaukee, WI 53226,
is associated with a higher risk of atheroscle- (fig. S1C and fig. S2). To further validate the
USA. 13Department of Medicine, Columbia University Vagelos rotic CVD (13–15), but whether decreased fi- role of hepatocyte tPA in determining plasma
College of Physicians and Surgeons, New York, NY 10032, USA.
14
brinolysis contributes to CVD in this setting apoB-lipoprotein levels, the hepatocyte tPA–
Department of Pathology and Cell Biology, Columbia
remains unknown. Another plausible mech- knockout mouse model was generated by
University Irving Medical Center, New York, NY 10032, USA.
15
Department of Physiology and Cellular Biophysics, Columbia anism linking low tPA to CVD is elevated plas- administering Plat fl/fl mice with an AAV8
University Irving Medical Center, New York, NY 10032, USA. ma apoB-lipoprotein cholesterol, which is seen expressing a Cre recombinase driven by the
16
Department of Physiology, Medical College of Wisconsin, in humans with decreased tPA activity (16–18). thyroxine-binding globulin (TBG) promoter,
Milwaukee, WI 53226, USA.
*Corresponding author. Email: zzheng@mcw.edu (Z.Z.); However, whether apoB-lipoprotein metabolism AAV8-TBG-cre (Cre). The WD-fed hepatocyte
iat1@columbia.edu (I.T.); wdai@versiti.org (W.D.) and tPA are causally linked and, if so, the mech- tPA–knockout mice showed a 30% higher plasma

Dai et al., Science 381, eadh5207 (2023) 1 September 2023 1 of 12

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

Western diet-fed Ldlr-/- mice treated with AAV8-H1-shPlat or -scrambled control

A scr scr scr

Plasma cholesterol (mg/dL)

Plasma apoB-100 ( g/mL)

* * * sh-tPA sh-tPA sh-tPA
Liver tPA normalized by

6 2500 2000 20 VLDL 20 20 VLDL

Cholesterol (mg/dL)

Triglyceride (mg/dL)
total protein (ng/mg)

ApoB-100 ( g/mL)
VLDL
2000 1500 15 15 15 LDL
4
1500 LDL
1000 10 10 10
2
1000
500 500 5 5 5
HDL
0 0 0 0 0 0
scr sh-tPA scr sh-tPA scr sh-tPA
15 18 21 24 27 30 33 15 18 21 24 27 30 33 15 18 21 24 27 30 33
Elution volume (mL) Elution volume (mL) Elution volume (mL)
Western diet-fed Apoe-/- mice treated with AAV8-H1-shPlat or -scrambled control
B scr scr scr
Plasma cholesterol (mg/dL)

Plasma apoB-100 ( g/mL)

* * * sh-tPA sh-tPA sh-tPA
Liver tPA normalized by

10 10 20

Cholesterol (mg/dL)
5 1000 800 LDL

Triglyceride (mg/dL)
total protein (ng/mg)

VLDL VLDL

ApoB-100 ( g/mL)
4 800 700 8 8 15 VLDL
3 600 600 6 LDL 6
10
2 400 500 4 4
1 2 2 5
200 400
HDL

p
0 0 300 0 0 0
scr sh-tPA scr sh-tPA scr sh-tPA 15 18 21 24 27 30 33 15 18 21 24 27 30 33 15 18 21 24 27 30 33
Elution volume (mL) Elution volume (mL) Elution volume (mL)
Western diet-fed Platfl/fl mice treated with AAV8-TBG-cre or -GFP
C GFP GFP GFP
Plasma cholesterol (mg/dL)

Plasma apoB-100 ( g/mL)

* * * Cre Cre Cre

Liver tPA normalized by

8 300 200 8 0.8 5

Cholesterol (mg/dL)

Triglyceride (mg/dL)
total protein (ng/mg)

ApoB-100 ( g/mL)
0.4
HDL VLDL LDL

g
6 150 6 0.3 LDL 0.6 4
0.2
200 0.1
0.0 3
4 100 4 16 18 20 0.4
2
100 VLDL
2 50 2 0.2

y
1
VLDL
0 0 0 0 0.0 0
GFP Cre GFP Cre GFP Cre 15 18 21 24 27 30 33 15 18 21 24 27 30 33 15 18 21 24 27 30 33
Elution volume (mL) Elution volume (mL) Elution volume (mL)
Human primary hepatocytes McA-RH7777 cells
D E
VLDL-triglyceride (mg/dL)

VLDL-triglyceride (mg/dL)
VLDL-cholesterol (mg/dL)

VLDL-cholesterol (mg/dL)

* * * *
0.6 2.0 0.8 4
Cell medium Cell medium
1.5 scr si-tPA 0.6 3 scr si-tPA
0.4
ApoB-100 ApoB-100

y g
1.0 0.4 2 ApoB-48
0.2 Albumin Albumin
0.5 0.2 1

0.0 0.0 0.0 0

p
scr si-tPA scr si-tPA scr si-tPA scr si-tPA

Fig. 1. Silencing hepatocyte tPA increases apoB-lipoprotein cholesterol and for tPA protein, and plasma samples were assayed for total cholesterol and
apoB independently of LDLR or apoE. (A) Ldlr−/− mice were treated with apoB-100 concentrations and for FPLC profiles of cholesterol, triglyceride, and

,
AAV8-H1-shPlat (sh-tPA) or AAV8-H1-scrambled control (scr) and then fed the apoB-100. The cholesterol in the VLDL fractions is shown in a zoomed-in smaller
WD for 8 weeks. The livers were assayed for tPA protein, and plasma samples were graph (n = 6 mice per group). (D) Human primary hepatocytes were treated
assayed for total cholesterol and apoB-100 concentrations and for FPLC profiles with siRNA against tPA mRNA (si-tPA) or scrambled RNA for 24 hours. Cell
of cholesterol, triglyceride, and apoB-100 (n = 9 to 10 mice per group). (B) Apoe−/− culture medium apoB was quantified by immunoblot. VLDL fractions were isolated
mice were treated with AAV8-H1-shPlat (sh-tPA) or AAV8-H1-scrambled control by ultracentrifugation, and cholesterol and triglyceride concentrations in VLDL
(scr) and then fed the WD for 8 weeks. The livers were assayed for tPA fractions were assayed. (E) McA-RH7777 cells were treated with siRNA against tPA
protein, and plasma samples were assayed for total cholesterol and apoB-100 mRNA (si-tPA) or scrambled RNA for 24 hours. VLDL was isolated from the
concentrations and for FPLC profiles of cholesterol, triglyceride, and apoB-100 medium by ultracentrifugation, and cholesterol and triglyceride concentrations in
(n = 5 mice per group). (C) Platfl/fl mice were treated with AAV8-TBG-cre (Cre) VLDL were assayed. Cell culture medium apoB was assayed by immunoblot. Data
or AAV8-TBG-GFP (GFP) and then fed the WD for 8 weeks. The livers were assayed are shown as means ± SEMs; *P < 0.05 by two-tailed Student’s t test.

total cholesterol level (P < 0.01) and 33% higher TBG-GFP [green fluorescent protein (GFP)] (Fig. plasma apoB was not because of increased
plasma apoB-100 level (P < 0.05), accompanied 1C and fig. S1D). apoB synthesis. Moreover, the plasma apoB
by higher cholesterol and apoB-100 in the VLDL Silencing tPA did not change Apob mRNA data were not a result of changes in apoE- or
and LDL fractions and higher triglyceride in in the livers of Ldlr−/−, Apoe−/−, or WT mice LDLR-mediated pathways because plasma
VLDL, compared with mice receiving AAV8- (fig. S3, A to C), which suggests that increased apoE levels and liver LDLR expression remained

Dai et al., Science 381, eadh5207 (2023) 1 September 2023 2 of 12

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

unchanged (fig. S4, A and B). Collectively, these
data show that silencing hepatocyte tPA raises
plasma apoB lipoprotein-cholesterol through
a mechanism independent of the LDL recep-
tor or apoE.
To examine hepatocyte-intrinsic effects and
to link this finding to human relevance, tPA
expression was silenced in primary human
hepatocytes using small interfering RNA
(siRNA) against PLAT mRNA (si-tPA). The si-
tPA treated cells had higher levels of apoB-100
in their media compared with control (scr)
hepatocytes (fig. S5A and Fig. 1D), whereas
hepatocyte APOB mRNA levels were similar in
the two groups of cells (fig. S5B). Further, the
contents of cholesterol and triglyceride were
higher in the VLDL fractions isolated from the
culture medium of tPA-silenced human hepa- cyte tPA silencing on VLDL secretion in mice
tocytes (Fig. 1D). Similar findings were ob- injected with the nonionic surfactant deter-
served in cultured McA-RH7777 cells (Fig. 1E gent poloxamer 407 (P407). The P407 injection
and fig. S5, C and D), a rat hepatoma cell line. blocks VLDL clearance by inhibiting lipo-
The McA-RH7777 cell is an established model protein lipase activity and VLDL lipolysis (22).
system to study VLDL production because it Hepatocyte tPA–silenced mice had a faster
synthesizes VLDL-sized apoB particles similar rise in triglycerides (Fig. 2A), which suggests
to human hepatocytes (21). These data show that increased VLDL production. Moreover, plas-
the ability of tPA to regulate apoB-lipoprotein ma apoB-100, which is derived only from
secretion is a cell-intrinsic property of hepato- hepatocytes, increased to a greater degree than
cytes and is applicable to humans. plasma apoB-48, which is produced by both
hepatocytes and intestinal epithelial cells in
tPA limits apoB lipidation in the ER mice (23) (Fig. 2B). These data, when combined
The data in Ldlr−/− and Apoe−/− mice suggest that with those presented above, suggest that si-
tPA limits the production—not the clearance— lencing hepatocyte tPA increases plasma apoB-
of apoB-lipoproteins. To examine this point containing lipoprotein levels by promoting
further, we investigated the effect of hepato- VLDL production.

Fig. 2. tPA limits apoB lipidation Western diet-fed wild type mice treated with AAV8-H1-shPlat or -scrambled control

p
in the ER. (A) WT mice were A B
Plasma triglyceride (mg/dL)

scr * * scr Plasma

*
rising rate (mg/dL/min)
2000 sh-tPA 40 sh-tPA
treated with AAV8-H1-shPlat Plasma triglyceride
scr sh-tPA
(sh-tPA) or AAV8-H1-scrambled 1500 30
460 kDa ApoB-100
control (scr) and then fed the WD *
1000 20 268 kDa ApoB-48
for 14 weeks. Mice were injected
500 10
with P407 intraperitoneally (i.p.) to 50 kDa
Albumin

g
assess VLDL secretion. Plasma 0 0
60-90 90-120
triglyceride concentration was 0 30 60 90 120
Time (mins) mins mins
measured (n = 4 to 5 mice per VLDL fractions from Ldlr mice
-/- Normal chow diet-fed holo-tPA-KO mice
group). (B) WT mice were treated C D E F
Percentage (%) Percentage (%)

scr 40 scr

LDL cholesterol ( g/mL)

Plasma apoB-100 ( g/mL)

with AAV8-H1-shPlat (sh-tPA) or VLDL * * * * *

y
VLDL cholesterol (mg/dL)
Triglyceride/apoB-100

30 80 2.5 10 70
4
VLDL Z-average

AAV8-H1-scrambled control (scr) 20

in VLDL ( g/ g)
diameter (nm)

10 2.0 8 60
and then fed the WD for 14 weeks. 60 3
0 1.5 6 50
Mice were injected with P407 i.p. to 20 40 60 80 100 40 2
40 1.0 4
assess VLDL secretion. Plasma sh-tPA sh-tPA 40
30 20 0.5 1 2
apoB concentration was measured 20
by ELISA (n = 4 to 5 mice per group). 10 0 0.0 0 0 0
VLDL scr sh-tPA scr sh-tPA LacZ tPA LacZ tPA LacZ tPA
0
(C) Ldlr−/− mice were treated with 20 40 60 80 100
AAV8-H1-shPlat (sh-tPA) or AAV8-H1- Diameter (nm)
Human primary hepatocytes
scrambled control (scr) and then
G H I J K
apoB-100 in VLDL ( g/ g)

fed the WD for 8 weeks. VLDL * * * * Endoplasmic reticulum

[3H]-apoB in cell medium

[3H]-apoB in cell medium

y g
Ratio of triglyceride to

2800 4000 60 1.0 protein extractions fraction

was isolated by ultracentrifugation
(dpm/mg protein)

(dpm/mg protein)

VLDL Z-average

0.8
diameter (nm)

and visualized by transmission 2400 3000 ApoB-100

40 si-tPA
0.6
electron microscopy. VLDL (n = 100 2000
2000
scr ApoB-100
for each group) diameter was 1600 0.4
20

p
1000 0.2 1 2 3 4 5 6
measured and analyzed using
Image-Pro Plus 10.0. Scale bars, 0 0 0 0.0 Sucrose density gradient
GFP tPA-HA scr si-tPA scr si-tPA scr si-tPA
100 nm. (D) Ldlr−/− mice were

,
treated with AAV8-H1-shPlat
(sh-tPA) or AAV8-H1-scrambled control (scr) and then fed the WD for 8 weeks (H) Human primary hepatocytes were treated with siRNA against tPA mRNA
(n = 9 to 10 mice per group). VLDL was isolated by ultracentrifugation, and VLDL (si-tPA) or scrambled RNA for 24 hours. apoB secretion was measured using
diameter was measured by dynamic light scattering. (E) Ldlr−/− mice were [3H]-labeling, as in (G). Radioactivity associated with apoB in cell medium was
treated with AAV8-H1-shPlat (sh-tPA) or AAV8-H1-scrambled control (scr) and quantified by scintillation counting. (I) Human primary hepatocytes were treated
then fed the WD for 8 weeks. VLDL was isolated by ultracentrifugation and with siRNA against tPA mRNA (si-tPA) or scrambled RNA for 24 hours. VLDL was
assayed for the ratio of triglyceride to apoB-100 (n = 9 to 10 mice per group). isolated by ultracentrifugation, and VLDL diameter was measured by dynamic light
(F) Whole-body tPA knockout mice (holo-tPA-KO) were treated with AAV8-TBG-Plat scattering. (J) Human primary hepatocytes were treated with siRNA against tPA
(tPA) or AAV8-TBG-lacZ (LacZ) and then fed a normal chow diet for 8 weeks. mRNA (si-tPA) or scrambled RNA for 24 hours. VLDL was isolated by ultracentrifugation
Plasma samples were assayed for VLDL cholesterol, LDL cholesterol, and apoB-100 and assayed for the ratio of triglyceride to apoB-100. (K) Human primary hepatocytes
concentrations (n = 6 mice per group). (G) Human primary hepatocytes were were treated with siRNA against tPA mRNA (si-tPA) or scrambled RNA for 24 hours.
transduced with a plasmid encoding tPA with C-terminal HA tag (tPA-HA) or GFP. The ER fraction was isolated, and proteins from the ER were extracted. apoB-
After 48 hours, apoB secretion was measured using a [3H]-labeling method as lipoproteins extracted from the ER were further separated by density gradient
follows: hepatocytes were incubated for 20 min in [3H]-leucine–containing medium ultracentrifugation and divided into six fractions of increasing density from fraction 1
and chased for 3 hours in [3H]-leucine–free medium, and then radioactivity to 6. apoB from each fraction was measured by immunoblot. Data are shown as
associated with apoB in the cell medium was quantified by scintillation counting. means ± SEMs; *P < 0.05 by two-tailed Student’s t test.

Dai et al., Science 381, eadh5207 (2023) 1 September 2023 3 of 12

RES EARCH | R E S E A R C H A R T I C L E
Lipidation is a key factor determining the
fate of intrahepatic apoB. Poorly lipidated apoB
undergoes intracellular degradation, whereas
fully lipidated apoB is efficiently secreted as
VLDL particles with larger size and lower
density (24). Electron microscopic scanning
of plasma VLDL particles isolated by density
ultracentrifugation showed a shifted distrib-
ution of the particles to larger diameters in the
hepatocyte tPA–silenced Ldlr−/− mice (Fig. 2C).
Moreover, dynamic light-scattering analyses
(25–27) revealed that VLDL from sh-tPA mice
had a larger hydrodynamic diameter than VLDL
from control mice (Fig. 2D). These combined data
suggest that the VLDL of the tPA-silenced mice
have a higher lipid content, which is further sup-
ported by the finding that the VLDL of the tPA-
silenced mice had a higher triglyceride/apoB
ratio (Fig. 2E).
To explore the impact of increasing hepatocyte
tPA expression on plasma apoB-lipoproteins
in vivo, whole-body tPA knockout mice (holo-
tPA-KO) were injected with AAV8-TBG-Plat
(tPA), to restore tPA expression specifically in
hepatocytes, or with AAV8-TBG-LacZ control
(LacZ) (19, 20, 28–30). As predicted, plasma
apoB-100 and VLDL- and LDL-cholesterol were
lower in the mice treated AAV8-TBG-Plat versus
AAV8-TBG-LacZ (Fig. 2F).
To show relevance to lipoprotein production
by human hepatocytes, apoB secretion was
measured using a 3[H]-leucine pulse-chase
method (31) in human primary hepatocytes.
Transfecting these cells with a plasmid en-
coding tPA led to lower apoB-associated radio-
activity in the cell medium, indicating lower
secretion of newly synthesized apoB (Fig. 2G).
By contrast, silencing tPA increased the se-
cretion of newly synthesized apoB (Fig. 2H).
Similar results were observed in tPA-silenced
McA-RH7777 cells (fig. S6). However, incubat-
ing human primary hepatocytes with recom-
binant human tPA did not change culture
medium apoB-100 levels and cholesterol and
triglyceride contents in VLDL (fig. S7), which
suggests that the hepatocyte tPA limits apoB
lipidation and production intracellularly.
Consistent with the mouse plasma data
above, silencing tPA increased the diameter
and triglyceride/apoB ratio of VLDL isolated
from the cell medium (Fig. 2, I and J). These
data suggest that silencing tPA increases apoB
lipidation. ER is the major site of apoB lip-
idation and VLDL assembly (32). The degree
of apoB lipidation can be ascertained by the
density of apoB lipoprotein particles—namely,
greater lipidation of apoB results in lower
density of the particle (33, 34). In this con-
text, tPA-silenced human primary hepatocytes
had overall higher ER-associated apoB and a
higher proportion of apoB in lower-density
fractions (fractions 1 and 2, Fig. 2K) than
control hepatocytes. Because density and lipid-
ation are inversely related, these findings fur-
Dai et al., Science 381, eadh5207 (2023) 1 September 2023
ther support the hypothesis that tPA limits
apoB lipidation in the ER.
tPA blocks MTP-mediated VLDL assembly by
interacting with apoB
MTP is a chaperone that promotes intra-
hepatic apoB lipidation by transferring and
incorporating neutral lipids, notably triglyc-
eride and cholesteryl esters, to apoB to as-
semble VLDL in the ER (10, 11). MTP-mediated
lipid transfer to apoB involves its direct bind-
ing to apoB (10, 35) because inhibiting the
apoB-MTP interaction decreases apoB lipida-
tion and secretion (11). We found that tPA
silencing in human primary hepatocytes did
not alter the level of MTP protein (Fig. 3A,
input). However, MTP immunoprecipitated
from tPA-silenced cells showed a higher con-
tent of apoB compared with immunoprecipi-
tates of MTP from control hepatocytes (Fig. 3A).
These data suggest that silencing tPA increases
apoB-MTP interaction, which would be ex-
pected to increase lipid transfer to apoB (11).
Consistent with this idea, microsomal frac-
tions isolated from tPA-silenced hepatocytes
had a twofold higher neutral lipid transfer
activity (36) compared with microsomes from
control cells (Fig. 3B, groups 1 and 2). This
lipid transfer activity was due to MTP, as the
MTP inhibitor CP-346086 (37) completely abol-
ished the elevated neutral lipid transfer activity
in microsomes from tPA-silenced hepatocytes
(Fig. 3B, group 3). Similar findings were ob-
served in McA-RH7777 cells (fig. S8). Conversely,
when tPA was expressed in human primary
hepatocytes, apoB-MTP interaction and neutral
lipid transfer activity were decreased (Fig. 3, C
and D). Moreover, purified recombinant tPA
protein reduced MTP-mediated neutral lipid
transfer to LDL in a dose-dependent manner
(Fig. 3E). Thus, tPA inhibits apoB-MTP inter-
action and reduces MTP-mediated lipid trans-
fer activity.
To examine tPA interaction with apoB, we
analyzed primary human hepatocytes trans-
fected with tPA with a C-terminal HA-tag and
found that apoB eluted from the anti-HA pre-
cipitates (Fig. 3F). This finding was further
supported by the results of a tPA-apoB proximity-
ligation assay in hepatocytes using labeled anti-
tPA and anti-apoB. The data showed intracellular
punctate fluorescence (Fig. 3G), which indi-
cates that endogenous tPA and apoB interact
in these cells. We also found, using confocal
immunofluorescence microscopy, that tPA and
apoB colocalize in the ER (Fig. 3H). Moreover,
expression of a form of tPA that is retained in
the ER because of a KDEL sequence (38, 39) at
the C terminus of tPA (tPA-KDEL) reduced
secretion of newly synthesized apoB (Fig. 3I).
Collectively, these data suggest that tPA inter-
acts with apoB in the ER, which, as described
below, is relevant to the lowering of apoB-
VLDL assembly and secretion by tPA.
The protease activity of tPA depends on the
serine residue at position 513, and substituting
serine to alanine (S513A) completely abolishes
the serine protease activity of tPA (40). We
found that the serine protease mutant tPA
(S513A) reduced both apoB secretion by hu-
man primary hepatocytes and MTP-mediated
lipid transfer activity to a similar degree as WT
tPA (Fig. 3, I and J), indicating that this action
of tPA is independent of its protease activity.
We also showed that purified recombinant WT
and S513A-tPA directly interacted in a dose-
dependent manner with either LDL or apoB-
100 immobilized to a microtiter plate surface
(Fig. 4A and figs. S9 and S10). By contrast,
purified tPA did not interact with solid-phase
immobilized MTP (Fig. 4B). Most importantly,
preincubating LDL with tPA inhibited the
binding of LDL to MTP (Fig. 4C). Finally, sur-
face plasmon resonance (SPR) studies provided
p
further evidence of noncovalent interaction be-
tween tPA and LDL (Fig. 4D). These combined
data suggest that tPA binds apoB in a tPA-
protease–independent manner and that this
process blocks the interaction of apoB with
MTP, thereby decreasing apoB lipidation and
g
VLDL assembly and secretion.
The Kringle 2 (K2) domain of tPA interacts
with the N terminus of apoB
The K2 domain of tPA has a lysine binding
y
site, which is required for its interaction with
fibrin (41). Within the K2 domain of human
tPA, negatively charged aspartic acid residues
at 236 and 238 are responsible for the binding
of tPA to positively charged lysine residues
(42). The surface-exposed N terminus of apoB
(43) has a positively charged lysine-rich region
that is responsible for its binding to MTP (44).
The MTP-interacting region (amino acids 430
to 570) (45) of the apoB N terminus can be
y g
partially blocked using a monoclonal antibody
1D1, developed by immunizing mice with a hu-
man apoB fragment comprising amino acids
474 to 539 (46). In this context, we show that
p
the 1D1, but not an epitope control antibody
raised against amino acids 4031 to 4080 (a3
domain) of human apoB, inhibited the interac-
,
tion between tPA and solid surface-immobilized
LDL (Fig. 4E). These data suggest that tPA
interacts with the N terminus of apoB.
Expressing tPA mutants lacking either the
K2 domain (tPA-D-K2) or in which aspartic acid
residues at positions 236 and 238 were sub-
stituted with asparagine (tPA-D236, 238N)
blocked the ability of tPA to lower apoB sec-
retion as measured by pulse-chase assay (Fig.
4F). Similarly, in the solid-phase protein bind-
ing assay, an antibody against the K2 domain
of tPA inhibited the binding of tPA to solid
surface-immobilized LDL (Fig. 4G). Further-
more, tranexamic acid (TXA), a lysine analog
that inhibits tPA binding to fibrin, partially re-
duced the interaction between tPA and solid
4 of 12
RES EARCH | R E S E A R C H A R T I C L E

Fig. 3. tPA blocks apoB-VLDL Human primary hepatocyte

assembly. (A) Human primary A B C D
Cell lysate Microsomal fraction Cell lysate Microsomal fraction
hepatocytes were treated with
* * *

Relative neutral lipid

Relative neutral lipid

siRNA against tPA mRNA (si-tPA) 2.5 GFP tPA-HA 1.5

(ex/em, 465/535 nm)

(ex/em, 465/535 nm)

scr si-tPA

transfer activity

transfer activity
or scrambled RNA for 24 hours. IB: apoB
2.0 IB: apoB
IP: MTP 1.0
Cell lysates (input) and anti-MTP IP: MTP 1.5
IB: MTP
immunoprecipitates (IP: MTP) IB: MTP 1.0
0.5
were assayed for apoB and MTP IB: MTP 0.5 IB: MTP
Input Input
by immunoblot. (B) Human IB: apoB 0.0 IB: apoB 0.0
primary hepatocytes were treated scr si-tPA si-tPA GFP tPA-HA
DMSO DMSO MTPi
with siRNA against tPA mRNA
(si-tPA) or scrambled RNA (group 1) E Lipid transfer activity assay F Cell lysate
G Proximity ligation assay
or tPA mRNA (si-tPA) (groups 2 Anti- A+anti-apoB No primary abs

and 3) for 24 hours. The microsomal *

GFP tPA-HA
fraction was isolated and assayed * *
Relative neutral lipid

(ex/em, 465/535 nm)

10 IB: apoB
transfer activity

for neutral lipid transfer activity

8 IP: HA
with DMSO (groups 1 and 2) or IB: tPA
6
without with CP-346086 (10 nM), IB: tPA
4 Input
an MTP inhibitor (group 3).
2 IB: apoB
(C) Human primary hepatocytes

p
0
were transduced with a plasmid BSA, g/mL 1 - - -
encoding tPA with C terminal MTP, g/mL - 1 1 1
tPA , g/mL - - 1 5
HA tag (tPA-HA) or GFP for Anti-tPA Anti-apoB
48 hours. Cell lysates (input) and
anti-MTP precipitates (IP: MTP)
H Confocal microscopy immunofluorescence imaging
I J Lipid transfer activity assay
were assayed for apoB and MTP by ApoB tPA Calnexin ApoB + tPA
*

g
immunoblot. (D) Human primary * *

[3H]-apoB in cell medium

* * *
hepatocytes were transduced with a

Relative neutral lipid

(ex/em, 465/535 nm)

4000 10

(dpm/mg protein)

transfer activity
plasmid encoding tPA-HA or GFP 8
3000
for 48 hours. The microsomal 6
fraction was isolated and assayed 2000

y
4
for neutral lipid transfer activity. 1000 2
(E) The effect of recombinant
0 0
human tPA on lipid transfer from BSA + - - -

-S A

-K A
-tP trl

EL
donor vesicles to human LDL was tP A-H - + + +

tP 513
MTP
c

D
tPA-WT - - + -
assayed. (F) Human primary
A
A
ApoB + Calnexin tPA + Calnexin ApoB + Calnexin ApoB + Calnexin
ex

tPA-S513A - - - +
hepatocytes were transduced with a + tPA + tPA + DAPI
plasmid encoding tPA-HA or GFP
for 48 hours. Cell lysates (input) and anti-HA immunoprecipitates (IP: HA) enzymatically inactive mutant of tPA (tPA-S513A), tPA with an ER retention
were assayed for apoB and tPA. (G) A proximity ligation assay was used to signal sequence (tPA-KDEL), or GFP for 48 hours. apoB secretion was measured
measure apoB-tPA interaction in human primary hepatocytes. Scale bars, 10 mm. by [3H]-labeling, as in Fig. 2. (J) The effect of recombinant WT tPA (tPA-WT) or

y g
(H) Confocal microscopy immunofluorescence imaging was used to measure the enzymatically inactive mutant of tPA (tPA-S513A) on lipid transfer from donor
subcellular localization of tPA and apoB in human primary hepatocytes. Scale vesicles to human LDL was assayed. Data are shown as means ± SEMs; *P <
bars, 10 mm. DAPI, 4′,6-diamidino-2-phenylindole. (I) Human primary hepato- 0.05 by two-tailed Student’s t test (D) or by one-way ANOVA followed by
cytes were transduced with a plasmid encoding WT tPA (tPA-WT), an Dunnett’s test [(B), (E), (I), and (J)].

p
surface-immobilized LDL (Fig. 4H and fig. and plasma (19). PAI-1 covalently binds to tPA, observed increased tPA–PAI-1 complex and re-
S11). Taken together, these data support the forming a stable complex (48). Therefore, we duced free tPA as early as 1 hour after oleate

idea that the K2 domain of tPA, which contains
lysine-binding sites at aspartate residues 236
and 238, binds to the lysine-rich region of the
N terminus of apoB (Fig. 4I).
PAI-1 sequesters tPA from apoB in hepatocytes
Increased circulating intestinal apoB-48–
lipoproteins with oral feeding enhances he-
patic production of VLDL in humans (47), yet
the regulatory mechanisms are not clearly
understood. Obesity, often associated with
dyslipidemia, increases tPA synthesis in he-
patocytes (19). However, the increased tPA is
overcompensated by a larger increase in the
serpin inhibitor of tPA, PAI-1, resulting in de-
creased net functional free tPA in the livers
hypothesized that postprandial lipid loading of
hepatocytes would increase the intracellular
interaction of tPA and PAI-1, leading to reduced
free tPA. Reduced free tPA would then promote
apoB-MTP interaction, leading to increased
VLDL assembly and secretion.
We first immunoblotted tPA in a PAI-1–
immunoprecipitate from primary human
hepatocytes and found a band at ~120 Kd, rep-
resenting a covalently bound, sodium dodecyl
sulfate (SDS)–stable complex of tPA (~70 kD)
and PAI-1 (~50 kD) (Fig. 5A). This interaction
was further validated by a proximity-ligation
assay using antibodies against tPA and PAI-1
(Fig. 5B). Oleate is a known stimulus for apoB
lipidation and VLDL production (49, 50). We
,
treatment of human primary hepatocytes, which
is the time needed to observe a stimulatory
effect of oleate on VLDL production (50), the
tPA–PAI-1 complex increased, and free tPA
was reduced (Fig. 5, C and D). Therefore, the
tPA–PAI-1 complex forms rapidly in the hepa-
tocytes after oleate treatment. With prolonged
oleate treatment for 6 and 24 hours, free tPA
was further decreased, accompanied by in-
creased tPA–PAI-1 complex formation (Fig. 5D).
These results suggest a type of regulation of
apoB lipidation and VLDL production, involv-
ing PAI-1 interacting with tPA, when hepato-
cytes are loaded with lipids.
Silencing PAI-1 in human primary hepato-
cytes led to higher free tPA and lower apoB

Dai et al., Science 381, eadh5207 (2023) 1 September 2023 5 of 12

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

Fig. 4. The K2 domain of tPA Solid-phase protein binding assay

interacts with the N terminus A tPA-WT+LDL B C
tPA-S513A+LDL
of apoB. (A) A solid-phase binding tPA-WT

Absorbance (OD=450 nm)

Absorbance (OD=450 nm)

Absorbance (OD=450 nm)

assay was used to measure the 1.5 tPA-S513A 1.5
1.5
interaction between immobilized
LDL and recombinant human WT 1.0 1.0 1.0
tPA (tPA-WT) or enzymatically
inactive tPA-S513A. The binding of 0.5 0.5 0.5
tPA-WT or tPA-S513A to wells
without LDL is also measured under 0.0 0.0
the same conditions as the control. 0 100 200 300 400 0 100 200 300 400 0 100 200 300 400
OD, optical density. (B) A solid- tPA concentration, nM tPA concentration, nM tPA concentration, nM
phase binding assay was used
to measure the interaction
between recombinant human tPA D E F
Surface plasmon resonance Solid-phase protein binding assay Plasmid transfected human
and immobilized purified human
primary hepatocytes
MTP complex. (C) A solid-phase n.s.
binding assay was used to mea- Epitope control n.s.
tPA, nM

Absorbance (OD=450 nm)

1D1 mAb *
sure the ability of tPA to inhibit

[3 H]-apoB in cell medium

800 2.0 4000
Response units

the binding of MTP to immobilized 1000 3500

(dpm/mg protein)
600 500 1.5

p
LDL. (D) SPR was used to mea- 250
3000
sure the interaction between 400 1.0 2500
human recombinant tPA and LDL. 2000
200 100 0.5
(E) A solid-phase binding assay 50
10
was used to test whether anti- 0 0.0 0
apoB N-terminal antibody (1D1) 0 200 400 600 800 0 100 200 300 400

tP FP

6, 2
8N
23 -K
tP tP -W
Time (seconds) tPA concentration, nM

g
G

23
versus control IgG (raised against

A
-D A-
the b3 domain of apoB) blocks
the binding between human

A
recombinant tPA and immobilized Solid-phase protein binding assay
LDL. (F) Human primary hepato- G H tPA+Glycine (100 M) I

y
tPA+TXA (1 M)
cytes were transduced with Mouse IgG control tPA+TXA (10 M)
Absorbance (OD=450 nm)

Absorbance (OD=450 nm)

the plasmid encoding WT tPA Anti-tPA Kringle 2 tPA+TXA (100 M)

1.5 2.0
(tPA-WT), a tPA mutant without
1.5
the K2 domain (tPA-D-K2), or tPA 1.0
mutated in the K2 domain lysine 1.0
binding site (tPA-D236, 238N). 0.5
0.5
apoB secretion was measured by
3
[ H]-labeling, as in Fig. 2. (G) A 0.0 0.0
solid-phase binding assay was 0 100 200 300 400 0 100 200 300 400
used to test whether an antibody tPA concentration, nM tPA concentration, nM

y g
against tPA-K2 domain interferes
with the interaction between human recombinant tPA and immobilized purified LDL. (H) A solid-phase binding assay was used to measure whether tranexamic
acid (TXA) interferes with the interaction between recombinant human tPA and immobilized purified LDL. (I) A schematic diagram depicting the interaction between
the N terminus of apoB and the K2 domain of tPA. The diagram was generated using biorender.com. Data are shown as means ± SEMs; *P < 0.05 by one-way

p
ANOVA followed by Dunnett’s test. n.s., not significant (P ≥ 0.05).

,
secretion, as measured by pulse-chase analysis ing, more tPA is PAI-1–bound, and silencing tPA complex was unable to bind to LDL and did
(Fig. 5, E and F). The inhibitory impact of si- PAI-1 leads to a robust increase in free tPA not inhibit MTP-mediated neutral lipid trans-
lencing PAI-1 on apoB secretion was more and a reduction in apoB secretion. Further- fer activity (Fig. 5, H and I).
prominent under oleate treatment conditions, more, compared with silencing tPA alone, si- To test this finding in vivo, we delivered
with an ~400% increase in free tPA and an lencing both tPA and PAI-1 at the same time olive oil to WT mice by oral gavage. Two and
~60% decrease in apoB secretion. By contrast, did not lower apoB secretion (Fig. 5G). These 6 hours after oral gavage, liver free tPA was
without oleate treatment, there was only an data support the idea that PAI-1 facilitates decreased compared with baseline (Fig. 5J)
~67% increase in free tPA and an ~25% de- apoB lipidation by sequestering tPA from apoB. without alteration of liver total tPA and PAI-1
crease in apoB secretion. Similar results were The formation of a complex between tPA levels (fig. S13). These data suggest that the
observed in the McA-RH7777 hepatocytes (fig. and PAI-1 changes tPA’s conformation (51, 52), physiological function of hepatocyte tPA–PAI-1
S12). These observations are consistent with causing tPA to lose its ability to bind fibrin, interaction in fine-tuning apoB lipidation after
the hypothesis that, under basal conditions, which is mediated by the lysine binding site in lipid loading to hepatocytes. As expected, he-
most tPA is free and not bound by PAI-1, and its K2 domain (53). We therefore reasoned that patocyte PAI-1 knockout mice (19, 54) had
thus PAI-1 silencing leads to only a moderate the binding of PAI-1 to tPA blocks the lysine higher liver free tPA and lower plasma apoB,
increase in free tPA and a modest reduction binding site in K2 and thereby blocks tPA-LDL total cholesterol, and cholesterol in VLDL and
in apoB secretion. However, with oleate load- interaction. Consistent with this idea, the PAI-1– LDL fractions compared with their littermate

Dai et al., Science 381, eadh5207 (2023) 1 September 2023 6 of 12

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

Human primary hepatocytes and BMI-matched control individuals from

A B Proximity ligation assay C Ctrl Oleate the same community (55). None of the indi-
Anti-tPA+anti-PAI-1 No primary abs
116 Kd viduals were taking lipid-lowering agents or
IgG Ctrl Anti-PAI-1 tPA-PAI-1
had a known history of CVDs. Plasma from
IP: PAI-1 IB: tPA 100 Kd 66 Kd Free tPA the PAI-1–deficient subjects had 22% lower
Input IB: PAI-1 50 Kd
40 Kd
LDL-cholesterol (P < 0.05), 17% lower apoB-100
-actin
(P < 0.05), and 18% lower VLDL-cholesterol (P =
D E F 0.06) (Fig. 6G). Compared with control indi-
* *
* *
* * viduals, the PAI-1–deficient individuals had

[3H]-apoB in cell medium

complex (ng/mg protein)

* * * * * *
Hepatocyte tPA-PAI-1

higher tPA in VLDL (Fig. 6H), and VLDL-

Hepatocyte PAI-1 free

Hepatocyte PAI-1 free

4 0.25 0.4 3000

tPA (ng/mg protein)

tPA (ng/mg protein)

(dpm/mg protein)
3 0.20
0.3 associated tPA levels were inversely associated
2000
2
0.15
0.2
with VLDL diameter (r = −0.59, P < 0.01) (Fig.
0.10
1
1000 6I). These data, together with those from hepa-
0.05 0.1
tocyte PAI-1–deficient mice and PAI-1–silenced
0 0.00 0.0 0
Ctrl 1 hr 6 hrs 24 hrs Ctrl 1 hr 6 hrs 24 hrs human primary hepatocytes (Fig. 5F and Fig.

r
I1

si c r

I1
r
I1

s i cr
1
Oleate Oleate

sc

sc
AI
6D), are consistent with the hypothesis that

A
s

s
-P

-P

-P
-P
si

si
Ctrl Oleate Ctrl Oleate the PAI-1 deficiency leads to higher free tPA in
G H I J hepatocytes, which enables more tPA to inter-
Solid-phase protein binding assay Lipid transfer activity assay Normal chow-fed wild type mice
*
act with apoB and thereby limits apoB lip-
n.s.
tPA * idation and VLDL production.
Relative neutral lipid

(ex/em, 465/535 nm)

n.s. *
[3 H]-apoB in cell medium

* * *
Absorbance (OD=450 nm)

tPA (ng/mg protein)

p
tPA-PAI-1
transfer activity

0.4
3000 2.0 10 Taken together, our findings show that tPA

Liver PAI-1 free

(dpm/mg protein)

8
2000
1.5 0.3 directly interacts with apoB within the ER of
6
1.0 0.2 hepatocytes and that this interaction reduces
4
1000
0.5 0.1 MTP-mediated VLDL assembly. Lipid load-
2
0 0.0 0 0.0
ing to hepatocytes induces tPA–PAI-1 complex
scr si-tPA si-tPA 0 100 200 300 400 BSA + - - - 0 1 2 6
Time post-oral gavage formation, sequestering tPA from apoB, and
+si-PAI1 MTP - + + +

g
tPA concentration, nM
tPA-WT - - + -
with olive oil (hours) thereby facilitating apoB lipidation and VLDL
tPA-PAI-1 - - - + assembly (Fig. 6J).
Fig. 5. PAI-1 sequesters tPA away from apoB, leading to increased VLDL assembly in hepatocytes. Discussion
(A) Human primary hepatocytes lysates (input) and anti–PAI-1 immunoprecipitates (IP: PAI-1) were
Decades of research have established the role

y
assayed for tPA and PAI-1 by immunoblot. (B) A proximity ligation assay was used to measure tPA–PAI-1
for circulating tPA in initiating the lysis of
interaction in human primary hepatocytes. Scale bars, 10 mm. (C) Human primary hepatocytes were treated
blood clots (56, 57). Intravenous infusion of
for 6 hours with 0.4 mM oleate complexed with fatty acid–free BSA (oleate group) or fatty acid–free BSA
recombinant tPA is approved as a thrombolytic
alone (vehicle control). Cell lysates were assayed for tPA by immunoblot using the Jess Simple Western
therapy to restore blood flow in atherothrom-
system. (D) Human primary hepatocytes were treated with 0.4 mM oleate complexed with fatty acid–free
botic diseases, including ischemic stroke and
BSA (oleate group) or fatty acid–free BSA alone (vehicle control). Cell lysates were assayed for tPA–PA-1
myocardial infraction (58, 59). Conversely, plas-
complex and PAI-1–free tPA concentrations by ELISA. (E) Human primary hepatocytes were treated
ma levels of the tPA inhibitor PAI-1 are positively
with siRNA against PAI-1 mRNA (si-PAI1) or scrambled RNA and then incubated in medium containing either
associated with atherothrombotic disease (60).
0.4 mM oleate complexed with fatty acid–free BSA (oleate group) or fatty acid–free BSA alone (vehicle
High plasma levels of apoB-lipoproteins pro-
control). Cell lysates were assayed for PAI-1–free tPA concentration by ELISA. (F) Human primary hepatocytes
mote the arterial retention of these lipopro-

y g
were treated with siRNA against PAI-1 mRNA (si-PAI1) or scrambled RNA and then incubated in medium
teins, forming atherosclerotic lesions (61, 62).
containing either 0.4 mM oleate complexed with fatty acid–free BSA (oleate group) or fatty acid–free BSA
The atherosclerotic lesions can then progress
alone (vehicle control). apoB secretion was measured by [3H]-labeling, as in Fig. 2. (G) Human primary
to a state in which plaque rupture or erosion
hepatocytes were treated with siRNA against tPA mRNA (si-tPA) or against PAI-1 mRNA (si-PAI1) or
occurs, leading to acute thrombotic vascular
scrambled RNA. apoB secretion was measured by [3H]-labeling, as in Fig. 2. (H) A solid-phase binding assay

p
events, such as myocardial infarction and stroke
was used to measure the interaction between immobilized LDL and recombinant human tPA or tPA–PAI-1
(63). This study establishes a link between tPA,
complex. (I) The effect of recombinant human tPA and tPA–PAI-1 complex on lipid transfer from donor
PAI-1, and plasma levels of atherogenic apoB
vesicles to human LDL was assayed. (J) Normal chow diet–fed WT mice had their food withdrawn for 5 hours
,
lipoproteins, beyond the well-known roles of
and were then euthanized at 0, 1, 2, and 6 hours after oral gavage with olive oil. Liver lysates were assayed
tPA and PAI-1 in the occlusive thrombus that
for PAI-1–free tPA concentration by ELISA. Data are shown as means ± SEMs; *P < 0.05 by one-way ANOVA
is the ultimate manifestation of atherosclerotic
followed by Dunnett’s test [(D) to (G), (I), and (J)]. n.s., not significant (P ≥ 0.05).
vascular disease. The mechanism involves tPA–
PAI-1 interaction–mediated regulation of VLDL
controls (Fig. 6, A to D). In contrast to the teracting with tPA and sequestering tPA away assembly in the liver through protein-protein
higher VLDL production observed in hepatocyte from apoB, facilitates the physiologic increase interactions rather than through regulation of
tPA–silenced mice (Fig. 2A), hepatocyte PAI-1 in hepatic VLDL production that occurs in re- thrombolysis per se. These findings provide a
knockout mice had a slower rise in triglycer- sponse to fat ingestion. possible mechanistic explanation behind the
ides after detergent P407 injection, suggesting correlation between high apoB cholesterol
lower VLDL production (Fig. 6E). Moreover, PAI-1–deficient humans have lower and low tPA in CVD patients (13–15). More-
2 hours after an oral gavage of olive oil, plasma apoB-cholesterol over, the work has the potential for suggesting
apoB-100 was increased by 46% in control We collected plasma samples from humans LDLR-independent therapeutic strategies for
mice but by only 13% in hepatocyte–PAI-1– with a homozygous loss-of-function mutation lowering cardiovascular risk.
KO mice (Fig. 6F). These combined data sug- in SERPINE1, the gene encoding for PAI-1, and tPA is a multidomain protein consisting of
gest that endogenous hepatocyte PAI-1, by in- compared them with plasma from age-, gender-, finger, growth factor, Kringle 1 (K1), K2, and

Dai et al., Science 381, eadh5207 (2023) 1 September 2023 7 of 12

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

Fig. 6. PAI-1 deficiency leads to Serpine1fl/fl mice treated with AAV8-TBG-cre or -control
lower concentrations of plasma A B C D

Plasma cholesterol(mg/dL)
Ctrl
apoB and apoB-cholesterol in * Ctrl Cre Cre *
80 5 HDL 2.0

Cholesterol (mg/dL)

tPA (ng/mg protein)

mice and humans. (A) Serpine1fl/fl

Liver PAI-1 free

ApoB-100 0.4
60 4 0.3
mice were treated with AAV8-TBG-Cre 0.2
1.5
ApoB-48 3 0.1 LDL
(Cre) or control AAV8-TBG-LacZ 40 0.0 1.0
2 16 18 20
(Ctrl) and then fed a high-fat diet
20 Albumin 0.5
for 8 weeks. Plasma total cholesterol 1
VLDL
concentration was measured (n = 6 0 0 0.0
Ctrl Cre 15 18 21 24 27 30 33 Ctrl Cre
per group). (B) Serpine1fl/fl mice Elution volume (mL)
were treated with AAV8-TBG-Cre E Ctrl
F
Ctrl
Plasma triglyceride (mg/dL) Ctrl
(Cre) or control AAV8-TBG-LacZ * *

Plasma apoB-100 ( g/mL)

Cre Cre Cre

rising rate (mg/dL/min)

1500 40 300

Plasma triglyceride
(Ctrl) and then fed a high-fat diet *
*
for 8 weeks. Plasma apoB was 30 *
1000 200 *
measured by immunoblot (n = 6 * 20
per group). (C) Serpine1fl/fl mice 500 * 100
10
were treated with AAV8-TBG-Cre
(Cre) or control AAV8-TBG-LacZ 0 0 0
0 30 60 90 120 60-90 90-120 0 2 4
(Ctrl) and then fed a high-fat diet Time (mins) mins mins Time post-oral gavage
for 8 weeks. Plasma samples were with olive oil (hours)

p
subjected to FPLC fractionation Plasma from humans with SERPINE1-/- and unaffected age/gender/BMI matched individuals

and assayed for cholesterol G P = 0.06 * * H * I

concentration. The cholesterol in
cholesterol (mg/dL)

15 150 600 4 80
cholesterol (mg/dL)

VLDL-associated
Plasma apoB-100
r= -0.59, P < 0.01

VLDL Z-average
Plasma VLDL-

VLDL fractions is shown in the

diameter (nm)
tPA (ng/mL)
Plasma LDL-

3 60
smaller graph in the inset (n = 6 per 10 100 400
( g/mL)

group). (D) Serpine1fl/fl mice were 2 40

5 50 200

g
treated with AAV8-TBG-Cre (Cre) or 1 20
control AAV8-TBG-LacZ (Ctrl) and
0 0 0 0 0
then fed a high-fat diet for 8 weeks. 0 1 2 3 4
-/-

-/-
R ted

PI d
R ted

ed
E1

VLDL-associated tPA (ng/mL)

E1

E1
E1

Liver lysates were assayed for

R te

ct
SE fec

SE ffec
SE fec

N
N

ffe
PI

PI

PI
PAI-1–free tPA concentration by ELISA
f

y
na

na

R
na

na
SE
U

U
U

U
(n = 6 per group). (E) Serpine1fl/fl
mice were treated with AAV8-TBG- J
Cre (Cre) or control AAV8-TBG-GFP
(Ctrl) and then fed a normal chow
diet for 4 weeks. Mice were injected
with P407 i.p. to assess VLDL
secretion (n = 10 per group).
(F) Serpine1fl/fl mice were treated
with AAV8-TBG-Cre (Cre) or control
AAV8-TBG-GFP (Ctrl) and then

y g
fed a normal chow diet for
6 weeks. The mice had their food
withdrawn for 5 hours and then
were euthanized at 0, 2, and

p
4 hours after oral gavage with olive
oil. Plasma apoB-100 concentration was measured by ELISA. (G) Plasma samples (J) A schematic diagram depicting how tPA–PAI-1 interaction in hepatocytes
from SERPINE1-deficient humans and unaffected age-, gender-, and BMI-matched determines VLDL assembly. Without lipid stimulation, tPA interacts with apoB and

,
individuals from the same community were assayed for VLDL cholesterol, LDL inhibits MTP-apoB interaction in the ER, thereby limiting MTP-mediated apoB
cholesterol, and apoB-100 concentrations (n = 10 per group). (H) tPA concentration lipidation and VLDL assembly. When hepatocytes are loaded with lipid, PAI-1
was measured in VLDL isolated by ultracentrifugation from the plasma of the sequesters free tPA away from apoB and increases VLDL assembly. The diagram
subjects in (G) (n = 10 per group). (I) VLDL from the plasma of the subjects in (G) was generated using biorender.com. Data are shown as means ± SEMs; P values
was analyzed by dynamic light scattering (n = 10 per group). The correlation were calculated by two-tailed Student’s t test [(A), (D), (E), and (F)], paired
between VLDL-associated tPA and VLDL diameter was calculated (n = 10 per group). Student’s t test [(G) and (H)], or Pearson’s correlation analysis (I). *P < 0.05.

serine protease domains (64). The extracellu- by inhibiting MTP-apoB interaction. The inhib- actions were found to occur, additional studies
lar function of tPA in fibrinolysis is linked to ition of MTP binding to apoB by tPA is likely to would then focus on the possible impact of
its serine protease activity, which mediates be competitive because both tPA and MTP these interactions on fibrinolysis and the non-
the enzymatic conversion of plasminogen to bind to the lysine-rich domains of the apoB fibrinolytic functions of these lysine-binding
plasmin, which carries out fibrinolysis (64). N terminus. However, whether apoB interacts proteins (65–67).
Our data suggest that the K2 domain, inde- with other lysine-binding proteins, such as plas- PAI-1 is a rapidly acting serine protease in-
pendent of its proteolytic activity, confers tPA’s minogen, in the hepatocyte ER, and whether hibitor (serpin) of tPA (68). Upon binding to tPA,
regulatory function in the assembly of VLDL apoB and tPA interact in the circulation re- PAI-1 inactivates the serine protease activity of
by direct binding to the apoB N terminus, there- present topics for future studies. If these inter- tPA and inhibits tPA’s other protease-independent

Dai et al., Science 381, eadh5207 (2023) 1 September 2023 8 of 12

RES EARCH | R E S E A R C H A R T I C L E
functions, such as tPA’s ability to bind fibrin
(53) and tPA’s receptor-mediated activities (65).
Previous studies have mostly focused on the
tPA–PAI-1 interaction in the blood (69) and on
the vascular endothelial surface (70), which
are essential to maintain a balanced fibrino-
lytic potential. Our findings show that tPA and
PAI-1 also interact within hepatocytes, fine-
tuning apoB-VLDL assembly. The intracellular
interaction between PAI-1 and tPA has also been
observed in other cell types, such as colonic
epithelial cells (71). Gerard et al. observed that
PAI-1–deficient mice have an increase in free
tPA in the colon (71). These data are consistent
with our findings that lowering hepatocyte
PAI-1 expression leads to an increase in free tPA
levels in both human primary hepatocytes and
mouse livers. The impact of the intracellular
interaction of tPA and PAI-1 in other cell types
is worthy of investigating in future studies.
Our findings provide possible mechanistic
explanations for the associations of PLAT and
SERPINE1 polymorphisms with circulating lipid
levels and atherosclerosis (72–75). Congenital
tPA deficiency in humans has not been re-
ported, which suggests that null mutations
in the tPA gene may be lethal in utero. An
insertion/deletion (I/D) polymorphism, rs4646972,
is located in the eighth intron of the tPA gene
(76). Homozygotes for the deletion (DD) showed
lower plasma tPA levels compared with those
carrying an insertion locus (DI and II) (77).
Moreover, homozygotes for the deletion (DD)
had a threefold higher risk of intracranial artery
atherosclerosis (72) and higher plasma total
cholesterol and LDL cholesterol (73) compared
with the homozygotes for the insertion (II).
Subjects with the DD genotype have a lower
forearm vascular release of tPA compared with
the individuals with the II genotype (78). Two
SERPINE1 single-nucleotide polymorphisms
(SNPs) have also been shown to be related to
plasma lipid levels. The rs6950982 is associ-
ated with elevated plasma PAI-1 levels (79) and
higher total cholesterol, LDL cholesterol, and
triglyceride (74). Similarly, the rs2227674 is asso-
ciated with higher plasma PAI-1 levels (79) and
increased triglyceride (75). Genome-wide associ-
ation studies showed that genetic variants har-
boring SERPINE1 loci, leading to elevated PAI-1
levels, are associated with a higher risk of
coronary artery disease (80, 81). Heterozygous
PAI-1 deficiency is associated with a lower com-
posite coronary artery disease risk score (55).
Although further investigation is needed, these
studies raise the possibility that targeting
hepatocyte PAI-1 may have promise in treat-
ing atherosclerotic CVDs.
Moreover, because our data suggest that he-
patic PAI-1–free tPA limits VLDL assembly and
reduces blood apoB cholesterol, future studies
should consider the genetic interaction of tPA
and PAI-1 SNPs as well as SNPs that affect tPA
and PAI-1 expression in hepatocytes. Two SNPs,
Dai et al., Science 381, eadh5207 (2023) 1 September 2023
rs9399599 and rs7301826, are associated with
liver expression of STXBP5 and STX2, both
of which are exocytosis mediators (82). These
SNPs are also associated with circulating tPA
levels in humans (82, 83). Future population
genetic studies are required to investigate the
correlation of these SNPs—their associated
expression quantitative trait loci (eQTLs) in
hepatocytes—with atherogenic lipoprotein and
lipid levels.
We found that the formation of the tPA–
PAI-1 complex in the hepatocyte is increased
in response to fatty acid loading. Postprandial
lipemia results in an influx of fatty acids into
hepatocytes and stimulates the lipidation of
apoB and subsequent VLDL secretion, which
serves as a physiological response to an in-
creased lipid burden within the hepatocyte
(50). Fatty acid ingestion increases the degree
and duration of the association between MTP
and apoB in hepatocytes (35), which could
contribute to fatty acid–induced VLDL se-
cretion. However, the mechanisms by which
fatty acids promote the association of apoB
and MTP have remained unclear. Our in vivo
and in vitro studies indicate an acute increase
in the formation of the tPA–PAI-1 complexes
within the hepatocyte after fatty acid stimu-
lation without an alteration of either total tPA
or PAI-1 protein levels in the cells. Moreover,
hepatocyte PAI-1 deficiency leads to increased
free tPA and reduced fatty acid–induced VLDL
secretion. These data offer a mechanistic ex-
planation behind the observed increase in
VLDL secretion after postprandial lipid load-
ing of hepatocytes. Additionally, apoB and
MTP have similar functions in both hepato-
cytes and intestinal epithelial cells, where they
are involved in chylomicron assembly (84), but
whether tPA in the intestine affects chylomi-
cron assembly requires further study.
Obesity increases the risk of atherothrom-
botic events, such as myocardial infarction and
stroke. Elevated PAI-1 in obesity could be a
mechanistic link between obesity and increased
CVD risk by both reducing fibrinolytic poten-
tial and promoting dyslipidemia. Both plasma
and liver PAI-1 levels are increased in obesity
(19, 85), and hepatocytes are an important re-
source of plasma PAI-1 (19, 86). Approximately
60 to 70% of obese people have dyslipidemia
with increased VLDL production (87). This
study provides possible links among obesity,
increased VLDL production, and CVD through
hepatic PAI-1.
Previous studies have shown that inhibiting
VLDL production by using MTP inhibitors in-
creases the risk of steatosis in humans, limit-
ing the wide use of the agents. However, there
are no reports to date that PAI-1 deficiency in
humans is associated with liver steatosis. Addi-
tionally, global PAI-1 inhibition or PAI-1 deficiency
in mice attenuates hepatic steatosis (88–90).
Notably, our study used a mouse model with
hepatocyte-specific PAI-1 deficiency (Fig. 6, A
to F), whereas the studies mentioned above
used mice with global PAI-1 deficiency (88–90).
This distinction could be significant because
PAI-1 is expressed in other cell types, including
adipocytes (91). Moreover, mice with global PAI-1
deficiency exhibit resistance to diet-induced
obesity and display decreased adipocyte size
compared with WT control mice (90, 92). Thus,
the attenuated hepatic steatosis seen in global
PAI-1 deficiency mice might be the result of
overall metabolic changes, including body weight
loss, which is not seen in mice with hepatocyte
PAI-1 deficiency. Therefore, targeting tPA–PAI-1
balance in the hepatocytes may be an ideal
therapeutic strategy for dyslipidemia and ath-
erosclerotic CVDs, but this possibility requires
more extensive human-based studies.
This study reveals that the tPA–PAI-1 inter-
action determines VLDL assembly in hepato-
p
cytes. These findings provide mechanistic
insight into the physiologic response to lipid
ingestion and to the atherosclerosis-relevant
process of excessive apoB-lipoprotein produc-
tion by the liver. The latter insight may suggest
ideas to therapeutically lower atherogenic
g
lipoproteins to prevent the development of
atherothrombotic vascular disease.
Materials and methods summary
Please refer to the supplementary materials for
y
the complete materials and methods.
Plasma from humans with PAI-1 deficiency
Plasma samples were collected from members
of the Berne Amish community, who harbor a
frameshift mutation in SERPINE1 (SERPINE1−/−;
n = 10) (55) and their age-, gender-, and BMI-
matched control individuals from the same
community (n = 10). The institutional review
boards (IRBs) at the Indiana Center for Hem-
y g
ophilia and Thrombosis and Medical College
of Wisconsin (MCW) approved the study pro-
tocols. The study participants provided written
informed consent.
p
Mice
Platfl/fl mice were generated using WT C57BL6/J
,
mice through the homologous recombina-
tion in embryonic stem cell–based approach
(Biocytogen, Wakefield, MA). Briefly, a tar-
geting construct is designed to insert loxP sites
into the introns 3 and 6, to flox the exons 4 to
6 of Plat gene. This design is to conditionally
knock out the exons 4 to 6 of Plat by Cre-loxP
system. Hepatocyte tPA knockout mice were
generated by administering Platfl/fl mice with
an AAV8 expressing a Cre recombinase driven
by the thyroxine-binding globulin (TBG) pro-
moter, AAV8-TBG-cre (Cre), and Platfl/fl mice
receiving AAV8-TBG-GFP (GFP) were used as
controls. To silence tPA expression in hepto-
cytes, mice were intravenously injected with
AAV8 virus containing shPlat (AAV8-H1-shPlat)
9 of 12
RES EARCH | R E S E A R C H A R T I C L E
(19). Age-matched control mice were injected
with AAV8-H1-scramble silencing control. Ldlr−/−,
Apoe−/−, and WT C57BL/6J mice, used for si-
lencing hepatocyte tPA, were purchased from
Jackson Laboratory (JAX) (cat. nos. 002207,
002052, and 000664, respectively). Ldlr−/− and
Apoe−/− mice were fed with WD (Teklad, cat.
no. TD 88137). WT mice were on a standard
chow (Lab diet, cat. no. 5053), DIO diet (Re-
search Diets, cat. no. 12492), or WD (Teklad, cat.
no. TD 88137). To express tPA in hepatocytes,
WT C57BL/6J mice and holo-tPA-KO mice
(Jax, cat. no. 002508), on a standard chow diet,
received intravenous injection of AAV8- TBG-
Plat. For all experiments, mice were main-
tained on a 12-hour light–12-hour dark cycle
with free access to normal chow, WD, or DIO
diet and water. Mice of the same age and
similar weight were randomly assigned to ex-
perimental and control groups. Plasma lipids
were assayed in blood collected after a 5-hour
withdrawal of food. We use power calculations
to determine the number of mice for each ex-
periment. We include both male and female
mice. Mice of the same age and weight are ran-
domly assigned to groups; exclusion criteria are
death, injury requiring euthanasia, or weight
loss >10%, assuming these are rare events that
are not statistically different between groups.
All endpoint assays and analyses were con-
ducted by researchers who were blinded to the
identity of the cohort. All mouse experiments
were performed with the approval of the In-
stitutional Animal Care and Use Committee
(IACUC) of Biomedical Resource Center at
MCW and the Institutional Animal Care and
Use Committee of Columbia University Irving
Medical Center.
Vector constructs
AAV8-TBG-cre and AAV8-TBG-GFP were pur-
chased from Addgene. As previously described
(19, 20), AAV8-H1–short hairpin RNA (shRNA)
construct targeting murine Plat was made by
annealing complementary oligonucleotides
and then ligating them into the pAAV-RSV-
GFPH1 vector. AAV8-TBG-Plat was purchased
from Vector Biolabs. Plasmid expressing WT
human tPA, pCMV3-tPA-HA, was purchased
from Sino Biologic (Beijing, China). The plas-
mid constructs to express human tPA mutants
were generated by Versiti BRI Core based on
the pCMV3-tPA-HA. Constructed tPA mutants
include tPA-S513A, tPA-KDEL, tPA-D-K2-HA,
and tPA-D236, 238N. tPA-D-K2 refers to the
tPA mutant with the replacement of K2 with
K1, leading to the presence of two copies of K1
but no K2. As previously described (93), the pur-
pose of this design is to create a mutant tPA that
lacks K2 but mimics normal tPA structure.
Human primary hepatocyte experiments
Human primary hepatocytes were obtained
from the Liver Tissue Cell Distribution System
Dai et al., Science 381, eadh5207 (2023) 1 September 2023
at the University of Pittsburgh (Pittsburgh,
Pennsylvania, USA). All cells were cultured in
Williams’ Medium E supplemented with He-
patocyte Maintenance Supplement Pack (Thermo
Fisher Scientific, cat. no. CM4000). Experiments
were conducted as described in the figure leg-
ends. Cells were harvested, and culture media
were collected, snap-frozen in liquid nitrogen,
and stored at –80°C until processing. The age
and gender information of human donors are
listed in table S3.
Pulse-chase assay for apoB secretion
Hepatocyte apoB-100 secretion was assayed
using the 3[H] labeling method as previously
described (31). Human primary hepatocytes or
McA-RH7777 cells were washed with leucine-free
media and pulsed with 3[H] leucine (80 uCi/ml;
160 Ci/mmol, Perkin Elmer, cat. no. NET1166005MC)
for 20 min. The 3[H] leucine–containing medium
was removed, and cells were incubated with
fresh Dulbecco’s minimum essential medium
(DMEM) for an additional 0.5, 1, or 3 hours. apoB
was immunoprecipitated from cell homogenates
and media using anti-apoB antibodies (Sigma-
Aldrich, cat. no. AB742). Nonlabeled apoB-100
standards were added to the precipitates, and the
samples were separated by SDS–polyacrylamide
gel electrophoresis (SDS-PAGE) gel. Gels were
silver stained, the bands corresponding to apoB-
100 were excised, and radioactivity associated
with apoB-100 was quantified by a scintillation
counter.
Neutral lipid transfer activity assay
Neutral lipid transfer activity in cultured he-
patocyte microsomal fractions was measured
using a commercially available kit (Sigma-
Aldrich, MAK110) as previously (36). Specif-
ically, hepatocytes were homogenized in low
hypotonic buffer (10 mM Tris-HCl, 1 mM EGTA,
and 1 mM MgCl2; pH 7.4) using a Polytron ho-
mogenizer. Microsomes were isolated by ultra-
centrifugation (SW55 Ti rotor, 50,000 rpm,
1 hour). Neutral lipid transfer activity was as-
sayed using the kit according to the manu-
facturer’s manual. Isolated microsomes were
incubated with donor vesicles containing fluo-
rescent lipids and the acceptor vesicles–LDL.
The fluorescence signal was self-quenched
when labeled lipids existed within the donor
vesicles but would be detected after lipids trans-
ferred to acceptor vesicles.
Solid-phase protein binding assay
Solid-phase binding was performed in poly-
styrene microtiter plates using an enzyme-
linked immunosorbent assay (ELISA). Microtiter
plate wells were coated with LDL at 5 mg/ml in
coating buffer [tris-buffered saline (TBS)] over-
night at 4°C. Unbound sites were blocked with
3% nonfat milk in TBS for 1 hour at 37°C. After
washing with TBS containing 0.05% Tween 20
(TBS-Tween), tPA was added to the wells at
concentrations from 0 to 20 mg/ml in TBS-Tween.
After a 1-hour incubation at 37°C, the wells were
washed with TBS-Tween. Bound proteins were
reacted with anti-tPA [1 mg/ml immunoglobulin
G (IgG) in TBS-Tween in the presence of 3%
nonfat milk] followed by goat anti-rabbit IgG
conjugated to horseradish peroxidase. Trime-
thylboron (TMB) substrates were added. After
stopping the reaction, the absorbance at 450 nm
was measured.
SPR
Studies of the binding of recombinant tPA to
purified LDL were performed with a Biacore
S200 SPR instrument (Biacore) using a CM5
sensor chip (Cytiva, cat. no. 29149603) (94). LDL
was attached to the chip using amine coupling
chemistry, according to the manufacturer’s in-
structions. In brief, the chip surface was pre-
pared by exposing the carboxylated dextran
p
matrix to an aqueous solution containing 0.4 M
1-ethyl-3-(3-dimethylaminopropyl) carbodiimide
and 0.1 M N-hydroxysuccinimide (10 ml/min for
7 min). Then LDL (500 mg/ml in 10 mM sodium
acetate buffer, pH 5.5) flowed across the chip
surface at the same rate for 7 min, followed by
1 M ethanolamine-HCl (pH 8.5) at 19 ml/min for
g
7 min to deactivate excessive reactive groups and
remove any noncovalently bound LDL. This
procedure led to ~6000 response units (RUs)
of LDL immobilized. To monitor LDL associ-
y
ation with tPA, tPA solutions in HBS-E buffer
(Biacore) (0.01 M Hepes, 0.15 M NaCl, 3 mM
EDTA, pH 7.4) in the concentration range of
10 to 500 mg/ml were flowed across the chip
at 20 ml/min for 6 min at room temperature.
The dissociation of tPA was then monitored
by washing the surface for 6 min with HES
buffer alone. tPA flowed across an activated,
but uncoated CM5 chip under the same condi-
tions as the control for nonspecific binding.
y g
VLDL particle diameters analysis
VLDL particles (the particle density is < 1.006 g/ml)
were isolated by KBr density ultracentrifuga-
p
tion (95) followed by two methods to measure
their diameters. First, isolated VLDL particles
were negatively stained with 20 g/L phospho-
,
tungstic acid (pH 7.0) for 2 min and then viewed
under a Philips CM10 electron microscope.
The mean diameters of VLDL particles were
determined using Image-Pro Plus 5.0 image
analysis software. Second, the hydrodynamic
diameters of isolated VLDL particles were
measured using a Zetasizer mV dynamic laser
light scattering instrument (Malvern Instru-
ments) at 633 nm. VLDL samples were trans-
ferred to a quartz cuvette, and light scatter
readings were performed at 20°C (25–27).
Statistics analysis
The data described in the study were gener-
ated from biological replicates. The number
for human participants research and mouse
10 of 12
RES EARCH | R E S E A R C H A R T I C L E
experiments are described in the figure leg-
ends, where applicable. The in vitro cell ex-
periments were repeated at least three times.
All results are presented as means ± SEMs.
P values were calculated using two-tailed Stu-
dent’s t test for data that passed the normality
test or the Mann-Whitney rank-sum U test for
data that were not normally distributed. One-
way analysis of variance (ANOVA) with post hoc
Tukey’s test was used to evaluate differences
among groups when three or more groups were
analyzed.
Study approval
All mouse experiments were conducted with
the approval of the IACUC and Institutional
Biosafety Committee of MCW and the IACUC
of Columbia University Irving Medical Center.
The use of human cells and plasma samples
in this study was approved by the IRB at the
MCW and Indiana Hemophilia and Throm-
bosis Center. All participants provided written
informed consent.
RE FE RENCES AND N OT ES
1. F. J. Brunner et al., Application of non-HDL cholesterol for
population-based cardiovascular risk stratification: Results
from the Multinational Cardiovascular Risk Consortium. Lancet
394, 2173–2183 (2019). doi: 10.1016/S0140-6736(19)32519-X;
pmid: 31810609
2. N. A. Marston et al., Association of Apolipoprotein B–Containing
Lipoproteins and Risk of Myocardial Infarction in Individuals
With and Without Atherosclerosis: Distinguishing Between
Particle Concentration, Type, and Content. JAMA Cardiol. 7,
250–256 (2022). doi: 10.1001/jamacardio.2021.5083;
pmid: 34773460
3. N. J. Stone et al., 2013 ACC/AHA guideline on the treatment
of blood cholesterol to reduce atherosclerotic cardiovascular
risk in adults: A report of the American College of Cardiology/
American Heart Association Task Force on Practice Guidelines.
J. Am. Coll. Cardiol. 63, 2889–2934 (2014). doi: 10.1016/
j.jacc.2013.11.002; pmid: 24239923
4. H. Mabuchi et al., Effect of an inhibitor of 3-hydroxy-3-
methyglutaryl coenzyme A reductase on serum lipoproteins
and ubiquinone-10 levels in patients with familial
hypercholesterolemia. N. Engl. J. Med. 305, 478–482 (1981).
doi: 10.1056/NEJM198108273050902; pmid: 7254297
5. J. D. Horton, J. C. Cohen, H. H. Hobbs, PCSK9: A convertase
that coordinates LDL catabolism. J. Lipid Res. 50, S172–S177
(2009). doi: 10.1194/jlr.R800091-JLR200; pmid: 19020338
6. C. Reith, J. Armitage, Management of residual risk after statin
therapy. Atherosclerosis 245, 161–170 (2016). doi: 10.1016/
j.atherosclerosis.2015.12.018; pmid: 26722833
7. N. D. Wong, Residual Risk After Treatment of Patients With
Atherosclerotic Cardiovascular Disease With Proprotein
Convertase Subtilisin-Kexin Type 9 Monoclonal Antibody
Therapy (from the Further Cardiovascular Outcomes Research
With PCSK9 Inhibition in Subjects With Elevated Risk Trial).
Am. J. Cardiol. 120, 1220–1222 (2017). doi: 10.1016/
j.amjcard.2017.06.063; pmid: 28802511
8. A. D. Sniderman et al., Apolipoprotein B Particles and Cardiovascular
Disease: A Narrative Review. JAMA Cardiol. 4, 1287–1295 (2019).
doi: 10.1001/jamacardio.2019.3780; pmid: 31642874
9. A. Sirwi, M. M. Hussain, Lipid transfer proteins in the assembly
of apoB-containing lipoproteins. J. Lipid Res. 59, 1094–1102
(2018). doi: 10.1194/jlr.R083451; pmid: 29650752
10. N. O. Davidson, G. S. Shelness, Apolipoprotein B: mRNA
editing, lipoprotein assembly, and presecretory degradation.
Annu. Rev. Nutr. 20, 169–193 (2000). doi: 10.1146/
annurev.nutr.20.1.169; pmid: 10940331
11. M. M. Hussain, J. Shi, P. Dreizen, Microsomal triglyceride
transfer protein and its role in apoB-lipoprotein assembly.
J. Lipid Res. 44, 22–32 (2003). doi: 10.1194/
jlr.R200014-JLR200; pmid: 12518019
12. E. A. Fisher, The degradation of apolipoprotein B100: Multiple
opportunities to regulate VLDL triglyceride production by
Dai et al., Science 381, eadh5207 (2023) 1 September 2023
different proteolytic pathways. Biochim. Biophys. Acta 1821,
778–781 (2012). doi: 10.1016/j.bbalip.2012.02.001;
pmid: 22342675
13. S. Munkvad, J. Gram, J. Jespersen, A depression of active
tissue plasminogen activator in plasma characterizes patients
with unstable angina pectoris who develop myocardial
infarction. Eur. Heart J. 11, 525–528 (1990). doi: 10.1093/
oxfordjournals.eurheartj.a059745; pmid: 2112469
14. J. Gram, J. Jespersen, A selective depression of tissue plasminogen
activator (t-PA) activity in euglobulins characterises a risk group
among survivors of acute myocardial infarction. Thromb. Haemost.
57, 137–139 (1987). doi: 10.1055/s-0038-1651081; pmid: 3110993
15. J. Gram, J. Jespersen, C. Kluft, D. C. Rijken, On the usefulness
of fibrinolysis variables in the characterization of a risk group for
myocardial reinfarction. Acta Med. Scand. 221, 149–153 (1987).
doi: 10.1111/j.0954-6820.1987.tb01259.x; pmid: 3591453
16. J. H. Jansson, B. Johansson, K. Boman, T. K. Nilsson,
Hypo-fibrinolysis in patients with hypertension and elevated
cholesterol. J. Intern. Med. 229, 309–316 (1991). doi: 10.1111/
j.1365-2796.1991.tb00352.x; pmid: 1902868
17. R. Yamada, S. Yamada, A. Ishii, M. Sasamata, S. Watanabe,
Association between tissue plasminogen activator and serum
lipids in healthy volunteers. Ann. Med. 22, 313–318 (1990).
doi: 10.3109/07853899009147913; pmid: 2127223
18. C. J. Glueck et al., Endogenous testosterone, fibrinolysis, and
coronary heart disease risk in hyperlipidemic men. J. Lab. Clin.
Med. 122, 412–420 (1993). pmid: 8228555
19. Z. Zheng et al., Interacting hepatic PAI-1/tPA gene regulatory
pathways influence impaired fibrinolysis severity in obesity.
J. Clin. Invest. 130, 4348–4359 (2020). doi: 10.1172/
JCI135919; pmid: 32657780
20. Z. Zheng et al., An ATF6-tPA pathway in hepatocytes contributes to
systemic fibrinolysis and is repressed by DACH1. Blood 133, 743–753
(2019). doi: 10.1182/blood-2018-07-864843; pmid: 30504459
21. J. Yamaguchi, M. V. Gamble, D. Conlon, J. S. Liang,
H. N. Ginsberg, The conversion of apoB100 low density
lipoprotein/high density lipoprotein particles to apoB100 very
low density lipoproteins in response to oleic acid occurs
in the endoplasmic reticulum and not in the Golgi in McA
RH7777 cells. J. Biol. Chem. 278, 42643–42651 (2003).
doi: 10.1074/jbc.M306920200; pmid: 12917397
22. J. S. Millar, D. A. Cromley, M. G. McCoy, D. J. Rader,
J. T. Billheimer, Determining hepatic triglyceride production in
mice: Comparison of poloxamer 407 with Triton WR-1339.
J. Lipid Res. 46, 2023–2028 (2005). doi: 10.1194/
jlr.D500019-JLR200; pmid: 15995182
23. M. E. Hinsdale, P. M. Sullivan, H. Mezdour, N. Maeda, ApoB-48
and apoB-100 differentially influence the expression of type-III
hyperlipoproteinemia in APOE*2 mice. J. Lipid Res. 43, 1520–1528
(2002). doi: 10.1194/jlr.M200103-JLR200; pmid: 12235184
24. S. Tiwari, S. A. Siddiqi, Intracellular trafficking and secretion of
VLDL. Arterioscler. Thromb. Vasc. Biol. 32, 1079–1086 (2012).
doi: 10.1161/ATVBAHA.111.241471; pmid: 22517366
25. D. Cheng et al., Very Low Density Lipoprotein Assembly Is
Required for cAMP-responsive Element-binding Protein H
Processing and Hepatic Apolipoprotein A-IV Expression. J. Biol.
Chem. 291, 23793–23803 (2016). doi: 10.1074/jbc.
M116.749283; pmid: 27655915
26. R. B. Weinberg, J. W. Gallagher, M. A. Fabritius, G. S. Shelness,
ApoA-IV modulates the secretory trafficking of apoB and the
size of triglyceride-rich lipoproteins. J. Lipid Res. 53, 736–743
(2012). doi: 10.1194/jlr.M019992; pmid: 22257482
27. T. Sakurai et al., Measurement of lipoprotein particle sizes using
dynamic light scattering. Ann. Clin. Biochem. 47, 476–481 (2010).
doi: 10.1258/acb.2010.010100; pmid: 20736248
28. L. Ozcan et al., Hepatocyte DACH1 Is Increased in Obesity via
Nuclear Exclusion of HDAC4 and Promotes Hepatic Insulin
Resistance. Cell Rep. 15, 2214–2225 (2016). doi: 10.1016/
j.celrep.2016.05.006; pmid: 27239042
29. D. S. Ghorpade et al., Hepatocyte-secreted DPP4 in obesity
promotes adipose inflammation and insulin resistance. Nature 555,
673–677 (2018). doi: 10.1038/nature26138; pmid: 29562231
30. X. Wang et al., Cholesterol Stabilizes TAZ in Hepatocytes
to Promote Experimental Non-alcoholic Steatohepatitis.
Cell Metab. 31, 969–986.e7 (2020). doi: 10.1016/
j.cmet.2020.03.010; pmid: 32259482
31. D. Sahoo et al., ABCA1-dependent lipid efflux to apolipoprotein
A-I mediates HDL particle formation and decreases VLDL
secretion from murine hepatocytes. J. Lipid Res. 45, 1122–1131
(2004). doi: 10.1194/jlr.M300529-JLR200; pmid: 14993246
32. S. Rustaeus et al., Assembly of very low density lipoprotein:
A two-step process of apolipoprotein B core lipidation.
J. Nutr. 129, 463S–466S (1999). doi: 10.1093/jn/129.2.463S;
pmid: 10064310
33. B. Lewis, Classification of lipoproteins and lipoprotein
disorders. J. Clin. Pathol. 5, 26–31 (1973). doi: 10.1136/
jcp.s1-5.1.26; pmid: 4354845
34. M. Pan, J.-S. Liang, E. A. Fisher, H. N. Ginsberg, The late addition of
core lipids to nascent apolipoprotein B100, resulting in the
assembly and secretion of triglyceride-rich lipoproteins, is
independent of both microsomal triglyceride transfer protein activity
and new triglyceride synthesis. J. Biol. Chem. 277, 4413–4421
(2002). doi: 10.1074/jbc.M107460200; pmid: 11704664
35. X. Wu, M. Zhou, L. S. Huang, J. Wetterau, H. N. Ginsberg,
Demonstration of a physical interaction between microsomal
triglyceride transfer protein and apolipoprotein B during the
assembly of ApoB-containing lipoproteins. J. Biol. Chem. 271,
10277–10281 (1996). doi: 10.1074/jbc.271.17.10277; pmid: 8626595
36. H. Athar, J. Iqbal, X. C. Jiang, M. M. Hussain, A simple,
rapid, and sensitive fluorescence assay for microsomal
triglyceride transfer protein. J. Lipid Res. 45, 764–772 (2004).
doi: 10.1194/jlr.D300026-JLR200; pmid: 14754905
37. C. E. Chandler et al., CP-346086: An MTP inhibitor that
lowers plasma cholesterol and triglycerides in experimental
animals and in humans. J. Lipid Res. 44, 1887–1901 (2003).
doi: 10.1194/jlr.M300094-JLR200; pmid: 12837854
38. D. A. Andres, I. M. Dickerson, J. E. Dixon, Variants of the
carboxyl-terminal KDEL sequence direct intracellular retention.
J. Biol. Chem. 265, 5952–5955 (1990). doi: 10.1016/
S0021-9258(19)39273-7; pmid: 2318841
p
39. A. Gerondopoulos et al., A signal capture and proofreading
mechanism for the KDEL-receptor explains selectivity and
dynamic range in ER retrieval. eLife 10, e68380 (2021).
doi: 10.7554/eLife.68380; pmid: 34137369
40. H. Pu et al., Protease-independent action of tissue
plasminogen activator in brain plasticity and neurological
recovery after ischemic stroke. Proc. Natl. Acad. Sci. U.S.A.
116, 9115–9124 (2019). doi: 10.1073/pnas.1821979116;
g
pmid: 30996120
41. G. A. W. De Munk et al., Binding of tissue-type plasminogen
activator to lysine, lysine analogues, and fibrin fragments.
Biochemistry 28, 7318–7325 (1989). doi: 10.1021/
bi00444a026; pmid: 2510823
42. E. J. Weening-Verhoeff et al., Involvement of aspartic and
y
glutamic residues in kringle-2 of tissue-type plasminogen
activator in lysine binding, fibrin binding and stimulation of
activity as revealed by chemical modification and
oligonucleotide-directed mutagenesis. Protein Eng. Des. Sel. 4,
191–198 (1990). doi: 10.1093/protein/4.2.191; pmid: 1963688
43. J. P. Segrest, M. K. Jones, H. De Loof, N. Dashti, Structure
of apolipoprotein B-100 in low density lipoproteins. J. Lipid Res.
42, 1346–1367 (2001). doi: 10.1016/S0022-2275(20)30267-4;
pmid: 11518754
44. M. M. Hussain, A. Bakillah, H. Jamil, Apolipoprotein B binding
to microsomal triglyceride transfer protein decreases with
increases in length and lipidation: Implications in lipoprotein
biosynthesis. Biochemistry 36, 13060–13067 (1997).
y g
doi: 10.1021/bi971395a; pmid: 9335568
45. M. M. Hussain, A. Bakillah, N. Nayak, G. S. Shelness, Amino
acids 430-570 in apolipoprotein B are critical for its binding to
microsomal triglyceride transfer protein. J. Biol. Chem. 273,
25612–25615 (1998). doi: 10.1074/jbc.273.40.25612;
pmid: 9748226
p
46. X. Wang, R. Pease, J. Bertinato, R. W. Milne, Well-defined
regions of apolipoprotein B-100 undergo conformational change
during its intravascular metabolism. Arterioscler. Thromb. Vasc.
Biol. 20, 1301–1308 (2000). doi: 10.1161/01.ATV.20.5.1301;
,
pmid: 10807746
47. J. S. Cohn, D. A. Wagner, S. D. Cohn, J. S. Millar, E. J. Schaefer,
Measurement of very low density and low density lipoprotein
apolipoprotein (Apo) B-100 and high density lipoprotein Apo A-I
production in human subjects using deuterated leucine. Effect of
fasting and feeding. J. Clin. Invest. 85, 804–811 (1990).
doi: 10.1172/JCI114507; pmid: 2107210
48. D. A. Lawrence et al., Serpin-protease complexes are trapped as
stable acyl-enzyme intermediates. J. Biol. Chem. 270, 25309–25312
(1995). doi: 10.1074/jbc.270.43.25309; pmid: 7592687
49. A. L. White, D. L. Graham, J. LeGros, R. J. Pease, J. Scott,
Oleate-mediated stimulation of apolipoprotein B secretion from
rat hepatoma cells: A function of the ability of apolipoprotein B
to direct lipoprotein assembly and escape presecretory
degradation. J. Biol. Chem. 267, 15657–15664 (1992).
doi: 10.1016/S0021-9258(19)49586-0; pmid: 1639804
50. J. L. Dixon, S. Furukawa, H. N. Ginsberg, Oleate stimulates
secretion of apolipoprotein B-containing lipoproteins from Hep
G2 cells by inhibiting early intracellular degradation of
apolipoprotein B. J. Biol. Chem. 266, 5080–5086 (1991).
doi: 10.1016/S0021-9258(19)67758-6; pmid: 1848237
11 of 12
RES EARCH | R E S E A R C H A R T I C L E
51. M. J. Perron, G. E. Blouse, J. D. Shore, Distortion of the catalytic
domain of tissue-type plasminogen activator by plasminogen
activator inhibitor-1 coincides with the formation of stable serpin-
proteinase complexes. J. Biol. Chem. 278, 48197–48203 (2003).
doi: 10.1074/jbc.M306184200; pmid: 14500731
52. J. A. Huntington, R. J. Read, R. W. Carrell, Structure of a serpin-
protease complex shows inhibition by deformation. Nature 407,
923–926 (2000). doi: 10.1038/35038119; pmid: 11057674
53. M. Kaneko, Y. Sakata, M. Matsuda, J. Mimuro, Interactions
between the finger and kringle-2 domains of tissue-type
plasminogen activator and plasminogen activator inhibitor-1.
J. Biochem. 111, 244–248 (1992). doi: 10.1093/
oxfordjournals.jbchem.a123744; pmid: 1314812
54. C. Jiang et al., Serpine 1 induces alveolar type II cell
senescence through activating p53-p21-Rb pathway in fibrotic
lung disease. Aging Cell 16, 1114–1124 (2017). doi: 10.1111/
acel.12643; pmid: 28722352
55. S. S. Khan et al., A null mutation in SERPINE1 protects against
biological aging in humans. Sci. Adv. 3, eaao1617 (2017).
doi: 10.1126/sciadv.aao1617; pmid: 29152572
56. T. Astrup, A. Stage, Isolation of a soluble fibrinolytic activator
from animal tissue. Nature 170, 929 (1952). doi: 10.1038/
170929a0; pmid: 13013265
57. D. Collen, H. R. Lijnen, The tissue-type plasminogen activator
story. Arterioscler. Thromb. Vasc. Biol. 29, 1151–1155 (2009).
doi: 10.1161/ATVBAHA.108.179655; pmid: 19605778
58. GUSTO Angiographic Investigators, The effects of tissue
plasminogen activator, streptokinase, or both on coronary-
artery patency, ventricular function, and survival after acute
myocardial infarction. N. Engl. J. Med. 329, 1615–1622 (1993).
doi: 10.1056/NEJM199311253292204; pmid: 8232430
59. D. Z. Wang, J. A. Rose, D. S. Honings, D. J. Garwacki,
J. C. Milbrandt, Treating acute stroke patients with intravenous
tPA. The OSF stroke network experience. Stroke 31, 77–81
(2000). doi: 10.1161/01.STR.31.1.77; pmid: 10625719
60. B. E. Sobel, Increased plasminogen activator inhibitor-1 and
vasculopathy. A reconcilable paradox. Circulation 99, 2496–2498
(1999). doi: 10.1161/01.CIR.99.19.2496; pmid: 10330378
61. I. Tabas, K. J. Williams, J. Borén, Subendothelial lipoprotein
retention as the initiating process in atherosclerosis: Update and
therapeutic implications. Circulation 116, 1832–1844 (2007).
doi: 10.1161/CIRCULATIONAHA.106.676890; pmid: 17938300
62. J. Borén, K. J. Williams, The central role of arterial retention of
cholesterol-rich apolipoprotein-B-containing lipoproteins in
the pathogenesis of atherosclerosis: A triumph of simplicity.
Curr. Opin. Lipidol. 27, 473–483 (2016). doi: 10.1097/
MOL.0000000000000330; pmid: 27472409
63. P. Libby, The molecular mechanisms of the thrombotic
complications of atherosclerosis. J. Intern. Med. 263, 517–527
(2008). doi: 10.1111/j.1365-2796.2008.01965.x; pmid: 18410595
64. D. Pennica et al., Cloning and expression of human tissue-type
plasminogen activator cDNA in E. coli. Nature 301, 214–221
(1983). doi: 10.1038/301214a0; pmid: 6337343
65. S. L. Gonias et al., PAI1 blocks NMDA receptor-mediated
effects of tissue-type plasminogen activator on cell signaling
and physiology. J. Cell Sci. 131, jcs217083 (2018). doi: 10.1242/
jcs.217083; pmid: 29930084
66. E. Mantuano, M. S. Lam, M. Shibayama, W. M. Campana,
S. L. Gonias, The NMDA receptor functions independently and
as an LRP1 co-receptor to promote Schwann cell survival
and migration. J. Cell Sci. 128, 3478–3488 (2015).
doi: 10.1242/jcs.173765; pmid: 26272917
67. A. L. Samson et al., Tissue-type plasminogen activator requires
a co-receptor to enhance NMDA receptor function. J. Neurochem.
107, 1091–1101 (2008). doi: 10.1111/j.1471-4159.2008.05687.x;
pmid: 18796005
68. S. Thorsen, M. Philips, J. Selmer, I. Lecander, B. Astedt,
Kinetics of inhibition of tissue-type and urokinase-type
plasminogen activator by plasminogen-activator inhibitor type
1 and type 2. Eur. J. Biochem. 175, 33–39 (1988). doi: 10.1111/
j.1432-1033.1988.tb14162.x; pmid: 3136015
69. W. L. Chandler et al., Clearance of tissue plasminogen activator
(TPA) and TPA/plasminogen activator inhibitor type 1 (PAI-1)
complex: Relationship to elevated TPA antigen in patients
with high PAI-1 activity levels. Circulation 96, 761–768 (1997).
doi: 10.1161/01.CIR.96.3.761; pmid: 9264480
70. Y. Sakata, M. Okada, A. Noro, M. Matsuda, Interaction of
tissue-type plasminogen activator and plasminogen activator
inhibitor 1 on the surface of endothelial cells. J. Biol. Chem.
263, 1960–1969 (1988). doi: 10.1016/S0021-9258(19)77972-1;
pmid: 3123483
71. G. E. Kaiko et al., PAI-1 augments mucosal damage in colitis.
Sci. Transl. Med. 11, eaat0852 (2019). doi: 10.1126/
scitranslmed.aat0852; pmid: 30842312
Dai et al., Science 381, eadh5207 (2023) 1 September 2023
72. M. S. Babu et al., Association of genetic variants of fibrinolytic
system with stroke and stroke subtypes. Gene 495, 76–80
(2012). doi: 10.1016/j.gene.2011.12.046; pmid: 22240314
73. M. Karadeniz, M. Erdogan, A. Berdeli, F. Saygili, C. Yilmaz,
4G/5G polymorphism of PAI-1 gene and Alu-repeat I/D
polymorphism of TPA gene in Turkish patients with polycystic
ovary syndrome. J. Assist. Reprod. Genet. 24, 412–418 (2007).
doi: 10.1007/s10815-007-9160-7; pmid: 17661167
74. M. Sotos-Prieto et al., Association between the rs6950982
polymorphism near the SERPINE1 gene and blood pressure and
lipid parameters in a high-cardiovascular-risk population:
Interaction with Mediterranean diet. Genes Nutr. 8, 401–409
(2013). doi: 10.1007/s12263-012-0327-1; pmid: 23225235
75. A. R. Bentley et al., Gene-based sequencing identifies lipid-
influencing variants with ethnicity-specific effects in African
Americans. PLOS Genet. 10, e1004190 (2014). doi: 10.1371/
journal.pgen.1004190; pmid: 24603370
76. M. Ludwig, K. D. Wohn, W. D. Schleuning, K. Olek, Allelic
dimorphism in the human tissue-type plasminogen activator (TPA)
gene as a result of an Alu insertion/deletion event. Hum. Genet. 88,
388–392 (1992). doi: 10.1007/BF00215671; pmid: 1346771
77. F. L. Sciacca et al., Genetic and plasma markers of venous
thromboembolism in patients with high grade glioma. Clin.
Cancer Res. 10, 1312–1317 (2004). doi: 10.1158/1078-0432.
CCR-03-0198; pmid: 14977830
78. C. Jern, P. Ladenvall, U. Wall, S. Jern, Gene polymorphism of t-PA
is associated with forearm vascular release rate of t-PA.
Arterioscler. Thromb. Vasc. Biol. 19, 454–459 (1999). doi: 10.1161/
01.ATV.19.2.454; pmid: 9974431
79. S. Kathiresan et al., Comprehensive survey of common genetic
variation at the plasminogen activator inhibitor-1 locus
and relations to circulating plasminogen activator inhibitor-1
levels. Circulation 112, 1728–1735 (2005). doi: 10.1161/
CIRCULATIONAHA.105.547836; pmid: 16172282
80. CARDIoGRAMplusC4D Consortium, A comprehensive 1000
Genomes–based genome-wide association meta-analysis of
coronary artery disease. Nat. Genet. 47, 1121–1130 (2015).
doi: 10.1038/ng.3396; pmid: 26343387
81. C. Song, S. Burgess, J. D. Eicher, C. J. O’Donnell, A. D. Johnson,
CHARGE Consortium Hemostatic Factor Working Group,
ICBP Consortium, CHARGE Consortium Subclinical Working
Group, Causal Effect of Plasminogen Activator Inhibitor Type 1 on
Coronary Heart Disease. J. Am. Heart Assoc. 6, e004918 (2017).
doi: 10.1161/JAHA.116.004918; pmid: 28550093
82. J. Huang et al., Genome-wide association study for circulating
tissue plasminogen activator levels and functional follow-up
implicates endothelial STXBP5 and STX2. Arterioscler. Thromb.
Vasc. Biol. 34, 1093–1101 (2014). doi: 10.1161/
ATVBAHA.113.302088; pmid: 24578379
83. G. Temprano-Sagrera et al., Multi-phenotype analyses of
hemostatic traits with cardiovascular events reveal novel genetic
associations. J. Thromb. Haemost. 20, 1331–1349 (2022).
doi: 10.1111/jth.15698; pmid: 35285134
84. J. Iqbal, M. Boutjdir, L. L. Rudel, M. M. Hussain, Intestine-specific
MTP and global ACAT2 deficiency lowers acute cholesterol
absorption with chylomicrons and HDLs. J. Lipid Res. 55,
2261–2275 (2014). doi: 10.1194/jlr.M047951; pmid: 25030663
85. J. B. Meigs et al., Hyperinsulinemia, hyperglycemia, and impaired
hemostasis: The Framingham Offspring Study. JAMA 283,
221–228 (2000). doi: 10.1001/jama.283.2.221; pmid: 10634338
86. T. Kietzmann, A. Samoylenko, U. Roth, K. Jungermann,
Hypoxia-inducible factor-1 and hypoxia response elements
mediate the induction of plasminogen activator inhibitor-1 gene
expression by insulin in primary rat hepatocytes. Blood 101,
907–914 (2003). doi: 10.1182/blood-2002-06-1693; pmid: 12393531
87. H. E. Bays et al., Obesity, adiposity, and dyslipidemia:
A consensus statement from the National Lipid Association.
J. Clin. Lipidol. 7, 304–383 (2013). doi: 10.1016/j.jacl.2013.04.001;
pmid: 23890517
88. A. S. Henkel, S. S. Khan, S. Olivares, T. Miyata, D. E. Vaughan,
Inhibition of Plasminogen Activator Inhibitor 1 Attenuates
Hepatic Steatosis but Does Not Prevent Progressive
Nonalcoholic Steatohepatitis in Mice. Hepatol. Commun. 2,
1479–1492 (2018). doi: 10.1002/hep4.1259; pmid: 30556037
89. S. M. Lee et al., TM5441, a plasminogen activator inhibitor-1
inhibitor, protects against high fat diet-induced non-alcoholic
fatty liver disease. Oncotarget 8, 89746–89760 (2017).
doi: 10.18632/oncotarget.21120; pmid: 29163785
90. L. J. Ma et al., Prevention of obesity and insulin resistance in
mice lacking plasminogen activator inhibitor 1. Diabetes 53,
336–346 (2004). doi: 10.2337/diabetes.53.2.336;
pmid: 14747283
91. C. H. Lundgren, S. L. Brown, T. K. Nordt, B. E. Sobel, S. Fujii,
Elaboration of type-1 plasminogen activator inhibitor from
adipocytes. A potential pathogenetic link between obesity and
cardiovascular disease. Circulation 93, 106–110 (1996).
doi: 10.1161/01.CIR.93.1.106; pmid: 8616916
92. K. Schäfer, K. Fujisawa, S. Konstantinides, D. J. Loskutoff,
Disruption of the plasminogen activator inhibitor 1 gene
reduces the adiposity and improves the metabolic profile of
genetically obese and diabetic ob/ob mice. FASEB J. 15,
1840–1842 (2001). doi: 10.1096/fj.00-0750fje; pmid: 11481248
93. C. Longstaff et al., The interplay between tissue plasminogen
activator domains and fibrin structures in the regulation of
fibrinolysis: Kinetic and microscopic studies. Blood 117,
661–668 (2011). doi: 10.1182/blood-2010-06-290338;
pmid: 20966169
94. M. Migliorini et al., High-affinity binding of plasminogen-
activator inhibitor 1 complexes to LDL receptor–related protein
1 requires lysines 80, 88, and 207. J. Biol. Chem. 295, 212–222
(2020). doi: 10.1074/jbc.RA119.010449; pmid: 31792055
95. D. Basu et al., Novel Reversible Model of Atherosclerosis and
Regression Using Oligonucleotide Regulation of the LDL
Receptor. Circ. Res. 122, 560–567 (2018). doi: 10.1161/
CIRCRESAHA.117.311361; pmid: 29321129
AC KNOWLED GME NTS
We thank the Indiana Hemophilia and Thrombosis Center Research
staff. We thank M. Kipp at the Versiti Blood Research Institute
(BRI) for assistance with the mouse colonies; E. Michalski and
p
B. Best at Versiti BRI Vector Core for the vector construction;
C. Wells at MCW Electron Microscopy Core Facility for the
electronic scanning of VLDL particles; R. Bohnsack at the MCW
Department of Biochemistry for instructing the SPR; W. Huang at
the Versiti BRI for instructing proximity ligation assay; and Y. Chen
and G. Stuttgen at MCW, C. Kastrup at Versiti BRI, M. Flick at
University of Carolina-Chapel Hill, and G. Reyes-Soffer at Columbia
University for helpful discussions. Human primary hepatocytes
g
were obtained from the Liver Tissue Cell Distribution System
(University of Minnesota and the University of Pittsburgh) and
through the Human Hepatocyte Isolation Distribution (University of
Pittsburgh), part of the CBRPC, Pittsburgh, Pennsylvania. Funding:
This study was supported by National Institutes of Health grant
R01HL163516 (Z.Zhe.), American Heart Association Career
y
Development Award 19CDA34660043 (Z.Zhe.), an American Heart
Association Collaborative Sciences Award (Z.Zhe.), an American
Society of Hematology Fellow Scholar Award (Z.Zhe.), an American
Society of Hematology Supplement Award (Z.Zhe.), a Medical
College of Wisconsin and Versiti Blood Research Institute Startup
Fund (Z.Zhe.), National Institutes of Health grant T32HL007343
(Z.Zhe.; H.N.G., primary investigator), a Medical College of
Wisconsin Cardiovascular Center Research Pilot Award (W.D.),
National Institutes of Health grant P01HL087123 (I.T.), National
Institutes of Health grant R01HL138907 (M.G.S.-T.), American
Heart Association grant 19TPA34890023 (M.G.S.-T.), National
Institutes of Health grant R35HL135833 (H.N.G.), National
Institutes of Health grant HHSN276201200017C (Liver Tissue Cell
y g
Distribution System, University of Minnesota, and the University
of Pittsburgh), and Pittsburgh Liver Research Center grant P30
DK120531 (Human Hepatocyte Isolation Distribution, University of
Pittsburgh). Author contributions: Conceptualization: Z.Zhe.,
I.T., and W.D. Methodology: Z.Zhe., I.T., W.D., H.Z., D.S., H.N.G.,
p
R.L.S., J.A.L., M.M.H., M.G.S.-T., A.D.S., J.Z., and S.V. Investigation:
W.D., Z.Zhe., H.Z., H.L., Z.Zha., M.C., M.R., G.K., and H.R.P. Funding
acquisition: Z.Zhe., I.T., and W.D. Project administration: H.L.
Supervision: Z.Zhe. and I.T. Writing – original draft: W.D. and
Z.Zhe. Writing – review & editing: H.L., M.C., R.L.S., D.S., I.T., and
,
Z.Zhe. Competing interests: Z.Zhe., W.D., and H.L. have filed a
provisional patent application on recombinant tPA fragments.
All other 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 © 2023 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.adh5207
Materials and Methods
Figs. S1 to S13
Tables S1 to S4
References (96–100)
MDAR Reproducibility Checklist
Submitted 9 March 2023; accepted 13 July 2023
10.1126/science.adh5207
12 of 12
RES EARCH

◥ though there is little observational evidence for

RESEARCH ARTICLE these (9). The formation process of a MAD has
not been observed in either type of BH (super-
BLACK HOLES massive or stellar-mass). Accretion flows onto
stellar-mass BHs can evolve much faster than
Observations of a black hole x-ray binary indicate those around supermassive BHs. For example,
outbursts in BHXRBs last for months to years,
formation of a magnetically arrested disk which potentially allows multiwavelength obser-
vations to probe the formation of a MAD in
Bei You1*, Xinwu Cao2*, Zhen Yan3*, Jean-Marie Hameury4, Bozena Czerny5, Yue Wu1,6,7, those systems.
Tianyu Xia1,8,9, Marek Sikora10, Shuang-Nan Zhang11,12 Pu Du11, Piotr T. Zycki10 Outbursts of BHXRBs have been explained
with a disk-instability model (10). In this model,
Accretion of material onto a black hole drags any magnetic fields present inwards, increasing their mass accumulates during quiescence in a cold
strength. Theory predicts that sufficiently strong magnetic fields can halt the accretion flow, producing and geometrically thin accretion disk (hereafter
a magnetically arrested disk (MAD). We analyzed archival multiwavelength observations of an outburst referred to as the thin disk). The increasing mass
from the black hole x-ray binary MAXI J1820+070 in 2018. The radio and optical fluxes were delayed triggers a thermal-viscous instability when, at
compared with the x-ray flux by about 8 and 17 days, respectively. We interpret this as evidence for some point in the thin disk, the temperature
the formation of a MAD. In this scenario, the magnetic field is amplified by an expanding corona, forming becomes high enough to ionize hydrogen. This
a MAD around the time of the radio peak. We propose that the optical delay is due to thermal viscous causes the system to enter an x-ray outburst as
instability in the outer disk. the thin disk empties material into the BH.

p
During an outburst, BHXRBs are observed

I
n a binary system consisting of a black hole
(BH) and a normal star, gas can be stripped
off the star and drawn by the strong grav-
itational field toward the BH (known as
the accretion flow will halt and form a mag-
netically arrested disk (MAD) (3, 4).
Sufficiently strong magnetic fields producing
a MAD have been observationally inferred in
in two very different spectral states—known as
the hard state and the soft state—in which the
x-ray emission is dominated by either hard
(>10 keV) or soft (<10 keV) photons. The timing
and radio-emission properties also differ be-
tween the two states (11, 12). The period between

g
an accretion flow), where it forms an accre- accretion systems with supermassive BHs (5–8).
tion disk. In the disk, matter moves inwards MADs are also predicted to form in BHXRBs, the two states is known as the state transition.
under the effect of viscosity. The system emits
large amounts of radiation in the x-ray band,
which allows it to be observed as a BH x-ray

y
binary (BHXRB). Relativistic jets can be launched
near the BH because of magnetic fields carried
by the accretion flow in the vicinity of the BH
(1, 2). Numerical simulations have shown
that any large-scale magnetic field present in
the accreting material can be dragged inward
through the accretion flow, which increases its
strength. If the radial magnetic force becomes
sufficiently high that it equals the gravita-
tional force of the BH, theory predicts that

y g
1
Department of Astronomy, School of Physics and

p
Technology, Wuhan University, Wuhan 430072, China.
2
Institute for Astronomy, School of Physics, Zhejiang
University, Hangzhou 310058, China. 3Shanghai Astronomical
Observatory, Chinese Academy of Sciences, Shanghai
200030, China. 4Observatoire Astronomique de Strasbourg,
Université de Strasbourg and Centre National de la ,
Recherche Scientifique, Unite Mixte de Recherche 7550,
67000 Strasbourg, France. 5Center for Theoretical Physics,
Polish Academy of Sciences, 02-668 Warsaw, Poland.
6
School of Astronomy and Space Science, Nanjing University,
Nanjing 210023, China. 7Key laboratory of Modern
Astronomy and Astrophysics, Nanjing University, Ministry of
Education, Nanjing 210023, China. 8Key laboratory for
Research in Galaxies and Cosmology, Department of
Astronomy, University of Science and Technology of China,
Hefei 230026, China. 9School of Astronomy and Space Fig. 1. Insight-HXMT light curves of MAXI J1820+070 during the 2018 outburst. Orange, black, and
Sciences, University of Science and Technology of China, blue data represent the light curves from HE (27 to 150 keV), ME (10 to 30 keV), and LE (1 to 10 keV)
Hefei 230026, China. 10Nicolaus Copernicus Astronomical
instruments, respectively. Shaded regions indicate the periods during which the source was in the
Center, Polish Academy of Sciences, 00-716 Warsaw, Poland.
11
Key Laboratory for Particle Astrophysics, Institute of High rising hard state (white), the hard-to-soft transition state (light purple), the soft state (gray), and the
Energy Physics, Chinese Academy of Sciences, Beijing flare (orange), the latter of which includes the soft-to-hard transition state and the decaying hard
100049, China. 12University of Chinese Academy of Sciences, state. Error bars, which are too small to be visible for most data, indicate the 1s confidence interval.
Chinese Academy of Sciences, Beijing 100049, China.
*Corresponding author. Email: youbei@whu.edu.cn (B.Y.); [Modified from (19), which is distributed under a Creative Commons Attribution License 4.0 (CC BY)
xwcao@zju.edu.cn (X.C.); zyan@shao.ac.cn (Z.Y.) https://creativecommons.org/licenses/by-nc/4.0/]

You et al., Science 381, 961–964 (2023) 1 September 2023 1 of 4

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

In the hard state, hot gas near the BH forms a

corona—modeled as an advection-dominated
accretion flow (ADAF)—in which a large frac-
tion of the viscously released gravitational en-
ergy is advected along the flow and enters the
BH event horizon (13). Hard states are observed
during the initial rise and subsequent decay of
an outburst, which we refer to as the rising and
decaying hard states. The soft-to-hard state tran-
sition and the decaying hard state are accom-
panied by a hard x-ray flare. These flares have A
been detected in optical, infrared, and radio
bands, which is thought to relate with the
formation of a jet (14, 15). However, the phys-
ical mechanisms and processes are not fully
understood. Hard x-rays during the flare are
usually interpreted as inverse Compton scat-
tering in the jet and/or in the ADAF (11).

Outburst of MAXI J1820+070 in 2018

p
MAXI J1820+070 (also known as ASASSN-18ey)
is an x-ray binary that consists of a low-mass star B
orbiting a stellar-mass BH (with BH mass
MBH ≃ 8:5M⊙) with a period of 16.5 hours (16).
It is located at a distance of 2.96 ± 0.33 kpc
(17). The system was identified during an x-ray

g
outburst that occurred on 11 March 2018 (18);
the outburst was observed by several x-ray
telescopes (19–22). We analyzed archival ob-
servations of MAXI J1820+070 by Insight
Hard X-ray Modulation Telescope [Insight-

y
HXMT (23)]. The Insight-HXMT data cover
a broad energy band of 1 to 250 keV and
were observed between 14 March 2018 and
21 October 2018 [corresponding to modified
C
Julian dates (MJDs) 58191 to 58412] (19, 20).
The outburst of MAXI J1820+070 also ap-
peared as an optically bright transient (24),
which was monitored by the American Asso-
ciation of Variable Star Observers (AAVSO),
who produced a light curve (brightness as a Fig. 2. Multiwavelength light curves of the flare. A distance of d = 2.96 kpc (17) was adopted in

function of time) that spanned from 12 March
2018 to 21 December 2018 (MJD 58188 to 59184)
(25). Radio emission from MAXI J1820+070
was detected by multiple telescopes, which
y g
estimating the observed luminosity. (A) X-ray luminosity from the thermal emission of the thin disk (blue),
the Compton-scattered emission from the ADAF (orange), and the total (black), integrated over energies
0.1 to 100 keV as determined by our spectral fitting of the Insight-HXMT data (28). (B) The radio emission
at 15.5 GHz measured by AMI-LA (26). (C) Extinction-corrected V-band luminosities from AAVSO (gray

p
indicated the presence of a jet. We used the points) and an exponential model (cyan line) fitted to the data between MJD 58380 to 58390. Orange points
radio flux by Arcminute Microkelvin Imager represent the data after subtraction of the exponential model. Error bars, which are too small to be visible
Large Array (AMI-LA) from (26). for most data, are 1s confidence interval.

These datasets provide high-cadence monitor-
ing of the 2018 outburst of MAXI J1820+070
over a broad range of wavelengths. The prox-
imity of the source and its low extinction with
color excess [E(B–V)] of 0.20 to 0.26 (27) make
it sufficiently bright to investigate the accre-
tion process of the magnetized flow near a BH.
Figure 1 shows the Insight-HXMT light curves
from the low-energy (LE, 1 to 10 keV), medium-
energy (ME, 10 to 30 keV), and high-energy
(HE, 27 to 150 keV) instruments between MJD
58191 and 58412. We focused our analysis on the
radio, optical, and x-ray observations of MAXI
J1820+070 during the hard x-ray flare, including
the soft-to-hard state transition and decaying
hard state (MJD 58380–58412).
Lags between flares at different
wavelengths
Figure 2A shows the x-ray luminosity of MAXI
J1820+070 during the flare, which we deter-
mined from the Insight-HXMT data (28). We
fitted spectral models to these data to separate
the luminosities of the accretion disk and
Compton-scattered emission components, which
are also shown. A radio flare at 15.5 GHz from
AMI-LA is presented in Fig. 2B. The optical
(V-band) data from AAVSO are shown in Fig.
2C, which also exhibits a flare superimposed
on the exponential decay of the underlying
outburst.
,
We performed an interpolated cross-correlation
function (ICCF) analysis (28) to determine the
time lag between the radio and x-ray light curves
and found a radio lag of 8:06þ0:880:56 days (fig. S1,
uncertainties are 1s). ICCF analysis comparing
the radio and the V band shows an optical lag of
8:61þ1:06
0:83 days with respect to the radio, which
indicates negligible contribution from the jet
to the V-band flux (supplementary materials).
ICCF analysis comparing the x-ray and the V
band shows an optical lag of 17:61þ0:50
0:44 days with
respect to the x-ray (fig. S1). This lag is far longer
than the time taken for light to travel across the
binary system, which is several seconds for an

You et al., Science 381, 961–964 (2023) 1 September 2023 2 of 4

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

mechanisms compete, so a peak in the hard

x-ray emission occurred when the latter
mechanism becomes dominant around t1 =
MJD 58389.
Previous studies have suggested that a weak
and coherent external magnetic field can be
dragged inwards and amplified by an ADAF be-
cause of its high radial velocity (31, 32). The
external magnetic field is expected to be more
strongly amplified by larger ADAFs (33). We
therefore calculated the advection and diffu-
sion of magnetic fields in the ADAF of MAXI
J1820+070 (28). We found that the magnetic
field B at RISCO was 2 × 106 G at t0 ≃ MJD 58381
(the onset of the flare), 4.5 × 107 G at t1 ≃ MJD
58389 (the peak of the hard x-ray emission),
reached a maximum at 7 × 107 G around t2 ≃
MJD 58397, and then declined. The jet power
increased with the magnetic field strength near
the BH, so its radio emission also increased until

p
t2 = MJD 58397, which explains the observed
delay of ~8 days with respect to the x-ray emis-
sion. We estimate that the jet power at the
radio peak as 1037 erg s1 , which is consistent
with theoretical predictions given the magnetic
field of 7 × 107 G (28). This supports our in-

g
terpretation that the magnetic field is trans-
ported and amplified by the expanding ADAF.

Formation of a magnetically arrested disk

In our interpretation, the increasingly strong

y
magnetic field influences the subsequent accre-
tion flow because the magnetic pressure acts
against the gravity of the BH. Our calculations
of the radius-dependent magnetic fields within
Fig. 3. Schematic diagram of our proposed interpretation. Purple arrows indicate the magnetic field the ADAF (28) show that the BH gravity still
lines. The outer thin disk is truncated at a radius Rtr, within which an ADAF forms. (A) The configuration dominated over magnetic pressure at the time
at the onset of the flare. We assume that the thin disk extends to ISCO and the ADAF expands with increasing of the hard x-ray peak (t1 = MJD 58389). After
Rtr. (B) Configuration during the peak hard x-ray emission, when the outer thin disk brings both mass this point, the magnetic field inside the expand-
and magnetic field into the inner ADAF. (C) Configuration at the time of peak radio emission, when the ing ADAF continued to be amplified, eventually
becoming dominant at the inner edge. This

y g
magnetic field has been amplified by the ADAF. A MAD forms within the magnetospheric radius Rm, where the
magnetic pressure equals the gravitational force (4). Movie S1 shows an animated version of this figure. caused a MAD to form around the time of the
radio peak (t2 = MJD 58397), when the mag-
netic pressure became equal to the gravitation-
al force (fig. S5) (28).

orbital period of 16.5 hours; thus, the lag cannot
be due to reprocessed x-ray illumination of the
outer disk or the companion star. We used these
cross-correlations to investigate the origins of
the emission at each wavelength (28).
Interpretation of the radio time lag
Hard x-ray emission during the flare could be
due to soft x-ray photons inverse Compton–
scattered by hot electrons (Comptonization)
within the jet or the ADAF. Given that the radio
emission originated from the jet (26), but the
radio flare lagged the x-ray by ~8 days, we ruled
out the possibility that x-rays were emitted by
the jet. We therefore suggest that the hard x-rays
were most likely emitted by the ADAF.
In the hard state of a BHXRB, a geometrically
thick ADAF is thought to be located close to the
BH, extending outward and connecting to a thin
disk with a truncated inner radius Rtr (equal to
the ADAF radial extent) (29, 30). We assumed
that Rtr = RISCO at t0 = MJD 58381 when the
soft state ended, where RISCO is the radius of
the BH’s innermost stable circular orbit (ISCO).
On the basis of the derived disk luminosity,
which we estimated from the spectral fitting
(28), we found that the truncation radius in-
creased to Rtr = 6.79RISCO when the hard x-ray
emission peaked at t1 = MJD 58389, then to
Rtr = 32.83RISCO by the time the radio emis-
sion peaked at t2 = MJD 58397 (28). Therefore,
Rtr increases with decreasing mass-accretion
rate. In this case, the fraction of the total grav-
itational power dissipated in the whole ADAF
increases with Rtr, whereas the power in the
very inner part of ADAF—where the hard x-ray
is mostly emitted—decreased with decreas-
ing accretion rate [eq. S5; (28)]. These two
p
We conclude that the increase of the trun-
cated radius during the flare in MAXI J1820
+070 led to the observed 8-day radio lag relative
,
to the hard x-ray emission. A MAD was then
gradually established in the ADAF as it con-
tinued to extend radially. Figure 3 and movie S1
illustrate our proposed scenario for the physical
processes associated with the x-ray and radio
flare and the formation of the MAD.
Interpretation of the optical time lag
We considered whether the long delay time of
the optical flare can be produced by a viscously
heated thin disk. In this scenario, the x-ray
flare heats up the outer thin disk, reviving the
thermal-viscous instability in the thin disk, which
produces a delayed optical emission. We used a
disk instability code (34) to determine the time-
dependent evolution of the thin disk (Fig. 4) (28).

You et al., Science 381, 961–964 (2023) 1 September 2023 3 of 4

RES EARCH | R E S E A R C H A R T I C L E
Fig. 4. Observed V-band flare compared to our simulations. The observed V-band luminosity (gray data
points) is the same as in Fig. 2C. The curves represent the predictions from our DIM simulation, including
one with a disk wind (red line) and one without a disk wind (dashed black line).
We found that a small optical flare is produced,
starting around MJD 58390 but fading away
after a few days, which is much shorter than the
observed lag of ~17 days. Therefore, the standard
disk instability model (DIM) cannot explain this
optical lag.
A thin disk is expected to launch disk winds,
which have previously been studied by using
observations and simulations (35, 36). Previous
work has found that a disk wind was present
throughout the outburst of MAXI J1820+070
(37). We therefore postulate that the disk wind
was launched during the flare. Disk winds
affect the profile of the outburst light curves of
BHXRBs (38, 39) because the disk wind re-
moves angular momentum from the disk, which
is equivalent to increasing the viscosity. As a
result, a brighter and more delayed optical peak
is produced compared with the case of no disk
wind. Using a simple parametrization of the
disk wind, we performed additional simulations
of the DIM (28). We found that the resulting
V-band flare is consistent with the observed
light curve, lagging the hard x-rays by more
than 15 days (Fig. 4).
RE FE RENCES AND N OT ES
1. R. D. Blandford, R. L. Znajek, Mon. Not. R. Astron. Soc. 179,
433–456 (1977).
2. R. D. Blandford, D. G. Payne, Mon. Not. R. Astron. Soc. 199,
883–903 (1982).
3. I. V. Igumenshchev, R. Narayan, M. A. Abramowicz, Astrophys.
J. 592, 1042–1059 (2003).
4. R. Narayan, I. V. Igumenshchev, M. A. Abramowicz, Publ. Astron.
Soc. Jpn. 55, L69–L72 (2003).
5. M. Zamaninasab, E. Clausen-Brown, T. Savolainen,
A. Tchekhovskoy, Nature 510, 126–128 (2014).
You et al., Science 381, 961–964 (2023) 1 September 2023
6. Event Horizon Telescope Collaboration, Astrophys. J. Lett. 910,
L13 (2021).
7. Event Horizon Telescope Collaboration, Astrophys. J. Lett. 930,
L16 (2022).
8. F. Yuan, H. Wang, H. Yang, Astrophys. J. 924, 124 (2022).
9. S.-S. Weng et al., Astrophys. J. Lett. 915, L15
(2021).
10. J.-P. Lasota, Nature 45, 449 (2001).
11. R. A. Remillard, J. E. McClintock, Annu. Rev. Astron. Astrophys.
44, 49–92 (2006).
12. R. Fender, E. Gallo, Space Sci. Rev. 183, 323–337
(2014).
13. R. Narayan, I. Yi, Astrophys. J. 428, L13
(1994).
14. D. M. Russell, D. Maitra, R. J. H. Dunn, S. Markoff, Mon. Not. R.
Astron. Soc. 405, 1759 (2010).
15. S. Corbel et al., Mon. Not. R. Astron. Soc. 428, 2500–2515
(2013).
16. M. A. P. Torres et al., Astrophys. J. Lett. 893, L37
(2020).
17. P. Atri et al., Mon. Not. R. Astron. Soc. 493, L81–L86
(2020).
18. T. Kawamuro et al., The Astronomer’s Telegram 2018, 11399
(2018).
19. B. You et al., Nat. Commun. 12, 1025 (2021).
20. X. Ma et al., Nat. Astron. 5, 94–102 (2021).
21. D. J. K. Buisson et al., Mon. Not. R. Astron. Soc. 490,
1350–1362 (2019).
22. E. Kara et al., Nature 565, 198–201 (2019).
23. S.-N. Zhang et al., Sci. China Phys. Mech. Astron. 63, 249502
(2020).
24. C. Littlefield, The Astronomer’s Telegram 2018, 11421
(2018).
25. J. Patterson et al., The Astronomer’s Telegram 2018, 11756
(2018).
26. J. S. Bright et al., Nat. Astron. 4, 697–703
(2020).
27. J. Mikołajewska, A. A. Zdziarski, J. Ziółkowski,
M. A. P. Torres, J. Casares, Astrophys. J. 930, 9
(2022).
28. Materials and methods are available as supplementary
materials.
29. F. Yuan, R. Narayan, Annu. Rev. Astron. Astrophys. 52,
529–588 (2014).
30. B. F. Liu, E. Qiao, iScience 25, 103544 (2021).
31. R. Narayan, A. Sądowski, R. F. Penna, A. K. Kulkarni, Mon. Not.
R. Astron. Soc. 426, 3241 (2012).
32. M. C. Begelman, N. Scepi, J. Dexter, Mon. Not. R. Astron. Soc.
511, 2040–2051 (2022).
33. X. Cao, Astrophys. J. 737, 94 (2011).
34. J.-M. Hameury, K. Menou, G. Dubus, J.-P. Lasota, J.-M. Hure,
Mon. Not. R. Astron. Soc. 298, 1048–1060 (1998).
35. J. M. Miller et al., Nature 441, 953–955 (2006).
36. B. You et al., Astrophys. J. 821, 104 (2016).
37. J. Sánchez-Sierras, T. Muñoz-Darias, Astron. Astrophys. 640,
L3 (2020).
38. B. E. Tetarenko, J. P. Lasota, C. O. Heinke, G. Dubus,
G. R. Sivakoff, Nature 554, 69–72 (2018).
39. G. Dubus, C. Done, B. E. Tetarenko, J.-M. Hameury, Astron.
Astrophys. 632, A40 (2019).
40. B. You, X. Cao, Z. Yan, Bei-You-BH/MAD: Formation of a
magnetically arrested disk around a black hole, Zenodo
(2023); https://doi.org/10.5281/zenodo.8080948.
AC KNOWLED GME NTS
We thank W.-F. Yu, J.-F. Wu, F.-G. Xie, W.-M. Gu, and M.-Y. Sun for
useful discussions. B.Y. thanks Y.-K. Zhang (Jinan University) for
help in producing movie S1. We acknowledge with thanks the
variable star observations from the AAVSO International Database
contributed by observers worldwide, which we used in this
research. Funding: B.Y. is supported by the National Program on
p
Key Research and Development Project 2021YFA0718500, the
National Science Foundation of China (NSFC) grants 12273026 and
U1931203, the Natural Science Foundation of Hubei Province of
China 2022CFB167, the Fundamental Research Funds for the
Central Universities 2042022rc0002, and the Xiaomi Foundation/
Xiaomi Young Talents Program. X.C. is supported by the NSFC
grants 11833007, 12073023, 12233007, and 12147103; the China
Manned Space Project science research grant CMS-CSST- 2021-
g
A06; and the fundamental research fund for Chinese central
universities. Z.Y. is supported by NSFC grants U1838203 and
U1938114, the Youth Innovation Promotion Association of CAS ID
2020265, and the funds for key programs of Shanghai
astronomical observatory. S.-N.Z. is supported by the National
Program on Key Research and Development Project grant
y
2016YFA0400802 and by the International Partnership Program of
Chinese Academy of Sciences (grant 113111KYSB20190020). P.D.
is supported by NSFC grants 12022301 and 11991051. Author
contributions: B.Y. initiated the project, performed the data
analysis and model interpretation, and led the writing of the
manuscript. X.C. led the interpretation of the MAD and the writing
of the related text. Z.Y. contributed to the data analysis, model
discussion, and the writing of the text. J.-M.H. provided the DIM
code and contributed to the model discussion and to the writing of
the text. B.C. contributed to the discussion of the DIM. Y.W. and
T.X. contributed to the simulation of the DIM and the data analysis.
M.S., S.-N.Z., P.D., and P.Z. contributed to the model discussion
and the writing of the text. Competing interests: The authors
y g
declare no competing interests. Data and materials availability:
The Insight-HXMT data used in this work are available from http://
archive.hxmt.cn/ under Obs ID P0114661. The radio luminosities
were taken from previous work [Source Data file, (26)], then
corrected to our adopted distance (18). The optical data were
p
obtained from https://www.aavso.org/data-download; we used
object “MAXI J1820+070,” start date “03/01/2018,” and end date
“03/01/2019.” The reduced data (including LE, ME, and HE light
curves and x-ray, radio, and V-band luminosities), our scripts for
the ICCF analysis and figure plotting, and the numerical codes we
,
used to compute the magnetic field strength and DIM are all available
at https://github.com/Bei-You-BH/MAD and archived at Zenodo
(40). License information: Copyright © 2023 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.abo4504
Materials and Methods
Supplementary Text
Figs. S1 to S8
References (41–97)
MDAR Reproducibility Checklist
Movie S1
Submitted 5 February 2022; accepted 30 June 2023
10.1126/science.abo4504
4 of 4
RES EARCH

ORGANIC CHEMISTRY was to create an ML-guided tool that immedi-

ately provides predicted hits for a new pro-
A machine-learning tool to predict substrate-adaptive posed coupling (which could then be optimized
if necessary), offering more than empirical
conditions for Pd-catalyzed C–N couplings guides and avoiding an experimental cam-
paign from an empirical approach, thereby
N. Ian Rinehart1, Rakesh K. Saunthwal1, Joël Wellauer2, Andrew F. Zahrt1, Lukas Schlemper2, accelerating the routine application of B-H
Alexander S. Shved1, Raphael Bigler2*, Serena Fantasia2*, Scott E. Denmark1* couplings (Fig. 1). This goal is complementary
to optimization, and we foresee that the com-
Machine-learning methods have great potential to accelerate the identification of reaction conditions for bination of the two could potentially create an
chemical transformations. A tool that gives substrate-adaptive conditions for palladium (Pd)–catalyzed end-to-end artificial intelligence–driven proc-
carbon-nitrogen (C–N) couplings is presented. The design and construction of this tool required the ess like that shown in Fig. 1.
generation of an experimental dataset that explores a diverse network of reactant pairings across a set From an ML standpoint, there are crucial
of reaction conditions. A large scope of C–N couplings was actively learned by neural network models differences between an optimizer tool and the
by using a systematic process to design experiments. The models showed good performance in tool proposed in this work. A visual illustra-
experimental validation: Ten products were isolated in more than 85% yield from a range of couplings tion contrasting reaction optimization to a
with out-of-sample reactants designed to challenge the models. Importantly, the developed workflow tool based on substrate-adaptive models is
continually improves the prediction capability of the tool as the corpus of data grows. shown in Fig. 2. A three-dimensional plot rep-
resents a hypothetical reaction space where

T
any specific combination of reactant(s) and

he strategic value of carbon-nitrogen cou-
plings makes them important transforma-
tions in many domains of the chemical
enterprise. In particular, Buchwald-Hartwig
(B-H) couplings (1, 2) are among the most
important C–N bond-forming reactions and
pyrimidines, thiophenes, oxazoles, thiazoles,
and pyrazoles) on the basis of work from just
two references (9, 10) and otherwise focuses
on recommending ligands for specific nucleo-
phile types. For example, 2-amino oxazoles are
a challenging class of nucleophile that re-
p
condition(s) produce an unknown yield, and
the goal of using ML is to use relatively few
measured yields to predict the rest. After se-
lecting a specific coupling from all those pos-
sible [a slice of a hypothetical reaction space
along the reactant dimension(s); top of Fig. 2],

have revolutionized the practice of modern
synthetic organic chemistry (3). In this process,
palladium complexes catalyze the cross-coupling
of (hetero)aryl electrophiles with various nitro-
gen nucleophiles. Experimentalists routinely
quired a specific publication from Buchwald’s
group (11). More recently, high-throughput ex-
perimentation has been used to evaluate a
range of five-membered heteroaryl bromides
in B-H couplings, and that work highlights the
g
an optimizer directs the selection of experi-
ments within that slice of reaction space to
increase yield. Sophisticated optimizers are
still being developed to make iterative rounds
of experimentation as efficient as possible (14).

identify substrate-specific conditions for new
B-H couplings. The extensive scope of electro-
philes and nucleophiles competent in this trans-
formation required the development of many
catalysts and conditions to enable successful
couplings of diverse reaction partners (3, 4).
Selection of the appropriate palladium ligand
is particularly important because B-H couplings
are exceptionally sensitive to changes in ligand
structure (5).
difficulty of couplings between five-membered
heteroaryl bromides and aliphatic heterocy-
cles (12). Even with these published reports, the
chemical literature does not come close to de-
scribing the enormous scope of possible B-H
reactant pairings; thus, when a new (hetero)aryl
halide is used, experimentalists must rely on
intuition.
The use of B-H couplings often creates a bot-
tleneck in routine synthesis campaigns in both
y
An optimizer could, in principle, be used on
multiple reactants at once, but the reactants
must have related reactivity trends (i.e., the
slices in reaction space are close enough that
similar conditions have similar reactivity). Be-
cause B-H couplings are so sensitive to reactant
structure (vide supra), a new slice—even close in
the reactant dimension(s)—frequently requires
a fresh start. A complementary approach to
the optimizer, shown on the right in Fig. 2,

Empirical guides with selected examples from
the literature and heuristics on the basis of re-
ported couplings are available to help experi-
mentalists select appropriate ligands and
y g
academia and industry. An experimentalist
begins with a specific chemistry problem: cou-
pling of a new pair of reactants (Fig. 1). They
then identify a subset of conditions on the basis

p
conditions for a given coupling (“prior expe- of prior knowledge, amalgamating recommen-
rience” in Fig. 1) (6–8). Because these recom- dations from the B-H user guides and B-H cheat
mendations are derived from literature data, sheets described above (when recommendations

they are limited to previous experience (i.e.,
retrospective). Specifically, Wuitschik’s guide
(7) recommends the same conditions for all
heteroaryl bromides, though more granular,
heteroarene-specific conditions are often re-
quired. Buchwald’s original user guide rec-
ommends conditions for a limited range of
heteroaryl bromides (pyridines, pyrazines,
1
Roger Adams Laboratory, Department of Chemistry,
University of Illinois at Urbana-Champaign, Urbana, IL 61801,
USA. 2Pharmaceutical Division, Synthetic Molecules
Technical Development, Process Chemistry and Catalysis,
F. Hoffmann–La Roche, Ltd., Basel, Switzerland.
*Corresponding author. Email: sdenmark@illinois.edu (S.E.D.);
raphael.bigler@roche.com (R.B.); serena_maria.fantasia@roche.com
(S.F.)
,
are available), as well as personal experience,
intuition, and specific literature precedent. Those
prior knowledge–based recommendations serve
as a starting point for an experimental cam-
paign to survey catalyst-solvent-base com-
binations for experimental hits. For many
applications, a broad range of compounds must
be synthesized in a timely manner, and hits
are an end point. In practice, an experimen-
talist will invest time and resources into
an optimization campaign to fine-tune those
conditions only if necessary. Of note, the Doyle
group recently published a machine learning
(ML)–based tool that accelerates the optimiza- Fig. 1. The goal of this work. Identifying conditions
tion of yield from a user-defined reaction space that furnish synthetically useful yields for new
(“reaction optimizer” in Fig. 1) (13). Our goal couplings without experimental campaigns.

Rinehart et al., Science 381, 965–972 (2023) 1 September 2023 1 of 8

RES EARCH | R E S E A R C H A R T I C L E
Fig. 2. Defining substrate-adaptive models and contrasting them with ML-assisted optimization
models.
addresses this limitation because it involves train-
ing models on the entire reaction space before-
hand. Thus, when a new specific coupling is
selected, those models can immediately lever-
age prior learning to predict the yield patterns
of that new reaction without additional ex-
periments. This approach produces predicted
hits, which may still need validation, but cir-
cumvents an experimental campaign. That ap-
proach requires an appropriate experimental
dataset that provides sufficient prior experi-
ence to the substrate-adaptive models. Before
Rinehart et al., Science 381, 965–972 (2023) 1 September 2023
embarking on this enterprise, it was prudent
to evaluate state-of-the-art applications of ML
to predict the outcome of new B-H coupling
reactions and related work.
In 2018, Doyle and co-workers trained ML
models to predict the yields of B-H couplings
between one nucleophile and several similar
electrophiles (15). By the introduction of is-
oxazole additives (catalyst poisons) (16), the
authors simulated the process of gathering
data for many different reactant pairs while
avoiding the analytical challenge of having
many more distinct products. Models effec-
tively predicted the extent of catalyst poisoning
by the additives and were tested on out-of-
sample additives. Importantly, their models
were not applied to predicting couplings with
new nucleophiles or electrophiles and thus
do not address the problem outlined in Fig. 1.
Recent follow-up work using that same data-
set is subject to the same inherent limitations
and does not address the problem outlined
above (17, 18).
A year later, Li and Eastgate combined pro-
prietary and literature B-H data to create clas-
sification models to predict which ligand to
use for a specific B-H coupling for the desired
synthetic route (19). Although the models suc-
cessfully leveraged prior experiments to make
predictions, experimental validation is lim-
ited to reactants that furnish one specific di-
arylamine product. Wuitschik and co-workers
p
very recently reported their efforts using a
similar approach to that of Li and Eastgate
but with a robust experimental validation that
demonstrated poor model performance (20).
The authors attribute that to the limitations of
the dataset and suggest that new approaches
g
such as the workflow presented in this work
are needed for B-H couplings, citing the asso-
ciated preprint of this work (21).
In 2022, Zimmerman and co-workers pub-
lished an interesting approach to transfer
y
learning on B-H couplings, wherein models
were trained on one type of reactant (i.e., benz-
amide) and used to predict effective condi-
tions for a new substrate (i.e., pyrazole or an
aniline) (22). Their goal was to maximize the
value of limited data from one reaction to as-
sist in the development of a new reaction, ef-
fectively learning one slice of a reaction space
from Fig. 2 and applying that knowledge to
another. When two slices show positive trans-
y g
fer, this approach avoids training models on
the new reaction. In Zimmerman et al.’s work,
positive transfer was observed between ani-
line and benzamide, whereas pyrazole and
p
benzamide couplings showed poor transfer.
That observation supports the premise that
general conditions are unlikely to exist for B-H
,
couplings.
In a very recent disclosure, Burke, Grzybowski,
and co-workers reported general conditions
for Suzuki-Miyaura (S-M) cross-couplings (23).
Their approach bypasses using predictive
models that respond to substrate structure, as
described in Fig. 2. Instead, the goal of general
conditions is to work across the many slices of
the reaction space of S-M couplings, and their
work relies on the assumption that those exist.
Unfortunately, general conditions are unlike-
ly to exist for B-H couplings (vide supra). The
structural variation of diverse nitrogen nu-
cleophiles and resulting changes in reactivity
necessitate domain-specific conditions, and
decades of research have not solved all those
2 of 8
RES EARCH | R E S E A R C H A R T I C L E
challenges. Zimmerman et al.’s work demon-
strating poor transfer within B-H couplings
illustrates the distinct reactivity domains of
B-H coupling reaction space. As a case in point,
Burke, Grzybowski, and co-workers used their
workflow on a dataset of B-H couplings (24);
however, no model predictions or recommended
general conditions are presented.
Although also unrelated to B-H coupling re-
actions, Doyle and co-workers reported suc-
cessfully modeling of a noncatalytic functional
group interconversion of carbinols to alkyl
fluorides (25). The authors collected a dataset
capable of training ML models to learn the in-
herent reactivity patterns of many substrates
for that reaction. Those models then success-
fully predict substrate-specific conditions for
new reactants, as illustrated in Fig. 2. Doyle et al.’s
approach is proof of concept for designing a
prediction tool for substrate-adaptive con-
ditions. However, by virtue of having a single
reactant and no catalyst, that transformation
has a substantially lower complexity.
Research strategy
The reactant dimension for B-H coupling reac-
tion space like that in Fig. 2 actually comprises
multiple subdimensions, and the conditions
dimension also captures solvents, bases, and
catalysts. All those dimensions are indepen-
dent (nucleophile, electrophile, catalyst, solvent,
and base), and all affect yield. Importantly,
each reactant can show different preferences
with regard to catalysts, solvents, and bases
(6, 8). As a result, models must learn the pref-
erences of each reactant and the interaction
terms between various combinations of them
and then correctly weigh those to be useful.
The data used to train such models must ex-
plore these complex relationships, and no such
dataset of appropriate complexity exists. As a
case in point, Schwaller et al. show that mod-
eling on the US Patent and Trademark Office
(USPTO) dataset of B-H couplings, which ex-
plores a broad range of reactants in a noncom-
Fig. 3. Representative scope of
nitrogen nucleophiles for the
B-H coupling reaction used
in this work and comparison to
other validated ML studies on
B-H couplings. The process
for selection is depicted: (a) a
representative scope was curated,
and then (b) an algorithmic method
clustered them using the new
chemical descriptors developed in
this work. The structures shown are
some exemplars from the eight
clusters that were identified. Me,
methyl; t-Bu, tert-butyl.
Rinehart et al., Science 381, 965–972 (2023) 1 September 2023
binatorial fashion, did not produce predictive
models (17). A dataset with a similar reactant
diversity to that of the USPTO dataset is ideal,
but that exact dataset cannot support models
complex enough to address the problem out-
lined in Fig. 1.
To build such a dataset, the reactant dimen-
sions must be unbounded so that it is possible
to continue expanding to new reactant do-
mains without starting over. Zimmerman et al.’s
observation of poor transfer and good transfer
between different reactant domains of B-H
reaction space indicates that there may be types
of couplings that can be grouped and learned
together and others that must be separately
addressed. A new strategy for dataset design
that is founded on separating reactant domains
is proposed. By combining expert knowledge,
new chemical descriptors, and well-established
clustering techniques, representative neighbor-
hoods (subspaces) of the multidimensional B-H
coupling reaction space could be identified.
Then, in a subsequent experimental campaign,
new data in new subspaces could be iterative-
ly generated, and the applicability domain of
models expanded when models are updated
with that new data (vide infra).
Before running experiments, we had to de-
fine a starting point for exploring reactant
dimensions in B-H couplings. The reactivity
patterns and pitfalls of this important trans-
formation are immensely complex: Indoles
form constitutional isomers (8, 26), benzylic
and cyclic aliphatic amines are prone to an
unproductive electrophile reduction pathway
(8), and many heterocycles undergo unproduc-
tive, palladium-catalyzed C–H activation path-
ways, to name a few (27). The relative rates of
these unproductive pathways, catalyst deac-
tivation, and productive coupling all affect the
yield and vary from substrate to substrate. For
a model to predict useful conditions for new
B-H reactions, it must learn under which cir-
cumstances these side reactions appear as well
as the inherent structure-reactivity patterns
for each reactant, catalyst, and condition. Ac-
cordingly, to make this work as more than just
a proof of concept, problematic substrates like
those described above were included.
Figure 3 shows a representative list of 19 of
the 50 nitrogen nucleophiles used in this work.
In a similar manner, 50 (hetero)aryl bromides
were selected to broadly represent many build-
ing blocks of interest to pharmaceutical de-
velopment as well as a range of electronic and
steric properties [see supplementary mate-
rials (SM) for the full list]. The scope of both
Zimmerman et al.’s and Doyle et al.’s work on
B-H couplings is subsumed within and greatly
expanded upon by the scope of this study. The
boxed structures represent the good and poor
transfer learning, represented by colored ar-
rows, that was demonstrated by Zimmerman
and co-workers (22).
To realize effective couplings across such a
p
broad range of reactants, many catalysts were
necessary. Buchwald- and Beller-type phos-
phine ligands were identified because (i) they
share a similar backbone structure, and (ii)
more than 50 such ligands are commercially
available and were developed to address the
broad scope of C–N couplings (3, 28). To cap-
g
ture that range, 20 ligands were selected to
represent the crucial ligand dimension of the
reaction space with a combination of algorith-
mic selection and expert knowledge (see SM for
y
details). In the interest of practical applicabil-
ity, models were trained to make predictions
for solvents and bases used at the bench. Thus,
two inorganic bases—potassium carbonate and
sodium tert-butoxide—and one organic base—
1,8-diazabicyclo[5.4.0]undec-7-ene (DBU)—were
selected. Finally, 1,4-dioxane, toluene, and tert-
amyl alcohol were selected as representative
solvents because they broadly represent cyclic
ethers, arenes, and alcoholic solvents regularly
y g
used in B-H couplings (see SM for selection of
catalyst and conditions). Thus, the reaction
space for this study—with nucleophile, electro-
phile, ligand, solvent, and base dimensions—
p
,
3 of 8
RES EARCH | R E S E A R C H A R T I C L E
contains 450,000 possible reactions from 180
conditions (3 bases times 3 solvents times 20
ligands) and 2500 reactant pairs (50 amines
times 50 bromides). As mentioned, this space
already represents an intractable number of
experiments for combinatorial experimentation.
The first experiments were designed to rep-
resent the reaction space broadly with 23 dif-
ferent algorithmically selected reactant pairs
and a systematically varied set of conditions
for each (see SM). Extensive experimental de-
velopment identified reproducible conditions
on the 0.5-mmol scale in a 24-tube parallel re-
actor. Twenty-four conditions were evaluated
for each reactant pair out of the 180 possible
(20 catalysts times 3 solvents times 3 bases).
This approach balanced the need to rapidly
explore new reactant pairings (i.e., slices from
Fig. 2) with the need to explore enough condi-
tions to learn that slice of the reaction space
(i.e., points on each slice from Fig. 2). The data
showed 63% of experiments with 0% yield,
and 82% with less than 20% yield. The paucity
of hits from which models needed to learn was
a problem. To generate a higher fraction of
hits in the data, a new strategy was needed for
evaluating the condition component of the
B-H reaction space.
To increase the number of positive hits, ML
models were trained to recognize zero and
nonzero yield patterns. Deep feed-forward neu-
ral networks were trained to classify reactions
as zero- or nonzero-yielding using that first
dataset and showed an average accuracy of 87%.
Although unable to distinguish between a 1%-
yielding and a 99%-yielding reaction, such a
classifier could still increase the number of
nonzero-yielding reactions that are run. Eigh-
teen reactant pairs from the first 23 gave sig-
nificantly more nonzero data (the rest were
hypothesized to be chemically challenging using
expert knowledge). The initial results from those
Fig. 4. New, experimentalist-driven, active-learning workflow for exploration of reaction space.
Rinehart et al., Science 381, 965–972 (2023) 1 September 2023
18 reactant pairs showed 56% zero-yielding
reactions, and only 12% with a >80% yield.
Twenty-four new conditions were selected from
the 156 remaining conditions (180 total, with
24 already evaluated) using the classifier to
predict nonzero-yielding conditions. The new
results from those 24 classifier-selected con-
ditions contained just 22% zero-yielding reac-
tions, and 29% of the data showed >80% yield
across all 18 reactant pairs (see SM for details).
Thus, even a limited binary classifier could
be used to improve the yield distribution of
new data.
Those first models allowed us to connect
steps 1 to 4 in a new workflow depicted in Fig. 4.
This workflow begins and ends with the ex-
perimentalist. Thus, (i) a new reactant pair is
selected by the experimentalist, (ii) the tool
calculates the corresponding chemical descrip-
tors, (iii) the tool then uses models to predict
the yield of all 180 conditions, and (iv) the
experimentalist can decide which conditions
to evaluate on the basis of both the predictions
and their expert knowledge. The second half
of the workflow, steps 5 to 8, describes the
process of domain expansion by (v) including
new data, (vi) retraining models with that new
data, (vii) testing those models in control ex-
periments, and (viii) having the experimental-
ist evaluate model performance. At this point
in the cycle, the experimentalist again inter-
venes with expert knowledge and their evalu-
ation of model performance to select the next
reactant pair to target.
To build the desired dataset, the experimen-
talist directs what the model learns by their
selection of the next reactant pairs. This pro-
cess relies on defined reaction subspaces to tar-
get for domain expansion (vide supra). Figure
5A depicts the map of reaction subspaces used
to select the next reactant pairs, which was
constructed from the output of the process in
Fig. 3. The B-H reactant space map is an array
with amines organized by cluster on the ver-
tical axis and bromides organized in the same
manner on the horizontal axis. Each small
square represents a coupling between a spe-
cific bromide and a specific amine; each rec-
tangular subdomain therefore represents a
set of couplings between a cluster of amines
and a cluster of bromides. The varying size of
the clusters is a consequence of the reactants
selected in Fig. 3A. The entire array is orga-
nized through the lens of new chemical de-
scriptors. The new descriptors combine a radial
distribution of atoms within each reactant
around the reaction center with various atomic
properties relevant to reactivity, producing
what we call a radial distribution function (see
SM). The initial algorithmically selected re-
actant pairings that were used to build the
first classifier model are shown in light green.
p
The workflow in Fig. 4 was then used, and re-
actant selections were made in an approach
similar to the transfer learning conducted by
Zimmerman and co-workers but on a larger
scale; many such transfer steps were made
to expand the domain of models (by moving
g
to a new subspace). Then, for each new sub-
space, multiple reactant pairs were evaluated
to enable learning of substrate-level trends,
making models robust in that subspace. The
results of those many selections are shown
y
in dark green.
The dataset can be described as a network,
and the goal of this work is to explore enough
connections (reactant pairings) to make infer-
ences about the missing connections that are
possible. To visualize this, a structured chord
diagram is depicted in Fig. 5B, showing amine
and bromide nodes on either side with edges
connecting reactants that were coupled in the
dataset. Similar to the map in Fig. 5A, reac-
y g
tants are organized by cluster on either side,
and the edge bundling in the figure illustrates
that the data are distributed across exemplars
from each cluster. The first visualization em-
p
phasizes that 121 out of the possible 2500 com-
binations of substrates were evaluated in this
work. The sparsity of these data is a feature,
,
not a flaw; identifying 121 out of 2500 couplings
and 24 out of 180 conditions to evaluate re-
duced the experimental burden from 450,000
experiments (180 conditions and 2500 sub-
strate pairs) to about 3300 experiments. The
diversity of the dataset is best represented
by the reactants evaluated (see SM for the full
list). Using randomly partitioned data, models
achieved a mean absolute error (MAE) of 9%
in an external test set. However, the problem
outlined in Fig. 1 is best represented by a test
set of out-of-sample reactant pairs, which is a
more difficult test.
After evaluating multiple couplings in each
subspace, models predicted reactivity trends (but
not necessarily exact yields) for new reactants
4 of 8
RES EARCH | R E S E A R C H A R T I C L E

perimental limitation and continue domain

expansion before reaching yield accuracy in
every subspace. Although models sometimes
showed limited accuracy, they were still able
to address the key problem because they could
identify the yield trends for the various reac-
tion conditions out of the 180 available. Ac-
cepting this limitation, a pragmatic, ordinal
ranking of predictions was used. Predictions
were divided into quartiles (25% of the total
predicted range), and these quartiles were
labeled as “high,” “moderate,” “low,” and “poor.”
In practice, multiclass classification models (in-
stead of regression models with this ordinal
ranking of predictions) were slightly less per-
formant. To test this system as a tool, models
trained on the resulting dataset were experi-
mentally evaluated with new B-H couplings.

Experimental validation

p
To evaluate the performance of these models,
the tool was tested in a typical use case: New
couplings were selected in which one or both
reactants were not seen by the model (out-of-
sample) and were tested using an experimen-
tally tractable number of conditions (one to

g
five in most cases) with the highest predicted
yields. To ensure that this test was rigorous, the
number of catalysts evaluated for each cou-
pling was limited to 3 out of the 20 available.
This precaution minimized the possibility that

y
a selected catalyst would work well by chance
instead of models learning meaningful chem-
ical structure and reactivity relationships. When
the experiments were carried out, predictions
were ranked into quartiles, and, as they were
available, a few conditions in the top quartile(s)
were evaluated. A complementary set of con-
ditions predicted to give low yields were also
evaluated to test whether the models were
correctly learning reactivity trends as described

y g
above. Together, conditions with high and
low predicted yields bookend a representative
experimental yield range for each coupling (see
SM for details). Figure 6 presents the results of

p
experimental validation, including (i) a chem-
ical structure indicating out-of-sample reactants
in red, in-sample reactants in blue, and the

,
coupling locus highlighted by the bold bond;
(ii) the number of experiments evaluating dif-
ferent conditions; (iii) the subset of conditions
with a high predicted yield (hits) that were
tested, represented as circles; (iv) an indication
of success by the coloring of the circle (green,
Fig. 5. Visualizations of B-H reaction space reactant component. (A) Dataset distributed across reactant
experimental yield in the top quartile), failure
clusters in a two-dimensional grid. (B) A structured chord diagram showing the network connectivity of the
(red, experimental yield not in the top quar-
reactant pairs sampled in the dataset. One exemplar is shown from each cluster.
tile), or a special case (gray); (v) the highest
isolated yield obtained from those predicted
with the 180 conditions. For example, the slope in the range of 0.5 to 0.75. Once models hits; and (vi) a heatmap showing all 180 pre-
highest-yielding predictions may be about could predict reactivity trends, much more data dicted yields for use as a visual fingerprint of
50% and the lowest about 0%, whereas the was required to achieve accurate yields. Ulti- the response of the models to changing re-
experimental data might range from 85 to 0%. mately, the trade-off between accurate pre- actant structure.
However, the linear correlation of predicted dictions and solving the problem outlined in To interpret experimental validation re-
and observed yields would be high, with a Fig. 1 required that we accept this as an ex- sults, success and failure must be defined. The

Rinehart et al., Science 381, 965–972 (2023) 1 September 2023 5 of 8

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

Fig. 6. Experimental validation

of substrate-adaptive models
as condition recommenders.
Out-of-sample (red) and in-sample
(blue) reactant fragments are indi-
cated for all products. Circular
iconography indicates the number
of predicted hits that were tested;
green indicates a success, and
red indicates a failure. Gray color is
explained in the text. The highest
isolated yield is indicated below
the circular icons. The prediction
heatmaps share the legend
shown for compound m. The
scale indicates the highest yield
prediction for each coupling
(middle number) and color
scale. Cy, cyclohexyl; Ph, phenyl;
THP, tetrahydropyranyl.

p
g
y
y g
p
hypothesis behind the experimental validation observed yields, correlations between them, demonstrating that the models adapted to elec-
is that models can correctly recommend syn- and common error metrics are provided in trophile structure.
thetically useful hits for couplings with one the SM. Next, the coupling of azepane with a protected

or more new reactants by predicting reactivity
trends. A successful experiment was defined as
one in which the predicted hit furnished a
yield in the top 25% of those observed for that
coupling (a top quartile yield). For practical
reasons, not all 2340 experiments (13 reactant
pairs in Fig. 6 times 180 conditions) were eval-
uated; instead, 187 experiments (both hits and
zero-yield predictions) were performed across
all 13 reactant pairs. Note that although this
experimental design does not test whether all
high-yielding conditions were identified, in
10 out of 13 cases, the highest observed yields
were >85%, and all three of the remaining
cases were expected to furnish a lower yield
(vide infra). The full details of predicted and
Our claim that yield trends are sufficient to
gauge performance is illustrated by the first
five validation experiments in Fig. 6. The cou-
plings to form the products a to e showed very
good predicted-observed correlations [0.93,
0.89, 0.91, 0.77, and 0.75 coefficient of deter-
mination (R2); see SM]. Poor residual errors
[>25% MAE, >34% root-mean-square error
(RMSE)] were observed for a, c, and e be-
cause the highest yield predictions of only
33, 49, and 57% led to the highest observed
yields of 98 to 99%. Moderate-to-good residual
errors were observed for b (14% MAE, 18%
RMSE) and d (9% MAE, 18% RMSE). Couplings
a to c share the same nucleophile, and the best
conditions changed among those couplings,
,
2-bromobenzyl alcohol derivative (product f)
used two out-of-sample reactants. The results
showed very good accuracy, with 4% MAE and
7% RMSE and a good correlation of 0.80 R2
between predicted and observed yields. By
contrast, the coupling of piperidine with 3-
bromo-5-methylpyridine (product g) involves
two in-sample reactants (the only example here
of coupling two in-sample reactants) but the
combination was never tested in the dataset.
The results were satisfying, with an accuracy
of 12% MAE and 16% RMSE and a correla-
tion of 0.86 R2. Comparing prediction finger-
prints of f and g provides some insight into
the models’ ability to adapt predictions to
substrate structures: For the more sterically

Rinehart et al., Science 381, 965–972 (2023) 1 September 2023 6 of 8

RES EARCH | R E S E A R C H A R T I C L E
hindered bromide in coupling f, the less bulky
diarylphosphino-substituted ligand 4 (CPhos
hybrid) is predicted to be the best. For the un-
hindered bromide in coupling g, bulkier tri-
cyclohexylphosphine-substituted ligands 17,
20, and 21 (CPhos, RuPhos, and SPhos) are pre-
dicted to be superior.
In the coupling of indole and 2-bromothiazole
(product h), the bromide was new to the mod-
el. This simple coupling is reported using copper
catalysis only (29–32) and is a difficult cou-
pling for palladium catalysis. As a result, this
experiment tests the ability of the model to
identify couplings that will not work. Indeed,
the models predicted a maximum 23% yield,
with mostly near- and zero-yield predictions.
Experimental evaluation of the five-highest
predictions showed that one condition pro-
vided a 25% yield, and the rest furnished no
product. Within the confines of this experi-
mental design, it appears that the model suc-
cessfully identified a challenging coupling;
however, to fully substantiate that claim, all
180 conditions would need to be evaluated.
An experimentalist with those predictions
should pursue another synthetic method if it
is available.
By contrast, the coupling of tert-butyl pyro-
glutamate with 2-(3-bromophenyl)-1,3-dioxolane
(both of which are out-of-sample, product l)
represents an extrapolation because the pyro-
glutamate (i.e., contains an ester group) is
substantially different from anything in the
dataset. As was seen with product h, the four
top predictions ranging from 59 to 68% yield
were evaluated, yet the highest observed yield
was 17%, a clear failure on the part of the models
to recognize a challenging, extrapolative cou-
pling. We observed that the nucleophile was
converted to the carboxylate in situ by 1H quan-
titative nuclear magnetic resonance (qNMR)
and hypothesized that this by-product pre-
vented coupling by coordinating Pd. Those
complications are excellent examples of real
challenges in generalizing this reaction that
prevent the tool from accurately identifying a
low-yielding reaction because it had not learned
to predict such complicated behavior.
2,6-Dimethylaniline derivatives represent a
class of nucleophile that is new to the models
because all other primary (hetero)aromatic
amines included in the dataset have no or-
tho substitution. Twenty-four conditions were
evaluated for three separate pairings of 2,6-
dimethylanilines with one out-of-sample bro-
mide (4-phenoxybromobenzene, product i)
and two in-sample bromides (3-bromoquinoline
and 3-bromonitrobenzene, products j and k).
The fractions of top-quartile predictions that
furnished yields in the top quartile of those
observed were 7/7, 7/9, and 7/9, respectively.
That i to k represent extrapolations is more
evident in their range of poor to moderate R2
of 0.70, 0.05, and 0.40, respectively. The ob-
Rinehart et al., Science 381, 965–972 (2023) 1 September 2023
served yield ranges for the best predictions
evaluated were 76 to 99%, 83 to 99%, and 77
to 89% yields, respectively. These results sug-
gest that for a subclass of anilines not repres-
ented in the dataset, models were still able to
identify high-yielding conditions. Thus, a trac-
table number of experiments (7 to 9 out of the
24 evaluated) were predicted to furnish good
yields and seven did for each of the three
couplings.
Finally, 5-fluoroskatole was included as a
nucleophile for its similarity to indole but its
inability to undergo C(3) arylation (product
m). While acquiring the dataset, we regularly
observed the C(3) arylation products—often as
dominant species—effectively forcing models
to predict C–N couplings that are in competi-
tion with C(3) aryl coupling. As chemists with
an understanding of reaction mechanisms, it
is intuitive that a minor structural perturba-
tion such as adding a methyl group to the C(3)
position of indole will prevent competitive C(3)-
arylation. The models being evaluated do not
learn mechanisms or the implications of such
a minor structural change (and corresponding-
ly minor change in the descriptors) as adding
a methyl group to that position of an indole.
The experimental validation results show a poor
correlation between predicted and observed
yields and stochastic errors of predicting zero-
and nonzero-yielding conditions, suggesting a
different observed reactivity pattern than what
the models had learned. Despite this, out of
the four conditions with the highest predicted
yield (40 to 55%), two provided good yields of
84 and 85%. This example illustrates that the
tool can still be useful, even on a challenging
coupling that represents a substantial mecha-
nistic extrapolation.
Taken together, the results of the validation
experiments demonstrate that the performance
of the model exists on a gradient. For new re-
actants from a reaction subspace that is well
represented in the dataset, predictions are ro-
bust (i.e., a to c and g to f). For cases in which
reactants represent structural permutations
from those in the dataset (i.e., d and e), the
models correctly learned reactivity trends and
could predict hits. For new types of structures
that may have different reactivity patterns
than those in the dataset, performance ranged
from moderate (i to k) to poor (l and m).
However, even the lowest model performance
demonstrated here provided good yields (de-
pending on the chemical limitations of the
coupling in question, e.g., l). Most importantly,
poor model performance can be rescued by
domain expansion of the dataset (see fig. S22).
This approach to network exploration using
active learning has enabled us to create mod-
els that are useful over a broad applicability
domain by evaluating only 0.7% of the reaction
space. Although we will continue to explore this
space, more importantly, we have now created
a transportable blueprint for domain expan-
sion of a dataset.
Outlook
The dataset described herein comprises >120
reactant pairs that systematically explore a
microcosm of B-H coupling space. Models
trained on these data simultaneously learned
nonlinear reactivity trends for many different
classes of reactants. These models could pre-
dict the yield of reactions with a mean absolute
error of 9% using randomly partitioned data
and were performant at reactant generaliza-
tion, as demonstrated by out-of-sample sub-
strate validation.
Key to achieving this goal was an informatics-
guided strategy that reduced the experimental
impossibility of exploring a 450,000-member
reaction space to an experimentally tractable
problem of acquiring a dataset comprising only
p
3300 experiments. We present both this val-
idated tool for Pd-catalyzed C-N couplings as
well as an active-learning workflow, which, un-
like prior work, was used to build an expansive
dataset for the chemical community. The chem-
istry community can engage with this work on
g
four different levels. An experimentalist with
no interest in ML can use the snapshot of the
tool presented in this work without expertise
in ML or programming and expect performance
similar to experimental validation. We invite
y
any practitioner with an interest in ML to take
the tool and resume the workflow in Fig. 4,
honing the tool to new reactant domains of in-
terest or steadily improving prediction accuracy
on existing dataset domains. Furthermore, we
invite any practitioner with expertise in ML to
use the new active-learning framework on other
important reactions with expansive multire-
actant spaces. Finally, we offer this dataset for
focused development on modeling noncombi-
y g
natorial, diverse datasets, which are rare in the
chemistry domain.
p
REFERENCES AND NOTES
1. A. S. Guram, R. A. Rennels, S. L. Buchwald, Angew. Chem. Int. Ed.
34, 1348–1350 (1995).
2. J. Louie, J. F. Hartwig, Tetrahedron Lett. 36, 3609–3612
,
(1995).
3. P. Ruiz-Castillo, S. L. Buchwald, Chem. Rev. 116, 12564–12649
(2016).
4. J. F. Hartwig, K. H. Shaughnessy, S. Shekhar, R. A. Green,
in vol. 100 of Organic Reactions, S. E. Denmark, Ed. (Wiley,
2019), pp. 853–958.
5. R. J. Lundgren, M. Stradiotto, Chemistry 18, 9758–9769
(2012).
6. B. T. Ingoglia, C. C. Wagen, S. L. Buchwald, Tetrahedron 75,
4199–4211 (2019).
7. M. Fitzner et al., Chem. Sci. 11, 13085–13093
(2020).
8. D. S. Surry, S. L. Buchwald, Chem. Sci. 2, 27–50
(2011).
9. M. D. Charles, P. Schultz, S. L. Buchwald, Org. Lett. 7,
3965–3968 (2005).
10. D. Maiti, B. P. Fors, J. L. Henderson, Y. Nakamura, S. L. Buchwald,
Chem. Sci. 2, 57–68 (2011).
11. E. P. K. Olsen, P. L. Arrechea, S. L. Buchwald, Angew. Chem.
Int. Ed. 56, 10569–10572 (2017).
7 of 8
RES EARCH | R E S E A R C H A R T I C L E
12. A. C. Sather, T. A. Martinot, Org. Process Res. Dev. 23,
1725–1739 (2019).
13. B. J. Shields et al., Nature 590, 89–96 (2021).
14. F. Häse, L. M. Roch, C. Kreisbeck, A. Aspuru-Guzik, ACS Cent.
Sci. 4, 1134–1145 (2018).
15. D. T. Ahneman, J. G. Estrada, S. Lin, S. D. Dreher, A. G. Doyle,
Science 360, 186–190 (2018).
16. K. D. Collins, F. Glorius, Acc. Chem. Res. 48, 619–627 (2015).
17. P. Schwaller, A. C. Vaucher, T. Laino, J.-L. Reymond,
Mach. Learn. Sci. Technol. 2, 015016 (2021).
18. T. R. Gimadiev et al., Mol. Inform. 40, e2100119 (2021).
19. J. Li, M. D. Eastgate, React. Chem. Eng. 4, 1595–1607
(2019).
20. M. Fitzner, G. Wuitschik, R. Koller, J.-M. Adam, T. Schindler,
ACS Omega 8, 3017–3025 (2023).
21. N. I. Rinehart et al., chemRxiv chemrixiv-2022-hspvw-v2
[Preprint] (2023); https://doi.org/10.26434/chemrxiv-2022-
hspwv-v2.
22. E. Shim et al., Chem. Sci. 13, 6655–6668 (2022).
23. N. H. Angello et al., Science 378, 399–405 (2022).
24. A. Buitrago Santanilla et al., Science 347, 49–53 (2015).
25. M. K. Nielsen, D. T. Ahneman, O. Riera, A. G. Doyle, J. Am.
Chem. Soc. 140, 5004–5008 (2018).
26. M. Yamaguchi, K. Suzuki, Y. Sato, K. Manabe, Org. Lett. 19,
5388–5391 (2017).
27. S. Basak, S. Dutta, D. Maiti, Chemistry 27, 10533–10557
(2021).
28. D. S. Surry, S. L. Buchwald, Angew. Chem. Int. Ed. 47,
6338–6361 (2008).
29. H. Chen, M. Lei, L. Hu, Tetrahedron 70, 5626–5631
(2014).
30. S. A. Iqbal, K. Yuan, J. Cid, J. Pahl, M. J. Ingleson, Org. Biomol.
Chem. 19, 2949–2958 (2021).
31. J.-K. Kwon, J.-H. Lee, T. S. Kim, E. K. Yum, H. J. Park, Bull.
Korean Chem. Soc. 37, 1927–1933 (2016).
Rinehart et al., Science 381, 965–972 (2023) 1 September 2023
32. A. Shoberu, C.-K. Li, H.-F. Qian, J.-P. Zou, Org. Chem. Front. 8,
5821–5830 (2021).
33. N. I. Rinehart, Substrate-adaptive C–N coupling yield prediction.
Zenodo (2023); https://doi.org/10.5281/zenodo.8185014.
ACKN OWLED GMEN TS
We acknowledge the support services of the NMR, mass spectrometry,
and microanalytical laboratories of the University of Illinois at
Urbana-Champaign. Funding: We acknowledge generous financial
support from Hoffmann La-Roche. A.S.S. thanks the National
Science Foundation (NSF) (grant no. CHE 1900617) for financial
support. This work was also supported in part by the Molecule
Maker Lab Institute, an AI Research Institutes program supported
by the NSF under grant no. CHE 2019897. Any opinions, findings,
and conclusions or recommendations expressed in this material
are those of the author(s) and do not necessarily reflect those of
the NSF. Author contributions: N.I.R. contributed to the dataset,
wrote the main code package, designed new descriptors, tested
neural network architectures for modeling, conceptualized the
informatics-guided strategy to exploring the reaction space as a
network, designed and contributed to the validation experiments,
contributed to composing the manuscript, and wrote the
Computational and informatics methods, Experimental methods,
Workflow development, General references, Experimental
validation, and Master list of dataset products sections of the SM;
R.K.S. contributed experimentally to the dataset, contributed
to the validation experiments, made authentic products for
zero-yielding plates, and composed most of the Characterization
data, Characterization references, and NMR spectra sections of
the SM; J.W. contributed experimental results to the dataset and
prepared reference materials; A.F.Z. helped conceptualize the
scope of the project and design descriptors as well as initial
classification modeling efforts; L.S. prepared palladium complexes and
reference materials; A.S.S. shared code that facilitated the
calculation and testing of new descriptors; R.B. conceived of and
supervised the project, designed the experimental part of the
workflow, contributed experimental results to the dataset, and
contributed to composing the manuscript; and both S.F. and S.E.D.
conceptualized and supervised the project and contributed to
composing the manuscript. Competing interests: The authors
declare no competing interests. Data and materials availability:
Full experimental procedures, including validation runs,
characterization data, experimental apparatus, qHPLC analytical
methodology, and copies of 1H, 13C, 31P, and 19F spectra as well as
feature engineering, modeling details and model validation,
structures of each product made in the dataset, and predictions
for each condition with every reactant pair in the dataset can be
found in the SM. Code developed in this work is available at
Zenodo (33), and ongoing development through our laboratory’s
Github organization will be available at github.com/SEDenmarkLab/
Lucid_Somnambulist. License information: Copyright © 2023
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.adg2114
Materials and Methods
Supplementary Text
Figs. S1 to S22
p
Tables S1 to S4
Scheme S1
Charts S1 and S2
NMR Spectra
References (34–96)
Submitted 8 December 2022; resubmitted 26 April 2023
Accepted 1 August 2023
g
10.1126/science.adg2114
y
y g
p
,
8 of 8
RES EARCH

ARCTIC CLIMATE 2A), which generates a mean anticyclonic

(clockwise) circulation. This drives predominant
Fluctuating Atlantic inflows modulate features of sea-ice drift and upper-ocean cir-
culation known as the Beaufort Gyre in the
Arctic atlantification Amerasian Basin as well as the Transpolar
Drift flowing from the Siberian shelf toward the
Igor V. Polyakov1*, Randi B. Ingvaldsen2, Andrey V. Pnyushkov3, Uma S. Bhatt4, Jennifer A. Francis5, Canadian Archipelago and Fram Strait (Fig. 1G).
Markus Janout6, Ronald Kwok7, Øystein Skagseth2,8 The primary mode of variability of the pan-
Arctic sea level pressure is known as the Arctic
Enhanced warm, salty subarctic inflows drive high-latitude atlantification, which weakens oceanic Oscillation, and the related wind pattern ac-
stratification, amplifies heat fluxes, and reduces sea ice. In this work, we show that the atmospheric counts for the observed climatological features
Arctic Dipole (AD) associated with anticyclonic winds over North America and cyclonic winds over of the atmospheric circulation. Beginning in
Eurasia modulates inflows from the North Atlantic across the Nordic Seas. The alternating AD phases 2007, however, the secondary Arctic Dipole
create a “switchgear mechanism.” From 2007 to 2021, this switchgear mechanism weakened northward (AD) pattern, which features higher sea lev-
inflows and enhanced sea-ice export across Fram Strait and increased inflows throughout the Barents el pressure over the Beaufort Gyre and the
Sea. By favoring stronger Arctic Ocean circulation, transferring freshwater into the Amerasian Basin, Canadian Archipelago along with lower sea
boosting stratification, and lowering oceanic heat fluxes there after 2007, AD+ contributed to slowing level pressure over the Siberian Arctic, became
sea-ice loss. A transition to an AD− phase may accelerate the Arctic sea-ice decline, which would further dominant [Fig. 2C; see also (14–16)], whereas
change the Arctic climate system. the Arctic Oscillation remained close to neu-
tral (fig. S4). This shift is evident in the higher

T
he Arctic region is rightfully called a fron-
tier for global climate change. Linked to
atmospheric circulation, radiative forc-
ing, and a host of climate feedback mech-
anisms, Arctic surface air temperatures
inform a broad and comprehensive under-
standing of system function.
Sea-ice changes
Although the end-of-summer ice extent and
p
spatial correlation between the mean 2007
to 2021 sea level pressure and the AD (R = 0.59,
where R is the correlation coefficient), com-
pared to that with the Arctic Oscillation (R = 0.47).
The AD index, a characteristic of the wind
cyclonicity in the central and Siberian Arctic

are rising at least three times faster than global-
average air temperatures (1). The Arctic Ocean
is warming faster than the global ocean (2).
Sea-ice decline is a true indicator of climate
change, affecting all aspects of life in the north-
thickness are declining, our results indicate that
the rate of decline slowed after 2007 compared
with 1992 to 2006 (Fig. 1, A and B). Over the
satellite record, the trend in summer ice extent
during 2007 to 2021 (−0.07 ± 0.18 × 106 km2 per
g
(14), varies between positive (AD+) and nega-
tive (AD−) over ~15-year regimes (see wavelets
in figs. S2 and S3). During 1992 to 2006, both
the AD and Arctic Oscillation spring-summer
indices were slightly negative (fig. S4, D and E),

ern high-latitude regions (1). One of the rea-
sons for sea-ice loss is the warming Arctic
Ocean, which is caused in part by anomalous
inflows from the North Atlantic and North
Pacific (3–5). System-wide changes in Arctic
basins caused by anomalous inflows from
the Nordic Seas are referred to as atlantifi-
cation (5–7). One of the many manifestations
of atlantification in the Eurasian Basin of the
Arctic Ocean is decreased upper-ocean strat-
decade) is weak and not statistically significant
in contrast to the much larger negative trends
in 1979 to 2021 (−0.79 ± 0.13 × 106 km2 per
decade) and 1992 to 2006 (−0.99 ± 0.51 × 106 km2
per decade). Thus, a more stable regime of Arctic
sea ice appears to have begun in 2007. This
transition was abrupt, with 2007 setting a record
for a single-year sea-ice–extent decrease of −1.6 ×
106 km2 (compare with 2012’s second record-
year drop of −1.0 × 106 km2) (10).
y
whereas during 2007 to 2021, the AD index be-
came increasingly positive (fig. S4E).
The AD+ drives an enhanced anticyclonic
Beaufort Gyre and Transpolar Drift (Fig. 2) (17).
Distinct from the Arctic Oscillation (18, 19),
the across-pole AD pattern results in increased
heat advection into the Arctic, especially along
the Siberian shores (Fig. 2B), and contributes
to higher surface air temperatures (fig. S3). The
AD was a major driver of the second record-

ification and enhanced heat release from the
warm intermediate (150- to 800-m depth)
Atlantic water layer, which results in accel-
erated loss of sea ice (5, 8, 9). However, these
Similarly, the composite record of mean winter
(February to March) ice thickness in the cen-
tral Arctic, now close to ~2 m, has not changed
markedly since 2007 (Fig. 1A), even though the
y g
low sea-ice extent in summer 2007 during the
satellite record (20). The Fram Strait sea-ice
export is correlated with the AD index during
1979 to 2014 [significant R = 0.45, (17); see Fig.

p
changes are complex, and their driving forces multidecadal decline has been significant. In 2D], with a stronger link to the AD than to the
and interactions with the Arctic atmosphere– addition, the mean thickness in fall (October Arctic Oscillation (21).
ice–ocean system are not well understood. to November) has remained above the 1-m low Most important for this study, the alternat-

This study identifies important mechanisms
that steer high-latitude atlantification, which
1
International Arctic Research Center and College of Natural
Science and Mathematics, University of Alaska Fairbanks,
Fairbanks, AK 99775, USA. 2Institute of Marine Research,
Bergen, Norway. 3International Arctic Research Center,
University of Alaska Fairbanks, Fairbanks, AK 99775, USA.
4
Geophysical Institute and College of Natural Science and
Mathematics, University of Alaska Fairbanks, Fairbanks, AK
99775, USA. 5Woodwell Climate Research Center, Falmouth,
MA 02540, USA. 6Alfred-Wegener-Institute, Helmholtz Centre
for Polar and Marine Research, D-27570 Bremerhaven,
Germany. 7Polar Science Center, Applied Physics Laboratory,
University of Washington, Seattle, WA 98105, USA. 8Bjerknes
Centre for Climate Research, Bergen, Norway.
*Corresponding author. Email: ivpolyakov@alaska.edu
established after the end of the summer of 2007.
Between 2003 and 2007, the thinning was
notable and occurred with the loss of a large
fraction of thick multiyear sea ice (11). The
overall thinning has slowed since 2007, when
satellite-based records of ice thickness began
(ICESat, ICESat-2, and CryoSat-2, 2003 to 2021).
The Arctic is now dominated by the behavior
of thinner seasonal ice, which now solely con-
trols the variability of ice thickness in the cen-
tral Arctic (12, 13).
Atmospheric changes
The atmosphere over the Arctic Ocean is dom-
inated by high pressure (known as the Polar
High) centered over the western Arctic (Fig.
,
ing AD phases were pivotal for a switchgear
mechanism that modulates the relative strength
of the Fram Strait and Barents Sea branches
transporting Atlantic water into the Arctic Ocean.
For example, the anomalous atmospheric forcing
during the AD+ in 2007 to 2021 was favorable for
reduced flows into the Arctic through the Fram
Strait along with enhanced inflows through the
Barents Sea Opening (Fig. 2, B, E, and F). The
Arctic Oscillation pattern contributes to large-
scale cyclonicity in Arctic atmosphere, ocean,
and ice circulation (22) but not the finer de-
tails suggested by the switchgear mechanism
discussed here (fig. S4). The AD-driven forcing
is best developed in spring and summer (figs.
S2 and S5) but can affect air temperatures,

Polyakov et al., Science 381, 972–979 (2023) 1 September 2023 1 of 8

RES EARCH | R E S E A R C H A R T I C L E
Fig. 1. Loss of Arctic sea-ice thickness and extent. (A) Arctic sea ice
thickness changes for autumn (red and dotted red) and winter (blue and dotted
blue). Shadings (blue and red) show 1 standard error (S.E.) ranges from the
regression analysis of submarine ice thickness and expected uncertainties in
satellite ice thickness estimates. Data release area of submarine data ice
thickness data are shown in the inset. Satellite ice thickness estimates are for
the Arctic south of 88°N. Thickness estimates from more localized airborne/
ground electromagnetic surveys near the North Pole (diamonds) and from
Operation IceBridge (circles) are shown within the context of the larger scale
changes in the submarine and satellite records (52). EM, electromagnetic.
(B) September Northern Hemisphere (NH) sea-ice extent (blue line), the long-
term trend (green line), and different regimes of sea-ice extent change in 1992 to
2006 and 2007 to 2021 (red segments); the inset shows available potential
energy (APE) anomalies in the upper ocean (surface mixed layer and halocline)
Polyakov et al., Science 381, 972–979 (2023) 1 September 2023
p
g
y
y g
p
,
for 2007 to 2017 relative to 1992 to 2006 (increasing APE signifies stronger
stratification suppressing mixing). (C and D) Maps of trends for the sea-ice
concentration (percentage per year) and the summer biweekly (time-integrated)
TI-NDVI (decade−1) over Arctic tundra for 1992 to 2006 (C) and 2007 to 2021
(D). The NDVI is a proxy for vegetation productivity and is derived from remotely
sensed products. A positive NDVI trend means that the vegetation has more biomass
and more photosynthetic productivity; this process is called greening. A negative
NDVI trend means that the vegetation has less biomass and is less healthy, and
it is called browning. (E and F) September sea-ice extent trends (E) and the
TI-NDVI (F) for 1992 to 2006 (blue) and 2007 to 2021 (red) as a function
of longitude. (G) Diagram of sea ice drift and upper ocean circulation (blue arrows),
as well as Atlantic water circulation (red arrows) (53). FJL-NZ, TPD, and BG
indicate Franz Joseph Land–Novaya Zemlya pass, TransPolar Drift, and
Beaufort Gyre, respectively.
2 of 8
RES EARCH | R E S E A R C H A R T I C L E

Fram Strait clearly show this alternating pat-

tern, with 5% increased annual mean currents
in the Barents Sea Opening and 15% decrease
in the Fram Strait (Fig. 3, C and D). Summer
and fall ORAS5 transports were the most not-
able contributors to anomalous Barents Sea
Opening inflows (15% increase), whereas spring
and summer processes dominated Atlantic
water inflows through the Fram Strait (28%
decrease in current speed) (fig. S7). At the 95%
confidence level, each of the aforementioned
estimates of anomalous transports is statisti-
cally significant. In comparison to the Barents
Sea Opening, changes in the upper 50-m water
volume transports were ~2.2 times as strong
in the Fram Strait. As a result, changes in trans-
ports through these gateways do not counter-
vail, and the importance of other gateways in
establishing the Arctic Ocean’s water balance
must not be underestimated. On weekly to

p
monthly timescales, increased eastward flow
in the northern Barents Sea and weakened in-
flow across the Fram Strait are associated with a
northward shift in atmospheric cyclones over
the Barents Sea (23).
Amplified Barents Sea Opening inflows in

g
2007 to 2021 resulted in increased transports
across the Franz Joseph Land–Novaya Zemlya
pass by 23%, thus providing an enhanced
inflow from the Barents Sea into the Arctic
Ocean (Fig. 3G and fig. S8). These changes

y
were crucial for the new state of the eastern
Arctic Ocean brought on by atlantification (5).
Tracer experiments showed a doubled prob-
ability that water parcels in the upper 50 m
crossed the Barents Sea in 2007 to 2021 com-
pared with 1992 to 2006, in contrast with the
Fram Strait, where the number of tracers de-
creased over the same time by a factor of four
(Fig. 4, D and E). Stronger impacts of the local
winds on the Barents Sea compared with Fram

y g
Strait inflows are consistent with stronger to-
pographic steering and more complex flow in
the Fram Strait (24, 25).
Fig. 2. Atmospheric forcing governing the switchgear mechanism. (A) Annual sea level pressure These findings are strongly supported by

p
(SLP)(hPa) averaged over 1992 to 2006. (B) Annual pressure anomalies (hPa, shading) in 2007 both in situ and satellite observations. For ex-
to 2021 relative to 1992 to 2006. Vectors show corresponding anomalous geostrophic winds. (C) April to ample, mooring and reanalysis records are posi-
July AD (hPa) pattern, which is correlated with the 2007 to 2021 pressure anomalies at R = 0.59. (D to tively correlated and show consistent increasing
F) Time series of atmospheric parameters. (D) April to July AD (blue) and March to August ice area transport
,
(decreasing) trends of currents across the Barents
across Fram Strait [green, from (19)] are reduced to anomalies by subtracting means (Mn) and normalized Sea Opening (Fram Strait) over the past 15 years
by SDs; AD and ice export are correlated at R = 0.44. (E and F) Annual wind across Fram Strait (E) and (fig. S9). A modest correlation R = 0.34 be-
Barents Sea Opening (BSO) (F). In (D) to (F), red horizontal segments show 1992 to 2006 and 2007 tween the Barents Sea Opening mooring and
to 2021 means. reanalysis time series is likely because moor-
ings do not cover the northern regions where
surface energy fluxes, storm tracks, sea-ice Reanalysis System 5 (ORAS5) reanalysis data the reanalysis shows the greatest increase in
drift and exports, and upper-ocean circulation (see the supplementary materials for details) flow. Moreover, anomalies in the satellite-based
in all seasons (fig. S3). suggest that these exchanges in the upper 50 m sea surface height provide further confirmation
were amplified across the northern Barents Sea for the switchgear mechanism modulating
Changes in the Arctic Ocean Opening and in the northern and central the inflows through the Barents Sea and Fram
Switchgear mechanism between the Fram Strait Barents Sea while being reduced across the Strait during the current AD+ (Fig. 4G). Geo-
inflow and Barents Sea throughflow Fram Strait in the past 15 years (Fig. 3), with strophic currents forced by the anomalous
Water exchanges between the Nordic Seas and similar patterns but weaker anomalies in the 2007 to 2021 sea surface height were ampli-
the Arctic Ocean are critically important for 50- to 200-m layer (fig. S6). A time series of fied in the northern Barents Sea Opening and
the state of the Arctic climate system. Ocean currents across the Barents Sea Opening and in the central Barents Sea but showed little

Polyakov et al., Science 381, 972–979 (2023) 1 September 2023 3 of 8

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

difference in the southern Barents Sea Open-

ing. However, the sea surface height indicates
an anomalous flow from the Lofoten Basin to
the Barents Sea in 2007 to 2020 (Fig. 4G). The
possibility of the Lofoten Basin feeding the
Barents Sea has been suggested by (26), at-
tributed to the vigorous eddy activity in the
Lofoten Basin (27, 28), and is further linked to
the atmospheric wind forcing (29). In contrast
to the Barents Sea Opening, the geostrophic
flow across the Fram Strait weakened (Fig. 4G).

Imprints of alternating AD patterns on the

Arctic Ocean circulation
The Arctic basins responded to the AD+ atmo-
spheric regime in 2007 to 2021 with basin-wide
changes in the upper Arctic Ocean circulation
associated with an amplified Beaufort Gyre,
stronger boundary currents along the Siberian
slope, and a shifted Transpolar Drift from the

p
Amerasian Basin toward the Lomonosov Ridge
(Fig. 3B and Fig. 4, A and B).
Since 2007, a smaller but more intense Arctic
high, associated wind-driven circulation, and
convergence of the upper Beaufort Gyre have
resulted in enhanced freshening and thicken-

g
ing of the surface fresh layer in the Amerasian
Basin (22, 30). The Beaufort Gyre mooring
record provides confirmation of these findings
(Fig. 5, A and B). This evidence is further sup-
ported by observed sea-ice melt, redirected

y
Siberian riverine waters into the Beaufort Gyre,
increasing inflow of relatively fresh Pacific Water
through the Bering Strait, and strengthened strat-
ification between the Amerasian Basin’s surface
and deep layers (22, 30–32). At the same time,
reanalysis and mooring observations showed
contrasting changes in the Eurasian Basin, with
increased salinification and weakened stratifica-
tion in the halocline, along with amplified up-
ward heat fluxes (Fig. 5, C and D) (5, 33).

y g
Our tracer experiments support these find-
ings (Fig. 4, A to C, and fig. S11). For example,
the 2007 to 2021 cyclonic atmospheric forcing
Fig. 3. Changes in upper 50 m oceanic circulation from ORAS5 reanalysis between alternate AD
widely dispersed transports of freshwater from

the Siberian shelves (where the Transpolar Drift
originates) and drove a substantial portion (17%
of all trajectories) of the Siberian freshwater
into the Beaufort Gyre. By contrast, during the
anticyclonic 1992 to 2006 AD− phase, not a
single water parcel ended up trapped in the
Beaufort Gyre, and they instead left the central
Arctic through the straits of the Canadian Ar-
chipelago. Changes in the spatial distributions
of meteoric water (i.e., precipitation including
water from lakes and rivers) provide confir-
mation of the diversion of freshwater from the
Eurasian Basin to the Amerasian Basin, with
meteoric water content decreasing in the
Eurasian Basin and increasing in the Amer-
asian Basin, consistent with (34, 35). These
changes in meteoric water content are accom-
panied by a general increase in net sea-ice melt-
water (fig. S12).
p
phases. (A and B) Maps of (A) 1992 to 2006 annual mean current speed |U| and (B) 2007 to 2021 |U|
anomalies (relative to the 1992 to 2006 mean) of the Arctic Ocean and Nordic Seas region with removed
low-frequency (1/30 years cut-off frequency) components by using running mean filtering. (C to E) Time
series of the ocean annual mean currents inflowing into the Arctic through Fram Strait (80°N, 14°W to 10°E)
,
(C), the Barents Sea Opening (71°N to 77°N, 20°E) (D), and the FJL–NZ passage (76.7°N to 80.6°N,
60.5°E to 64.3°E) (E). Red horizontal lines show means over 1992 to 2006 and 2007 to 2021. Transfer
coefficients from current in centimeters per second to water transport in sverdrups (Sv) (1 Sv = 106 m3/s)
are given in each time series panel. Note the reduced (enhanced) inflow through Fram Strait and
enhanced (reduced) Barents Sea throughflow in the years 2007 to 2021 (1992 to 2006).
In the Eurasian Basin, the local effects of al- convective entrainment in winter (5, 9). A stron-
ternating AD patterns and the remote effects ger coupling between atmosphere, ice, and ocean
of atlantification owing to changing influxes in the eastern Arctic in the recent decade and
across Fram Strait and Barents Sea are inter- intensified upper-ocean currents play critical
connected in an ice/ocean heat feedback mech- roles in developing this feedback (Fig. 3B) (36).
anism. Weakened stratification and increased
oceanic heat fluxes associated with atlantifica- Changes in the Nordic Seas
tion drive sea-ice melt, which is amplified by Anomalous anticyclonic winds over the Nordic
increased oceanic heat fluxes through increased Seas, as evident during the AD+ in 2007 to 2021

Polyakov et al., Science 381, 972–979 (2023) 1 September 2023 4 of 8

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

Fig. 4. Difference of the oceanic circulation patterns between alternate
AD phases. (A to G) A shift in the Transpolar Drift [(A) to (C)], inflows of Atlantic
water from the Nordic Seas across the Barents Sea Opening and Fram Strait
[(D) to (F)], and anomalous (2007 to 2021 minus 1992 to 2006) sea surface
height (SSH) (cm) and corresponding geostrophic currents (G). (A to C) ORAS5-
based trajectories of parcels exiting the central Siberian shelf in 1992 to 2006 (A)
and 2007 to 2021 (B) and the probability of finding a parcel within the polygon
p
g
y
y g
p
(indicated with the green line) north of the Canadian Archipelago (C). (D to F)
ORAS5-based trajectories of parcels released in the Nordic Seas along the
sections indicated with the black line in 1992 to 2006 (D) and 2007 to 2021 (E)
and the probability of finding a parcel traveling across the Barents Sea Opening, ,
Fram Strait, and FJL–NZ [indicated with the green lines in (D) and (E)]. (G)
Satellite-based anomalous SSH and geostrophic currents showing the switchgear
mechanism from Fram Strait to Barents Sea Opening inflow after 2007.

(Fig. 2B), weakened the poleward Atlantic wa-
ter flow from south of the subpolar gyre and
through the Nordic Seas (25, 37). Attributed
to a weak reinforcement of the subpolar gyre
(fig. S13) that altered the properties of the
inflowing Atlantic water (38), the Atlantic
water inflow from the northern North At-
lantic became colder and fresher in the late
2000s (Fig. 5, Q and R) (39). The Atlantic water
temperatures and salinities in the Nordic Seas
and the Barents Sea Opening responded to this
change by peaking in the late 2000s (Fig. 5, K to P).
Consistent with these changes, the salinity
trends in the Nordic Seas after 2006/2007 are
negative and similar in magnitude from the
North Atlantic up to the Barents Sea Opening
(Fig. 5, R, P, N, J, and I). Atmospheric cir-
culation over the Nordic Seas associated with
the AD+ in 2007 to 2021 (Fig. 2B) weakened
the northward Norwegian Atlantic Current
(Fig. 3B) (40), which is consistent with a neg-
ative trend of −3 cm/s per decade shown by
the mooring record from the Svinøy section, the
gateway for the Atlantic water into the Nordic
Seas (fig. S9). Despite these changes in the
Nordic Seas, the salinity in the eastern Eurasian
Basin halocline and Atlantic water increased
(Fig. 5D), driven by salinification in the up-
stream northern Barents Sea due to lack of
meltwater input from seasonal ice melt (41).

Polyakov et al., Science 381, 972–979 (2023) 1 September 2023 5 of 8

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

Fig. 5. Contrasting trends of ocean temperature and salinity in the region
spanning from the northern North Atlantic to the Arctic Ocean during the
positive (2007 to 2021) and negative (1992 to 2006) phases of the AD.
(A to R) Annual time series of water temperature and salinity from repeated
observations from the Atlantic water layer (for depth ranges, see supplementary
p
g
y
and methods). Observations with substantial gaps were complemented by ORAS5
reanalysis time series [gray lines; for the eastern Eurasian Basin (EEB), we plotted
ORAS5 temperature, adding 1°C]. Ninety-five percent statistically significant trends are
shown by solid red lines and black font; otherwise, dashed red lines and gray font
are used. The Beaufort Gyre (BG), Barents Sea Opening (BSO), Fram Strait (FS),

y g
materials; blue lines) and halocline (150 m, black lines) (see details in materials South Cape (SC), and North Eastern (NE) Barents Sea are indicated on the map.

The cooling trend in 2007 to 2021 is reduced 2007, with temperatures at the Barents Sea ern high-latitude climate system. Switchgear is

along stream the Atlantic water from the
northern North Atlantic into the Barents Sea
and Fram Strait (−0.55°C in the North Atlantic/
Rockall, −0.36°C at Svinøy, and insignificant
farther north; Fig. 5). These changes in the
Nordic Seas are supported by the reduced
oceanic heat loss over the past decades (42).
Thus, the traditional paradigm of Atlantic
water variability associated with either cold
and fresh or warm and saline Atlantic water
has changed (Fig. 5, E to N). Moreover, our
analysis confirms a northward amplification
of oceanic warming (7), with the origin in the
Nordic Seas, which makes this a source region
of atlantification and Arctic amplification.
The observed temperature and salinity trends
are also consistent with the switchgear-driven
stronger flow through the Barents Sea Opening
compared with Fram Strait flow after 2006/
Opening and northeastern Barents Sea show-
ing little or no decrease (Fig. 5, G and K), where-
as the Fram Strait temperatures fell rapidly (Fig.
5, E and I). Since the mid-2000s, the along-track
heat loss of the Fram Strait inflow has increased
(43), whereas the heat loss of the Barents Sea
throughflow has decreased or remained sta-
ble (43, 44), which thereby implies a stronger
warming of the Atlantic water entering the
Arctic from the Barents Sea as compared with
Fram Strait. Thus, a combination of the switch-
gear mechanism and regional changes in heat
loss make the Barents Sea a key contributor to
the Arctic Ocean atlantification.
Discussion
This research identifies the mechanisms driv-
ing atlantification and paints a broad and
comprehensive picture of changes in the north-
p
one of these mechanisms resulting from alter-
nating AD atmospheric regimes. We discovered
increased Atlantic water inflows throughout
,
the Barents Sea and reduced inflow across Fram
Strait, which resulted in a stronger warming of
the Barents Sea as compared with Fram Strait
during 2007 to 2021. This regime was also as-
sociated with the amplified Beaufort Gyre,
stronger boundary currents at the Siberian
slope, and a shifted Transpolar Drift from the
Amerasian Basin toward the Lomonosov Ridge.
One of the most notable changes associated
with AD variability is the 2007 to 2021 hiatus
of Arctic summer sea-ice loss, which, we ar-
gue, is a response to enhanced redistribution
of freshwater into the Amerasian Basin caused
by anomalous winds and increased stratifica-
tion that suppress oceanic heat fluxes. This
process is regionally limited to the Amerasian

Polyakov et al., Science 381, 972–979 (2023) 1 September 2023 6 of 8

RES EARCH | R E S E A R C H A R T I C L E
Basin, where the sea-ice area has actually in-
creased since 2007 (Fig. 1, D and E). Thus,
although variations in atmospheric forcing
may affect the ice-loss slowdown since 2007
(45), the shutting down of oceanic heat fluxes
by increasing Amerasian Basin stratification
(Fig. 1B, inset) may help drive the ice-loss
hiatus. To validate this hypothesis, we used the
winter survival of the near-surface tempera-
ture maximum created by the summer trap-
ping of solar radiation below the surface mixed
layer (10- to 30-m depth) (46). By using an ex-
tensive 2007 to 2020 archive of Ice-Tethered
Profiler observations, we found that 65% of
vertical temperature profiles showed the pres-
ence of the near-surface temperature maximum
(and therefore negligible upper Amerasian
Basin ventilation) during October to March (fig.
S14; see materials and methods for details)—a
convincing argument that enhanced sea-ice
winter growth owing to reduced ocean heat
fluxes contributed to the observed 2007 to 2021
ice-loss hiatus. Sea-ice dynamics should not be
disregarded, however. For example, despite in-
creased ice-area export (Fig. 2D), decreased
ice thickness in the Fram Strait (10) may cause
a decline in ice volume export, which would
offset the decline in net ice production in high-
latitude regions (47). In contrast to the Amer-
asian Basin, ventilation of the upper Eurasian
Basin in the 2010s is well documented (5, 9, 36),
and sea-ice loss in this basin continued through
the 2010s (Fig. 1, D and E).
There are numerous ongoing and potential
ecological consequences of the observed physical
changes. For example, the summer-integrated
normalized difference vegetation index (TI-
NDVI), a remotely sensed proxy for Arctic vege-
tation productivity, shows markedly different
behavior during AD+ and AD−. The TI-NDVI
trends at all longitudes were primarily posi-
tive in 1992 to 2006, which suggests that the
vegetation gained biomass and photosynthe-
tic productivity increased; the vegetation was
“greening” (Fig. 1, C and F). In 2007 to 2021,
negative trends dominated the area between
210°E to 300°E (corresponding to the Amer-
asian Basin sector where sea ice increased
during 2007 to 2021), which suggests that the
vegetation lost biomass and was possibly less
vigorous, often called “browning” (Fig. 1, D and
F). Thus, sea-ice variability can influence Arctic
vegetation productivity on numerous time-
scales, which is consistent with a correlation
between the TI-NDVI and spring sea-ice var-
iations (48).
Furthermore, the import of subarctic waters
has profound impacts on Arctic marine life
(49), and both the Fram Strait and the Barents
Sea Opening branches are potential pathways
of subarctic-boreal organisms into the eastern
Eurasian Basin (50). Our results suggest that
organisms drifting in the upper 50 m of the
Fram Strait branch had a fundamentally dif-
Polyakov et al., Science 381, 972–979 (2023) 1 September 2023
ferent fate if entering during AD− as compared
with AD+ (Fig. 4, D and E). During AD−, most
organisms entered the western Eurasian Basin
via the Transpolar Drift, whereas during AD+,
they were kept at the shelf break and trans-
ported into the eastern Eurasian Basin. In ad-
dition, during AD+, more organisms entered
the eastern Eurasian Basin from the Barents
Sea (Fig. 4E). An increased influence of the
Barents Sea may cause the eastern Eurasian
Basin to be more productive and provide more
suitable living conditions for subarctic-boreal
species than its western part (51), which is con-
sistent with recent observations (50). Improved
knowledge of asymmetric conditions in the
pelagic ecosystems of the western and eastern
Eurasian Basin is imperative to properly un-
derstand and manage the central Arctic Ocean
fisheries agreement established in 2021.
In addition, we note that recent atmospheric
changes as indicated with a shifting AD phase
have been important drivers of the regional
patterns of sea-ice and oceanic responses. There
are, however, indications that the Arctic system
may be entering another regime (see wavelets
in figs. S3, S7, and S8), with potential conse-
quences for the state of the physical, chemical,
and biological components. The transition may
be abrupt, similar to the rapid changes in 2007.
The trajectory of the Arctic climate system
into the future is further complicated by the
existence of large-amplitude, multidecadal varia-
bility (fig. S15). Thus, accurate future projec-
tions require a comprehensive understanding of
complex air-ice-ocean interactions and associ-
ated feedback mechanisms on broad spatio-
temporal scales through advancement of the
observing system and modeling capabilities.
RE FERENCES AND NOTES
1. IPCC, Climate Change 2022: Impacts, Adaptation, and
Vulnerability. Contribution of Working Group II to the Sixth
Assessment Report of the Intergovernmental Panel on
Climate Change, H.-O. Pörtner, D. C. Roberts, M. Tignor,
E. S. Poloczanska, K. Mintenbeck, A. Alegría, M. Craig,
S. Langsdorf, S. Löschke, V. Möller, A. Okem, B. Rama, Eds.
(Cambridge Univ. Press, 2022).
2. Q. Shu et al., Sci. Adv. 8, eabn9755 (2022).
3. R. A. Woodgate, T. Weingartner, R. Lindsay, Geophys. Res. Lett.
37, L01602 (2010).
4. K. Shimada et al., Geophys. Res. Lett. 33, L08605
(2006).
5. I. V. Polyakov et al., Science 356, 285–291 (2017).
6. I. V. Polyakov et al., Front. Mar. Sci. 7, 491 (2020).
7. R. B. Ingvaldsen et al., Nat. Rev. Earth Environ. 2, 874–889
(2021).
8. V. Ivanov et al., J. Phys. Oceanogr. 46, 1437–1456
(2016).
9. I. V. Polyakov et al., J. Clim. 33, 8107–8123 (2020).
10. H. Sumata, L. de Steur, D. V. Divine, M. A. Granskog, S. Gerland,
Nature 615, 443–449 (2023).
11. R. Kwok et al., J. Geophys. Res. 114, C07005 (2009).
12. R. Kwok, Environ. Res. Lett. 13, 105005 (2018).
13. S. Kacimi, R. Kwok, Geophys. Res. Lett. 49, e2021GL097448
(2022).
14. B. Wu, J. Wang, J. E. Walsh, J. Clim. 19, 210–225 (2006).
15. J. Zhang, R. Lindsay, M. Steele, A. Schweiger, Geophys. Res.
Lett. 35, L11505 (2008).
16. J. E. Overland, M. Wang, Tellus Ser. A Dyn. Meterol. Oceanogr.
62, 1–9 (2010).
17. L. H. Smedsrud, M. H. Halvorsen, J. C. Stroeve, R. Zhang,
K. Kloster, Cryosphere 11, 65–79 (2017).
18. J. E. Overland, M. Wang, Geophys. Res. Lett. 32, L23808
(2005).
19. J. E. Overland, M. Wang, S. Salo, Tellus Ser. A Dyn. Meterol.
Oceanogr. 60, 589–597 (2008).
20. J. Wang et al., Geophys. Res. Lett. 36, L05706 (2009).
21. M. Tsukernik, C. Deser, M. Alexander, R. Tomas, Clim. Dyn. 35,
1349–1360 (2010).
22. J. Morison et al., J. Phys. Oceanogr. 51, 1053–1075
(2021).
23. V. S. Lien, F. B. Vikebø, O. Skagseth, Nat. Commun. 4, 1488
(2013).
24. M. Muilwijk, L. H. Smedsrud, M. Ilicak, H. Drange, J. Geophys.
Res. Oceans 123, 8159–8179 (2018).
25. M. Muilwijk et al., J. Geophys. Res. Oceans 124, 6286–6322
(2019).
26. S. Broomé, L. Chafik, J. Nilsson, J. Geophys. Res. Oceans 126,
e2021JC017248 (2021).
27. D. L. Volkov, T. V. Belonenko, V. R. Foux, Geophys. Res. Lett.
40, 738–743 (2013).
28. P. K. Jakobsen, M. H. Ribergaard, D. Quadfasel, T. Schmith,
C. W. Hughes, J. Geophys. Res. 108, 3251 (2003).
29. L. Chafik, J. Nilsson, Ø. Skagseth, P. Lundberg, J. Geophys.
Res. Oceans 120, 7897–7918 (2015).
30. A. Proshutinsky et al., J. Geophys. Res. Oceans 124,
p
9658–9689 (2019).
31. R. A. Woodgate, Prog. Oceanogr. 160, 124–154 (2018).
32. S. B. Hall, B. Subrahmanyam, M. Steele, J. Geophys. Res.
Oceans 128, e2022JC019247 (2023).
33. I. V. Polyakov, M. Mayer, S. Tietsche, A. Y. Karpechko, J. Clim.
35, 2849–2865 (2022).
34. J. Morison et al., Nature 481, 66–70 (2012).
35. M. B. Alkire, J. Morison, R. Andersen, J. Geophys. Res. Oceans
g
120, 1573–1598 (2015).
36. I. V. Polyakov et al., Geophys. Res. Lett. 47, e2020GL089469
(2020).
37. L. H. Smedsrud et al., Rev. Geophys. 60, e2020RG000725
(2022).
38. H. Hátún, A. B. Sandø, H. Drange, B. Hansen, H. Valdimarsson,
y
Science 309, 1841–1844 (2005).
39. N. P. Holliday et al., J. Geophys. Res. Oceans 120, 5945–5967
(2015).
40. K. A. Orvik, Geophys. Res. Lett. 49, e2021GL096427
(2022).
41. B. I. Barton, Y.-D. Lenn, C. Lique, J. Phys. Oceanogr. 48,
1849–1866 (2018).
42. K. A. Mork et al., Geophys. Res. Lett. 41, 6221–6228
(2014).
43. G. W. K. Moore, K. Våge, I. A. Renfrew, R. S. Pickart, Nat. Commun.
13, 67 (2022).
44. Ø. Skagseth et al., Nat. Clim. Chang. 10, 661–666 (2020).
45. J. A. Francis, B. Wu, Environ. Res. Lett. 15, 114034
y g
(2020).
46. J. M. Jackson, W. J. Williams, E. C. Carmack, Geophys. Res.
Lett. 39, L03603 (2012).
47. J. Zhang, Geophys. Res. Lett. 48, e2021GL094780
(2021).
p
48. U. S. Bhatt et al., Environ. Res. Lett. 16, 055019
(2021).
49. B. A. Bluhm et al., Front. Mar. Sci. 7, 544386 (2020).
50. P. Snoeijs-Leijonmalm et al., Sci. Adv. 8, eabj7536
(2022).
,
51. R. B. Ingvaldsen et al., Sci. Rep. 13, 1000 (2023).
52. S. K. Gulev et al., in Climate Change 2021 – The Physical
Science Basis Working Group I Contribution to the Sixth
Assessment Report of the Intergovernmental Panel on Climate
Change (Cambridge Univ. Press, 2023), pp. 287–422.
53. Q. Wang et al., Ocean-Land-Atmos. Res. 2, 0013
(2023).
AC KNOWLED GME NTS
Funding: This study was supported by NSF grant no. 1724523
(A.V.P. and I.V.P.) and ONR grant no. N00014-21-1-2577 (I.V.P.).
J.A.F. was supported by funding from the Woodwell Climate
Research Center. Support for R.B.I. was provided by the Research
Council of Norway through the Nansen Legacy project (276730)
and Institute of Marine Research, Norway. U.S.B. was supported by
NASA's Arctic Boreal Vulnerability Experiment initiative under grant
80NSSC22K1257. Ø.S. received support from RCN grant no.
295962 NorEMSO. Author contributions: All authors participated
in data processing and preliminary analysis; A.V.P. and I.V.P.
carried out statistical analysis of reanalysis and mooring data from
7 of 8
RES EARCH | R E S E A R C H A R T I C L E

the eastern Eurasian Basin, R.K. provided sea-ice analysis and
processing, M.J. provided help interpreting Fram Strait data and
formulating objective of the study, J.A.F. supervised processing
and analysis of atmospheric data, Ø.S. and R.B.I. provided
processing and analysis of data from the Nordic and Barents seas,
and U.S.B. added multidisciplinary data and discussion. All authors
contributed to interpreting the data and writing the paper.
Competing interests: The authors declare no competing interests.
Data and materials availability: All data used in the analysis SUPPLEMENTARY MATERIALS
are available as described in the supplementary materials. science.org/doi/10.1126/science.adh5158
Correspondence and requests should be addressed to I.V.P. Materials and Methods
License information: Copyright © 2023 the authors, some rights Figs. S1 to S15
reserved; exclusive licensee American Association for the References (54–73)
Advancement of Science. No claim to original US government
works. https://www.science.org/about/science-licenses-journal- Submitted 9 March 2023; accepted 19 July 2023
article-reuse 10.1126/science.adh5158

p
g
y
y g
p
,

Polyakov et al., Science 381, 972–979 (2023) 1 September 2023 8 of 8

RES EARCH

HUMAN EVOLUTION However, these formulae may not achieve the

required accuracy because of propagation and
Genomic inference of a severe human bottleneck accumulation of numerical errors resulting from
their dependence on the joint probability den-
during the Early to Middle Pleistocene transition sity function of coalescent times (16–18).
In this study, to circumvent this numerical
Wangjie Hu1,2†‡, Ziqian Hao3†, Pengyuan Du1,3, Fabio Di Vincenzo4, Giorgio Manzi5, Jialong Cui2, problem, we developed the fast infinitesimal
Yun-Xin Fu6,7, Yi-Hsuan Pan2*, Haipeng Li1,8* time coalescent process (FitCoal) (Fig. 1) that
analytically derives expected branch length
Population size history is essential for studying human evolution. However, ancient population size history for each SFS category under arbitrary demo-
during the Pleistocene is notoriously difficult to unravel. In this study, we developed a fast infinitesimal time graphic models. FitCoal does not need phased
coalescent process (FitCoal) to circumvent this difficulty and calculated the composite likelihood for present- haplotype data and prior information on de-
day human genomic sequences of 3154 individuals. Results showed that human ancestors went through a mography. The effects of sequencing errors or
severe population bottleneck with about 1280 breeding individuals between around 930,000 and 813,000 years hitchhiking due to positive selection can be
ago. The bottleneck lasted for about 117,000 years and brought human ancestors close to extinction. This circumvented, largely by focusing on a subset
bottleneck is congruent with a substantial chronological gap in the available African and Eurasian fossil record. of SFS that are less influenced by those factors.
Our results provide new insights into our ancestry and suggest a coincident speciation event. We used FitCoal to analyze a large number of
present-day human genomic sequences from 10
African and 40 non-African populations. Results

A
showed that our ancestors experienced a sev-

lthough the lineage of humans is esti-
mated to have separated from that of
chimpanzees and bonobos more than
6 million years ago, anatomically mod-
ern humans (Homo sapiens) are estimated
to have originated around 300 thousand to
proach is needed to improve the inference ac-
curacy of population size history.
Population size changes that occurred hun-
dreds of thousands of years ago affected the
rates of coalescence and thus have left their
signatures in the site frequency spectrum (SFS)
p
ere population bottleneck between about 930
and 813 kyr BP, most likely because of climatic
changes. The average number of breeding indi-
viduals was only about 1280 during the bottle-
neck period. Our findings indicate that the
severe bottleneck brought the ancestral human

200 thousand years before the present (kyr BP)
in Africa (1–3). On the basis of present-day hu-
man genomic sequences, the recent population
size histories (i.e., the dynamics of population
size since the emergence of modern humans)
g
of genomic sequences. The SFS is the distri- population close to extinction and completely
bution of allele frequencies in the sequences, reshaped present-day human genetic diversity.
randomly collected from the present-day hu-
man population. Each SFS category contains Results
a certain number of mutations of the same FitCoal

y
have been intensively studied, revealing the size. Because SFS is crucial for demographic We developed FitCoal to determine the expected
worldwide spread of our ancestors (4–8). How- inference (6, 8–12) and construction of key branch lengths for an SFS (Fig. 1). During
ever, ancient population size history of the genus summary statistics (13), many efforts have been FitCoal analysis of a sample, the time period
Homo during the Pleistocene is still poorly devoted to deriving its analytical formulas (14–17). in which the most recent common ancestor
known, although it is essential for understand-
ing the origin of the human lineage. It is likely
to be very difficult or impossible to obtain an-
cient DNA from African Homo samples dated
before the emergence of H. sapiens. It would
be particularly notable if present-day human

y g
genomic sequences could be used to robustly
infer both the recent and ancient population
size histories of humankind. Thus, a new ap-

p
1
CAS Key Laboratory of Computational Biology, Shanghai

,
Institute of Nutrition and Health, University of Chinese Academy
of Sciences, Chinese Academy of Sciences, Shanghai, China.
2
Key Laboratory of Brain Functional Genomics of Ministry of
Education, School of Life Science, East China Normal University, Fig. 1. Illustration of FitCoal. (Left) The backward process in which four lineages (represented by the
Shanghai, China. 3College of Artificial Intelligence and Big Data four solid black circles at the bottom) coalesce into one (represented by the single solid black circle at
for Medical Sciences, Shandong First Medical University &
Shandong Academy of Medical Sciences, Jinan, China. 4Natural
the top) after passing through millions of infinitesimal time intervals (Dt). The area highlighted in blue shows the
History Museum, University of Florence, Florence, Italy. backward transformation process of different coalescent states with tiny probability changes in an infinitesimal
5
Department of Environmental Biology, Sapienza University of time interval. Thick arrows indicate high transformation probabilities, and thin arrows indicate low transformation
Rome, Rome, Italy. 6Department of Biostatistics and Data
probabilities. The blue and purple arrows correlate to the two events in the middle pane represented by
Science, School of Public Health, University of Texas Health
Science Center at Houston, Houston, TX, USA. 7Key Laboratory blue- and purple-colored lines. Each state is indicated with a box, in which one circle indicates one lineage.
for Conservation and Utilization of Bioresources, Yunnan The boxes with solid black circles represent the states with the probability of 1. The boxes with empty
University, Kunming, China. 8Center for Excellence in Animal circles represent the states with the probability of 0. The probabilities between 0 and 1 are represented
Evolution and Genetics, Chinese Academy of Sciences,
Kunming, China. by gray circles. (Middle) Hypothetical coalescent trees with branches of different states, indicating the
*Corresponding author. Email: yxpan@sat.ecnu.edu.cn (Y.-H.P.); number of lineages. Blue branches represent a transformation from four to three lineages. Purple branches
lihaipeng@sinh.ac.cn (H.L.) indicate that no coalescent event occurred. (Right) The size of a theoretical population over time. The
†These authors contributed equally to this work.
‡Present address: Department of Genetics and Genomic Sciences, width of shadowed area denoted as N(t) indicates the effective population size (i.e., the number of breeding
Icahn School of Medicine at Mount Sinai, New York, NY, USA. individuals) at time t. It is assumed that the effective population size remains unchanged within a Dt.

Hu et al., Science 381, 979–984 (2023) 1 September 2023 1 of 6

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

A Constant size B Instantaneous increase C Bottleneck

n = 30, L = 10 Mb n = 30, L = 10 Mb n = 170, L = 30 Mb
Effective population size

100
100 100
(in thousands)

10 10 True Model
10
1
1 FitCoal

10 100 1000 10 100 1000 10 100 1000

PSMC
D Exponential growth I E Exponential growth II F Exponential growth III
n = 30, L = 10 Mb n = 30, L = 10 Mb n = 170, L = 10 Mb
Effective population size

10,000 10,000 500

Stairway Plot
(in thousands)

100

p
100 100
SMC++
10
1
1

g
10 10 100 10 100 1000
Time (thousand years ago) Time (thousand years ago) Time (thousand years ago)
Fig. 2. Population size histories inferred by FitCoal, PSMC, Stairway Plot, and histories; thin red lines represent 2.5 and 97.5 percentiles of FitCoal-inferred population

SMC++ with simulated samples. (A) Constant size model. (B) Instantaneous
increase model. (C) Bottleneck model. (D) Exponential growth I model. (E) Exponential
growth II model. (F) Exponential growth III model. In all panels, thin black lines indicate
the true models. Thick red lines indicate the medians of FitCoal-inferred population size
y
size histories. Yellow, green, and blue lines indicate results obtained with PSMC, Stairway
Plot, and SMC++, respectively. The mutation rate is assumed to be 1.2 × 10–8 per
base per generation, and a generation time is assumed to be 24 years. n is the number of
simulated sequences, and L is the length in Mb of each simulated sequence.

originated was partitioned into as many time
intervals as needed, such that each time inter-
val (Dt) was very small (e.g., 1 month or 1 year).
During each time interval, the population size
was assumed to be constant. The probabilities
the likelihood of the SFS was first maximized
with the constant size model, followed by re-
peated maximization of the likelihood with
increased number of inference time intervals
until the best model was found.
that FitCoal could distinguish between instan-
taneous and exponential changes (table S1).
Overall, our results confirmed that SFSs could
be used to estimate population size histories (22).
It has been suggested that a population size

of all coalescent states (i.e., all possible ances-
tral lineages) were calculated backward in
time. For each state, the branch length during
a time interval was calculated by multiplying
Demographic inference from simulated data
The accuracy of FitCoal demographic infer-
ence was evaluated by simulation and com-
y g
history could be inferred by using a subset of
SFS or a collapsed SFS (6, 19); the latter is an
SFS with high frequency mutations combined
into one category. Results of simulations showed

p
its probability with population size and then parison of results with those of PSMC (pairwise that the FitCoal could still accurately determine
transformed to determine the expected branch sequentially Markovian coalescent), Stairway a population size history even when a portion
lengths. Because the expected branch length Plot, and SMC++ methods (6–8) (Fig. 2). To (10 to 90%) of an SFS was truncated (i.e., ex-

of an SFS category during a time interval was
precalculated, FitCoal could be very fast.
FitCoal demographic inference
After the expected branch lengths were de-
termined, the composite likelihood of an SFS
(6, 9, 19) was calculated. FitCoal is effective
for a wide range of sample sizes in the calcu-
lation of the composite likelihood of a given
SFS and is much more accurate than simu-
lation approaches (fig. S1). When inferring
population size history, the likelihood was
maximized in a wide range of demographic
scenarios. Moreover, both instantaneous and
long-term exponential population changes were
considered. Similar to previous studies (6, 19),
ensure fair comparisons, we tested six demo-
graphic models by simulating 200 independent
datasets for each model, as described previ-
ously (6), with the assumption that a generation
time is 24 years (6, 20) and that the mutation
rate is 1.2 × 10–8 per site per generation (6, 21).
Results showed that the medians of FitCoal-
inferred population size histories were almost
identical to the true models, and the 95% con-
fidence intervals of FitCoal inference were nar-
rower than those of PSMC, Stairway Plot, and
SMC++ (Fig. 2). FitCoal inference accuracy
could be further improved by increasing sam-
ple size and lengths of sequences (fig. S2). The
proportion of the most recent changes in popu-
lation size inferred from the six models showed
,
cluded for analysis) (figs. S3 to S5), thus reduc-
ing the impact of confounding factors, such as
hitchhiking effect due to positive selection (fig.
S6) or sequencing errors on FitCoal analyses.
Demographic inference of African populations
To infer population size histories of African
populations, seven African populations in the
1000 Genomes Project (1000GP) (23) and three
African populations in the Human Genome
Diversity Project–Centre d’Etude du Polymor-
phisme Humain (HGDP-CEPH) panel (24) were
analyzed by the FitCoal (tables S2 to S4). Only
autosomal noncoding regions were used to
partially avoid the effect of purifying selection.
To avoid hitchhiking effect due to positive

Hu et al., Science 381, 979–984 (2023) 1 September 2023 2 of 6

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

Fig. 3. Histories of human populations in 1000GP and HGPD-CEPH genomic
datasets inferred by FitCoal, SMC++, Stairway Plot, PSMC, and Relate. The
mutation rate is assumed to be 1.2 × 10–8 per base per generation, and a generation
time is assumed to be 24 years. (A) Inferred population size histories of 26
populations in 1000GP. (B) Linear-scaled estimation of sizes over time of populations
in 1000GP during the severe bottleneck period. (C) Inferred population size histories
of 24 populations in the HGPD-CEPH panel. (D) Linear-scaled estimation of sizes
over time of populations in the HGPD-CEPH panel during the severe bottleneck period.
(E and F) Comparison of population size histories of African YRI and Yoruba
p
g
y
populations inferred by FitCoal, SMC++, Stairway Plot, PSMC, and Relate. Only
the population size histories up to 200 kyr BP were analyzed by Stairway
Plot. In (A) to (D), colored lines indicate the following: red, African populations;
brown, Middle East populations; yellow, European populations; blue, East Asian
populations; green, Central or South Asian populations; and gray, American
populations. Black dashed circles with arrow heads represent the discrepancy in
the population size between African agriculturalist and non-African populations.
In (E) and (F), red, blue, yellow, green, and gray lines indicate results obtained
with FitCoal, SMC++, PSMC, Stairway Plot, and Relate, respectively.

y g
selection (25), high-frequency mutations were the 21 non-African populations in the HGDP- but in none of the 40 non-African populations.
excluded from the analysis. Results showed that CEPH panel (Fig. 3, A to D; figs. S7 and S8; and To investigate this discrepancy, we performed
all 10 African populations went through a sev- tables S2, S5, and S6). The average ancestral simulations with three demographic models,

p
ere bottleneck (Fig. 3 and figs. S7 and S8). The population sizes of the populations in the two designated bottlenecks I, II, and III (Fig. 4, A
bottleneck was estimated to persist for 117 kyr, datasets were 20,260 (range, 18,850 to 22,220) to C, and figs. S9 and S10). Bottleneck I sim-
from 930 ± 23.52 (SEM) (range, 854 to 1042) and 20,030 (range, 19,060 to 21,850), respec- ulated the population size history of African

to 813 ± 11.02 (SEM) (range, 772 to 864) kyr BP.
The average effective population size (i.e., the
number of breeding individuals) (26) during
the bottleneck period was determined to be
1280 ± 131 (SEM) (range, 770 to 2030), which
was only 1.3% of its ancestral size (98,130 ±
8720; range, 58,600 to 135,000). To evaluate the
impact of the bottleneck on current human
genetic diversity, we analyzed the expected
pairwise nucleotide diversity. Results showed
that 65.85% of current human genetic diver-
sity was lost because of the bottleneck.
Demographic inference of non-African populations
The severe bottleneck was not directly detected
in all 19 non-African populations in 1000GP and
tively, similar to those determined in previous
studies (7, 8, 24). The estimated population
size started to decline around 368 (range, 175 to
756) and 367 (range, 167 to 628) kyr BP, respec-
tively, which is consistent with previous find-
ings that African and non-African divergence
occurred much earlier than the out-of-Africa
dispersal (7, 8, 23, 24). The inferred out-of-
Africa dispersal and the recent population size
expansion and reduction are consistent with
those of previous studies (5–8, 23, 24).
Severe bottleneck during the Early to Middle
Pleistocene transition
The ancient severe bottleneck was directly de-
tected in each of the 10 African populations
,
agriculturalist populations with the ancient
severe bottleneck, and bottlenecks II and III
simulated that of non-African populations with-
out and with the ancient severe bottleneck,
respectively. Both bottlenecks I and II were
inferred precisely in all simulated data sets
(tables S7 to S9). However, no ancient severe
bottleneck was detected in bottleneck III sim-
ulations, which indicates that the out-of-Africa
dispersal hinders the chance of discovering the
ancient severe bottleneck. Furthermore, the
ancient severe bottleneck was found to cause a
discrepancy in the estimation of the population
size between the bottleneck III model and the
inferred population size history after the bottle-
neck was relieved, which suggests a hidden effect

Hu et al., Science 381, 979–984 (2023) 1 September 2023 3 of 6

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

Fig. 4. Verification of the severe bottleneck. (A) Bottleneck I model,
mimicking the true population size history of African agriculturalist populations.
(B) Bottleneck II model, mimicking the inferred population size history of non-
African populations. (C) Bottleneck III model, mimicking the true population
size history of non-African populations. The gap between actual and FitCoal
estimated population size is indicated by the black dashed circle and arrowhead.
(D) Bottleneck IV model, mimicking a population with an exponential reduction
in size 1.5 million years ago. (E) Bottleneck V model, mimicking a population
with a moderate bottleneck. (F) Bottleneck VI model, mimicking a population
p
g
y
with a weak bottleneck. Black lines represent the true models. Thick red lines
represent the medians of FitCoal estimated population sizes over time; thin
red lines represent 2.5 and 97.5 percentiles of FitCoal estimated population
sizes over time. Blue, green, yellow, and gray lines represent the medians of
10 runs each of SMC++, Stairway Plot, PSMC, and Relate. The mutation rate is
assumed to be 1.2 × 10–8 per base per generation, and a generation time is
assumed to be 24 years. Numbers of simulated sequences are 188 in bottlenecks
I, IV, V, and VI and 194 in bottlenecks II and III. The lengths of simulated
sequences are 800 Mb each.

of the ancient severe bottleneck on non-African
populations (Fig. 4C and figs. S9C and S10C).
After the bottleneck was relieved, the average
population size of non-African populations in
formed an extended FitCoal analysis. To eli-
minate noise effects resulting from problems
such as sequencing or overfitting errors on the
inference of population size history, we used
y g
that PSMC, SMC++, and Relate methods (7, 8, 27)
underestimated the severity of the ancient bot-
tleneck (Fig. 4 and figs. S14 to S17). More-
over, the inferred population declines shown

p
1000GP was 20,260, and that of those in HGDP- the FitCoal-inferred recent population size his- in Fig. 4A were more severe than those inferred
CEHP was 20,030. For African agriculturalist tory as a starting point for size estimation of from the real data because the simulated data
populations, the average population size in an ancient population. With this modifica- were generated under the assumption of a neu-

1000GP was 27,080, and that of those in HGDP-
CEHP was 27,440. This population size difference
of 7020 (Fig. 3, A and C) is likely due to the
hidden effect of the ancient severe bottleneck
on non-African populations. Because the out-of-
Africa dispersal existed in non-African popu-
lations but not in African populations, African
populations had more lineages remaining to
be traced back to the ancient severe bottle-
neck (fig. S11). In the analysis of the African YRI
(Yoruba in Ibadan, Nigeria) population, the
minimum sample size was three individuals
for detection of the severe bottleneck (fig. S12).
Because the signal for the existence of the
severe bottleneck was too weak to be detected
in non-African populations by FitCoal, we per-
tion, all 19 non-African populations in 1000GP
were found to have gone through the severe
bottleneck with approximately 1450 individ-
uals between 921 and 785 kyr BP (fig. S13 and
table S10). This result is consistent with that
obtained with African populations.
To further examine the severe bottleneck, we
simulated a slow population reduction start-
ing 1.5 million years ago (Fig. 4D). The FitCoal-
inferred population size histories were different
from those observed in 1000GP and HDGP-
CEHP populations, which supports the hypoth-
esis that a sudden size reduction occurred at
the beginning of the bottleneck. Results of sim-
ulations were similar to that of the observed
cases (Fig. 3, E and F) in that they also showed
,
trally evolved single population and a homoge-
neous recombination rate, whereas humans
evolved with subpopulations and a heterogeneous
recombination rate (28, 29). FitCoal was not
found to overestimate such severity (Fig. 4) or
to falsely detect a bottleneck when examining
the effects of continuous and pulsed introgres-
sions among existing and ghost populations
(figs. S18 to S21). Therefore, the discovery of
the ancient severe bottleneck was not due to
overfitting the data in FitCoal analyses.
Discussion
In this study, we developed FitCoal, a model-
flexible method for demographic inference.
One key feature of FitCoal is that the expected

Hu et al., Science 381, 979–984 (2023) 1 September 2023 4 of 6

RES EARCH | R E S E A R C H A R T I C L E
Fig. 5. Schematic diagram of
human population size history.
Both African (light green) and non-
African (light blue) populations
are presented. The width of the boxes
represents the effective population
size (i.e., the number of breeding
individuals) with naturally occurred
fluctuations. The occurrence time of
the out-of-Africa dispersal and the
divergence between African and
non-African populations are indicated.
The gray-shaded time duration
indicates the Early to Middle Pleis-
tocene transition between 1250 and
700 kyr BP. The red arrow indicates
the peak of glaciation during the
transition (i.e., the 0.9 Ma event).
The ancient severe bottleneck
inferred in this study is highlighted. The gap in the available African hominin fossil record and an indicative
chronology for H. erectus, the LCA, and H. sapiens are shown. The estimated time period in which two
ancestral chromosomes (chromosome, Chr.) fused to become one is also shown on the right.
branch lengths can be accurately determined
for an SFS under an arbitrary demographic
model, which allows precise calculation of the
likelihood. Analyses by FitCoal are, in most
cases, less time consuming than those by other
methods such as PSMC, SMC++, Stairway Plot,
and Relate (28/32 = 87.5%) (tables S11 and S12).
By discarding rare and high-frequency muta-
tions, FitCoal can avoid the effects of sequencing
errors or hitchhiking due to positive selection
without losing its inference accuracy. Because
both instantaneous and exponential changes
are allowed within each inference time inter-
val, FitCoal can reveal the dynamic of population
size precisely. Because coalescent events be-
come rare when tracing backward in time, the
length of inference time interval is usually set
to increase progressively (6–8). Although this
strategy can capture recent population size
changes, it may miss ancient ones. Therefore,
FitCoal inference time intervals are allowed to
vary during demographic inference, and FitCoal
can make a fast and accurate inference of recent
and ancient population size histories.
The most important discovery with FitCoal
is that human ancestors went through a se-
vere bottleneck in the late Lower Pleistocene
(Fig. 5). This ancient severe bottleneck was
directly found in all 10 African populations,
but only a weak signal of the existence of such
was detected in all 40 non-African popula-
tions. This observation is consistent with the
coalescent theory and the occurrence of the
out-of-Africa dispersal. Results of our large-
scale simulations demonstrated that FitCoal
did not falsely infer the bottleneck because of
positive selections (figs. S6 and S22) or pop-
ulation structure (fig. S18 to S21). Because we
observed no overfitting cases and results ob-
tained by examining different sets of genomic
regions (Fig. 3, A and C, and fig. S23) were sim-
Hu et al., Science 381, 979–984 (2023) 1 September 2023
ilar, the existence of the ancient severe bottle-
neck was ascertained.
Our results indicate that the ancient severe
bottleneck lasted for approximately 117 kyr
(Fig. 5), and that about 98.7% of human an-
cestors were lost at the beginning of the bot-
tleneck, thus threatening our ancestors with
extinction. The estimated effective population
size during the bottleneck period was only
1280 breeding individuals, which was compa-
rable to the effective population sizes of other
endangered mammals (30, 31). This size (1280)
might have been overestimated because of hid-
den population structure (32). Naturally occur-
ring population size fluctuations might have
further increased the extinction risk for our
ancestors during the bottleneck. The bottleneck
could also have increased the inbreeding level
of our ancestors, thus contributing to the 65.85%
loss in present-day human genetic diversity.
The ancient population size reduction that
occurred around 930 kyr BP was likely driven
by climatic changes during the Early to Mid-
dle Pleistocene transition (33, 34). During this
transition period known as the “0.9 Ma event”
(Ma, million years ago), glaciations were changed
from predominantly short-term to long-term
events with more extreme thermic intensity,
especially at the peak of glaciation. This event
resulted in a decrease in marine surface tem-
perature to the lowest that occurred during the
entire transition period (33), with an inferred
long period of drought and extensive wildlife
turnover in Africa and Eurasia (35).
The existence of the ancient severe bottle-
neck could explain the extreme scarcity of the
available hominin fossil record in Africa and
Eurasia between 950 and 650 kyr BP (Fig. 5 and
fig. S24). In Africa, only a few fossil specimens
dated in this time period have been found, in-
cluding the cranial fragments from Gombore
in Ethiopia and the fossil samples from Tighenif
in Algeria (36, 37). Although the taxonomic sta-
tuses of these fossils are still not clear, they have
features resembling those of later fossils at-
tributed to Homo heidelbergensis. They are dif-
ferent from the coeval Homo antecessor from a
paleoanthropological site in Spain (Atapuerca,
Gran Dolina), and some scholars considered
H. antecessor as a possible alternative for the last
common ancestor (LCA) (38). During the same
chronological interval, the East Asian fossil
record contains specimens identified as Homo
erectus (39). It does not appear that East Asian
H. erectus is connected to the ancient severe
bottleneck because it is unlikely to have contrib-
uted to the lineage leading to modern humans
(38). In addition, coincident with this bottleneck,
two ancestral chromosomes are believed to have
fused to form chromosome 2 in humans around
900 to 740 kyr BP (40, 41). Therefore, the ancient
p
severe bottleneck possibly marks a speciation
event leading to the emergence of the LCA
shared by Denisovans, Neanderthals, and mod-
ern humans, whose divergence has been dated
to about 765 to 550 kyr BP (38, 42, 43).
A rapid population recovery was detected in
g
all 10 African populations with a 20-fold in-
crease in size around 813 kyr BP. Control of
fire could be part of the explanation for this
population expansion, which is shown by the
early archaeological evidence found in Israel
y
dated about 790 kyr BP (44). Other factors, such
as climatic changes (33, 34), might also be a
driving force for this rapid population recovery.
The ancient severe bottleneck was not de-
tected in previous SFS-based analyses (6, 10, 12, 14).
This failure might be due to the use of prede-
fined demographic models. In this study, we
found that the likelihood must be accurately
calculated to detect the severe bottleneck (fig.
S1). The use of other methods such as Stairway
y g
Plot—which may not have sufficient resolu-
tion power for estimation of ancient popula-
tion size history—is another possible reason for
the failure (6).
p
Our study revealed that an extremely small
human population lasted for about 117 kyr
around 930 to 813 kyr BP. Many questions
,
remain unanswered, such as where these indi-
viduals lived, how they overcame the cata-
strophic climate changes, and how the ancient
population remained so small for so long. Fur-
ther studies are warranted to investigate these
matters to obtain a more detailed picture of
human evolution during the Early to Middle
Pleistocene transition.
REFERENCES AND NOTES
1. J. Galway-Witham, C. Stringer, Science 360, 1296–1298 (2018).
2. C. M. Schlebusch et al., Science 358, 652–655 (2017).
3. A. Meneganzin, T. Pievani, G. Manzi, Evol. Anthropol. 31, 199–212 (2022).
4. M. Stoneking, J. Krause, Nat. Rev. Genet. 12, 603–614 (2011).
5. S. Ramachandran et al., Proc. Natl. Acad. Sci. U.S.A. 102,
15942–15947 (2005).
6. X. Liu, Y. X. Fu, Nat. Genet. 47, 555–559 (2015).
7. H. Li, R. Durbin, Nature 475, 493–496 (2011).
5 of 6
RES EARCH | R E S E A R C H A R T I C L E
8. J. Terhorst, J. A. Kamm, Y. S. Song, Nat. Genet. 49, 303–309 (2017).
9. H. Li, W. Stephan, PLOS Genet. 2, e166 (2006).
10. L. Excoffier, I. Dupanloup, E. Huerta-Sánchez, V. C. Sousa,
M. Foll, PLOS Genet. 9, e1003905 (2013).
11. R. C. Griffiths, S. Tavaré, Math. Comput. Model. 23, 141–158 (1996).
12. R. N. Gutenkunst, R. D. Hernandez, S. H. Williamson,
C. D. Bustamante, PLOS Genet. 5, e1000695 (2009).
13. Y. X. Fu, Theor. Popul. Biol. 145, 95–108 (2022).
14. J. Jouganous, W. Long, A. P. Ragsdale, S. Gravel, Genetics 206,
1549–1567 (2017).
15. Y. X. Fu, Theor. Popul. Biol. 48, 172–197 (1995).
16. A. Polanski, M. Kimmel, Genetics 165, 427–436 (2003).
17. D. Zivković, T. Wiehe, Genetics 180, 341–357 (2008).
18. A. Polanski, A. Bobrowski, M. Kimmel, Theor. Popul. Biol. 63,
33–40 (2003).
19. A. Bhaskar, Y. X. R. Wang, Y. S. Song, Genome Res. 25,
268–279 (2015).
20. A. Scally, R. Durbin, Nat. Rev. Genet. 13, 745–753 (2012).
21. A. Kong et al., Nature 488, 471–475 (2012).
22. A. Bhaskar, Y. S. Song, Ann. Stat. 42, 2469–2493 (2014).
23. The 1000 Genomes Project Consortium, Nature 526, 68–74 (2015).
24. A. Bergström et al., Science 367, eaay5012 (2020).
25. J. C. Fay, C. I. Wu, Genetics 155, 1405–1413 (2000).
26. H. C. Harpending et al., Proc. Natl. Acad. Sci. U.S.A. 95,
1961–1967 (1998).
27. L. Speidel, M. Forest, S. Shi, S. R. Myers, Nat. Genet. 51,
1321–1329 (2019).
28. E. M. L. Scerri et al., Trends Ecol. Evol. 33, 582–594 (2018).
29. Z. Hao, P. Du, Y. H. Pan, H. Li, Hum. Genet. 141, 273–281 (2022).
30. H. Li et al., Proc. Natl. Acad. Sci. U.S.A. 113, 14079–14084 (2016).
31. Y. X. Chen et al., Zool. Res. 43, 523–527 (2022).
32. J. Wakeley, Genetics 153, 1863–1871 (1999).
33. M. J. Head, P. L. Gibbard, in Early-Middle Pleistocene
Transitions: The Land–Ocean Evidence, M. J. Head, P. L. Gibbard,
Eds. (Geological Society of London, 2005), pp. 1–18.
34. P. U. Clark et al., Quat. Sci. Rev. 25, 3150–3184 (2006).
Hu et al., Science 381, 979–984 (2023) 1 September 2023
35. M. J. Head, B. Pillans, S. A. Farquhar, Episodes 31, 255–259 (2008).
36. A. Profico, F. Di Vincenzo, L. Gagliardi, M. Piperno, G. Manzi,
J. Anthropol. Sci. 94, 41–63 (2016).
37. A. Mounier, F. Marchal, S. Condemi, J. Hum. Evol. 56, 219–246
(2009).
38. A. Bergström, C. Stringer, M. Hajdinjak, E. M. L. Scerri,
P. Skoglund, Nature 590, 229–237 (2021).
39. J.-J. Bahain et al., Anthropologie 121, 215–233 (2017).
40. B. Poszewiecka, K. Gogolewski, P. Stankiewicz, A. Gambin,
BMC Genomics 23, 616 (2022).
41. T. R. Dreszer, G. D. Wall, D. Haussler, K. S. Pollard, Genome Res. 17,
1420–1430 (2007).
42. K. Prüfer et al., Nature 505, 43–49 (2014).
43. D. Reich et al., Nature 468, 1053–1060 (2010).
44. N. Goren-Inbar et al., Science 304, 725–727 (2004).
45. H. Li, Genomic inference of a severe human bottleneck during
the Early to Middle Pleistocene transition, version 3, Mendeley
Data, 2023; https://doi.org/10.17632/xmf5r8nzrn.3.
46. W. Hu et al., FitCoal v1.1., Zenodo, 2023; https://doi.org/
10.5281/zenodo.7857456.
ACKN OWLED GMEN TS
We thank D. Živković for sharing his codes to calculate the expected
branch lengths, X. Liu for providing simulation results, and Bio-Med
Big Data Center at Shanghai Institute of Nutrition and Health for
providing a computing facility. Funding: National Natural Science
Foundation of China grants 32270674, 91131010, and 91731304
(H.L.); National Natural Science Foundation of China grant 31100273
(Y.H.P.); National Natural Science Foundation of China grants
91631304 and 32130011 (Y.X.F.); National Natural Science
Foundation of China grant 82171801 (Z.H.); Strategic Priority Research
Program of the Chinese Academy of Sciences grant XDPB17 (H.L.);
National Key Research and Development Project grant 2022YFF1203202
(H.L.); National Institutes of Health grant R01HG009524 (Y.X.F.);
Education Bureau of Jinan and Shandong First Medical University grant
JNSX2021046 (Z.H.); Key Laboratory of Brain Functional Genomics at
East China Normal University grant 20SKBFGK2 (Y.H.P. and H.L.);
Shanghai Institute of Nutrition and Health grant JBGSRWBD-SINH-2021-
10 (H.L.); China Postdoctoral Science Foundation grant
2022M711978 (Z.H.); Shandong Provincial Natural Science
Foundation grant ZR2022QC062 (Z.H.); and Shandong Provincial
Postdoctoral Innovation Talent Support Program grant
SDBX2022012 (Z.H.). Author contributions: Conceptualization:
W.H., Z.H., F.D.V., G.M., Y.X.F., Y.H.P., and H.L.; Methodology: W.H.,
Z.H., Y.H.P., and H.L.; Software: W.H., Z.H., and H.L.; Investigation:
W.H., Z.H., P.D., F.D.V., G.M., J.C., Y.X.F., Y.H.P., and H.L.; Visualization:
W.H., Z.H., P.D., J.C., F.D.V., and G.M.; Funding acquisition: Z.H., Y.X.F.,
Y.H.P., and H.L.; Supervision: G.M., Y.H.P., and H.L.; Writing: W.H.,
Z.H., P.D., F.D.V., G.M., Y.X.F., Y.H.P., and H.L. Competing interests:
The authors declare that they have no competing interests. Data
and materials availability: Raw data are deposited at Mendeley (45),
and FitCoal is archived at Zenodo (46). The download links for
1000GP and HGDP-CEPH data are available in table S2. License
information: Copyright © 2023 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.abq7487
p
Materials and Methods
Supplementary Text
Figs. S1 to S60
Tables S1 to S23
References (47–96)
MDAR Reproducibility Checklist
Submitted 28 April 2022; resubmitted 7 February 2023
Accepted 11 July 2023
g
10.1126/science.abq7487
y
y g
p
,
6 of 6
RES EARCH

NATURAL HAZARDS The second earthquake (Mw 7.7) occurred on

the Savrun-Çardak Fault (SCF), extending
The complex dynamics of the 2023 Kahramanmaraş, ~150 km along the east-west direction (Fig. 1).
The SCF has been relatively quiescent, with
Turkey, Mw 7.8-7.7 earthquake doublet only two moderate (Mw <6) events recorded in
the past 100 years (25).
Zhe Jia1*, Zeyu Jin1, Mathilde Marchandon2, Thomas Ulrich2, Alice-Agnes Gabriel1,2, Wenyuan Fan1, We constrained the rupture geometry on
Peter Shearer1, Xiaoyu Zou1, John Rekoske1, Fatih Bulut3, Aslı Garagon3, Yuri Fialko1 the basis of surface traces mapped using Syn-
thetic Aperture Radar data (26) and precisely
The destructive 2023 moment magnitude (Mw) 7.8-7.7 earthquake doublet ruptured multiple segments of the relocated aftershocks (27, 28). We found that
East Anatolian Fault system in Turkey. We integrated multiscale seismic and space-geodetic observations the Kahramanmaraş doublet ruptured at least
with multifault kinematic inversions and dynamic rupture modeling to unravel the events’ complex rupture six major fault segments (Fig. 1). The epicenter
history and stress-mediated fault interactions. Our analysis reveals three subshear slip episodes during the of the Mw 7.8 earthquake is located on a sub-
initial Mw 7.8 earthquake with a delayed rupture initiation to the southwest. The Mw 7.7 event occurred 9 hours sidiary fault, the Nurdağı-Pazarcık (Narlı) Fault
later with a larger slip and supershear rupture on its western branch. Mechanically consistent dynamic models (NPF) (Fig. 1A, fault 1) (20), from which the rup-
accounting for fault interactions can explain the unexpected rupture paths and require a heterogeneous ture propagated to the EAF, and then ruptured
background stress. Our results highlight the importance of combining near- and far-field observations with along the EAF to both the northeast and south-
data-driven and physics-based models for seismic hazard assessment. west (Fig. 1A, faults 2 and 3), for a total length
of about 300 km. Unlike the historical Mw ~7
events, the Mw 7.8 earthquake propagated

he moment magnitude (Mw) 7.8 and 7.7
Kahramanmaraş earthquakes in Turkey
on 6 February 2023 caused enormous de-
struction and tens of thousands of ca-
sualties from collapsed structures and
together were one of the deadliest natural
applied to complex ruptures on multiple faults,
conventional earthquake source imaging often
involves oversimplified assumptions, yielding
stark differences in source models and their
interpretations (14, 15). Initial studies of the
Kahramanmaraş earthquakes presented a wide
p
across at least four possible geometric barriers,
including fault bends and stepovers.
The static slip distribution (Fig. 1B) obtained
from inversions of Synthetic Aperture Radar
(SAR) and Global Navigation Satellite System
(GNSS) data (figs. S1 to S7) shows that the
largest slip in the Mw 7.8 event is on the EAF at

disasters for Turkey and Syria over the past
millennium (1). The Kahramanmaraş sequence
is the first great earthquake doublet with a
combined moment magnitude of 8 recorded
in a continental strike-slip fault system. Unlike
range of earthquake models and interpreta-
tions (16–21), likely from focusing on particular
datasets and aspects of the rupture process.
These differences motivate unified and self-
consistent approaches that integrate diverse
g
its junction with the NPF, near the towns of
Kahramanmaraş and Pazarcık, with a peak
slip in excess of 8 m. Most of the coseismic
slip is in the upper 20 km of the seismogenic

regular aftershocks that are more than one
order of magnitude smaller than their main-
shock, doublet events pose a greater hazard
because they can cause more severe damage
by striking already weakened buildings and
structures. We show that the Kahramanmaraş
earthquake doublet involved a remarkable se-
quence of subevents that occurred with vary-
ing rupture velocities, geometries, and time
delays on branched fault segments, which chal-
datasets with state-of-the-art rupture mod-
els to advance our understanding of the earth-
quake dynamics.
We performed a comprehensive investiga-
tion of the Mw 7.8-7.7 Kahramanmaraş doublet
using data-driven and physics-based analyses
applied to near- and far-field seismic and ge-
odetic observations. Our results reveal that
the earthquakes followed unexpected rupture
trajectories, which included delayed backward
y
layer (Fig. 1B). Slip at the surface is highly
heterogeneous, which is consistent with field
observations (18), but on average increases
from the southwest to the northeast ends of
the Mw 7.8 rupture (fig. S8). The area of sub-
stantial slip extends to the northeast from
the junction for about 150 km to the western
tip of the 2020 Mw 6.7 Elazığ rupture (Fig. 1A)
(29). South of the junction, the Mw 7.8 rup-
ture extends to the southern end of the EAF.

lenge our understanding of earthquake interac-
tions and the dynamics of rupture propagation.
Seismologists commonly approximate earth-
quakes as point sources or as slip along a sin-
branching, statically and dynamically aided
triggering, and a combination of subshear and
supershear rupture episodes. These discoveries
call for reevaluating the role of cascading fail-
y g
The average coseismic slip on the southwest
branch of the Mw 7.8 rupture is smaller than
the average slip on the northeast branch (Fig.
1B and fig. S2).

gle fault with fixed rupture velocity. However,
large earthquakes often rupture multiple fault
segments within a complex network (2–6).
Occasionally, events of a comparable magni-
tude occur within minutes to hours of the ini-
tial event, resulting in earthquake doublets
(7–9). Branching faults may further compli-
cate rupture dynamics (10–12). Whether rup-
ture stops or continues propagating at fault
junctions can determine earthquakes’ even-
tual size and destructive potential (13). When
1 quakes occurred on the EAF historically, but
Scripps Institution of Oceanography, University of California,
San Diego, La Jolla, CA 92093, USA. 2Department of Earth
and Environmental Sciences, Ludwig-Maximilians-Universität
München, 80539 Munich, Germany. 3Geodesy Department,
Bogazici University Kandilli Observatory and Earthquake
Research Institute, Istanbul 34342, Turkey.
*Corresponding author. Email: z5jia@ucsd.edu
ure mechanisms when assessing the destructive
potential of large earthquakes within complex
fault networks.
The geometrically complex Mw 7.8-7.7
earthquake doublet
On 6 February 2023, two major (Mw > 7) earth-
quakes ruptured several previously recognized
fault systems within 9 hours (Fig. 1). The East
Anatolian Fault (EAF) is a mature transform
fault accommodating up to 10 mm/year of left-
lateral motion between the Arabian and Anato-
lian plates (Fig. 1) (22). Several Mw ~7 earth-
none ruptured the entire southern section of
the EAF (23). The estimated dimensions of the
historic events suggest that geometric com-
plexities such as fault bends and step-overs
may have controlled the event sizes (23, 24).
p
We resolved the spatiotemporal rupture
process with a subevent inversion method by
using both near- and far-field seismic obser-
vations (30, 31). The Mw 7.8 earthquake had
,
six subevents that altogether spanned ~90 s
(Fig. 2A). The Mw 6.8 subevent E1 that rup-
tured the NPF was followed 18 s later by the
largest subevent E2 (Mw 7.5) at the NPF-EAF
intersection. The earthquake then ruptured
northeastward along the EAF for about 130 km
(Mw 7.5 subevent E3), as well as, after a short
delay, backward from the NPF junction for
about 150 km along the southwestern segment
of the EAF, with an integrated slip equivalent
to a Mw 7.4 earthquake (subevents E4 to E6).
Teleseismic P wave back-projection (32) con-
firmed the rupture process, with imaged high-
frequency radiation peaks outlining the major
subevents (Fig. 2A) and indicating an average

Jia et al., Science 381, 985–990 (2023) 1 September 2023 1 of 6

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

rupture velocity of 3 km/s. To further con-
strain the slip history, we performed a joint
kinematic slip inversion of the Mw 7.8 earth-
quake constrained by far- and near-field seis-
mic and geodetic data (26, 33). Our kinematic
inversion results agree with the static and
subevent models (Fig. 2B). The best-fit kine-
matic slip model images 10-s-delayed backward
branching at the NPF-EAF intersection, toward
the southwest (Fig. 2B), constrained by the
strong-motion data (fig. S17). It also indi-
cates average rupture velocities of 3.2 km/s and
2.8 km/s for the northeastern and southwest-
ern branches, respectively (fig. S18). Tracking
ground motion pulses at near-fault strong

A Mw 7.7 B
50 km Slip (m)
6 0 5 10
Savrun-Çardak Fault 4
lt
38
F au
l ian 2
5 n ato 2020 Mw 6.7
stA
Ea
1
Nurdaǧı-Pazarcık Turkey
Fault

37
Black Sea
0
5

p
Depth, km
NAF
Mw 7.8
3 AS AT F 10
EA EU
Syria 15
Mediterranean
20
36 AF AR 25
0 0.5 1

g
36 37 38 39 Normalized slip

Fig. 1. A multifault earthquake doublet. (A) Tectonic background and aftershock blue beachball denote the rupture extent and focal mechanism of the 2020 Mw 6.7
seismicity of the study area near Kahramanmaraş, Turkey. Red and purple stars Elazığ earthquake (29). (Inset) The regional tectonics and major plate boundary
indicate the Mw 7.8 and 7.7 earthquake epicenters according to the Turkey Disaster faults (solid black lines). Red outline denotes the study area. (B) Finite-fault model of

and Emergency Management Authority (53), and red and purple beachballs indicate
focal mechanisms from the Global Centroid Moment Tensor catalog, respectively.
Red and purple lines indicate surface ruptures identified from SAR data (26). Yellow
dots indicate aftershocks for the period between the Mw 7.8 and 7.7 earthquakes,
and black dots indicate aftershocks after the Mw 7.7 event (28). The blue line and
y
the 2023 doublet derived from inversions of space geodetic (InSAR and GNSS)
data. Fault segment numbers correspond to those shown in (A), in order of their
rupture time: 1, Nurdağı-Pazarcık Fault; 2 and 3; East Anatolian Fault; 4 to 6,
Savrun-Çardak Fault. (Inset) The along-strike averaged coseismic slip normalized by
the maximum slip amplitude, as a function of depth (49).

A B C Distance along EAF (km)

Subevents NE −150 −100 −50 0 50 100 150

NPF-EAF junction
E2
Hypocenter
E3 Subevent
0-11 s

y g
E4E5 E1 BP
E1 E6
Finite Fault
0 25 50 75 NE 80
38˚ Time (s) E3 Mw 7.5 E6
11-21 s E2

p
E4 Mw 7.0
E2 Mw 7.5 60 E5
SW
Time (s)

yed NE, 3.2km/s

E5 Mw 7.2 Dela
E4

,
37˚ E1 Mw 6.8 21-50 s
SW

E4 40
E6 Mw 6.9 E3
E3
/s
8 km
, 2. E2
BP SW 20
50-90 s Delay
NE

0−16 s 16−30 s
36˚ 30−38 s 38−44 s E5
km 44−62 s 62−82 s E6 50 km
0 50 82−95 s 0 E1
36˚ 37˚ 38˚ 39˚
Fig. 2. Complex slip evolution of the Mw 7.8 earthquake, including delayed intervals from our kinematic slip inversion of far- and near-field seismic and
initiation of slip. (A) Subevent model from near- and far-field seismic geodetic data. We infer rupture velocities of 3.2 and 2.8 km/s for the northeast
observations and back-projection results, suggesting that the Mw 7.8 earthquake and southwest episodes, respectively, and a 10-s delay in the onset of the
initiated on the NPF-1 (Fig. 1B, fault 1), then propagated bilaterally, northeast southwest rupture along EAF-3 with respect to the NE rupture along EAF-2. The
along the EAF-2 (Fig. 1B, fault 2) and southwest along the EAF-3 (Fig. 1B, fault 3). slip distribution within each time interval agrees with the subevent (black circles)
The rupture of fault 2 terminates around 50 s, whereas rupture of fault 3 inversion. (C) Subevents, back-projection locations and times, and finite-fault
continues for an additional 30 s. (B) Rupture history within different time velocities [in (B)] consistently indicate delayed initiation of slip on branch EAF-3.

Jia et al., Science 381, 985–990 (2023) 1 September 2023 2 of 6

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

motion stations along the southwestern seg-
ment also yielded a rupture velocity of ~3 km/s
(fig. S18), confirming an overall subshear na-
ture. All our kinematic models consistently
reveal a ~300-km-long complex bilateral multi-
segment rupture, subshear rupture velocities,
and delayed triggering of the southwest seg-
ment of the Mw 7.8 event (Fig. 2C).
The subsequent Mw 7.7 earthquake ruptured
a 150-km-long section of the west-trending
SCF, within 90 km of the Mw 7.8 earthquake
hypocenter. The aftershock distribution and
surface offsets indicate branching and abrupt
changes in strike at both the eastern and west-
ern ends of the Mw 7.7 rupture (Fig. 1). Geo-
detic data and our associated static slip model
(Fig. 1B) suggest rupture along an 80-km-
long segment of the SCF system (Fig. 1, faults
4 and 5), but not along the eastern end of the
Sürgü fault that connects to the EAF. In-
bilateral rupture in the central SCF. The focal
mechanism (strike of 237°) and location of
the last subevent (E4; Mw 7.1) agree with the
static slip model on the Doğanşehir branch
(Fig. 1B). All subevents of both earthquakes
have almost pure double-couple mechanisms
(Figs. 2A and 3A), suggesting that the strong
non–double-couple components in the Global
Centroid-Moment-Tensor solutions (Fig. 1A) (34)
are due to highly variable rupture geometries.
The overall shorter duration and smaller rup-
ture extent of the Mw 7.7 event make back-
projection analysis less effective for resolving
rupture details, but our kinematic finite-slip
inversion can still be applied.
The kinematic finite-fault model of the Mw
7.7 earthquake also indicates a compact slip
distribution. In addition, it indicates a west-
ward rupture velocity of ~4.5 km/s (Fig. 3B),
exceeding the shear-wave speed in the crust.
arrest. Unlike purely data-driven kinematic slip
inversions, such models predict the evolution
of slip, seismic waves, and surface deformation
in a physically self-consistent manner. Detailed,
physics-based interpretations can help verify
whether inferred rupture scenarios are me-
chanically plausible but are computationally
challenging and typically take years to devel-
op [for example, (10, 12, 13)].
We present data-informed dynamic rupture
simulations of the 2023 Kahramanmaraş earth-
quakes that illuminate complex details of the
rupture process. Our three-dimensional (3D)
dynamic rupture models include stress changes
computed from the slip distribution of the
static slip model (37), large-scale variability in
fault loading inferred from regional seismo-
tectonics, and the relative effects of the static
and dynamic stresses of the Mw 7.8 event on
the faults hosting the second earthquake (fig.

stead, the Mw 7.7 rupture diverted sharply
onto the Doğanşehir branch, which angles
to the northeast (Fig. 1, fault 6). The Mw 7.7
event shows a concentrated slip distribution
with >10 m peak slip around its hypocenter,
suggesting a substantially higher stress drop
than that of the initial Mw 7.8 earthquake,
The waveforms recorded at the westward seis-
mic stations strongly constrain this supershear
rupture episode (Fig. 3C and fig. S19), which is
consistent with analysis of high-rate GNSS data
(20). By contrast, the eastward rupture likely
propagated at a slower velocity of 2.5 km/s.
p
S20) (26). The dynamic rupture models inde-
pendently reproduce the main features of the
kinematic models (Fig. 4 and fig. S21), pro-
viding a physics-based validation of the in-
ferred rupture histories.
Our forward simulations use the complex

which spread a lower-amplitude slip over a
larger region.
Our analysis of the rupture history of the
Mw 7.7 event identified four major subevents,
The intriguing supershear rupture episode
may imply locally higher prestress (35) and
high stress drop (36) as in our dynamic rup-
ture models.
g
fault geometries of both earthquakes informed
from geodetic analysis (Fig. 1) to spontaneously
replicate the moment rate release, magnitude,
rupture velocity and delays, as well as the lack
of instantaneous dynamic triggering of the

lasting for about 30 s (Fig. 3A). The first three
subevents, E1 to E3, all cluster near the epi-
center and account for more than 80% of the
total seismic moment, suggesting a compact
Dynamics, triggering, and stress interaction
of the doublet
Dynamic rupture modeling involves simulat-
ing how earthquakes nucleate, propagate, and
y
Mw 7.7 event. The dynamic rupture synthetics
produce surface displacements and slip histor-
ies that compare well with the high-resolution
geodetic data (fig. S22), kinematic rupture

A B C
Westward Supershear (4.5 km/s)
W E
39˚ 0-8 s E−W N−S
E2 Mw 7.6 E1 4612
E2

y g
E1 Mw 7.1
E3 Mw 7.4 E4 Mw 7.1 0131

0127
Supershear Subshear
4612 (4.5 km/s) (2.5 km/s)

p
8-20 s 0 40 80 0 40 80
38˚ 0131
0127 Time (sec) Time (sec)
E3
Westward Subshear (3 km/s)
E−W N−S ,

Time (s) NE 4612

37˚
km 20-35 s
0 10 20 30
0 50 Subevents 0131
E2 E4
0127
E1 E3
E4
30 km 0 40 80 0 40 80
36˚ 37˚ 38˚ Time (sec) Time (sec)

Fig. 3. Asymmetric kinematics of the Mw 7.7 earthquake. (A) Three subevents
close to the hypocenter suggest a bilateral rupture. The fourth event images the
rupture of the Doğanşehir branch (Fig. 1B, fault 6). (B) Asymmetric bilateral
rupture velocities of the Mw 7.7 event. The westward rupture has an inferred
supershear velocity of 4.5 km/s, whereas a subshear velocity is seen toward the
east (2.5 km/s). Subevent locations are based on their seismic moment
centroids. The slip may not be the largest at the centroid location, specifically for
bilateral ruptures. For example, E3 (10 to 30 s) averages slip pulses of both the
westward supershear and the eastward subshear rupture. (C) A westward
supershear rupture velocity (red waveforms) better explains observed waveforms
(black) at near-fault strong motion stations to the west [triangles in (A)] than a
subshear rupture (blue).

Jia et al., Science 381, 985–990 (2023) 1 September 2023 3 of 6

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

A C

stress and strength

25
shear stress

(MPa)
8s
fault strength
20

13s delayed backward

rupture 0 20 40 60
time (s)
b.
16s

32s
peak dynamic dCFS (MPa)

44s 0.0 3.5 7.0

D
68s

absolute slip rate (m/s)

B North supershear rupture 0 2 4 6 8

p
East

dCFS (MPa)

−1.5 0.0 1.5

rupture speed (m/s)

g
0 3000 6000

Fig. 4. 3D dynamic rupture scenarios and stress-mediated interactions of the scenario (fig. S21 and movie S2). (C) Peak absolute dynamic shear stress

Mw 7.8 and 7.7 earthquakes. (A) Snapshots of absolute slip rate evolution in the
Mw 7.8 dynamic rupture scenario (movie S1). (B) Modeled rupture speeds in linked
dynamic rupture simulations (54) of both earthquakes indicating dominantly
subshear rupture speeds but sustained westward supershear during the Mw 7.7
y
perturbation reaching up to 7 MPa measured in the direction of maximum initial
traction. (Inset) Evolution of dynamic shear stress and fault strength at the Mw 7.7
hypocenter (black star). (D) Static Coulomb failure stress changes (DCFS)
assuming a static friction coefficient of 0.6.

representations (Fig. 4 and fig. S21), and ob- event experienced an increase in static Coulomb pared with the initial rupture on the NPF. In
served ground motions (Fig. 5 and figs. S23 stress of several hundred kilopascals because addition, the Mw 7.8 earthquake increased the
to S25). The modeled Mw 7.8 earthquake dy- of the Mw 7.8 earthquake, resulting from both Coulomb stress on the central part of the SCF,
namics are illustrated in Fig. 4A. The NPF-EAF an increase in shear stress and a decrease in which may have aided the nucleation of the

intersection slows subshear rupture on the
NPF that then branches with dynamically fa-
vorable forward directivity (38) northeastward
along the EAF. The large fault branching angle
fault-normal compression (fig. S27). It also ex-
perienced a much larger transient increase
in the Coulomb stress of a few megapascals
owing to passing seismic waves (Fig. 4C), which
y g
Mw 7.7 earthquake 9 hours later. The entire
process highlights the additional hazard brought
by rupture triggering across a network of faults,
challenging earthquake hazard assessments

poses a strong dynamic barrier in backward-
directivity (39), leading to substantially delayed
EAF rupture toward the southwest. Continuous
dynamic unclamping, transient shear stress-
ing, and static stress buildup at the fault in-
tersection due to the unilaterally propagating
northeast rupture allowed the rupture to even-
tually fracture the EAF bilaterally (fig. S26).
Rupture speed remained overall subshear dur-
ing the earthquake (Fig. 4B).
Dynamic rupture modeling of the Mw 7.7
earthquake features bilateral rupture with un-
equal rupture speeds, confirming dominant
supershear westward and subshear eastward
propagation. Our Mw 7.8 dynamic rupture
model predicts a highly variable pattern of
static and dynamic stresses resolved on the
faults that hosted the Mw 7.7 earthquake (Fig. 4,
C and D). The hypocentral area of the Mw 7.7
nevertheless did not result in instantaneous
triggering.
Discussion and conclusions
Our analyses reveal unexpected rupture paths.
The Kahramanmaraş doublet originated as a
moderate event on the NPF branch fault with
a magnitude of only 6.8, yet the rupture was
able to successfully cross the junction of the
NPF and EAF, which would usually be con-
sidered a geometric barrier that conditionally
gates the rupture propagation (40, 41). As a
result, the earthquake intensified with the
northeastward propagation along the EAF then
dynamically triggered backward rupture toward
the southwest by continuously unclamping and
stressing from the forward branch, eventually
culminating in a Mw 7.8 event, with total seis-
mic moment increased by a factor of 30 com-
p
that typically do not consider such multifault-
triggering scenarios.
The Mw 7.8 earthquake involved backward
,
fault branching, which is highly unfavorable
from a dynamic perspective, thus commonly
neglected in hazard studies. Several previous
continental earthquakes—including the 1992
Landers, the 1999 Hector Mine, and the 2002
Denali earthquakes—have also exhibited local-
ized backward branching (10). Existing explan-
ations of this phenomenon include backward
rupture jumping induced by sudden rupture
arresting or nonuniform prestress fields caused
by earthquake cycles (39, 42). Our dynamic
rupture models indicate that backward branch-
ing during the Mw 7.8 event does not necessarily
require a complex arrangement of the receiver
fault (42) or triggering of supershear rupture
(43). Instead, the progressive build-up of slip on

Jia et al., Science 381, 985–990 (2023) 1 September 2023 4 of 6

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

A B field seismic and geodetic observations and

101 investigating data-derived models and physics-
based rupture simulations, we show that stress
interactions and static and dynamic triggering
100 worked together across a complex fault sys-
tem, resulting in a cascade of rupture with a
PGV (m/s)

larger than usual total rupture length and mo-

10-1 ment magnitude. Our study shows that com-
plementary data-driven and physics-based
analyses, which in isolation often lead to non-
10-2 unique or even contradictory results, can jointly
and efficiently unravel highly complex earth-
quake dynamics based on dense near-field ob-
100 101 102 103 100 101 102 103
RJB (km) RJB (km)
servations. The unusual static and dynamic
interactions during and between the events of
Fig. 5. Peak ground velocities (PGVs) plotted against Joyner-Boore distance (RJB) for the Mw 7.8 the Kahramanmaraş doublet call for reassess-
and Mw 7.7 earthquakes. (A) The Mw 7.8 earthquake. (B) The Mw 7.7 earthquake. Observed PGVs from ment of common assumptions built into seis-
strong motion accelerometers are indicated with open black circles, and simulated PGVs from the dynamic mic hazard assessments.
rupture simulations are indicated with open blue squares. We bin the PGV data by RJB and plot the medians
REFERENCES AND NOTES
for each distance bin (solid markers). The red curve indicates PGV predicted by a ground motion model

p
1. M. Edrik, M. B. D. Tümsa, A. Pınar, E. Altunel, A. C. Zülfikar,
(55), assuming an average shear wave velocity for the top 30 m of soil (VS30) of 760 m/s, with the shaded “A preliminary report on the February 6, 2023 earthquakes in
area denoting its uncertainty. All PGV are rotationally independent geometric mean values (GMRotD50). Türkiye” (Temblor, 2023).
2. T. Lay, L. Ye, Y. Bai, K. F. Cheung, H. Kanamori, Geophys. Res. Lett.
We include simulated and observed data at the same locations, respectively.
45, 9542–9551 (2018).
3. L. Meng et al., Science 337, 724–726 (2012).
4. E. Hauksson, L. M. Jones, K. Hutton, Bull. Seismol. Soc. Am. 92,
the forward branch of the EAF continuously 1B), which is well resolved in our inverse mod- 1154–1170 (2002).
5. E. Hauksson, L. M. Jones, K. Hutton, D. Eberhart‐Phillips,

unclamps and stresses the backward branch
of the EAF, eventually leading to a delayed and
self-sustained branching toward the southwest,
which is a simple yet effective mechanism.
One of the unexpected aspects of the Mw 7.7
els (fig. S28). Subsequent observations will show
whether this reduction can be compensated
by shallow afterslip or constitutes a long-term
shallow slip deficit (49), implying widespread
off-fault yielding (47, 50).
g
J. Geophys. Res. 98 (B11), 19835–19858 (1993).
6. Z. E. Ross et al., Science 366, 346–351 (2019).
7. A. Ghods et al., Geophys. J. Int. 203, 522–540 (2015).
8. C. J. Ammon, H. Kanamori, T. Lay, Nature 451, 561–565 (2008).
9. W. Fan, P. M. Shearer, Science 353, 1133–1136 (2016).

earthquake is that it did not rupture through
the eastern Sürgü segment and arrive at the
EAF, contrary to earlier suggestions (20), but
instead deviated to the Doğanşehir branch.
The Interferometric Synthetic Aperture Radar
(InSAR), aftershock, and seismic data clearly
show such a deviation (figs. S1 and S14). The
straightforward rupture path along the Sürgü
fault was encouraged by the static stress changes
from the Mw 7.8 event (fig. S27), unlike the sharp
We also found intriguing variations in rup-
ture velocity across segments of the EAF-SCF
fault network. Although the Mw 7.8 event pro-
duced extreme shaking with peak ground ac-
celerations (PGAs) exceeding 1 g for near-fault
stations, the observed and simulated Mw 7.7
ground motions are similar or larger when com-
pared at the same distance (Fig. 5 and figs. S23
and S24), which is consistent with a potentially
larger stress drop of the Mw 7.7 event. The
y
10. D. D. Oglesby, S. M. Day, Y.-G. Li, J. E. Vidale, Bull. Seismol.
Soc. Am. 93, 2459–2476 (2003).
11. H. S. Bhat, M. Olives, R. Dmowska, J. R. Rice, J. Geophys. Res.
112 (B11), B11309 (2007).
12. T. Ulrich, A.-A. Gabriel, J.-P. Ampuero, W. Xu, Nat. Commun.
10, 1213 (2019).
13. R. Douilly, H. Aochi, E. Calais, A. Freed, J. Geophys. Res. Solid
Earth 120, 1108–1128 (2015).
14. S. Minson, M. Simons, J. Beck, Geophys. J. Int. 194, 1701–1726
(2013).
15. K. Wang, D. S. Dreger, E. Tinti, R. Bürgmann, T. Taira, Bull. Seismol.
Soc. Am. 110, 1603–1626 (2020).
16. S. Barbot et al., Seismica 10.26443/seismica.v2i3.502 (2023).

deviation to the Doğanşehir fault, which was
actually unloaded by the Mw 7.8 event (Fig. 4D
and movie S3). Possible explanations, which may
be tested by future geodetic and seismological
modeled and observed Mw 7.7 event shaking
shows less distance dependence, which may be
due to the effects of supershear rupture.
The western branch of the SCF experienced a
y g
17. D. E. Goldberg et al., Seismic Record 3, 156–167 (2023).
18. V. Karabacak et al., J. Geol. Soc. London 180, jgs2023–jgs2020
(2023).
19. P. M. Mai et al., Seismic Record 3, 105–115 (2023).
20. D. Melgar et al., Seismica 10.26443/seismica.v2i3.387 (2023).
21. R. Okuwaki, Y. Yagi, T. Taymaz, S. P. Hicks, Geophys. Res. Lett.

observations, include velocity-strengthening be-
havior of the eastern Sürgü segment or local
stress heterogeneity, for example, because of
supershear rupture episode, whereas the east-
ern SCF branch and the EAF hosted subshear
ruptures with considerable delays. In general,
p
50, e2023GL103480 (2023).
22. N. Lyberis, T. Yurur, J. Chorowicz, E. Kasapoglu, N. Gundogdu,
Tectonophysics 204, 1–15 (1992).
23. S. E. Güvercin, H. Karabulut, A. Ö. Konca, U. Doğan, S. Ergintav,

past earthquakes (44). Considerable regional
stress heterogeneity, as is required by our dy-
namic rupture models (fig. S20), is implied by
extremely complex rupture geometries that
involve changes in the strike angle of up to 90°
(Fig. 1 and fig. S27) (45). Some faults in the
study area, including the EAF, exhibit shallow
creep (46); however, creep has to be pervasive
to potentially suppress an incoming dynamic
rupture. Observations spanning all phases of
the earthquake cycle are needed to constrain
the velocity- and depth-dependent frictional
properties of active faults (47, 48). Shallow
creep might be responsible for a substantial
reduction in the amplitude of coseismic slip in
the top few kilometers of the upper crust (Fig.
our modeling shows that the pre-event stress
heterogeneities, dynamic and static redistribu-
tion of stress, and the geometry of the faults may
control these diverse rupture characteristics.
The Kahramanmaraş doublet ruptured mul-
tiple faults in distinct slip episodes, likely in-
volving complex stress-triggering processes
across different temporal and spatial scales.
Such processes resulted in the increased rupture
length and seismic moment of the Turkey earth-
quake doublet, and a substantially larger de-
structive potential compared with the “typical”
Mw ~7 historical earthquakes in the region (23).
Such a variability might be interpreted in terms
of the supercycle model (51). By using inte-
grated methods that combine near- and far-
,
Geophys. J. Int. 230, 50–69 (2022).
24. T. Y. Duman, Ö. Emre, Spec. Publ. Geol. Soc. Lond. 372, 495–529
(2013).
25. A. Koç, N. Kaymakcı, J. Geodyn. 65, 292–307 (2013).
26. Materials and methods are available as supplementary materials.
27. Z. Jin, Y. Fialko, Bull. Seismol. Soc. Am. 110, 1660–1679 (2020).
28. A. Lomax, Precise, NLL-SSST-coherence hypocenter catalog
for the 2023 Mw 7.8 and Mw 7.6 SE Turkey earthquake
sequence. Zenodo (2023); https://doi.org/10.5281/
zenodo.8089273.
29. A. Doğru, F. Bulut, C. Yaltırak, B. Aktuğ, Geophys. J. Int. 224,
389–400 (2021).
30. Z. Jia et al., Earth Planet. Sci. Lett. 531, 115997 (2020).
31. Z. Jia, Z. Zhan, H. Kanamori, Geophys. Res. Lett. 49,
e2021GL097104 (2022).
32. M. Ishii, P. M. Shearer, H. Houston, J. E. Vidale, Nature 435,
933–936 (2005).
33. Z. Jia, X. Wang, Z. Zhan, Geophys. Res. Lett. 47,
e2020GL089802 (2020).
34. G. Ekström, M. Nettles, A. Dziewoński, Phys. Earth Planet. Inter.
200-201, 1–9 (2012).

Jia et al., Science 381, 985–990 (2023) 1 September 2023 5 of 6

RES EARCH | R E S E A R C H A R T I C L E
35. C. Liang, J. P. Ampuero, D. Pino Muñoz, Geophys. Res. Lett. 49,
e2022GL099749 (2022).
36. E. M. Dunham, J. Geophys. Res. Solid Earth 112, JB004717
(2007).
37. E. Tinti et al., Earth Planet. Sci. Lett. 576, 117237 (2021).
38. N. Kame, J. R. Rice, R. Dmowska, J. Geophys. Res. Solid Earth
108, JB002189 (2003).
39. S. Fliss, H. S. Bhat, R. Dmowska, J. R. Rice, J. Geophys. Res.
Solid Earth 110, JB003368 (2005).
40. G. King, J. Nabecaronlek, Science 228, 984–987 (1985).
41. D. Andrews, J. Geophys. Res. 94, 9389–9397 (1989).
42. B. Duan, D. D. Oglesby, J. Geophys. Res. Solid Earth 112,
JB004443 (2007).
43. A. Rosakis, M. Abdelmeguid, A. Elbanna, Evidence of early
supershear transition in the Mw 7.8 Kahramanmaraş
earthquake from near-field records. arXiv:2302.07214
[physics.geo-ph] (2023).
44. S. S. Nalbant, J. McCloskey, S. Steacy, A. A. Barka, Earth Planet.
Sci. Lett. 195, 291–298 (2002).
45. Y. Fialko, J. Geophys. Res. Solid Earth 126, e2021JB022000
(2021).
46. Z. Cakir et al., Earth Planet. Sci. Lett. 608, 118085 (2023).
47. Y. Kaneko, Y. Fialko, D. T. Sandwell, X. Tong, M. Furuya,
J. Geophys. Res. Solid Earth 118, 316–331 (2013).
48. E. O. Lindsey, Y. Fialko, J. Geophys. Res. Solid Earth 121,
1097–1113 (2016).
49. Y. Fialko, D. Sandwell, M. Simons, P. Rosen, Nature 435,
295–299 (2005).
50. Z. Jin, Y. Fialko, Geophys. Res. Lett. 48, e2021GL095213 (2021).
51. L. Dal Zilio, J.-P. Ampuero, Commun. Earth Environ. 4, 71 (2023).
52. L. Luzi et al., Bull. Earthquake Eng. 18, 5533–5551 (2020).
53. Disaster and Emergency Management Authority (AFAD)
(1973); https://deprem.afad.gov.tr/event-catalog.
54. T. Taufiqurrahman et al., Nature 618, 308–315 (2023).
55. S. Akkar, M. A. Sandıkkaya, J. J. Bommer, Bull. Earthquake Eng.
12, 359–387 (2014).
56. Z. Jia et al, 2023_Turkey_doublet_Data_Model. Zenodo (2023);
https://doi.org/10.5281/zenodo.8128343.
Jia et al., Science 381, 985–990 (2023) 1 September 2023
ACKN OWLED GMEN TS
We thank two anonymous reviewers for constructive comments.
The facilities of IRIS Data Services, and specifically the IRIS Data
Management Center, were used for access to waveforms, related
metadata, and/or derived products used in this study. The GNSS
data are provided by the General Directorate of Land Registry and
Cadastre and the General Directorate of Mapping, Turkey. IRIS
Data Services are funded through the Seismological Facilities for
the Advancement of Geoscience (SAGE) Award of the National
Science Foundation under Cooperative Support Agreement EAR-
1851048. We thank the Engineering Strong-Motion (ESM) Database
by ORFEUS (52) for providing access to the data from the Turkish
National Strong Motion Network, which is operated by Turkey’s
Disaster and Emergency Management Authority (AFAD). The
computations in this study were supported by the supercomputer
SuperMUC-NG. Funding: Funding for this work was provided by
the National Science Foundation (EAR-2123529, EAR-1841273
to Y.F., EAR-2022441, EAR-2143413 to W.F., and EAR-2121568 to
A.-A.G), the National Aeronautics and Space Administration
(80NSSC22K0506 to Y.F., 80NSSC20K0495 to A.A.G.), the United
States Geological Survey (G22AP00011 to P.S. and W.F.), and the
Cecil and Ida Green Foundation. M.M., T.U., and A.-A.G. were
supported by the European Union’s Horizon 2020 Research and
Innovation Programme (TEAR grant 852992) and Horizon Europe
(ChEESE-2P grant 101093038, DT-GEO grant 101058129, and
Geo-INQUIRE grant 101058518). J.R. was supported by the National
Science Foundation Graduate Research Fellowship Program
(grant DGE-2038238). A.-A.G. gratefully acknowledges the Gauss
Centre for Supercomputing e.V. (https://www.gauss-centre.eu) for
providing computing time on the GCS Supercomputer SuperMUC-NG
at Leibniz Supercomputing Centre (https://www.lrz.de), in project
pn49ha. Author contributions: Z. Jia led the study, performed the
subevent and kinematic slip inversions, and wrote the initial draft
of the manuscript. Z. Jin, X.Z., and Y.F. processed the InSAR
data, quantified the fault geometry, and performed static slip
inversion. M.M., T.U., and A.-A.G. performed the dynamic rupture
simulations and the Coulomb failure stress analysis. W.F. performed
back projection analysis and coordinated the collaboration. Y.F.
conducted static Coulomb stress calculations. X.Z. contributed to the
fault geometry determination and figure design. J.R., T.U., and A.-A.G.
conducted the ground motion analysis. F.B. and A.G. processed
the GNSS data. W.F., P.S., A.-A.G., and Y.F. supervised the modeling
and contributed to interpretation of the results. A.A.G., Y.F., and
W.F. provided access to computational resources. All authors
discussed the results and contributed to writing and revising the
manuscript. Competing interests: The authors declare no competing
interests. Data and materials availability: The Sentinel-1 SAR
data are provided by the European Space Agency (ESA) and mirrored
at the Alaska SAR Facility (ASF). The ALOS-2 PALSAR data are owned
by the Japanese Space Agency (JAXA) and provided to Y.F. under
a Research User Agreement. The global seismic data are publicly
available from the IRIS-DMC. The regional strong motion data are
publicly available from the EMS Database by ORFEUS. The geodetic
and seismic data, the data-driven slip models and subevent
models, and the data required to reproduce the dynamic rupture
earthquake sequence scenarios can be downloaded from (56). We
provide detailed README files summarizing the data and data
formats provided. License information: Copyright © 2023 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
p
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.adi0685
Materials and Methods
Figs. S1 to S28
Tables S1 and S2
References (57–112)
Movies S1 to S3
g
Submitted 1 April 2023; accepted 19 July 2023
Published online 3 August 2023
10.1126/science.adi0685
y
y g
p
,
6 of 6
RES EARCH

EVOLUTIONARY BIOLOGY jected to overexploitation on a global scale,

and determining the magnitude of that de-
Wild pedigrees inform mutation rates and historic pletion is critical for understanding the current
status of whale populations and the carrying
abundance in baleen whales capacity of the historic oceans. The values of
mPHY that were used to estimate preexploita-
Marcos Suárez-Menéndez1*†, Martine Bérubé1,2†, Fabrício Furni1, Vania E. Rivera-León1, tion abundance, based on current levels of ge-
Mads-Peter Heide-Jørgensen3, Finn Larsen4, Richard Sears5, Christian Ramp5,6, netic diversity in some whale species (18, 19),
Britas Klemens Eriksson1, Rampal S. Etienne1, Jooke Robbins2‡, Per J. Palsbøll1,2*‡ yielded abundance estimates that were 4 to
10 times higher than results obtained by non-
Phylogeny-based estimates suggesting a low germline mutation rate (m) in baleen whales have influenced genetic approaches (20). The high genetic es-
research ranging from assessments of whaling impacts to evolutionary cancer biology. We estimated m directly timates of preexploitation whale abundance
from pedigrees in four baleen whale species for both the mitochondrial control region and nuclear genome. imply less population recovery than expected
The results suggest values higher than those obtained through phylogeny-based estimates and similar to and historic oceans that supported a much
pedigree-based values for primates and toothed whales. Applying our estimate of m reduces previous genetic- larger biomass of whales and prey.
based estimates of preexploitation whale abundance by 86% and suggests that m cannot explain low
Nuclear mutation rate
cancer rates in gigantic mammals. Our study shows that it is feasible to estimate m directly from pedigrees in
natural populations, with wide-ranging implications for ecological and evolutionary research. Estimates of mPED in the nuclear genome have
so far been obtained from 84 vertebrate spe-

A
cies, including 36 mammals (11). Most non-

daptive evolution is ultimately driven by
the emergence of novel genotypes, which
in turn are a result of de novo germline
mutations and recombination. The rate
of de novo mutations (or m, the proba-
bility of a nucleotide substitution per site per
records and unknown mutation model pa-
rameters), aspects that also vary among spe-
cies groups (5–10). However, m can now also be
inferred directly from the pedigrees of individ-
uals (mPED) through inexpensive whole-genome
sequencing (11). Direct estimation of mPED re-
p
human mammal estimates of mPED have been
based on known sire-dam-offspring trios (hence-
forth referred to as trios) from captive pop-
ulations. In contrast, estimates of mPED in wild
populations are rare and confined to isolated,
easily sampled groups of parents and offspring

generation) varies considerably among taxo-
nomic groups. Multiple causes have been pro-
posed to explain observed differences in m
(1), such as selection on m itself or differences
in physiological and cellular processes (e.g.,
lies on far fewer assumptions and depends not
on fossil dating or access to ancient samples
but rather on contemporary parameters, such
as generation time. Perhaps most importantly,
mPED estimates yield generational values of m
g
in only five species (table S1). Bergeron and
colleagues (11) noted that estimates of mPED in
captive animals are likely not representative
of wild conspecific populations, which differ
in terms of generation times as well as the

generation time, metabolic rates, or DNA re-
pair mechanisms). The germline mutation rate
is also a central parameter in evolution and
population genetic inference and is neces-
sary to convert genetic estimates into intuitive
quantities, such as years and abundance. Most
estimates of m are inferred from the number of
observed substitutions among homologous
DNA sequences in phylogenies (mPHY) with
dated branching points (e.g., using fossils or
that are directly comparable among species,
thus providing a means to test current hypothe-
ses about the cause(s) of differences in m and
to revisit key findings and hypotheses based
on mPHY.
Baleen whales (Mysticeti, Cetacea) include
the largest and longest-lived mammals and
are at the center of ongoing questions that
pertain to m. Their gigantic body sizes and rel-
atively low metabolic rates (compared with
y
nature and level of natural selection. Here, we
show that estimating mPED in wild, difficult-to-
sample populations is feasible and straight-
forward. Toward this end, we identified trios
among baleen whale samples of unknown an-
cestries (table S2) and sequenced the genome
of each member in multiple trios, from which
we estimated mPED.
We used genetic parentage exclusion anal-
ysis (based on microsatellite genotypes; table

carbon dating). In a few species, m has been
inferred from comparisons of ancient and
contemporary DNA sequences from samples
across a wide temporal range (mANC), typically
smaller-bodied mammals) are thought to re-
sult in less intracellular DNA damage and
thus a lower expected m (12). Indeed, prior es-
timates of mPHY in both the nuclear and mito-
y g
S3) with molecular sex determination to iden-
tify trios among skin samples collected from
free-ranging North Atlantic blue (Balaenoptera
musculus), fin (Balaenoptera physalus), bow-

p
yielding higher estimates than mPHY (2–4). The chondrial (mt) genome of baleen whales were head (Balaena mysticetus), and humpback
observed discrepancies among different estima- lower than similar estimates for smaller-bodied (Megaptera novaeangliae) whales. Despite sam-
tions of m may reflect either real biological dif- mammals with similar generation times, such pling a relatively low proportion of the over-

ferences or experimental issues. Both approaches
are subject to multiple sources of error (e.g.,
incorrect mutation model, postmortem DNA
damage, and sampling bias) and uncertainty
(e.g., incomplete and/or poorly dated fossil
1
Groningen Institute for Evolutionary Life Sciences, University
of Groningen, Groningen, Netherlands. 2Center for Coastal
Studies, Provincetown, MA, USA. 3Greenland Institute of
Natural Resources, Copenhagen, Denmark. 4National
Institute of Aquatic Resources, Kongens Lyngby, Denmark.
5 as more plausible explanations for lower can-
Mingan Island Cetacean Study Inc., St. Lambert, Quebec,
Canada. 6Scottish Oceans Institute, University of St. Andrews,
St. Andrews, UK.
*Corresponding author. Email: m.suarez.menendez@rug.nl
(M.S.-M.); p.j.palsboll@gmail.com (P.J.P.)
†These authors contributed equally to this work.
‡These authors contributed equally to this work.
as primates (13). This observation has been
the subject of considerable debate regarding
the cause and consequences of the lower mPHY
(13, 14). The lower mPHY observed in baleen
whales has been proposed as one of several
possible explanations for low cancer rates in
baleen whales, part of a phenomenon known
as Peto’s paradox (15, 16). Later work has
proposed other specific mechanisms, such as
duplication of cancer suppression genes (14, 17),
cer rates in baleen whales and other large mam-
mals, but the role of the lower mPHY remains
untested.
Another important implication of m relates
to historical whaling. Baleen whales were sub-
,
all population in some species (table S2), we
were able to identify trios. Available a priori
pedigree information (only for some humpback
whales) was not used in trio identification. We
subsequently estimated mPED in baleen whales
from the de novo germline mutations detected
in the whole-genome sequences in eight trios
(21 genomes in total): one each in blue, fin,
and bowhead whales, and five in humpback
whales.
The complete genome was sequenced in all
individuals to an average depth of ~32× and
aligned against the blue whale genome (21). In
all eight trios, the genome sequences con-
firmed the relationships inferred from the
microsatellite data. An initial set of de novo

Suárez-Menéndez et al., Science 381, 990–995 (2023) 1 September 2023 1 of 5

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

Fig. 1. Summary of the

estimation of mPED in
the nuclear genome per
sire-dam-offspring trio
in humpback, blue, fin,
and bowhead whales. M. novaeangliae
Squares and circles
denote males and females,
respectively, and contain
the sample identification M
MnnS
S1
1 Mn
MnDD1
1 Mn
MnSS2
2 Mn
MnDD2
233 Mn
MnSS3
344 Mn
MnDD4
4
number (top), the first *1980 *1980 1979 1990 1989 *1985
~ 30x ~ 30x ~ 29x ~ 29x ~ 29x ~ 29x
sighting year (middle), 8 2 13 2 6 11 8 2
and the mean genome
sequence coverage 31 DNMs 30 DNMs 26 DNMs 24 DNMs
(bottom). The first sighting
Mn
MnCC1
1 Mn
MnCC2
2 Mn
MnCC3
3 Mn
MnCC4
455 Mn
MnSS5
5
was known to be the birth *1982
2000 2008 2004 1998
year or, where marked ~ 31x ~ 30x ~ 30x ~ 29x ~ 29x
3 13
with an asterisk, the year
first sighted before sampling.
A question mark denotes µPED=1.16x10-8 µPED =1.20x10-8 µPED=1.03x10-8 µPED =0.95x10-8 32 DNMs
(0.76 - 1.56x10 -8) (0.78 - 1.62x10 -8) (0.64 - 1.42x10 -8) (0.58 - 1.33x10 -8)

p
individuals with no previous
sightings. Values along
M nC 5
the vertical lines are the total
2011
number of detected de novo ~ 30x
mutations (DNMs, in bold)
as well as the number B. musculus µPED =1.27x10-8

g
of de novo mutations (0.84 - 1.70x10 -8)
originating from the sire
(number to the left of the Bm
B mS
S Bm
B mD
D
? ? B. physalus
vertical line) and dam
~ 33x ~ 32x
(number to the right of 17 3

y
the vertical line). The
37 DNMs Bp
B pS
S Bp
B pD
D
parameter mPED denotes
the estimate obtained
? ? B. mysticetus
Bm
B mC
C ~29x ~ 45x
from each trio, with the 11 3
?
95% confidence intervals ~ 32x 30 DNMs Bw
B wS
S Bw
B wD
D
in parentheses. [Whale
? ?
illustrations by Frédérique
µPED =1.24x10-8 Bp
B pC
C ~33x
9 2
~34x
Lucas (relative sizes are ?
(0.85-1.63x10 -8)
not to scale)] ~ 49x 30 DNMs

y g
µPED =0.88x10-8 Bw
B wC
C
(0.57 - 1.20x10 -8) ?
~ 30x

p
µPED =1.20x10-8
(0.78 – 1.61x10 -8)

mutations was identified in each offspring’s
genome by means of a modified version of the
bioinformatic pipeline by Bergeron et al. (22).
In total, 240 de novo mutations were identi-
fied in the eight offspring, among which 113
were phased, of which 90 (80%) were of pa-
ternal origin (tables S4 and S5). De novo mu-
tations transmitted across two generations in
humpback whales (Fig. 1) segregated accord-
ing to Mendelian expectations. No de novo
mutations were shared between half-siblings
(Fig. 1), suggesting that most de novo muta-
tions arose during meiosis. The majority of
the de novo mutations comprised substitu-
tions from a strong to a weak base, with no
functional impacts (fig. S6). In this study, age
was only known for humpback whale individ-
uals sighted earlier, either as a calf during their
birth year (i.e., absolute age) or in a later year
as adults (i.e., minimum age). However, the
humpback whale mPED estimates suggested a
positive correlation between the number of
detected paternal de novo mutations and pa-
ternal age at conception, in agreement with
earlier studies (11, 22–24).
De novo mutations were callable in ~60% of
the autosomal genome, yielding estimates of
mPED at 1.12 × 10−8 [95% confidence interval
,
(CI): 0.94 × 10−8 to 1.30 × 10−8] in the hump-
back whales and 1.11 × 10−8 (95% CI: 0.97 × 10−8
to 1.25 × 10−8) in all baleen whales combined
(Fig. 1 and table S4). Paternal age has a large
impact on the number of de novo mutations in
offspring and consequently on individual trio
estimates of mPED (11). Humpback whale sires
were selected in a uniform manner (Fig. 1) cen-
tered around the generation time in humpback
whales [21.5 years (25)]. Although ages were
unknown for the other baleen whale species,
the overall baleen whale and humpback whale
estimates of mPED were very similar to each
other. They were also similar to estimates of

Suárez-Menéndez et al., Science 381, 990–995 (2023) 1 September 2023 2 of 5

RES EARCH | R E S E A R C H A R T I C L E
A B
2.0×10–8
1.5×10–8
Mutation rate
1.0×10–8
5.0×10–9
Chimpanzee Human Common
bottlenose
dolphin
Fig. 2. Comparison of estimates of nuclear mPED and preexploitation humpback
whale abundance. (A) Estimates of nuclear mPED for baleen whales (this study), toothed
whales, chimpanzees, and humans (table S6) ordered by body mass. (B) Genetic
mPED in humans, large primates, and toothed
whales (11, 23, 26) (Fig. 2A and table S6), spe-
cies with similar generation times and pater-
nal ages. Our results therefore contradict the
current notion that m is lower in baleen whales
than in smaller-bodied mammals with similar
generation times. Given the crucial role of pa-
ternal age in terms of the number of de novo
mutations in the offspring, we note that there
is currently only one method to determine chro-
nological age from tissue samples—DNA meth-
ylation (27)—but the development costs are
high, and samples from known-age individ-
uals are required for calibration. However, a
simpler and more straightforward way to ob-
tain an estimate of mPED, which reflects the
average in wild populations, would be to con-
secutively add estimates of mPED from addi-
tional trios until stationarity of mPED is reached,
suggesting that the estimate of mPED has
converged on the population mean. It is worth
noting that population generation times and
reproductive maturity are nonstatic entities that
change with the underlying population dynam-
ics. Most contemporary baleen whale populations
are recovering from historic overexploitation,
and thus the demographic structure may still
be skewed toward younger individuals when
compared with stable populations. In addition,
current generation time estimates for baleen
whales are exclusively informed by female re-
productive histories (25), ignoring the sires from
which ~80% of de novo mutations were inherited.
Thus, if anything, our estimate of mPED could be
biased downward relative to stable populations.
Previously reported low mPHY estimates seemed
to be in agreement with the hypothesized ef-
fects of a reduced metabolic rate in gigantic
mammals (500× to 2000× by weight relative
Suárez-Menéndez et al., Science 381, 990–995 (2023)
Orca Humpback Bowhead Fin Blue
whale whale whale whale
(dark gray and black) and nongenetic (light gray) estimates of preexploitation and
contemporary abundance of North Atlantic humpback whales; genetic estimates based
on mt mPHY scaled to a generation time at 18 years used in the original analysis.
to humans), resulting in lower levels of intracel-
lular free radicals and consequently less DNA
damage and mutations (13). However, our esti-
mate of mPED in baleen whales is similar to that
observed in primates, which have much higher
cancer rates. Barring the possibility that DNA
damage repair mechanisms differ substantially
between somatic and germline cells in baleen
whales (28), our result thus negates the notion
that the lower observed cancer rates in gigantic
mammals are due solely to lower mutation rates,
i.e., the cause of Peto’s paradox (14, 29).
Mitochondrial mutation rate
In addition to the nucleus, mammal cells also
contain maternally transmitted mitochon-
dria, which have their own genomes. The fast-
evolving control region in the mt genome is
one of the most common genetic markers em-
ployed in population genetics. Contemporary
levels of genetic diversity in baleen whales have
been used to infer unexpectedly high preex-
ploitation abundances of North Atlantic baleen
whales (18). These genetic assessments of the
impacts of whaling were based on mPHY. The
much smaller size of the mt genome (~16,500
base pairs) implies few, if any, de novo muta-
tions in each transmission, and thus the above
approach to infer mPED is infeasible for the mt
genome in most species. However, mPED for the
mt genome can be inferred from the frequency
of heteroplasmy and the change in the ratio of
heteroplasmy between females to their off-
spring (3, 30), an approach so far applied to a
few model species (31–33) and only a single
wild population of penguins (3).
We assessed mt heteroplasmy using data
from long-term research on North Atlantic
humpback whales in the Gulf of Maine since the
1 September 2023
–6
250,000 0.36×10
North Atlantic humpback whale
200,000
abundance
150,000 0.93×10–6
100,000
50,000
4.3×10 –6
Phylogenetic Pedigree Historical 1992-93
p
1970s. Among 850 sampled humpback whales,
g
we identified 141 distinct maternal lineages,
characterized by 20 different mt haplotypes.
Each maternal lineage contained between 2
and 49 whales and spanned from one to four
generations. De novo mutations in the mt
y
control region were evident as heteroplasmy,
detected initially by Sanger sequencing (34)
and subsequently confirmed by massive par-
allel sequencing (Fig. 3). Heteroplasmy was
detected in the oldest known female in nine
maternal lineages and in a minimum of one
descendant, confirming the transmission of
heteroplasmy along the maternal germline. In
total, heteroplasmy was detected in one, three,
and five lineages at nucleotide positions 235,
y g
258, and 82, respectively (table S8). The hetero-
plasmy detected at position 82 was identified
in three different mt haplotypes, supporting
the well-known presence of mutational hot-
p
spots with elevated mutation rates in the mt
control region (19, 35). The heteroplasmy ob-
served in seven lineages gave rise to mt control
,
region haplotypes already present in the pop-
ulation (table S8). These de novo mutations
could go unaccounted for in estimates of mPHY
(19). The mt control region haplotypes result-
ing from the heteroplasmy in the remaining
two lineages were novel in this population.
We inferred mPED from the generational
change in the frequency of each variant at the
heteroplasmic site, as per Hendy and colleagues
(30) (Fig. 3), from massive parallel sequenc-
ing data with an average read depth of 6900×.
In total, 47 maternal transmissions of a germ-
line mt heteroplasmy yielded a median esti-
mate of the number of segregating units at
10.4 [interquartile range (IQR): 5 to 28.9; fig.
S7 and table S9] during the oocyte bottleneck.
3 of 5
RES EARCH | R E S E A R C H A R T I C L E
The number of segregating units was very sim-
ilar to the range of values reported in other
mammals (3, 31–33, 36). The final estimate of
mPED for the mt control region was 4.3 × 10−6
(IQR: 1.55 × 10−6 to 8.85 × 10−6; table S6), sim-
ilar to comparable estimates of mPED for the
same mt region in humans (3.8 × 10−6; IQR:
1.18 × 10−6 to 1.16 × 10−5) (31). The single mANC
estimate from baleen whales [bowhead whale
(4)] was 7 × 10−6 to 1.05 × 10−5 (depending on
the assumed generation time; table S6), but a
subsequent assessment by the authors deemed
the estimate unreliable (37).
We assessed the effect of our estimated val-
ue of mPED in the mt control region on previous
estimates of preexploitation abundance of
North Atlantic humpback whales inferred from
contemporary levels of genetic diversity in the
mt control region (18, 19, 38). Insights into the
preexploitation abundance of baleen whales
are not only important when assessing current
levels of recovery but also provide fundamen-
tal insights into the overall state of the historic
oceans and the biomass that the oceans sup-
ported before overexploitation. To evaluate
the effect of using our estimate of mPED, we
replicated the approach and the parameter
values used in the previous studies (18, 19),
with the exception of mPED. The resulting esti-
mate of the effective female population size
was 2561 (IQR: 1245 to 7128), yielding an over-
all effective population size of 5122, which, in
turn, translated to an overall census abun-
dance of 20,488 humpback whales (IQR: 9964
to 57,030). This newly calculated estimate is
86% lower than the lowest mPHY estimate
based on mt control region sequence diversity
L115
Fig. 3. Changes in the frequencies of a mitochondrial point heteroplasmy
after the transmission bottleneck during oocyte development. Heteroplasmy
present in a humpback whale and three of her offspring detected by Sanger
sequencing (chromatogram peak height, left-hand square) and massive parallel
Suárez-Menéndez et al., Science 381, 990–995 (2023)
[150,000, credited to (19) by (38)]. The abun-
dance based on our estimate of mPED is sim-
ilar to nongenetic estimates of preexploitation
humpback whale abundance [17,151 to 22,647
(20, 38)] (Fig. 2B). Our result demonstrates
how a downward-biased estimate of m, inferred
from mPHY, leads to an overestimate of preex-
ploitation abundance. The revised mPED-based
estimate is much closer to the most recent
estimate of the abundance of North Atlantic
humpback whales [10,752 (39)]. Accordingly,
our revised estimate implies that the pre-
exploitation oceans likely supported a lower
biomass of baleen whales than suggested by
the earlier genetic studies. However, estimat-
ing historic whale abundance in this manner
is subject to several uncertainties, such as mi-
gration, population size changes, and generation
time (5). In its simplest form, the contempo-
rary level of genetic diversity represents the
temporal harmonic mean of the effective pop-
ulation size and thus not necessarily the actual
abundance at a specific, narrowly defined time
period—in this case, immediately before the
onset of whaling (5). Therefore, we also applied
a linkage disequilibrium–based approach to
the data obtained from whole genomes of the
unrelated individual humpback whales to es-
timate the effective population size before the
start of commercial whaling (30 generations
ago). This estimation yielded an average effec-
tive population size of 5700 individuals [stan-
dard deviation (SD): 685], which was stable at
an average of 5263 (SD: 1104) between 30 and
200 generations ago and thus very similar to
our revised estimate based on mPED estimated
for the mt control region. However, and per-
L115_1
MtDNA
bottleneck
L115_2
L115_3
sequencing (read frequencies, right-hand square). The relative frequencies of the
two mt haplotypes were reversed in offspring 2 compared with the mother. In
offspring 3, the heteroplasmy was undetectable by Sanger sequencing but detected
among the reads. [Whale illustrations by Frédérique Lucas]
1 September 2023
haps most importantly, our reanalysis demon-
strates how sensitive these inferences are to
changes in the underlying parameter values
(5), which also applies to nongenetic estimates
of historic abundance (20).
Conclusions
The rate and nature of de novo mutations are at
the center of evolution itself and key to many
ecological and evolutionary processes. Differ-
ences in m among taxonomic groups have fos-
tered an array of intriguing hypotheses aimed
at understanding the underlying causes (1).
However, the technical caveats in the estima-
tion of mPHY and mANC render direct compar-
isons among different species and studies
difficult. In contrast, estimates of mPED are di-
rectly comparable among species, and conse-
quently, this approach holds great potential to
facilitate and improve such comparisons. Our
p
study expands on recent reports of pedigree-
based estimation of m in two crucial aspects.
First, we show that it is feasible to estimate
mPED in wild populations in a cost-effective
manner by conducting an initial screening of
samples using inexpensive microsatellite data.
g
We show that trios can be reliably detected
and confirmed using whole-genome sequenc-
ing in the absence of known pedigrees, even in
notoriously difficult-to-study large mammals
with exceptionally wide ranges. Second, our
y
estimates of mPED in the gigantic baleen whales
were similar to estimates of mPED in the com-
paratively smaller-bodied primates and toothed
whales, both for the nuclear and mt genome,
thus negating a major effect of gigantism on m.
We note that mPED estimates are not necessarily
y g
p
4 of 5
RES EARCH | R E S E A R C H A R T I C L E
directly applicable as a substitute for m in all
instances. This is particularly the case when
results are based on analytical methods that
integrate over extensive time periods, when
populations are subject to changes in abun-
dance and connectivity, selection, and multi-
ple substitutions at the same nucleotide sites.
As an example, in some cases, robust infer-
ences require applying a mutation model that
accounts for multiple substitutions, which
may result in a “time dependence” of mutation
rates (40). However, as sample sizes increase
in studies aiming to detect de novo mutations
in pedigrees, a much more detailed insight will
be gained into both the nature and distribu-
tion of de novo mutations across the genome.
In turn, these insights will enable refinement
of current mutation models and thereby aid in
improving evolutionary and population ge-
netic inference methods in general. We also
note that while the generational mutation
rate might be similar among larger mammals,
generation times can differ substantially among
species, resulting in lower annual mutation
rates in long-lived species. This generation
time effect comes into play when inferences
based on mPED require conversion into years
(e.g., historic abundances and population di-
vergence times). Given the relative ease with
which trios can be identified in large sample
sizes by inexpensive microsatellite genotyp-
ing and given the low costs of whole-genome
sequencing, robust estimates of mPED are likely
to become available across a wide range of
biodiversity, facilitating novel insights into
many different fundamental and applied as-
pects that relate to m.
RE FE RENCES AND N OT ES
1. L. Bromham, Philos. Trans. R. Soc. Lond. Ser. B 366,
2503–2513 (2011).
2. D. M. Lambert et al., Science 295, 2270–2273 (2002).
3. C. D. Millar et al., PLOS Genet. 4, e1000209 (2008).
4. S. Y. W. Ho, S.-O. Kolokotronis, R. G. Allaby, Biol. Lett. 3,
702–705 (2007).
5. P. J. Palsbøll, M. Zachariah Peery, M. T. Olsen, S. R. Beissinger,
M. Bérubé, Mol. Ecol. 22, 22–40 (2013).
Suárez-Menéndez et al., Science 381, 990–995 (2023)
6. B. C. Emerson, M. J. Hickerson, Mol. Ecol. 24, 702–709
(2015).
7. A. D. Foote, M. Hofreiter, P. A. Morin, Ann. Anat. 194, 112–120
(2012).
8. J. P. Herrera, L. M. Dávalos, Syst. Biol. 65, 772–791
(2016).
9. J. Felsenstein, J. Mol. Evol. 17, 368–376 (1981).
10. G. P. Tiley, J. W. Poelstra, M. Dos Reis, Z. Yang, A. D. Yoder,
Trends Genet. 36, 845–856 (2020).
11. L. A. Bergeron et al., Nature 615, 285–291 (2023).
12. A. P. Martin, S. R. Palumbi, Proc. Natl. Acad. Sci. U.S.A. 90,
4087–4091 (1993).
13. J. A. Jackson et al., Mol. Biol. Evol. 26, 2427–2440
(2009).
14. M. Tollis et al., Mol. Biol. Evol. 36, 1746–1763 (2019).
15. R. Peto, F. J. Roe, P. N. Lee, L. Levy, J. Clack, Br. J. Cancer 32,
411–426 (1975).
16. A. M. Leroi, V. Koufopanou, A. Burt, Nat. Rev. Cancer 3,
226–231 (2003).
17. D. Tejada-Martinez, J. P. de Magalhães, J. C. Opazo, Proc. Biol.
Sci. 288, 20202592 (2021).
18. J. Roman, S. R. Palumbi, Science 301, 508–510 (2003).
19. S. E. Alter, S. R. Palumbi, J. Mol. Evol. 68, 97–111
(2009).
20. A. E. Punt, N. A. Friday, T. D. Smith, J. Cetacean Res. Manag. 8,
145–159 (2006).
21. Y. V. Bukhman et al., Research Square [Preprint] (2022).
https://doi.org/10.21203/rs.3.rs-1910240/v1.
22. L. A. Bergeron et al., Gigascience 10, giab029 (2021).
23. S. Besenbacher, C. Hvilsom, T. Marques-Bonet, T. Mailund,
M. H. Schierup, Nat. Ecol. Evol. 3, 286–292 (2019).
24. H. Jónsson et al., Nature 549, 519–522 (2017).
25. B. L. Taylor, S. J. Chivers, J. Larese, W. F. Perrin, “Generation
length and percent mature estimates for IUCN assessments
of cetaceans,” Administrative Report LJ-07-01, NOAA
National Marine Fisheries Service, Southwest Fisheries
Science Center (2007).
26. M. D. Kessler et al., Proc. Natl. Acad. Sci. U.S.A. 117,
2560–2569 (2020).
27. C. Z. Li et al., bioRxiv 2021.05.16.444078 [Preprint] (2021).
https://doi.org/10.1101/2021.05.16.444078.
28. A. Cagan et al., Nature 604, 517–524 (2022).
29. L. Nunney, Evol. Appl. 13, 1581–1592 (2020).
30. M. D. Hendy, M. D. Woodhams, A. Dodd, Biol. Lett. 5, 397–400
(2009).
31. B. Rebolledo-Jaramillo et al., Proc. Natl. Acad. Sci. U.S.A. 111,
15474–15479 (2014).
32. J. N. Wolff, D. J. White, M. Woodhams, H. E. White,
N. J. Gemmell, PLOS ONE 6, e20522 (2011).
33. B. Arbeithuber et al., Proc. Natl. Acad. Sci. U.S.A. 119,
e2118740119 (2022).
34. F. Sanger, S. Nicklen, A. R. Coulson, Proc. Natl. Acad. Sci. U.S.A.
74, 5463–5467 (1977).
35. M. Stoneking, Am. J. Hum. Genet. 67, 1029–1032
(2000).
36. C. Haag-Liautard et al., PLOS Biol. 6, e204 (2008).
37. S. Y. W. Ho et al., Syst. Biol. 60, 366–375 (2011).
38. K. Ruegg et al., Conserv. Genet. 14, 103–114 (2013).
1 September 2023
39. P. Stevick et al., Mar. Ecol. Prog. Ser. 258, 263–273
(2003).
40. S. Y. W. Ho, M. J. Phillips, A. Cooper, A. J. Drummond, Mol. Biol.
Evol. 22, 1561–1568 (2005).
41. M. Suárez-Menéndez et al., Dryad (2023); https://doi.org/10.
5061/dryad.9kd51c5pq.
AC KNOWLED GME NTS
We are indebted to the many staff who undertook the field and
laboratory work over the years and to the long-term population
research programs that made this work possible. We thank
S. Beissinger and O. Gaggiotti, as well as two anonymous reviewers
for their valuable suggestions. The Center for Information Technology
of the University of Groningen provided access and support to its
Peregrine High Performance Computing cluster. We are grateful to
F. Lucas for providing whale drawings. Funding: M.S.-M. was
supported by a doctorate fellowship from the University of
Groningen. P.J.P. and M.B. acknowledge support over the years by
the University of Copenhagen, University of California, Berkeley,
and Stockholm University. Additional funding was provided by
the Aage V. Jensen Foundation (P.J.P.), US National Marine Fisheries
Service (P.J.P. and M.B.), the International Whaling Commission
Scientific Committee (M.B. and P.J.P.), WWF-DK (P.J.P.), the
Commission for Scientific Research in Greenland (P.J.P.), the
Greenland Home Rule (P.J.P.), the European Union’s Horizon 2020
Research and Innovation Programme under Marie Skłodowska-Curie
p
grant no. 813383 (F.F. and P.J.P.), and Consejo Nacional de Ciencia y
Tecnología (V.E.R.-L.). Author contributions: M.S.-M., M.B.,
V.E.R.-L., J.R., and P.J.P. conceived of and designed the study. J.R.,
R.S., C.R., and M.-P.H.-J. provided samples and field data. M.B.
and M.S.-M. conducted the laboratory work. M.S.-M.,V.E.R.-L., and
F.F. performed the bioinformatic and data analyses under the
supervision of P.J.P. M.S.-M., M.B., and P.J.P. interpreted the results,
with contributions from all authors. M.S.-M. and P.J.P. wrote
g
the manuscript, with input from all authors. All authors approved
the final version of the manuscript. Competing interests: The
authors declare no competing interests. Data and materials
availability: Raw reads for whole genomes and mitogenomes are
available at the National Center for Biotechnology Information under
BioProject ID PRJNA990679. The rest of the data and code are
y
available in Dryad (41). License information: Copyright © 2023
the authors, somerights 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.adf2160
Materials and Methods
Figs. S1 to S7
Tables S1 to S10
y g
References (42–103)
MDAR Reproducibility Checklist
Submitted 12 October 2022; resubmitted 14 April 2023
Accepted 25 July 2023
10.1126/science.adf2160
p
,
5 of 5
RES EARCH

CALCULUS INSTRUCTION al instruction, and it was not logically possible

for the treatment condition to remain hidden
Establishing a new standard of care for calculus from students or faculty after the treatment be-
gan. Random assignment of faculty to control
using trials with randomized student allocation or treatment conditions would not be possible
because an individual faculty member’s knowl-
Laird Kramer1,2*, Edgar Fuller1,3,4*, Charity Watson1,3,4, Adam Castillo1,3,4†, edge, philosophy, and experience with a variety
Pablo Duran Oliva1,3,5, Geoff Potvin1,2 of classroom strategies and instructional prac-
tices may intersect with the features of the treat-
Calculus, the study of change in processes and systems, serves as the foundation for many STEM disciplines. ment or control conditions. The experimental
Traditional, lecture-based calculus instruction may present a barrier for students seeking STEM degrees, protocol thus included a group of instructors
limit their access to STEM professions, and block their potential to address society’s challenges. A large-scale willing to adopt the instructional methods in
pragmatic trial with randomized student allocation was conducted to compare two calculus instruction the treatment condition. This comprehensive
styles: active student engagement (treatment condition) versus traditional, lecture-based instruction (control experimental approach was intended to secure
condition). A sample of 811 university students were studied across 32 sections taught by 19 instructors the strongest possible evidence for critical stake-
over three semesters at a large, US-based Hispanic-serving institution. Large effect sizes were consistently holders to sustain the treatment beyond the trial.
measured for student learning outcomes in the treatment condition, which demonstrates a new standard The treatment condition used the modeling
for calculus instruction and increased opportunities for completion of STEM degrees. practices in calculus (MPC) curriculum and
pedagogy, and the control condition represented

C
the preexisting, traditional instructional prac-

alculus instruction needs substantial trans-
formation because it is often a barrier to
STEM degree attainment, especially for
traditionally underrepresented groups
(1–3), depriving both individuals and so-
ciety of the potential benefits of their inclu-
STEM pipeline, as noted in 1988 by Robert
White, president of the National Academy of
Engineering (13). Handlesman et al. (14) re-
cently argued that “we must fix the classrooms
where many students from historically excluded
communities are discouraged from pursuing
STEM” and that “… the continued exclusive use
p
tices at the study institution. MPC integrates
the practices of mathematicians as a central
design tenet throughout the course. Instructors
facilitate students’ application of mathematical
“habits of mind” (21) that foster deeper under-
standing of calculus concepts, including the

sion. National calls for calculus transformation
are numerous (4, 5), because failing calculus
can contribute to a student’s departure from
STEM degree programs. Only ~40% of stu-
dents entering universities with STEM degree
of lectures is malpractice at best, or an act of
discrimination at worst.” Thus, it is imperative
that substantial transformation in calculus in-
struction takes place to promote more equita-
g
identifying of patterns, hypothesis development
and testing, making connections, and communi-
cating ideas precisely to learn calculus through-
out the course. Class time is devoted to students
working collectively in small groups on prede-

intentions actually graduate with a STEM de-
gree (6). More concerning is that the odds of
female students switching out of a STEM pro-
gram after a calculus course is ~1.5 times higher
than that of comparable male students (3). Fur-
thermore, Hispanic and Black/African American
students had >50% higher failure rates than
white students in calculus (7, 8).
Evidence-based instruction, which is imple-
mented in many STEM disciplines, has reliably
ble learning environments for all students.
We present a large-scale trial featuring ran-
domized student allocation into treatment or
control conditions to rigorously compare an
evidence-based, active student engagement
calculus course against traditional, lecture-based
calculus instruction. The work extends prior cal-
culus research investigations (15–17) by includ-
ing random assignment of students to treatment
and control sections, as well as anonymized
y
signed notes and learning activities developing
their calculus understanding with minimal lec-
turing. Treatment included learning assistants
(22) who were undergraduate peers integrated
within the instructional team to facilitate student
learning and promote culturally responsive ins-
truction. The curriculum promotes mathemati-
cal practices (sense-making, problem solving,
argumentation, etc.) and established strategies
to optimize student engagement, including coop-

led to profound improvement in student suc-
cess (9–11). However, common approaches to
calculus instruction continue to rely on tradi-
tional, lecture-based practices in which students
analysis of the identical end-of-semester learn-
ing outcomes. The study uses a pragmatic (18)
design with random allocation of students to
inform on the effectiveness of similar inter-
y g
erative learning, argumentation and metacogni-
tion, mathematical fluency, and a culturally
responsive environment (23) (described in the
supplementary materials section 2). The MPC

p
are passive learners in the classroom and are ventions at higher education institutions, re- design builds on the SCALE-UP calculus model
expected to construct their knowledge mostly flecting real-world classroom constraints. In (24) and intentionally embodies well-established
outside of the classroom by doing homework or these contexts, blinding of the treatment and recommendations for calculus instruction, in-

in recitation sessions (12). Mathematics as a dis-
cipline thus needs to embrace its role in enabling
STEM careers that will lead to prosperity for
both individuals and society at large. “Calculus …
must become a pump and not a filter” for the
1
STEM Transformation Institute, Florida International
University, Miami, FL 33199, USA. 2Department of Physics,
Florida International University, Miami, FL 33199, USA.
3
Center for Transforming Teaching in Mathematics, Florida
International University, Miami, FL 33199, USA. 4Department
of Mathematics and Statistics, Florida International
University, Miami, FL 33199, USA. 5Department of
Mathematics and Applied Mathematics, Virginia
Commonwealth University, Richmond, VA 23284, USA.
*Corresponding author. Email: Laird.Kramer@fiu.edu (L.K.);
Edgar.Fuller@fiu.edu (E.F.)
†Present address: Department of Mathematics, The University of
Texas at Arlington, Arlington, TX 76019, USA.
control conditions to both students and faculty
is not possible because blinding is only feasible
when the treatment and control conditions re-
main unknown to the participants during the
period of study (such as in a clinical trial drug
study). As with other public health or sociologi-
cal interventions, enrollment of participants
in this study revealed some aspects of a cohort
structure, but it was still possible to maintain
the essential aspects of random assignment by
following a modified protocol, as was done in
Zwarenstein et al. (18). The treatment condi-
tion integrated a suite of coherent strategies
that have been independently found (19, 20)
to improve student learning; thus, the treat-
ment was a distinct departure from tradition-
,
cluding ambitious teaching practices and strate-
gies promoted by national mathematics societies
and national reports (12, 20, 25–28).
The study was performed at Florida Interna-
tional University (FIU) in Miami, Florida, the
fourth-largest public research university in the
United States, with 58,787 students, of which
41,795 are undergraduates as of fall 2019 (29).
FIU is a Hispanic-serving institution, with 64%
of students identifying as Hispanic/Latino/a/.
Moreover, 79% of the students identify as mem-
bers of historically underrepresented racial/
ethnic minority groups, and 57% are women.
The institution’s size provided a unique oppor-
tunity to perform this study because there are
18 to 34 40-student sections of calculus 1 being

Kramer et al., Science 381, 995–998 (2023) 1 September 2023 1 of 4

RES EARCH | R E S E A R C H A R T I C L E
taught each semester and primarily serving
STEM majors. Furthermore, institutional con-
ditions created urgency to transform calculus
because historic pass rates in introductory cal-
culus averaged 55% (range 13 to 88%) over the
six semesters before the project’s pilot.
Research design
A pragmatic trial (30–32) of the MPC approach
was performed during the fall 2018, spring
2019, and fall 2019 semesters to rigorously test
student outcomes. Students were randomly
assigned individually to treatment and control
conditions at the beginning of the semester
after enrolling in sections on the basis of their
scheduling preferences using the institution’s
enrollment system. To accommodate the ran-
domized assignments, each of the experimen-
tal sections doubled in size from the usual 40
seats to 80 seats before enrollment opening.
Instructor names and section sizes were invis-
ible to students throughout the enrollment
phase. Just before each term, the 80-seat sec-
tions were split into two 40-seat sections by
assigning each student at random to either a
treatment or control section.
Once assigned, the treatment sections im-
plemented the MPC approach, whereas the
control sections were unchanged. After assign-
ment, students were free to change/drop/add
course sections up until the regular institutional
drop/add deadline (7 days after classes began).
To account for such changes, enrollments were
monitored, and only students who were ran-
domly assigned to either a treatment or control
section and remained in that section through
the regular, nonpenalty drop/add deadline were
included in the data for the experimental study
reported below. In total, 1019 students were
randomly assigned to either the treatment or
control groups. Of these, 516 students were
assigned to the treatment group, and 417 re-
mained in the section at the drop/add deadline.
At the same time, 503 students were assigned
to the control group and 394 students remained
in the section at the drop/add deadline. Our
study followed the Consolidated Standards
for Reporting Trials (CONSORT) guidelines
(18, 33, 34). The specifics of recruiting, enroll-
ment, assignment, and completion for the trial
are discussed in sections 1.3 and 3.1 of the
supplementary materials. The randomization
process produced comparable groups by mathe-
matical background and demographics; class
sizes were typical for the course (see section 3
of the supplementary materials).
Faculty participating in the study included
seven individuals teaching 16 treatment sections,
along with 12 individuals teaching 16 control
sections. Faculty recruited to teach the treat-
ment sections indicated a willingness to adopt
and implement the MPC approach, replicating
the authentic condition of faculty reforming
their classroom practice under the study design.
Kramer et al., Science 381, 995–998 (2023) 1 September 2023
To prepare for the new instructional approach,
faculty participated in a 2-day, pre-semester
professional development workshop and were
provided with the MPC curricular materials.
Consistency of the MPC treatment was moni-
tored through weekly preparation meetings in
which the course objectives and pacing were
discussed. In-class monitoring by the project
team was deemed overly intrusive and disrup-
tive to classroom engagement. Control section
faculty were not guided to use any particular
practices and chose their normal instructional
approach, best described as traditional lecture
format with at most limited student engage-
ment. Potential effects of instructor differences
on learning outcomes were investigated and
are presented in section 3 of the supplementary
materials and summarized below.
The student outcome measures reported in-
clude identical end-of-semester learning mea-
sures, as well as course success data (i.e., course
grades). The end-of-semester learning measures
focused on evaluating learning using a set of
identical assessment items (problems) developed
by instructors spanning all calculus sections and
spanning the major learning objectives of a cal-
culus 1 course. The aim was to determine how
well students understood essential elements of,
and exhibited fluency and technical competency
in, calculus at course end. Assessment items
aligned to both local and national standards (35)
were embedded in a cumulative final examina-
Effect Sizes for End of Semester Learning Measures by Group
Other
Sci./Math
Eng./Comp.
Sci.
Biology
Transfer
FTiC
Black
Hispanic
Female
Male
Overall
−0.25 0.00 0.25 0.50 0.75
Cohen's d
Fig. 1. Overall end-of-semester learning mea-
sures effect sizes (Cohen’s d) broken out by
major, race/ethnicity, and gender. Error bars
indicate the 95% CI for effect size for each group.
Effect Sizes for End of Semester Course Outcome by Group
Other
Sci./Math
Eng./Comp.
Sci.
Biology
Transfer
FTiC
Black
Hispanic
Female
Male
Overall
0.00 0.25 0.50
Cohen's d
Fig. 2. Overall course success (earned grades of
A, B, or C) effect sizes (Cohen’s d) broken out
by major, race/ethnicity, and gender. Error bars
indicate the 95% CI for effect size for each group.
tion and were administered to all students in
each treatment and control section. To ensure
fidelity and fairness to both treatment and
control sections, control and treatment faculty
collaboratively developed a set of items to be
administered to both conditions in identical for-
mat and wording. This set of identical items
formed roughly two-thirds of the total final exa-
mination content, with the remaining items
added by individual faculty in a separate section
of the examination, allowing them to address
their specific instructional goals. Furthermore,
the examinations and problems were formatted
identically and without course section identifiers
to allow completely anonymized evaluation dur-
ing the subsequent comparative analysis. The
identical items covered core calculus topics in-
cluding evaluating limits, identifying extrema,
curve sketching, related rates, and evaluating
indefinite integrals. For the second and third
p
semesters, additional items focusing on implicit
differentiation and optimization were added
to the identical set of items (for details, see sec-
tion 3.3 of the supplementary materials). Course
success data (grades) reflect the overall assess-
ment of students as assigned by each section’s
g
instructor. Course grade policies were estab-
lished by individual instructors following depart-
mental syllabus guidelines and were broadly
consistent across sections and semesters.
Analysis of the end-of-course learning mea-
y
sures used a rubric for each problem, with five
researchers testing the initial rubric on a subset
of examinations to establish interrater reliability.
The final rubric represented consensus on all
elements and accounted for initial ambiguity
or disagreement. The analysis was performed by
a team of 10 trained evaluators, each of whom
evaluated a completely anonymous set of stu-
dent solutions. An average of two evaluators
reviewed each solution for correctness on a
y g
scale from 0 to 100%. The evaluators were very
1.00 1.25 1.50
consistent and interrater reliability was high
(Cohen’s kappa 0.827 in fall 2018 and 0.797 in
spring 2019) (36, 37). The same rubric was ap-
p
plied to the fall 2019 data given its high degree
of agreement. Once all problems were evaluated,
the research team deanonymized and sorted the
,
results by treatment and control sections for
the comparative analysis.
Results
The results indicate significant improvements
in student learning for the MPC group across
all three semesters. Students in the treatment
group showed substantially higher scores on
the identical end-of-semester learning outcomes:
(fall 2018: d = 0.505, P < 0.01; spring 2019: d =
0.75 1.00 0.748, P < 0.001; fall 2019: d = 0.925, P < 0.001)
compared with the control group. Combining
results from all three semesters of trials (i.e.,
32 sections and 811 total students), the overall
standardized mean difference between treat-
ment and control was d = 0.774 [95% confidence
2 of 4
RES EARCH | R E S E A R C H A R T I C L E
interval (CI) = 0.618 to 0.930] at the individual
level, a medium/large effect size (36, 37). An ad-
justed effect size (38) was computed using section-
level cluster properties and the mixed-effects
model structure described below and was found
to be dT = 0.771 (95% CI = 0.468 to 1.073; see sec-
tion 3.4.2.1 of the supplementary materials).
The success of the MPC intervention could
be seen across racial and ethnic groups, majors
and academic pathways, and genders (Fig. 1).
Similar medium/large overall effect sizes were
observed for students in the treatment condi-
tion who identified as Black/African American
(d = 0.882, P < 0.001) or Hispanic/Latino/a/
(d = 0.772, P < 0.001) when directly comparing
the identical learning measures with their coun-
terparts in the control condition. Although all
STEM majors showed significantly improved
learning, there were larger effect sizes for biology
majors in the treatment group (d = 0.925, P <
0.001). Students matriculating onto campus as
both first time in college (FTiC) and transfer
students showed medium/large effect sizes and
most were FTiC. Overall, treatment group stu-
dents show more consistency in applying the
tools of calculus to optimization problems, using
derivatives to sketch graphs of functions, and
in the evaluation of limits and integrals.
Potential biases arising in the random stu-
dent assignment and faculty selections were
investigated for hidden-level effects or confound-
ers to establish limitations of the study (see sec-
tion 3 of the supplementary materials). Random
allocation of students provided equivariance in
the demographics of student populations. Analy-
ses showed that allowing students to drop/add
sections during the open registration period
after the initial assignment did not affect the
measured outcomes. Faculty characteristics were
compared and found to be similar in both back-
ground and prior course student grade distribu-
tions. A mixed-effects model with student fixed
effects and random cluster effects due to section
and instructor levels was fit with tests of fixed
effects computed using Satterthwaite approxi-
mations (see section 3.2.4.1 of the supplemen-
tary materials). The explanatory power of the
model was found to be high (conditional R2 =
0.39, marginal R2 = 0.31). The effect of treat-
ment was statistically significant (t(660) = 5.68,
P < 0.001, semipartial R2 = 0.119) (39), implying
an estimated effect size of Cohen’s f = 0.368
with covariates and cluster-level effects present.
Random effects correlated with 0.085 of outcome
variance (intraclass correlation coefficient = 0.13
with demographic covariates). A sensitivity anal-
ysis (see section 3.2.4.3 of the supplementary
materials) showed that unmeasured confound-
ers would need to be four times more powerful
than any measured covariate, including student
mathematics background, to be responsible for
the observed effect.
Furthermore, students in the MPC treatment
condition had improved course grades. Average
Kramer et al., Science 381, 995–998 (2023) 1 September 2023
grades were significantly higher by ~0.4 points
(4.0 grade point scale) in MPC sections across
all semesters of the study (P < 0.001, d = 0.295).
This translated to success rates (A, B, or C grades)
averaging 11% higher in MPC sections compared
with traditional sections (P < 0.001, d = 0.251;
Fig. 2). Outcomes were consistent across the
three semesters of the experiment (Fig. 3). More-
over, the MPC sections also had lower course
late drop rates (departure after the regular drop/
add period ends) across all three semesters
(P < 0.05, d = 0.141), suggesting that students
more clearly perceived that they were likely to
succeed in the course.
The trend of improved outcomes in course
success was also observed for demographic
subgroups, as can be seen in Fig. 2. A logistic re-
gression model of success using gender identifi-
cation, FTiC status, and Hispanic identification
as independent variables showed the odds of a
female-identified student in the treatment group
passing the course to be 58% higher than the
odds of a female-identified student in the control
group (b1 = 0.46, P < 0.05). Hispanic students’
odds of passing the course were almost double
that of their counterparts in the control group
(b1 = 0.70, P < 0.001). The likelihood of FTiC
students in the treatment group passing the
course increased by ~85% compared with FTiC
students in the control group (b1 = 0.61, P < 0.01)
(for details, see section 3.4 of the supplementary
materials).
Discussion and conclusions
This pragmatic trial demonstrates that stu-
dent learning outcomes were significantly im-
proved in the treatment condition. Contrary to
previous research (40), this study shows that
when students are expected to engage with
Drop by Term and Assignment Group with Std Error
Percent of Student Success (A/B/C), Fail, and Late
Fall 2018
75
50
25
0
Treatment Control
Fig. 3. Final course grade outcomes broken out by term and curriculum, including success (earned
grades of A, B, or C), fail (earned grades of D or F), and late drops (withdrawals or drops after the
institution’s drop/add deadline). Vertical scale is percentage by outcome. Error bars indicate the 95% CI
for the mean percentage of students in each outcome group over all sections in a term.
calculus concepts collaboratively using inten-
tional, evidence-based teaching strategies, they
develop a better understanding of calculus con-
cepts and techniques. The benefits of the MPC
curriculum and pedagogy were realized regard-
less of racial/ethnic group, gender, or major/
academic pathway. These trends suggest that
the treatment includes culturally responsive and
equitable strategies. Specifically, the MPC learn-
ing environment is designed to promote learn-
ing communities that provide ongoing support
for learning mathematics through collaborative
engagement and ongoing formative feedback.
This aims to promote inclusion and increases
access for students with different mathematical
backgrounds, cultural identities, and life expe-
riences by allowing them to use their mathe-
matics skills in a supportive, nonthreatening
environment.
The improved learning and course success
p
for modeling practices in calculus reported in
this study have profound implications for cal-
culus instruction. This study demonstrates the
substantial benefit to students of the MPC ap-
proach designed around established, evidence-
based principles, and should motivate educators
g
in mathematics and other STEM disciplines to
adopt the same or similar approaches and
conduct similar studies to replicate these find-
ings. Improved student success also leads to
more efficient student progress to graduation
y
and boosts institutional effectiveness. Applying
this study’s 11% average improvement in pass
rate to all 2000 first-time calculus students at
FIU would translate to 220 additional students
succeeding in calculus annually and reducing
the instructional load by five sections annually.
Extending this strategy to the ~300,000 stu-
dents across the nation taking calculus 1 each
y g
Spring 2019 Fall 2019
p
Outcome
Success ,
Fail
Late Drop
Treatment Control Treatment Control
3 of 4
RES EARCH | R E S E A R C H A R T I C L E
year, these results translate to the potential for
an additional 33,000 students passing calculus
each year, saving students an estimated $23.9
million on tuition [based on a three-credit course
at the average public college/university tuition
rate of $242/credit (41, 42)]. Pragmatic trials
provide guidance on what can be achieved by
engaging faculty willing to change their ins-
truction. These results potentially represent a
lower bound on the long-term effects because
faculty likely develop additional expertise
through continued instruction and realize im-
proved outcomes. The measured effect size pro-
vides the rationale to stop the control due to
treatment benefit if one follows medical re-
search protocols (43, 44).
The experimental methodology described here
establishes a new standard of care for calculus
instruction and a high standard of evidence to
bear on understanding the impacts on student
learning. Improved learning of calculus aims to
foster higher success in future STEM courses
and develop the STEM “habits of mind” that stu-
dents take with them into their future careers.
Further, MPC shows the potential to address
the disparities that differentially affect histori-
cally underrepresented groups, thus offering a
mechanism to address Handelsman et al.’s
(14) call to promote the success of historically
excluded communities. We envision a mathe-
matics experience for all students built on this
approach and recommend that active student
engagement must be deployed across all STEM
disciplines to improve our development of fu-
ture STEM professionals from all backgrounds.
RE FE RENCES AND N OT ES
1. U. Treisman, Coll. Math. J. 23, 362–372 (1992).
2. B. B. Alexander, A. C. Burda, S. B. Millar, J. Women Minor. Sci.
Eng. 3, 145–159 (1997).
3. J. Ellis, B. K. Fosdick, C. Rasmussen, PLOS ONE 11, e0157447
(2016).
4. S. L. Ganter, D. J. Lewis, D. Hughes-Hallett, Science 279,
2019–2025 (1998).
5. S. Olson, D. G. Riordan, Engage to Excel: Producing One Million
Additional College Graduates with Degrees in Science,
Technology, Engineering, and Mathematics. Report to the
President (Executive Office of the President, 2012).
6. E. Seymour, N. M. Hewitt, Talking About Leaving (Westview
Press, 1997).
7. E. Seymour, A.-B. Hunter, Talking About Leaving Revisited: Persistence,
Relocation, and Loss in Undergraduate STEM Education (Springer,
2019); http://link.springer.com/10.1007/978-3-030-25304-2.
Kramer et al., Science 381, 995–998 (2023) 1 September 2023
8. A. K. Koch, B. Drake, Digging into the Disciplines: The Impact of
Gateway Courses in Accounting, Calculus, and Chemistry on
Student Success (John N. Gardner Institute for Excellence in
Undergraduate Education, 2018).
9. M. J. Graham, J. Frederick, A. Byars-Winston, A. B. Hunter,
J. Handelsman, Science 341, 1455–1456 (2013).
10. M. Stains et al., Science 359, 1468–1470 (2018).
11. C. Henderson et al., Int. J. STEM Educ. 4, 14 (2017).
12. C. Rasmussen et al., J. Res. Math. Educ. 50, 98–111
(2019).
13. L. A. Steen, in Papers Presented at a Colloquium, Washington,
DC, October 28-29, 1987 (ERIC, 1988); MAA Notes Number 8.
14. J. Handelsman et al., Science 376, 1057–1059 (2022).
15. J. Bookman, C. P. Friedman, in Assessment Practices in
Undergraduate Mathematics, B. Gold, S. Keith, W. A. Marion, Eds.
(Mathematical Association of America, 1999), MAA Notes, p. 253.
16. M. Brown, in Calculus: The Dynamics of Change, W. Roberts,
Ed. (Mathematical Association of America, 1996), vol. 39 of
MAA Notes, pp. 52–58.
17. E. Tidmore, Focus on Calculus 7, 5–6 (1994).
18. M. Zwarenstein et al., BMJ 337, a2390–a2390 (2008).
19. S. Freeman et al., Proc. Natl. Acad. Sci. U.S.A. 111, 8410–8415
(2014).
20. M. Abell, L. Braddy, D. Ensley, L. Ludwig, H. Soto-Johnson, MAA
Instructional Practices Guide (Mathematical Association of
America, 2018).
21. K. Saxe et al., A Common Vision for Undergraduate Mathematical
Sciences Programs in 2025 (Mathematical Association of
America, 2015).
22. V. Otero, S. Pollock, N. Finkelstein, Am. J. Phys. 78, 1218–1224
(2010).
23. G. Ladson-Billings, Theory Pract. 34, 159–165 (1995).
24. L. Benson et al., in Proceedings of the 2009 Annual Conference &
Exposition, American Society of Engineering Education (ASEE,
2009); pp. 14.159.1–14.159.7.
25. A. Kezar, A. C. Chambers, J. C. Burkhardt, Higher Education for
the Public Good: Emerging Voices from a National Movement
(Wiley, 2015).
26. D. Bressoud, Math. Teach. 109, 179–185 (2015).
27. G. Sonnert, P. M. Sadler, S. M. Sadler, D. M. Bressoud, Int. J.
Math. Educ. Sci. Technol. 46, 370–387 (2015).
28. P. Zorn, E. Bressoud, W. David, M. Pearson, “Response to the
PCAST Report to the President, Engage to Excel” (MAA, 2012);
https://www.maa.org/sites/default/files/pdf/sciencepolicy/
MAA-ResponsePCAST-E2E.pdf.
29. D. Burrows, “Common data set 2019-2020” (Florida
International University CDS Archive, 2020); https://opir.fiu.
edu/CDS/CDS2019.pdf.
30. C. J. Torgerson, Education 37, 313–321 (2009).
31. M. Roland, D. J. Torgerson, BMJ 316, 285 (1998).
32. B. Styles, C. Torgerson, Educ. Res. 60, 255–264 (2018).
33. K. F. Schulz, D. G. Altman, D. Moher; CONSORT Group, Trials
11, 32 (2010).
34. D. Moher et al., Int. J. Surg. 10, 28–55 (2012).
35. M. A. Tallman, M. P. Carlson, D. M. Bressoud, M. Pearson,
International Journal of Research in Undergraduate
Mathematics Education 2, 105–133 (2016).
36. J. Cohen, Statistical Power Analysis for the Behavioral Sciences
(Lawrence Erlbaum, ed. 2., 1988).
37. J. C. Valentine, H. Cooper, “Effect size substantive
interpretation guidelines: Issues in the interpretation of effect
sizes” (US Department of Education, 2003); https://wmich.
edu/sites/default/files/attachments/u58/2015/Effect_Size_
Substantive_Interpretation_Guidelines.pdf.
38. L. V. Hedges, in The Handbook Of Research Synthesis and
Meta-Analysis, H. Cooper, L. V. Hedges, J. C. Valentine, Eds.
(Russell sage Foundation, 2009); 399–416.
39. B. C. Jaeger, L. J. Edwards, K. Das, P. K. Sen, J. Appl. Stat. 44,
1086–1105 (2017).
40. M. T. Hora, Proc. Natl. Acad. Sci. U.S.A. 111, E3024
(2014).
41. D. M. Bressoud, M. P. Carlson, V. Mesa, C. Rasmussen, Int. J.
Math. Educ. Sci. Technol. 44, 685–698 (2013).
42. T. Snyder, C. de Brey, S. Dillow, “Digest of education statistics,
2018” (NCES, 2019); https://nces.ed.gov/programs/digest/
d19/tables/dt19_330.10.asp.
43. V. M. Montori et al., JAMA 294, 2203–2209
(2005).
44. S. J. Pocock, JAMA 294, 2228–2230 (2005).
45. Data for: L. Kramer et al., Establishing a new standard of care for
calculus using trials with randomized student allocation, Dryad
(2023); https://doi.org/10.5061/dryad.kkwh70s95.
AC KNOWLED GME NTS
We thank the FIU administration and the State of Florida for initial
support that made this project possible. The work would not be
possible without the faculty, undergraduate learning assistants,
and students participating in the study, who have earned our deep
gratitude. We thank K. Rambo-Hernandez for insightful statistical
p
analysis discussions, and we are indebted to the reviewers and
editors for their insight and guidance. The research for this study
was conducted under an institutional review board protocol
approved by the Florida International University IRB (approval no.
IRB-18-0211-a.m.02). Students participating in the study consented
to participation at the beginning of each semester. Only students
aged 18 or over were included. Any opinions, findings, and
conclusions or recommendations expressed in this material are
those of the author(s) and do not necessarily reflect the views of
g
the National Science Foundation. Funding: This work was
supported by the National Science Foundation (grant DUE 1832450
to L.K., E.F., G.P., A.C., and C.W.). Author contributions:
Conceptualization: L.K., E.F., G.P., A.C., C.W.; Funding acquisition:
L.K., E.F., G.P.; Investigation: L.K., E.F., G.P., A.C., C.W., P.D.O.;
y
Methodology: L.K., E.F., G.P., A.C., C.W., P.D.O.; Project administration:
L.K., E.F.; Writing – original draft: L.K., E.F., G.P., A.C., C.W., P.D.O.;
Writing – review and editing: L.K., E.F., G.P., A.C., C.W., P.D.O.
Competing interests: The authors declare no competing interests.
Data and materials availability: All experimental data are available
at Dryad (45). License information: This work is licensed under a
Creative Commons Attribution 4.0 International (CC BY 4.0) license,
which permits unrestricted use, distribution, and reproduction in any
medium provided the original work is properly cited. To view a copy of
this license, visit https://creativecommons.org/licenses/by/4.0/. This
license does not apply to figures/photos/artwork or other content
included in the article that is credited to a third party; obtain authorization
from the rights holder before using such material.
y g
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.ade9803
Materials and Methods
Figs. S1 to S10
p
Tables S1 to S17
References (46–109)
MDAR Reproducibility Checklist
,
Submitted 20 September 2022; accepted 28 July 2023
10.1126/science.ade9803
4 of 4
RES EARCH

NEUROSCIENCE across 1000 trials. Details about model archi-

tecture and hyperparameters are given in
A principal odor map unifies diverse tasks in the supplementary methods. When properly
hyperparameter-tuned, performance was found
olfactory perception to be robust across many model architectures.
We present results for the model with the high-
Brian K. Lee1†, Emily J. Mayhew2,3†, Benjamin Sanchez-Lengeling1, Jennifer N. Wei1, est mean area under the receiver operating
Wesley W. Qian4,1,5, Kelsie A. Little2, Matthew Andres2, Britney B. Nguyen2, Theresa Moloy2, characteristic curve (AUROC) on the cross-
Jacob Yasonik4,1, Jane K. Parker6, Richard C. Gerkin4,1,7, Joel D. Mainland2,8*, Alexander B. Wiltschko4,1* validation set (AUROC = 0.89) (17).
How does the POM compare to perceptual
Mapping molecular structure to odor perception is a key challenge in olfaction. We used graph odor maps and conventional structure-based
neural networks to generate a principal odor map (POM) that preserves perceptual relationships and maps of odorants? Empirical perceptual space
enables odor quality prediction for previously uncharacterized odorants. The model was as reliable as a (Fig. 1D) intuitively represents perceptual dis-
human in describing odor quality: On a prospective validation set of 400 out-of-sample odorants, the tances (e.g., two molecules that smell of jas-
model-generated odor profile more closely matched the trained panel mean than did the median mine should be nearer to each other than to a
panelist. By applying simple, interpretable, theoretically rooted transformations, the POM outperformed beefy-smelling molecule) and hierarchies (e.g.,
chemoinformatic models on several other odor prediction tasks, indicating that the POM successfully jasmine and lavender are subtypes of the floral
encoded a generalized map of structure-odor relationships. This approach broadly enables odor odor family). We show that whereas this struc-
prediction and paves the way toward digitizing odors. ture is lost in Morgan fingerprint-based maps
of odorant space (Fig. 1E), the POM preserves

A
fundamental problem in neuroscience
is mapping the physical properties of a
stimulus to perceptual characteristics. In
vision, wavelength maps to color; in audi-
tion, frequency maps to pitch. By con-
neural network (MPNN) (7), which is a spe-
cific type of graph neural network (GNN) (8),
to map chemical structures to odor percepts.
Each molecule was represented as a graph,
with each atom described by its valence, de-
p
relative perceptual distances and hierarchies
(Fig. 1F and figs. S1 to S3).
Our model outperformed the median panelist
on the prospective validation task
To test whether our model extends to novel

trast, the mapping from chemical structures to
olfactory percepts is poorly understood. Detailed
and modality-specific maps such as the Com-
mission Internationale de l’Elcairage (CIE) color
space (1) and Fourier space (2) led to a better
gree, hydrogen count, hybridization, formal
charge, and atomic number. Each bond was
described by its degree, its aromaticity, and
whether it is in a ring. Unlike traditional fin-
gerprinting techniques (9), which assign equal
g
odorants, we designed a prospective valida-
tion challenge (18) in which we benchmarked
model predictive performance against indi-
vidual human raters. In olfaction, no reliable
instrumental method of measuring odor percep-

understanding of visual and auditory coding.
Similarly, to better understand olfactory cod-
ing, the field of olfaction needs a better map.
Pitch increases monotonically with frequen-
cy. By contrast, the relationship between odor
percept and odorant structure is riddled with
discontinuities; this is exemplified by Sell’s
triplets (3), which are trios of molecules in
which the structurally similar pair is not the
perceptually similar pair (Fig. 1A). These dis-
weight to all molecular fragments within a
set bond radius, a GNN can optimize fragment
weights for odor-specific applications. Neural
networks have unlocked predictive modeling
breakthroughs in diverse perceptual domains
[e.g., natural images (10), faces (11), and sounds
(12)] and naturally produce intermediate rep-
resentations of their input data that are func-
tionally high-dimensional, data-driven maps.
We used the final layer of the GNN (henceforth,
y
tion exists, and trained human sensory panels
are the gold standard for odor characterization
(19). Odor perception is variable across indi-
viduals (20, 21), but group-averaged odor rat-
ings are stable across repeated measurements
(22) and represent our best avenue to estab-
lish the ground-truth odor character for novel
odorants. We trained a cohort of subjects to
describe their perception of odorants using
the rate-all-that-apply method (RATA) and a

continuities in the structure-odor relation-
ship suggest that standard chemoinformatic
representations of molecules—functional group
counts, physical properties, molecular finger-
“our model”) to directly predict odor qualities,
and the penultimate layer of the model as a prin-
cipal odor map (POM). The POM (i) faithfully
represented known perceptual hierarchies and
y g
55-word odor lexicon. During training sessions,
each term in the lexicon was paired with visual
and odor references (fig. S4 and table S1). Only
subjects that met performance standards on

p
prints, and so on—that have been used in re- distances, (ii) extended to out-of-sample (here- the pretest of 20 common odorants (data S2;
cent odor modeling work (4–6) are inadequate after, “novel”) odorants, (iii) was robust to dis- individual test-retest correlation R > 0.35; rea-
to map odor space. continuities in structure-odor distances, and sonable label selection for common odorants)

The principal odor map represents perceptual
distances and hierarchies
To generate odor-relevant representations of
molecules, we constructed a message passing
1
Google Research, Brain Team, Cambridge, MA, USA. 2Monell
Chemical Senses Center, Philadelphia, PA, USA. 3Department of
Food Science and Human Nutrition, Michigan State University,
East Lansing, MI, USA. 4Osmo Labs, PBC, Cambridge, MA,
USA. 5Department of Computer Science, University of Illinois at
Urbana-Champaign, Champaign, IL, USA. 6Department of Food
and Nutritional Sciences, University of Reading, Reading,
Berkshire, UK. 7School of Life Sciences, Arizona State
University, Tempe, AZ, USA. 8Department of Neuroscience,
University of Pennsylvania, Philadelphia, PA, USA.
*Corresponding author. Email: jmainland@monell.org (J.D.M.);
alex@osmo.ai (A.B.W.)
†These authors contributed equally to this work.
(iv) generalized to other olfactory tasks.
We curated a reference dataset of ~5000 mol-
ecules, each described by multiple odor labels
(e.g., creamy, grassy), by combining the Good
Scents (13) and Leffingwell & Associates (14)
(GS-LF) flavor and fragrance databases (Fig.
1B). To train our model, we optimized model
parameters with a weighted-cross entropy loss
over 150 epochs using Adam (15) with a learn-
ing rate decaying from 5 × 10−4 to 1 × 10−5 and
a batch size of 128. The GS-LF dataset was split
80/20 training/test, and the 80% training set
further subdivided into five cross-validation
splits. These cross-validation splits were used
to optimize hyperparameters using Vizier (16),
a Bayesian optimization algorithm, by tuning
,
were invited to join the panel.
To avoid trivial test cases, we applied the fol-
lowing selection criteria for the set of 400 novel
odorants: (i) molecules must be structurally
distinct from each other (fig. S5), (ii) mole-
cules should cover the widest gamut of odor
labels (data S1), and (iii) molecules must be
structurally or perceptually distinct from any
training example (e.g., Fig. 1A and data S1).
Our prospective validation set consisted of 55–
odor label RATA data for 400 novel, intensity-
balanced odorants generated by our cohort of
≥15 panelists (two replicates). Summary sta-
tistics and the correlation structure of the hu-
man perceptual data are presented in figs. S6
to S8. Our panel’s mean ratings were highly

Lee et al., Science 381, 999–1006 (2023) 1 September 2023 1 of 8

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

A B
O

Structurally similar pair OH

O
O

cheesy, fruity S

roasted, creamy, cocoa

O

O
POM
green, fruity
Perceptually similar pair GoodScents Both Leffingwell

Molecular representations from

penultimate model layer
C GNN model training

citrus creamy
baked spicy
vanilla
clean alcoholic beefy

p
chocolate fruity

Odor Islands in PCA space for multiple representations

True Labels Fingerprints POM
D 3 E F floral
meaty
savory
floral 2 muguet beefy
lavender roasted

g
muguet meaty 2
2 lavender jasmine
savory
jasmine beefy
roasted 1
1 1
PC2 (0.60%)
PC2 (2.2%)

PC2 (12%)

y
0 0 0

−1
−1
−1

−2 ethereal
ethereal −2
cognac cognac
fermented −2 fermented
−3 alcoholic alcoholic

−3 −2 −1 0 1 2 3 −2 −1 0 1 2 −2 −1 0 1 2
PC1 (2.7%) PC1 (0.77%) PC1 (16%)

y g
Fig. 1. The POM preserves the structure of odor perceptual space. first and second principal components (PCs) of their (D) perceptual labels from
(A) Example triplet of molecules in which the structurally similar pair is not the the GS-LF training dataset (138 labels), (E) cFP structural fingerprints (radius
perceptually similar pair. (B) The GNN was trained on a curated dataset of four, 2048 bit), and (F) POM coordinates (256 dimensions). Areas dense with

p
~5000 semantically labeled molecules drawn from GS-LF. One square represents molecules that have the broad category labels floral, meaty, or alcoholic are shaded;
100 molecules. Three example training set molecules and their odor descriptions areas dense with narrow category labels are outlined. The POM recapitulates
are shown: 2-methyl-2-hexenoic acid (top), 2,5-dimethyl-3-thioisovalerylfuran the true perceptual map, but the fingerprint map does not; note that only relative

,
(middle), and 1-methyl-3-hexenyl acetate (bottom). (C) Schematic illustrating the (not absolute) coordinates matter. Additional labels are visualized for the POM
process of training a GNN to generate the POM. (D to F) Odorants plotted by the in fig. S1. PCA, principal components analysis.

stable (panel test-retest: R = 0.80, n = 15; fig.
S9) and more consistent than the DREAM Ol-
faction Prediction Consortium cohort’s ratings
(6) (figs. S10 and S11).
Of the 400 molecules characterized, 77 were
dropped from the final prospective validation
set because of low intensity (42) (fig. S12), re-
dundancy (2), mistaken inclusion (1), confirmed
contamination (19), or potential contamination
(13) (data S1). Model performance was eval-
uated on the remaining 323 molecules without
model retraining.
To measure the model’s performance, we
compared its normalized predictions with the
normalized panel mean rating (Fig. 2, A and
C). One example of raw ratings and predic-
tions for a single molecule, representative of
relative GNN and random forest (RF) per-
formance and panel ratings trends, is given
in Fig. 2; additional examples are provided
in the supplementary materials (fig S14). Al-
though there is considerable variation across
molecules in the ability of both individual
raters and the model to match the panel mean
ratings, the model output comes closer to the
panel mean than does the median panelist
for 53% of molecules (Fig. 2, E and F). No-
tably, panelists were able to smell each odorant
as they rated it, whereas the model’s predic-
tions were based solely on nominal molec-
ular structure.
As a baseline comparison, we trained a count-
based fingerprint (cFP)–based RF model, the pre-
vious state-of-the-art approach (6), on the same
dataset (Fig. 2B). This baseline model surpassed
the median panelist for only 41% of molecules.

Lee et al., Science 381, 999–1006 (2023) 1 September 2023 2 of 8

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

A 1.00 Top 5 predicted labels

p(label applies)
RGNN = 0.63
0.75 O All other labels
88th percentile GNN prediction
GNN

0.50 O
0.25

0.00

B 1.00
p(label applies)

Top 5 predicted labels

0.75 RRF = 0.45 All other labels
RF

0.50

0.25

0.00

C
Panel mean rating

1.5 Top 5 rated labels

All other labels
1.0

0.5

p
0.0

D
P6
P19 Rbest = 0.52 Rating
P25
P10 5
P28
Panelists

P26 4
P27
P23 Rmedian = 0.18 3

g
P32 2
P11
P5 1
P14 0
P2
P1
P9

Ch ous
Nu d

Me er

Wo ous

ho lic
F l o ery

ram ted

Co nut

Or ine
Ap ng

Co eesy

Pi e

Wa to
M ry

Ca Alc assy
Gr ody
S p oky

Ea al
Bu one

Me Mus l

F i sffee
Le ey

y
F r eet

Ho en

Ro ty
dic ty

To arp
Da s

rlic
Gr le
rm rry

Ro usk
Oz tty

He iry

An se

Me Fatty

lfu r

Sh ne
P e Mint
rba
Va i t y
Po i l l a

Fe Be l

Cit i c

mm l
Sm y

Co i c y

Tro a c h
W ical
cu y

lic

Ga y
Ja Hay

xy
Su Sou

y
te

R u ina

rth
ra

ru

g
mo

mb

im
a
p

ma
tte

Cu i n e

h
mp oho
ell

n
oli

an
e
en

tal

sm
wd

Ca as

co
re
Sw
u

r
n

E F 0.05 0.05

0.00 0.00
Difference in Median Correlation
Relative to Human Baseline
Proportion of Molecules

y g
-0.05 -0.05

-0.10 -0.10
-0.20 -0.20

p
-0.22 -0.22

-0.25 -0.25

-0.27 -0.27

-0.30 -0.30 ,
d P red _cFP rdred N
ffle cF GN
sh
u N_ ord RF o
N_ KN _M _M
Correlation to Panel Mean GN KN
N RF

Fig. 2. The GNN model displays human-level odor description performance.
(A) GNN model label predictions, (B) RF model label predictions, (C) panel mean
ratings with standard error bars, and (D) individual panelist ratings (where 0 means the
label does not apply and 5 means the label is very applicable), averaged over two
replicates, for the molecule 2,3-dihydrobenzofuran-5-carboxaldehyde. In (A) to (C), the
top-five-ranked descriptors are in orange (GNN), purple (RF), or green (panel).
Descriptors in (A) to (D) are ordered by panel mean ratings. (A) to (D) are annotated
with the Pearson correlation coefficient of their data to the panel mean rating shown in
(C). Included in (D) are the panelist-panel correlation coefficients for the panelist
that best matches the panel mean and for the panelist with the median match.
(E) Cumulative density plot showing the distribution of correlations between human
panelists and the panel mean (green) and between the GNN, RF, and GNN-shuffled
model predictions and the panel mean on a per molecule basis. Curves shifted
to the right are more strongly correlated to the panel mean. The dashed horizontal
line indicates the median molecule; the dotted vertical lines indicate the median
correlations. The purple and orange symbols indicate the increase and decrease in
correlation, respectively, of the GNN and RF models relative to median human
performance. (F) Difference in the median correlation to the panel mean relative to
the median human subject’s correlation to the panel mean for models trained using
k-nearest neighbor (KNN) and RF, models trained on cFPs or Mordred features,
and the GNN model. Only the GNN model has a median correlation to the panel
mean that is higher than that of the median panelist.

Lee et al., Science 381, 999–1006 (2023) 1 September 2023 3 of 8

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

The GNN model shows human-level perform-
ance in aggregate, but how does it perform
across perceptual and chemical classes? When
we disaggregated performance by odor label,
the model was within the distribution of hu-
man raters for all labels except musk and sur-
passed the median panelist for 30/55 labels
(55%; Fig. 3A). This per-label view indicates
that the GNN model is superior to the previous
state-of-the-art model trained on the same data
(paired two-tailed Student’s t test, p = 3.3 × 10−7).
Predictive performance for a given label de-
pends on the complexity of the structure-odor
mapping for that label. It is thus unsurprising
that our model performs best for labels such as
garlic and fishy that have clear structural de-
terminants (sulfur-containing for garlic; amines
for fishy) and worst for the label musk, which
includes at least five distinct structural classes
(macrocyclic, polycyclic, nitro, steroid-type, and
straight chain) (23, 24). By contrast, a panel-
ist’s performance for a given label depended
on their familiarity with the label in the con-
text of smell; consequently, we observed strong
panelist-panel agreement for labels that describe
common food smells such as nutty, garlic, and
cheesy and weak agreement for labels such as
musk and hay.
Model performance also depended on the
number of training examples for a given label;
with enough examples, models can learn even
complex structure-percept relationships. In gen-
eral, our model’s performance was high for labels

A Garlic GNN B Garlic

Meaty
Meaty RF Fruity
Camphoreous
Fruity panelist Fishy
Camphoreous Sulfurous
Fishy
Sulfurous

GNN Correlation with Panel Mean

Cooling Roasted
Pine
Cheesy Alcoholic Sweet
Apple
Floral

p
Cooling
Roasted
Sour Green
Mint
Sweet Medicinal
Floral
Alcoholic
Berry Musty
Nutty Sharp

g
Powdery Waxy
Smoky Fermented
Tropical
Ozone
Dairy
Woody
Peach

y
Spicy
Green Musk
Honey
Rose
Herbal
Coffee Training Data Counts
Tomato
Caramellic C Sulfur
Rummy GNN
Earthy Carboxylic acid
RF
Vanilla Alkyl 4-carbon chain panelist
Cucumber Ester
Winey
Medicinal Phenyl

y g
Citrus Carbonyl
Animal
Metallic 8 atoms or fewer
Orange Nitrile
Fatty 13 atoms or more
Lemon

p
Musty Amine
Sharp
Jasmine 0.00 0.25 0.50 0.75
Waxy Correlation to Panel Mean
Coconut
D ,
Grassy
Buttery
Fermented Odor caused by nominal compound
Ozone Minor impact on odor by contaminants
Hay
Musk Cause of odor cannot be determined
0.00 0.25 0.50 0.75 Major impact on odor by contaminants
Correlation to Panel Mean Odor not caused by nominal compound

Fig. 3. Model performance is robust across structural and perceptual
classes. (A) Correlation of GNN (orange) and RF (purple) model predictions
and panelist ratings (gray) to the panel mean for each of the 55 odor
labels. (B) GNN model correlation to the panel mean for each of the 55 odor
labels plotted against the number of molecules in the training data for which
the label applies. Circle size is proportional to the number of test-set molecules
for which the label applies. Selected data points are annotated. (C) Mean
correlation of GNN (orange) and RF (purple) model predictions and panelist
ratings (gray) to the panel mean for molecules belonging to 10 common
chemical classes. (D) Categorization of GC-olfactometry QC results for
50 test-set stimuli. In the (A) and (C) box plots, the center line represents
the median, box limits are upper and lower quartiles, and whiskers are minimum
and maximum values or 1.5 times the inner-quartile range, whichever is closer
to the median.

Lee et al., Science 381, 999–1006 (2023) 1 September 2023 4 of 8

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

p
g
Fig. 4. The POM is robust to discontinuities in structure-odor mapping. (C) Diagram of the psychophysical task in which panelists rated explicit

(A) Example triplet of molecules in which the structurally similar pair is
not the perceptually similar pair (i.e., “discordant”), according to the empirical
odor labels of each molecule. Training-set descriptors (anchor) and mean
panel ratings (novel odorants) are shown in colored text beneath the molecular
structure; model-predicted labels are listed in black text. Structural nodes
highlighted in darker red are those that are more important to model predictions.
(B) We selected 41 such triplets from the empirical label data without consulting
the model; by design, 100% of these are discordant and thus represent a difficult
test for a predictive perceptual model based on molecular structure. Each
colored line connects molecules in a triplet that share the same anchor, as in (C).
y
perceptual distances between molecules in triplets. (D) Experimentally measured
explicit perceptual distance ratings in the same triplets also show high discordance
with structural distance, that is, the molecule that is more structurally similar
to the anchor is usually (90% of the time) less perceptually similar. (E) The
GNN model–predicted labels agree with the counterintuitive-but-correct perceptual
relationship 50% of the time, that is, they correctly predict the empirical
discordance half of the time, as measured by the cosine distance of the
predicted, binarized labels. (F) A baseline model correctly predicts the empirical
discordance only 19% of the time. The models in (E) and (F) are the same as
those from Figs. 2 and 3.

y g
with many training examples (e.g, fruity, sweet, odor character was not due to the nominal sweet, and dairy, but this odor percept was at-
floral) (Fig. 3B), but performance for labels with compound. We selected the 50 stimuli to rep- tributed through QC to the contaminant diacetyl,
few training examples was either high (e.g., fishy, resent three quadrants of intrapanel agree- a well-known buttery odorant. In another case, the

p
camphoreous, cooling) or low (e.g, ozone, sharp, ment and model-panel agreement (high-high, purchased odorant, isobornyl methylacrylate,
fermented). Likewise, model performance was high-low, and low-high), anticipating that odor- was described by the panel and the model as
bounded above by panel test-retest correlation ous contaminants may explain cases of poor both piney and floral; however, the nominal

(fig. S15). When we disaggregated by chemical
classes (e.g., esters, phenols, amines), both pan-
elist and model performances were relatively
uniform (Fig. 3C), with sulfur-containing mole-
cules showing the strongest performance from
panelists and the model (R = 0.52).
Chemical materials are impure—a fact too
often unaccounted for in olfactory research (25).
To measure the contribution of impurities to
the odor percept of our stimuli, we applied a gas
chromatography–mass spectrometry (GC-MS)
and gas chromatography–olfactometry (GC-O)
quality control (QC) procedure to 50 stimuli
(data S1). This QC procedure matches an odor
percept to its causal molecule, which allowed
us to identify stimuli for which the primary
model-panel agreement. Our QC procedure led
to diverse conclusions: The nominal compound
caused the odor (11/50), contaminants con-
tributed to the odor in a minor (15/50) or major
(5/50) way, contaminants caused the odor
(15/50), or the cause of the odor could not be
determined (4/50) (Fig. 3D). Fishy, garlic, and
sulfurous were the most prevalent contami-
nant odor qualities (fig. S17); these labels were
overrepresented in the QC set, so we expected
that the rate of severe contamination is likely
lower in the full test set than in the QC set. In
some cases, although we purchased a novel
odorant, the dominant odorant was not novel;
for example, the stimulus 4,5-dimethyl-1,3-thiazol-
2-amine was described by the panel as buttery,
,
compound was floral only and the piney aroma
was due to the closely related compound, bor-
neol, which was detected as a contaminant in
the sample. Based on QC results, we removed
32 molecules known or suspected to have high
degrees of odorous contamination (data S1).
There were various QC results, each with dis-
tinct implications on model performance (data
S1). In some cases, the model performed well de-
spite the presence of odorous contaminants. We
estimate that if these contaminants were re-
moved from the rated samples, model perform-
ance would improve in 6 of 50 scenarios, would
degrade in another 6 of 50 scenarios, would re-
main neutral in 21 of 50 scenarios, and would be
indeterminate in 17 of 50 scenarios. We estimate

Lee et al., Science 381, 999–1006 (2023) 1 September 2023 5 of 8

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

A Principal Odor Map (POM)
B Y = POM X dissimilar odorant was the more perceptually
y2 O
similar of the two to the anchor. To visualize
the internal logic of the model, we made small
H
changes to each node in a molecule and ob-

Pe
served which perturbations had large effects

rc
ep
Descriptor 2 Applicability
on model predictions; perturbations in nodes

tu
al
known odorants
with a darker red highlight had a larger impact

D
ist
an
O (Fig. 4A and fig. S19). Explicit similarity ratings

e c
agreed with odor profile distances in 90% of
the triplets (Fig. 4D). The model correctly pre-
ility dicted this counterintuitive structure-odor rela-
tab
etec tionship in 50% of cases (Fig. 4E), whereas the
orD RF model failed in 81% of cases (binomial test
Od
of proportions, p < 0.01; Fig. 4F).
y1 A reliable structure-odor map allows us to
Descriptor 1 Applicability
explore odor space at scale. We compiled a list of
~500,000 potential odorants whose empirical
C Descriptor Applicability properties are presently unknown to science
Dravnieks (35) Keller et al. (6) Current data or industry; most have never been synthesized
before. Because a molecule’s coordinates in
Target Correlation (R)

p
the POM are directly computable from the
model, we can plot these potential odorants
in the POM (Fig. 5A), revealing a potential
space of odorous molecules that is much larger
than the space covered by current fragrance
catalogs (~5000 purchasable, characterized

D Odor Detectability
g
odorants; inset of Fig. 5A). These molecules
would take ~70 person-years of continuous
smelling time to collect using our trained hu-
man panel.
E Perceptual Distance We show that the POM has a meaningful

y
Abraham et al. (36) Snitz et al. (4)
interpretation by extracting intuitive, geomet-
Target Correlation (R)

Target Correlation (R)

ric measures and mapping them to several

olfactory prediction tasks (Fig. 5B). The appli-
cability of any set of odor descriptors corre-
sponds to a projection of the POM coordinates
onto axes corresponding to those descriptors;
odor strength (detectability) corresponds to
the magnitude of this projection (fig. S13), and
odor similarity corresponds to the distance
between such projections for different mole-

y g
cules. A simple linear model applied to the
Fig. 5. The POM solves a fundamental set of olfactory prediction tasks. (A) A two-dimensional trimap
POM that uses these geometric interpretations
embedding of 500,000 unique likely odorants that were previously uncharacterized. The position of each point
had comparable or superior performance to
(molecule) is determined by POM coordinates, and the RGB values of each point correspond to their coordinates
a chemoinformatic support vector machine

p
in the first three dimensions of a non-negative matrix factorization of the predicted odor labels. The inset plot
(SVM) model across multiple published data-
shows the known odorants from the GS-LF training set (~5000) in color superimposed over the likely odorants in
sets (Fig. 5, C to E), collectively representing
gray. (B) Intuitive geometric measures like vector length, vector distance, and vector projection correspond to the
some of the most thorough previous public
odor prediction tasks of odor detectability, similarity, and descriptor applicability. The equation shows that the
,
efforts to characterize these features of odor.
projected space Y represents the dot product between POM and a task-specific projection matrix X. (C) A linear
model using POM coordinates outperforms a chemoinformatic SVM baseline at predicting odor applicability Discussion
on two extant datasets, provided by Dravnieks (35) and the DREAM Olfaction Prediction Consortium (6), as well as
There is no universally accepted method for
the current data. Error bars represent the standard error of the mean over descriptors. (D) A linear model
quantifying and categorizing an odor percept.
using POM coordinates outperforms a chemoinformatic SVM baseline at predicting an odor detection
Several systems of odor classification have been
threshold using data from Abraham et al. (36). (E) A linear model using POM coordinates outperforms a
proposed: first, intuitive categorizations (26);
chemoinformatic SVM baseline at predicting perceptual similarity using data from Snitz et al. (4).
then empirically supported universal spaces
(27, 28); and later, attempts to incorporate
the overall rate of significant odorous contam- designed an additional challenge in which 41 receptor mechanisms (29, 30). However, these
ination in our stimulus set at 31.5% (95% con- new triplets (Fig. 4, A and B) were constructed systems do not tie stimulus properties to percep-
fidence interval: 27.4 to 35.6%) (fig. S18). and validated by the panel (as in Figs. 1A and tion, and none have reached broad acceptance.
4C). In each triplet, the anchor molecule was Here, we propose and validate a data-driven,
The POM generalizes to diverse olfactory tasks a known odorant and was matched with one high-dimensional map of human olfaction. This
To test whether the POM was robust to dis- structurally similar and one structurally dissim- map recapitulates the structure and relation-
continuities in structure-odor distances, we ilar novel odorant, where the more structurally ships of odor perceptual categories evoked by

Lee et al., Science 381, 999–1006 (2023) 1 September 2023 6 of 8

RES EARCH | R E S E A R C H A R T I C L E
single molecules. It achieves prospective pre-
dictive accuracy in odor description that ex-
ceeds that of the typical individual human, and
it is broadly transferrable to arbitrary olfactory
perceptual tasks using natural and interpret-
able transformations.
Nearly all of the published chemosensory
models were fit to the data used in their con-
struction. Even when using cross-validation,
the opportunity for overfitting is high because
the data come from a single distribution, task,
or experimental source. Prospective validation
on new data from a new source with no adjust-
ments represents a much more stringent test of
real-world utility. In this prospective context,
we found that our model performs roughly on
par with the median human panelist, beating a
chemoinformatic baseline. However, in a real-
world setting, models can and should be up-
dated as new data become available (31). A linear
model using the POM coordinates reaches an
even higher level of performance when the POM
is tuned to the new dataset (Fig. 5C).
The success of this model is not merely an
advance in predictive modeling. It offers a sim-
ple, contiguous, hierarchical, parseable map of
molecular space in terms of odor, much as color
spaces represent wavelengths of light in terms
of colors and color components. This model not
only enables human-level performance for odor
description but also generalizes to a gamut of
other olfactory tasks. It offers the opportunity
to reason, intuitively and computationally, about
the relationships within and between molec-
ular and odor spaces. Unlike well-known color
spaces, this model does not provide clear guid-
ance about how stimuli can be mixed to produce
new percepts nor does it use a biologically plau-
sible architecture. Its closest analogy in vision
is the Munsell color system, which is a princi-
pled way to describe a color in terms of coor-
dinates but lacks any specific guarantees about
mixture behavior. Nonetheless, the Munsell
system (and we hope the POM) is still consid-
ered to be a useful representation of sensory
perception. Further work can aim for a CIE-
like representation of odor—one that specifi-
cally predicts what odors can be made from
mixing specified components.
There are some practical considerations to
keep in mind when using this map. First, the
concentration of an odor influences odor char-
acter but is not explicitly included in the map.
Although it can predict a detection threshold,
which is a property of the odorant molecule, it
cannot predict suprathreshold intensity, which
is a function of the odorant and its concentra-
tion. Many molecules have no odor, which we
addressed by prescreening with a separate,
simpler model (32), and we diluted odorants
to standardize intensity. Second, predictive
performance is strong for organic molecules,
which are most of the odorants that we en-
counter, but we could not extend the predic-
Lee et al., Science 381, 999–1006 (2023) 1 September 2023
tions into halides or molecules that include
functional groups without extensive safety
data. Given the uniformly strong performance
across the broad chemical classes that we
tested in our prospective validation set (Fig. 3C),
we expect high accuracy on novel chemicals
within these chemical classes, but we would
not expect high performance for molecules
that have chemical motifs that were not rep-
resented in our training set. For instance, if
our training dataset did not contain any mol-
ecules with carbon macrocycles, we would not
expect the model to accurately predict the odor
of an unseen macrocyclic musk (Fig. 3A). Third,
many chemical stimuli have odorous contam-
inants (25), particularly those that have not
been developed for use in fragrance applica-
tions. Neural networks perform well, even with
substantial noise in the training and test sets,
which we see in the present work. Nonethe-
less, we recommend isolating the compound of
interest from odorous contaminants and/or
characterizing the perceptual quality of con-
taminants. Fourth, the model was designed to
predict the population average, much as color
maps predict average perception. It does not
yet account for individual differences in per-
ception. Finally, datasets in real-world settings
are not static but grow and shift in distribution;
thus, models should be periodically retrained
to incorporate new data. Model performance
tends to improve with increased training data
(Fig. 3B) and data quality (fig. S15), consistent
with machine-learning applications in other
areas (33, 34).
Progress in neuroscience is often measured
by the creation and discovery of new maps of
the world supported by neural circuitry—maps
of space in hippocampus, tonotopy in auditory
cortex, and retinotopy and Gabor filters in V1
visual cortex, among others. Each is only pos-
sible because scientists first possessed a map
of the external world and then measured how
responses in the brain varied with stimulus
position on the map. This study proposes and
validates a data-driven map of human olfac-
tion. We hope this map will be useful to re-
searchers in chemistry, olfactory neuroscience,
and psychophysics—first, as a drop-in replace-
ment for chemoinformatic descriptors and,
more broadly, as a new tool for investigating
the nature of olfactory sensation.
RE FERENCES AND NOTES
1. T. Smith, J. Guild, Trans. Opt. Soc. 33, 73–134 (1931).
2. E. F. Evans, Psychophys. Physiol. Hear. 79, 185–192 (1977).
3. C. S. Sell, Angew. Chem. Int. Ed. 45, 6254–6261 (2006).
4. K. Snitz et al., PLOS Comput. Biol. 9, e1003184 (2013).
5. A. Ravia et al., Nature 588, 118–123 (2020).
6. A. Keller et al., Science 355, 820–826 (2017).
7. J. Gilmer, S. S. Schoenholz, P. F. Riley, O. Vinyals, G. E. Dahl,
in Machine Learning Meets Quantum Physics, K. T. Schütt et al.,
Eds., vol. 968 of Lecture Notes in Physics (Springer, 2020),
pp. 199–214.
8. B. Sanchez-Lengeling, E. Reif, A. Pearce, A. Wiltschko, Distill
2021, 1 (2021).
9. H. L. Morgan, J. Chem. Doc. 5, 107–113 (1965).
10. A. Krizhevsky, I. Sutskever, G. E. Hinton, Commun. ACM 60,
84–90 (2017).
11. F. Schroff, D. Kalenichenko, J. Philbin, in Proceedings of the
IEEE Conference on Computer Vision and Pattern Recognition
(IEEE, 2015), pp. 815–823.
12. N. Jaitly, P. Nguyen, A. Senior, V. Vanhoucke, in Proceedings
of Interspeech 2012 (International Speech Communication
Association, 2012).
13. The Good Scents Company, The Good Scents Company
information system; http://www.thegoodscentscompany.com/.
14. Leffingwell & Associates, PMP 2001 - Database of perfumery
materials and performance; http://www.leffingwell.com/
bacispmp.htm.
15. D. P. Kingma, J. Ba, Adam: A method for stochastic
optimization. arXiv:1412.6980 [cs.LG] (2014).
16. D. Golovin et al., in Proceedings of the 23rd ACM SIGKDD
International Conference on Knowledge Discovery and Data
Mining (Association for Computing Machinery, 2017),
pp. 1487–1495.
17. B. Sanchez-Lengeling et al., Machine learning for scent:
Learning generalizable perceptual representations of small
molecules. arXiv:1910.10685 [stat.ML] (2019).
18. S. Kearnes, Trends Chem. 3, 77–79 (2021).
19. H. T. Lawless, H. Heymann, Sensory Evaluation of Food: Principles
and Practices, Food Science Text Series (Springer, 2010).
20. C. Trimmer et al., Proc. Natl. Acad. Sci. U.S.A. 116, 9475–9480
p
(2019).
21. A. Keller, M. Hempstead, I. A. Gomez, A. N. Gilbert, L. B. Vosshall,
BMC Neurosci. 13, 122 (2012).
22. A. Dravnieks, Science 218, 799–801 (1982).
23. K. J. Rossiter, Chem. Rev. 96, 3201–3240 (1996).
24. O. R. P. David, Chemistry 26, 7537–7555 (2020).
25. M. Paoli et al., eNeuro 4, ENEURO.0070-17.2017 (2017).
26. H. Zwaardemaker, Die Physiologie Des Geruchs (Рипол
Классик, 1895).
g
27. H. Henning, Der Geruch (J. A. Barth, 1916).
28. E. C. Crocker, L. F. Henderson, Analysis and Classification of
Odors: An Effort to Develop a Workable Method (Robbins
Perfumer Company, 1927).
29. M. Guillot, C. R. Hebd. Seances Acad. Sci. 226, 1307–1309
y
(1948).
30. J. E. Amoore, in Olfaction and Taste III: Proceedings, C. Pfaffmann,
Ed. (Rockefeller Univ. Press, 1969), pp. 158–171.
31. G. I. Parisi, R. Kemker, J. L. Part, C. Kanan, S. Wermter, Neural
Netw. 113, 54–71 (2019).
32. E. J. Mayhew et al., Proc. Natl. Acad. Sci. U.S.A. 119, e2116576119
(2022).
33. T. B. Brown et al., Language models are few-shot learners.
arXiv:2005.14165 [cs.CL] (2020).
34. G. Branwen, The scaling hypothesis (2020); https://gwern.
net/scaling-hypothesis.
35. A. Dravnieks, Atlas of Odor Character Profiles, ASTM Data
Series (ASTM, 1985).
y g
36. M. H. Abraham, R. Sánchez-Moreno, J. E. Cometto-Muñiz,
W. S. Cain, Chem. Senses 37, 207–218 (2012).
AC KNOWLED GME NTS
We acknowledge Z. Mariet for experimental design; Y. Halpern,
p
B. Datta, S. Kearnes, C. Zelano, A. Morcos, D. Bear, and A. Koulakov
for feedback on the draft; J. S. Elmore for contributions to the
GC-MS-O analyses; and C. Laudemiel for domain expertise.
Funding: This work was funded by Google Research, National
,
Institutes of Health grant F32 DC019030 (E.J.M.), and National
Institutes of Health grant T32 DC000014 (E.J.M.). Author
contributions: Conceptualization: A.B.W.; Methodology: B.K.L.,
E.J.M., R.C.G., J.D.M., B.S.-L., J.K.P.; Software: B.K.L., B.S.-L., R.C.G.,
J.N.W.; Validation: R.C.G., B.K.L.; Formal analysis: B.K.L., E.J.M., R.C.G.,
J.Y.; Investigation: E.J.M., K.A.L., M.A., B.B.N., T.M., J.K.P., J.D.M.,
B.S.-L., W.W.Q., J.N.W.; Data curation: B.K.L., B.S.-L., R.C.G., E.J.M.,
M.A., K.A.L., J.K.P.; Writing – original draft: E.J.M., B.K.L., A.B.W.,
J.D.M., R.C.G.; Writing – review and editing: E.J.M., B.K.L., A.B.W.,
J.D.M., R.C.G., J.K.P., B.S.-L., W.W.Q., J.N.W.; Visualization: E.J.M.,
R.C.G., B.K.L., B.S.-L.; Supervision: A.B.W., J.D.M., E.J.M., J.K.P.;
Funding acquisition: A.B.W., J.D.M., E.J.M.; Project administration:
A.B.W., J.D.M. Competing interests: The original work and funding
for this manuscript was provided by Google Research. B.K.L.,
J.N.W., B.S.-L., W.W.Q., R.C.G., and A.B.W. were employees of
Google at the time this study was conducted. During the review
process, A.B.W., R.C.G., and W.W.Q. joined Osmo Labs, PBC,
a new venture that is commercializing some of the technologies
described in this manuscript. Provisional patents have been filed
on some of these technologies and derivatives thereof based
on additional development at Osmo Labs, PBC. A.B.W., R.C.G., and
7 of 8
RES EARCH | R E S E A R C H A R T I C L E

W.W.Q. each have an ownership interest in Osmo Labs, PBC, and
receive a salary from the company. A.B.W. is an officer of the
company. J.D.M. received funding from Google and serves on the
scientific advisory board of Osmo Labs, PBC. E.J.M., K.A.L., M.A.,
and B.B.N. received funding from Google. Data and materials
availability: Human psychophysics data, model predictions, and
model embeddings for tested odorants are included in the
supplementary data files and will be deposited at the olfactory data
repository pyrfume.org upon publication. Lightweight reproduction
notebooks, scripts, and data are shared at https://github.com/
osmoai/publications. Datasets used for transfer learning analyses
(Fig. 5, C to E), which are published in cited references
(4, 6, 35, 36), are also available from pyrfume.org. All material
needed to critique and replicate this work can be obtained through SUPPLEMENTARY MATERIALS
a materials transfer agreement from Osmo Labs, PBC, for science.org/doi/10.1126/science.ade4401
noncommercial use. The code for implementing and training the Materials and Methods
GNN model can be obtained through a software license agreement Supplementary Text
from Osmo Labs, PBC, for noncommercial use. All code and Figs. S1 to S21
processed data that are necessary to replicate every figure that Table S1
does not require an internal model state are provided at the GitHub References (37–45)
link above. The model architecture and hyperparameters are MDAR Reproducibility Checklist
provided in the supplementary materials. License information: Data S1 to S7
Copyright © 2023 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/ Submitted 16 September 2022; accepted 3 August 2023
about/science-licenses-journal-article-reuse 10.1126/science.ade4401

p
g
y
y g
p
,

Lee et al., Science 381, 999–1006 (2023) 1 September 2023 8 of 8

RES EARCH

PLANT SCIENCE sites (Fig. 1A and fig. S1B). Whereas the pri-
mary roots of lzy2;3;4 triple mutants tended to
Cell polarity linked to gravity sensing is generated grow upward (2), the expression of LZY3 fused
with mCherry (LZY3-mCherry) harboring A-7Q,
by LZY translocation from statoliths to the A-10Q, and B-7Q substitutions, as well as wild
type (WT), rescued the primary root angle
plasma membrane phenotypes of lzy2;3;4 under the control of the
LZY3 promoter (Fig. 1B and fig. S1, D and E).
Takeshi Nishimura1,2†, Shogo Mori1†, Hiromasa Shikata1,2†, Moritaka Nakamura1, Plants expressing LZY3 B-11Q exhibited varia-
Yasuko Hashiguchi3‡, Yoshinori Abe4, Takuma Hagihara4, Hiroshi Y. Yoshikawa5, ble growth directions (Fig. 1B and fig. S1D).
Masatsugu Toyota4,6,7, Takumi Higaki8,9, Miyo Terao Morita1,2* The LZY3-mCherry signals of WT, A-7Q, and
B-7Q were faint in the living columella cells of
Organisms have evolved under gravitational force, and many sense the direction of gravity by means the primary roots. In contrast, the signals of
of statoliths in specialized cells. In flowering plants, starch-accumulating plastids, known as amyloplasts, A-10Q and B-11Q exhibited observably in-
act as statoliths to facilitate downstream gravitropism. The gravity-sensing mechanism has long creased fluorescence to different extents (Fig.
been considered a mechanosensing process by which amyloplasts transmit forces to intracellular 1C and fig. S1, C and G). The fluorescence sig-
structures, but the molecular mechanism underlying this has not been elucidated. We show here that nals of A-10Q accumulated at the lower side
LAZY1-LIKE (LZY) family proteins involved in statocyte gravity signaling associate with amyloplasts of the PM, whereas the signals of B-11Q were
and the proximal plasma membrane. This results in polar localization according to the direction of mainly detected in the cytosol (Fig. 1, C and D).
gravity. We propose a gravity-sensing mechanism by which LZY translocation to the plasma membrane Combination of B-7Q with A-7Q or A-10Q in-

p
signals the direction of gravity by transmitting information on the position of amyloplasts. creased the LZY3-mCherry levels and the rel-
ative fluorescence intensities in the cytosol,

P
lants sense their inclination relative to
the direction of gravity and reorient their
growth direction accordingly, a process
which is consistent with their lower degree of
In Arabidopsis, LZY proteins have been shown phenotypic rescue in gravitropism (Fig. 1, C
to play critical roles in the signaling process and D, and fig. S1, C, F, and G). LZY3-mCherry
controlling directional auxin flow after amylo- that has both A-10Q and B-11Q was mostly

called gravitropism (1, 2). The directional
change is sensed mainly by specialized
cells called statocytes. Statocytes contain amylo-
plasts, which are plastids filled with dense starch
granules (3, 4). Amyloplasts act as statoliths
plast displacement in statocytes in both shoots
and roots (16). Although there are no known
functional domains in LZY proteins, five con-
served regions with 10 to 20 amino acids were
found (fig. S1A) (2, 10). Mutational analyses
g
found in the cytosol, as was the version with
only B-11Q (Fig. 1C). These findings suggest
that sites B and A, including their surrounding
sequences, play major and minor roles, re-
spectively, in PM association of LZY3, possibly

within cells by displacing toward gravity. Amylo-
plast displacement is believed to promote auxin
transport in the direction of gravity, leading to
organ bending (5). Several hypotheses have been
proposed regarding the sensing mechanism by
which the physical process of amyloplast displace-
ment generates biochemical signals in stato-
cytes, including the force-sensing model (6) and
position sensor hypothesis (7). However, the mo-
lecular mechanisms of gravity sensing and sig-
have demonstrated the functional importance
of regions I, II (also known as IGT motif) (15, 17),
and V (CCL region) (16, 18). LZY1, rice LAZY1,
and maize ZmLA1 localize to the plasma mem-
brane (PM) and the nucleus in ectopic or tran-
sient expression systems. As a result, possible
functions in the nucleus in addition to the PM
have been suggested (14, 19, 20). It should be
noted that, under the control of its own pro-
moter, LZY3 is polarly localized to the PM in
y
through electrostatic interactions with anionic
membrane lipids, and that both regions help
maintain LZY3 at low levels (fig. S1G).
A putative PEST motif (i.e., a motif enriched
with Pro, Glu, Ser, and Thr), which may facil-
itate protein degradation via the proteasome
pathway, is found in the vicinity of site A, im-
plying that it may be involved in the regulation
of LZY3 levels (fig. S2A) (26, 27). The predicted
PEST score of A-10Q is reduced compared with

naling remain largely unknown.
LAZY1-LIKE (LZY) family genes, which are
conserved in most land plants, are involved
in gravitropism and whole-plant architecture
the direction of gravity and repolarizes in re-
sponse to the reorientation in statocytes in the
columella cells of lateral roots (21). This find-
ing prompted us to investigate the mechanism
y g
wild-type LZY3, whereas that in A-7Q is not
affected (fig. S2A). As expected, LZY3DPEST-
mCherry, which lacks the PEST motif without
impairing site A itself, displayed similarities to

through the control of branch angles in various
species in angiosperms, including trees (8–15).
1
Division of Plant Environmental Responses, National
Institute for Basic Biology, Okazaki 444-8585, Japan.
2
Course for Basic Biology, The Graduate Institute for
Advanced Studies, SOKENDAI, Hayama 240-0115, Japan.
3
Graduate School of Bioagricultural Sciences, Nagoya
University, Nagoya 464-8601, Japan. 4Department of
Biochemistry and Molecular Biology, Saitama University,
Saitama 338-8570, Japan. 5Department of Applied Physics,
Osaka University, Suita 565-0871, Japan. 6Suntory Rising
Stars Encouragement Program in Life Sciences (SunRiSE),
Suntory Foundation for Life Sciences, Kyoto 619-0284,
Japan. 7Department of Botany, University of Wisconsin,
Madison, WI 53706, USA. 8Faculty of Advanced Science and
Technology, Kumamoto University, Kumamoto 860-8555,
Japan. 9International Research Organization for Advanced
Science and Technology, Kumamoto University, Kumamoto
860-8555, Japan.
*Corresponding author. Email: mimorita@nibb.ac.jp
†These authors contributed equally to this work.
‡Present address: JOCAVIO Co. Ltd., Fukuoka 839-0864, Japan.
of LZY3 polarization to understand gravity sens-
ing and signaling in gravitropism.
Regions for functional regulation of LZY3
We determined the regions responsible for mem-
brane association in the LZY3 protein, which
lacked detectable transmembrane domains (by
InterProScan) (22). Basic-hydrophobic clusters
in proteins can facilitate their membrane asso-
ciation in eukaryotes, including plants (23–25).
A computational search indicated that all LZY
proteins in Arabidopsis have more than two
regions with a high basic-hydrophobic cluster
score (≥0.6) that have the potential to serve
as unstructured membrane-binding sites (fig.
S1, A and B) (23). Glutamine (Gln) substitu-
tions were introduced at two sites, designated
sites A and B, in LZY3 that decreased the basic-
hydrophobic cluster scores of the corresponding
p
A-10Q in terms of root growth angle, localiza-
tion pattern within columella cells, and accu-
mulation level within those cells (fig. S2, B to
,
D). These findings suggest that the accumula-
tion of LZY3-mCherry on the PM in A-10Q lines
resulted from impairment of the PEST motif,
thereby disturbing quantitative regulation of
LZY3 on the PM through proteasomal protein
degradation. The addition of a farnesylation site
[K8-far (28)] at the C termini of LZY3 mutant
proteins B-11Q and A-10Q;B-11Q overrode the
effect of their mutation on localization and ac-
tivity (fig. S3). These results suggest that the
B-11Q and A-10Q;B-11Q proteins retain all func-
tions of the wild-type LZY3 with the exception
of association with the PM. In addition, when
K8-far was present, the level of the B-11Q protein
was reduced, implying a link between the mem-
brane localization and quantitative regulation

Nishimura et al., Science 381, 1006–1010 (2023) 1 September 2023 1 of 5

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

A C WT #1-4 A-10Q #3-1 of LZY3. Taken together, these findings indi-

Root tip cate that the PM localization of LZY3 is re-
101 Site A 125
WT
QC quired for its function in gravity signaling.

columella
A-7Q T1

cells
T2
A-10Q T3 LZYs associate with amyloplasts and the
Site B T4
168 200 nearby PM
WT
B-7Q Meanwhile, we noticed faint ring-shaped fluo-
B-11Q
g rescence signals in the tier 3 columella cells
of wild-type LZY3-mCherry, which were ob-
B 180˚ 250 400 250 1000
served at the position of amyloplasts (Fig. 1C).
90˚ Given that LZY localizes to amyloplasts, LZY
n=105 n=108 n=19 B-11Q #19-3 A-10Q; B-11Q #17-3
g
0˚
0 might help link gravity sensing to the down-
0.2 stream signaling. However, the weak fluorescence
0.4
intensity of LZY3-mCherry made it difficult to
*
** lzy2;4 *
0.6
analyze the behavior of LZY3 in tier 2 cells,
0˚ Wild-type lzy2;3;4 0.8
1.0
(Col-0) which are the major contributors to gravitropism
(29). Therefore, we examined whether LZY4,
** ** *
#9-3 #14-1 #17-1 #12-2 which redundantly functions in root gravi-
n=27 n=27 n=22 n=25 tropism (30, 31), can be applied for live-cell im-
LZY3p::LZY3-mCherry / lzy2;3;4

aging analysis. There are four splicing variants

p
250 1279 250 4210 of LZY4 (annotated in the TAIR10 genome),
D
Ratio of fluorescence intensities

1.4 three of which lack the CCL region (fig. S4, A and
**
#1-4 #3-1
** #19-3
*
#17-3 B), which is important for LZY1/2/3 function
1.3
n=25 n=25 n=23 n=22
(16, 18, 21). We demonstrated that LZY4.4 with
(PM / whole cell)

1.2 a CCL-like sequence is the functional form in

1.1 gravitropism (fig. S4C). The CCL-like sequence of

g
LZY4.4 was capable of interacting with the
** ** * * 1.0 BREVIS RADIX (BRX) domain of RCC1-like do-
#17-2 #1-2 #5-2 #15-5 main proteins (RLDs) in the yeast two-hybrid
n=26 n=25 n=25 n=25
0.9 system (fig. S4D). RLD1 is an interacting part-
ner of LZY in the regulation of auxin transport

y
#14-1
#3-1
#1-2
#17-1
#19-3
#5-2
#18-1
#22-3
#7-1
#3-4
#4-5
#8-1
#7-3
#3-1
#9-3
#12-2
#17-3
#15-5
through membrane trafficking, and it is re-
A-10Q; cruited to the PM in a LZY2/3-dependent man-
A-10Q B-11Q A-7Q; A-10Q; A-7Q; A-10Q;
WT A-10Q B-11Q B-11Q B-7Q B-7Q B-11Q B-11Q ner (21, 32). LZY4.4 recruited RLD1 to the PM in
Arabidopsis protoplasts (fig. S5). In addition,
LZY4.4-mClover3 driven by the statocyte-specific
ADF9 promoter rescued the primary root angle
phenotype of lzy1;2;3;4 (fig. S6). These results
Fig. 1. A putative membrane association site in LZY3 is required for subcellular localization in the indicate that the molecular function of LZY4.4
columella cells and for root gravitropism. (A) Putative PM association sites in LZY3. The lines indicate (hereafter referred to as LZY4) is almost equiv-
sites with basic-hydrophobic scores ≥ 0.6 according to computational prediction. In A-7Q and A-10Q, alent to that of LZY3 in root statocytes.

y g
7 and 10 basic amino acid residues at and near site A, respectively, were replaced with glutamine Ring-shaped and linear fluorescence signals
residues. In B-7Q and B-11Q, 7 and 11 basic residues at and nearby site B, respectively, were replaced from functional LZY4p::LZY4-mScarlet were
with glutamine residues. The basic residues (K and R) and glutamine residue (Q) are colored magenta observed in the tier 2 columella cells of both
and green, respectively. Single-letter abbreviations for the amino acid residues are as follows: A, Ala; young lateral and primary roots (Fig. 2, A and

p
C, Cys; D, Asp; E, Glu; F, Phe; G, Gly; H, His; I, Ile; K, Lys; L, Leu; M, Met; N, Asn; P, Pro; Q, Gln; R, Arg; S, Ser; B, and fig. S7A). The linear fluorescence signal
T, Thr; V, Val; W, Trp; and Y, Tyr. (B) Absolute values of growth angle (q) of the primary root tips of merged with that from the PM-staining re-
5-day-old seedlings of wild-type, lzy2;4, lzy2;3;4, and transgenic lines expressing mCherry-tagged LZY3 agent (fig. S8), which confirms the PM local-

,
mutant proteins under the control of the LZY3 promoter (LZY3p::LZY3-mCherry) in the lzy2;3;4 mutant ization of LZY4. LZY4-mScarlet on the PM
background. The direction of each root tip was measured as the absolute value toward the direction of gravity was observed only proximal to amyloplasts,
(g). The frequency was calculated as the proportion of root numbers that fell within intervals of 15° to implying polarity in the direction of gravity
the total number of analyzed roots for each line (range: 0° to 180°). Asterisks indicate significant differences (Fig. 2A). LZY4 also has a basic amino acid–
[P < 0.0033 (0.05/15)] by Wilcoxon rank sum test with Bonferroni correction (*, compared with lzy2;4; rich region in the middle of the protein (basic-
**, compared with lzy2;3;4). (C) Representative confocal images of wild-type, A-10Q, B-11Q, and A-10Q;B-11Q hydrophobic cluster score ≥ 0.6; figs. S1A and
mutants of LZY3-mCherry in columella cells of primary roots. The area containing tier 1 to 3 columella S9A). LZY4-mScarlet carrying the B-6Q muta-
cells of the representative lines is shown. The seedlings were kept in a vertical position before and during tion in this region exhibited lower fluorescence
imaging. The roots of the B-11Q and A-10Q;B-11Q lines were straightened at least 3 hours before imaging at the PM, lack of polarity, and weaker capac-
to make the amyloplasts sediment toward the root tip, as those roots were meandering (fig. S1D). All images ity for phenotypic rescue, suggesting a PM-
were acquired with an identical optical setting, and the color-coded heatmaps for signal intensities appear binding mechanism and function similar to
below each image. Scale bars, 20 mm. QC, quiescent center. (D) Ratio of fluorescence intensities of LZY3 (Fig. 2, B and C, and figs. S7B and S9B). To
LZY3-mCherry at the PM to that of the whole cell. The ratios were calculated with mean values within a investigate the association of the ring-shaped
3-pixel-width line drawn on the PM region and within the area surrounded by the line. For each line, signal of LZY4-mScarlet with amyloplasts, we
19 to 41 cells with clear cell edges in 5 to 8 roots were measured. Wild-type, A-7Q, and B-7Q lines were conducted observations using green fluorescent
not analyzed because of low protein accumulation. protein–fused PERMEASE IN CHLOROPLASTS1

Nishimura et al., Science 381, 1006–1010 (2023) 1 September 2023 2 of 5

RES EARCH | R E S E A R C H A R T I C L E
(PIC1-GFP) and OUTER ENVELOPE MEM- expression levels of LZY4-mScarlet and the en-
BRANE PROTEIN 7 (OEP7-GFP), both of which velope markers, as well as technical limitations.
serve as plastid markers localized at an inner Next, we generated LZY4-mScarlet lacking
and outer envelope of plastids, respectively. the N-terminal domain I (LZY4-DI-mScarlet)
These observations were performed on both and found that the protein was present at the
living and fixed lateral root columella cells. PM and in the cytosol, but the ring-shaped sig-
LZY4-mScarlet was observed in close proxim- nal was hardly detected (Fig. 2E and fig. S7C).
ity to or partially overlapping with PIC1-GFP In addition, LZY4-DI-mScarlet was not fully
and OEP7-GFP, indicating that LZY4 likely lo- functional (Fig. 2B). Consistently, mCherry fused
calizes to the amyloplast envelopes or their im- with the N-terminal 54 amino acids of LZY3,
mediate vicinity (Fig. 2D and fig. S10). However, including the conserved domain I, displayed ob-
defining the exact localization of LZY4 in amylo- vious ring-shaped signals (fig. S11). These find-
plasts remains challenging because of the weak ings suggest that the N-terminal region including
A + BF LZY4-mScarlet B of LZYs to amyloplasts, possibly through their
WT lzy2;3 lzy2;3;4
n=64 n=56 n=64
g 1.0
LZY4-mScarlet LZY4(B-6Q)-mScarlet
LZY4(B-6Q)
C + BF -mScarlet
**
#2 **
#45 *
#7
n=85 n=90 n=57
g *
** ** * * *
#10 #12 #13
n=62 n=73 n=70
D E F by reorientation of plants on the localization
LZY4-mScarlet PIC1-GFP
+ BF
+ BF
g g
g servation, although amyloplasts were slightly
Merge Merge + BF
LZY4-mScarlet
/ lzy4 arl2
g and B; fig. S13C; and movie S2). Simultaneously,
Fig. 2. Amyloplast and basal plasma membrane localization of LZY4 in the columella cells. (A) Subcellular
localization of LZY4-mScarlet (#2)/lzy4 in lateral root columella cells. Arrows in the right panel indicate
the amyloplast-like ring shape localization of LZY4-mScarlet, and arrowheads indicate the localization to the
basal plasma membrane. BF, bright field. Scale bars, 10 mm. (B) Complementation test of LZY4-mScarlet/
lzy2;3;4, LZY4(B-6Q)-mScarlet/lzy2;3;4, and LZY4-DI-mScarlet/lzy2;3;4 in the direction of the primary root
tips of 5-day-old seedlings. Asterisks indicate significant differences [P < 0.0045 (0.05/11)] by Wilcoxon
rank sum test with Bonferroni correction (*, compared with lzy2;3; **, compared with lzy2;3;4). (C) Subcellular
localization of LZY4(B-6Q)-mScarlet (#7)/lzy2;3;4 in lateral root columella cells. Scale bars, 10 mm. (D) Localization
of LZY4-mScarlet and PIC1-GFP in amyloplasts. Scale bar, 10 mm. (E) Subcellular localization of LZY4-DI-
mScarlet (#12)/lzy2;3;4. Scale bars, 10 mm. (F) Subcellular localization of LZY4-mScarlet (#2) in the lzy4 arl2
mutant background. Scale bars, 10 mm.
Nishimura et al., Science 381, 1006–1010 (2023) 1 September 2023
domain I is involved in the amyloplast localiza-
tion of LZY3/4.
Several mutations in genes encoding com-
ponents of the translocon at the outer chloro-
plast membrane (TOC) complex enhance the
phenotype of altered response to gravity 1 (arg1)
and its paralog arg1-like2 (arl2) (33, 34). ARG1
and ARL2, encoding J-domain proteins, are in-
volved in gravity signaling in statocytes (35–37)
similarly to LZYs. We investigated the possible
involvement of ARG1 and ARL2 in the amylo-
plast localization of LZY4-mScarlet. In the arg1
and arl2 background, the LZY4-mScarlet signal
became undetectable in amyloplasts and at the
PM, and the signal was dispersed in the cyto-
sol (Fig. 2F and figs. S7, D and E, and S12). ARG1
and ARL2 may be involved in the recruitment
180˚
90˚ chaperone activity. We cannot rule out the
0˚ possibility that ARG1 and ARL2 might also be
0
p
0.2 involved in the PM localization of LZYs, be-
0.4 cause ARL2 is reported to localize at the PM
0.6
0.8 (36). Considering that precursor proteins tar-
geted to plastids interact with TOC through
their N-terminal transit peptides with the help
of molecular chaperones (38), mechanisms
g
* similar to plastid protein targeting could be in-
**
#8 volved in targeting LZYs to amyloplasts. Fur-
n=59
ther investigations into the precise localization
of LZYs in amyloplasts and their targeting
mechanisms are warranted to gain a deeper
y
understanding of the LZY-mediated interac-
tion between the amyloplast and PM.
**
#20 LZY4 polarity rapidly changes
n=65
upon gravistimulation
Next, we examined the effect of gravistimulation
of LZY4-mScarlet. We used live-cell imaging of
lateral root columella cells with a vertical stage
confocal microscope (39). Before gravistimu-
y g
lation, polar localization of LZY4-mScarlet on
the PM was maintained throughout the ob-
agitated (fig. S13B and movie S1). After gravi-
p
stimulation, LZY4-mScarlet appeared on the
PM of the new lower side of the cells at almost
the same time as amyloplast displacement, and
,
the signal gradually became stronger (Fig. 3, A
the LZY4-mScarlet signal on the PM in the orig-
inal direction of gravity gradually disappeared,
resulting in the generation of the new polarity
of LZY4. Such repolarization on the PM was not
observed with LZY4(B-6Q)-mScarlet (fig. S14 and
movie S3). The starchless phosphoglucomutase
(pgm) mutant has amyloplasts that hardly sedi-
ment, because of the lack of dense starch gran-
ules, and it exhibits reduced gravitropism (40–42).
In the pgm mutant, LZY4-mScarlet localized to
smaller starchless amyloplasts and to the PM.
LZY4-mScarlet failed to accumulate polarly on
the PM in the pgm mutant, probably because
of altered movement of amyloplasts (Fig. 3C,
3 of 5
RES EARCH | R E S E A R C H A R T I C L E

A B 0 min 15 min A Before 0 min 1 min B

Fluorescence intensity (a.u.)

135° 800

(red fluorescence)
LZY4-mEos2
600

400
g
C 0 min 15 min
200

0
0 min 1 min
g C 0s 29.5 s 92.8 s 156.0 s D 1.2
High

Bright field -mScarlet

F/F0 (LZY4-mScarlet)
1.0

LZY4
D 0.8
0.6
Weighted fluorescence

600 135° rotation

no rotation Low 0.4
intensity (a.u.)

500 0.2
0
400
-0.2
0 50 100 150 200 250
300 Time after manipulation started (s)
Fig. 4. Translocation of LZY4 from the amyloplast to the plasma membrane. (A and B) Observation of the
200

p
0 translocation of LZY4-mEos2 from the amyloplasts to the PM by the photoconversion method (A). Immediately after
0 5 10 15 photoconversion, there was almost no red fluorescence signal on the PM (white arrowhead at 0 min). At 1 min after
Time after rotation (min)
photoconversion, a linear red fluorescence signal was observed at the PM in close proximity to the amyloplasts
Fig. 3. Polar localization of LZY4 changes in (white arrowhead at 1 min). Scale bar, 10 mm. There was a significant difference between the findings at 0 and 1 min
response to gravistimulation. (A) Gravistimulation and (Wilcoxon rank sum test, P < 0.05). The increase in the red fluorescence signal on the PM at 1 min after
the site of observation. Stage 2 lateral roots (<3 mm) photoconversion was quantitatively analyzed and plotted as boxplots. Center lines show the medians; box limits
were used. (B and C) Representative confocal images indicate the 25th and 75th; whiskers extend 1.5 times the interquartile range from the 25th and 75th percentiles;

of the columella cells of LZY4p::LZY4-mScarlet in the lzy4
(B) or lzy4 pgm-1 (C) background immediately before
(0 min) or 15 min after gravistimulation. The seedlings
were kept in a vertical position before and during imaging.
g
outliers are represented by dots (B). The plots represent the results of eight independent experiments. See
also fig. S21B, showing images of mEos2 green fluorescence signals. (C and D) Temporal changes in the
localization of LZY4-mScarlet upon manipulating amyloplasts with an optical tweezer. Amyloplasts were fixed
with the optical tweezer and positioned near the PM (yellow arrowheads) by moving the roots immobilized on

mScarlet was excited at 561 nm, and the fluorescence
that passed through a 617/73 nm emission filter was
detected. Scale bars, 25 mm. (D) Temporal changes
in the localization of LZY4-mScarlet on the PM along
the direction of gravity. Images were taken at 20-s
intervals for 15 min. Eight cells from five of the nonrotated
lateral roots and 17 cells from 11 of the 135°-rotated
lateral roots were used for quantification. Error bars
indicate 95% confidence intervals. a.u., arbitrary units.
y
the stage. The representative change is shown in (C). Scale bar, 5 mm. The temporal changes in the signal of
mScarlet on the PM where the amyloplasts were positioned are quantitatively shown in (D). The plot represents
the mean ± standard error of three independent experiments.
a lower value means that LZY4-mScarlet polar- displacement occurred 3 min after gravistim-
izes in the direction of gravity. Without gravi- ulation in our observation system, as previously
stimulation, the weighted intensity gradually and reported (16). Polar localization of RLD1-GFP
slightly decreased, probably because of quench- merged with LZY4-mScarlet was observed at
ing during the 15-min observation. In contrast, 15 min after gravistimulation (fig. S18). There-

fig. S15, and movie S4), indicating that amylo-
plast sedimentation is required for the polar-
ization of LZY4 on the PM.
weighted intensity decreased after gravistimula-
tion, indicating repolarization of LZY4-mScarlet
in the direction of gravity. An incremental de-
crease in weighted intensity occurred immedi-
y g
fore, repolarization of LZY4-mScarlet is chrono-
logically consistent with a role in linking gravity
sensing to signaling. Although we tested relo-
cation of PIN3-GFP upon gravistimulation ac-

p
We quantified the temporal change in the lo- ately after gravistimulation, presumably linked cording to the measurement method reported
calization of the LZY4-mScarlet signal at the PM with amyloplast displacement. previously (21, 44), obvious relocation was not de-
along the direction of gravity (Fig. 3D and fig. To clarify the temporal relationship between tected within 15 min (fig. S19). As reported previ-

S16). Briefly, to focus on the change in the fluo-
rescence intensities on the PM along the vertical
line (y axis), the intensities on the PM were av-
eraged along the x axis (fig. S16, A to H). Then,
each averaged value along the y axis was multi-
plied by a value (between 1 and 100) assigned to
the same position on the y axis, which is linearly
distributed from the bottom to the top of the
intensity profile (fig. S16, I and J). The calculated
values were integrated for each time frame (fig.
S16K), and this weighting process was designed to
produce progressively lower integrated values
over time as LZY4-mScarlet moves toward the
direction of gravity. Thus, a higher value of
weighted intensity means that LZY4-mScarlet
polarizes at a higher position on the PM, whereas
the repolarization of LZY4 and other processes
in gravitropism, we examined LZY4-mScarlet
repolarization, auxin distribution, and organ
curvature with primary roots under as sim-
ilar conditions as possible (fig. S17). The be-
havior of LZY4-mScarlet in primary roots was
almost the same as that in lateral roots. Sig-
nificant repolarization of LZY4-mScarlet was
detected within 15 min after gravistimulation
(fig. S17, A to C, and movie S5). The asymmetric
distribution of auxin, detected using the ratio-
metric auxin biosensor R2D2 (43), appeared
16 min after gravistimulation, and it became sig-
nificant at 30 min (fig. S17, D and E). Significant
root curvature appeared 30 min after gravi-
stimulation (fig. S17F). In addition, amyloplast
,
ously, the polar pattern of PIN3-GFP in the root
cap columella after gravistimulation was observed
300 min after gravistimulation in the stage 2 lat-
eral root in our experimental conditions (21). This
implies that the LZY-RLD module may not direct-
ly regulate the trafficking of PIN3 but rather reg-
ulates other regulatory factors for auxin transport.
Amyloplast position determines LZY4 polarity
at the PM
Given that LZY interacts electrostatically with
negatively charged lipids, such lipids might
polarly distribute on the PM and contribute to
the polar localization of LZYs. Therefore, we
examined the distribution of several nega-
tively charged lipids in the statocytes of primary

Nishimura et al., Science 381, 1006–1010 (2023) 1 September 2023 4 of 5

RES EARCH | R E S E A R C H A R T I C L E
and lateral roots using fluorescent biosen-
sors (fig. S20). Only the phosphatidylinositol-
4,5-bisphosphate biosensor CITRINE-2xPHPLC
(45) tended to be polarly distributed toward the
root tip. However, clear repolarization of CITRINE-
2xPHPLC was not detected upon gravistimula-
tion in the examined time window. Therefore,
the distribution of these lipids is unlikely to be
related to the polarity of LZYs.
Next, to elucidate the relationship between
LZY4 residing in amyloplasts and its polariza-
tion at the PM, the behavior of LZY4 was ana-
lyzed using the LZY4p::LZY4-mEos2/lzy1;2;3;4
rescued line (fig. S21A). LZY4-mEos2 on the
amyloplasts was photoconverted from green
to red fluorescence by irradiation with a 405-nm
laser on a vertical stage confocal microscope.
The red fluorescence signal was observed only
in amyloplasts immediately after irradiation,
and a linear pattern emerged on the nearby PMs
(Fig. 4A; figs. S21, B and C, and S22; and movie
S6). The fluorescence intensity on the PM prox-
imal to amyloplasts significantly increased
1 min after irradiation (Fig. 4B). Although we
could not dismiss the possibility that a trace
amount of cytoplasmic LZY4-mEos2 was photo-
converted and accumulated at the PM, this re-
sult suggests that LZY4 is rapidly translocated
from amyloplasts to the PM. It is expected that
the position of amyloplasts determines the LZY4
accumulation site on the PM independent of
the direction of gravity. To test this possibility,
amyloplasts were manipulated by an optical
tweezer during observation by confocal laser
scanning microscopy (46). After trapping some
amyloplasts, the amyloplasts were manipulated
into close proximity of the PM (figs. S23 and S24
and movies S7 and S8). The fluorescence inten-
sity of LZY4-mScarlet significantly increased at
the PM region where the amyloplasts were new-
ly placed nearby (Fig. 4, C and D; fig. S23, G to L;
and movie S9), indicating that amyloplasts act
as determinants of LZY4 polarity on the PM.
Gravity sensing in gravitropism has long been
considered a mechanosensing mechanism
whereby the weight of statoliths exerts a force
on intracellular structures (6, 47). On the basis
of the experimental results using angiosperm
shoots, it has recently been suggested that the
sensor functions as a clinometer but not as a
force sensor (48). This supports the “position
sensor hypothesis,” which proposes that prox-
imity or contact between statoliths and the
membrane induces local auxin fluxes (7) and
Nishimura et al., Science 381, 1006–1010 (2023) 1 September 2023
that statocytes act as clinometers (49), although
there had been no known molecules support-
ing this hypothesis. We revealed in this study
that LZY exhibits behavior that fits the posi-
tion sensor hypothesis well. LZY polarity on the
PM formed according to the position of amylo-
plasts by translocation of LZY from amyloplasts
to the PM (fig. S25A). LZY appears to act as a
signal molecule that transmits the positional
information of statoliths to the PM, which di-
rectly links gravity sensing to subsequent signal-
ing processes. Polarly located LZY recruits RLD
to promote polar auxin flow in the direction of
gravity (fig. S25B) (21).
RE FERENCES AND NOTES
1. S.-H. Su, P. H. Masson, in Sensory Biology of Plants, S. Sopory,
Ed. (Springer, 2019), pp. 95–136.
2. N. Kawamoto, M. T. Morita, New Phytol. 236, 1637–1654 (2022).
3. F. D. Sack, Planta 203 (suppl. 1), S63–S68 (1997).
4. M. T. Morita, Annu. Rev. Plant Biol. 61, 705–720 (2010).
5. H. Han, M. Adamowski, L. Qi, S. S. Alotaibi, J. Friml, New Phytol.
232, 510–522 (2021).
6. G. Perbal, D. Driss-Ecole, Trends Plant Sci. 8, 498–504 (2003).
7. O. Pouliquen et al., Phys. Biol. 14, 035005 (2017).
8. P. Li et al., Cell Res. 17, 402–410 (2007).
9. T. Yoshihara, M. Iino, Plant Cell Physiol. 48, 678–688 (2007).
10. T. Yoshihara, E. P. Spalding, M. Iino, Plant J. 74, 267–279 (2013).
11. Z. Dong et al., Plant Physiol. 163, 1306–1322 (2013).
12. Y. Uga et al., Nat. Genet. 45, 1097–1102 (2013).
13. J. Salojärvi et al., Nat. Genet. 49, 904–912 (2017).
14. A. Ashraf et al., PLOS ONE 14, e0214145 (2019).
15. J. M. Waite, C. Dardick, Curr. Opin. Plant Biol. 59, 101983 (2021).
16. M. Taniguchi et al., Plant Cell 29, 1984–1999 (2017).
17. J. M. Guseman, K. Webb, C. Srinivasan, C. Dardick, Plant J. 89,
1093–1105 (2017).
18. T. Yoshihara, E. P. Spalding, Plant Physiol. 182, 1039–1051 (2020).
19. Z. Li et al., Mol. Plant 12, 1143–1156 (2019).
20. T. P. Howard III et al., PLOS ONE 9, e87053 (2014).
21. M. Furutani et al., Nat. Commun. 11, 76 (2020).
22. T. Paysan-Lafosse et al., Nucleic Acids Res. 51, D418–D427 (2023).
23. H. Brzeska, J. Guag, K. Remmert, S. Chacko, E. D. Korn, J. Biol.
Chem. 285, 5738–5747 (2010).
24. M. L. A. Simon et al., Nat. Plants 2, 16089 (2016).
25. I. C. Barbosa et al., Development 143, 4687–4700 (2016).
26. S. Rogers, R. Wells, M. Rechsteiner, Science 234, 364–368 (1986).
27. M. L. Spencer, M. Theodosiou, D. J. Noonan, J. Biol. Chem. 279,
37069–37078 (2004).
28. S. Clarke, Annu. Rev. Biochem. 61, 355–386 (1992).
29. E. B. Blancaflor, J. M. Fasano, S. Gilroy, Plant Physiol. 116,
213–222 (1998).
30. L. Ge, R. Chen, Nat. Plants 2, 16155 (2016).
31. T. Yoshihara, E. P. Spalding, Plant Physiol. 175, 959–969 (2017).
32. L. Wang et al., Nat. Commun. 13, 7 (2022).
33. J. P. Stanga, K. Boonsirichai, J. C. Sedbrook, M. S. Otegui,
P. H. Masson, Plant Physiol. 149, 1896–1905 (2009).
34. A. K. Strohm, G. A. Barrett-Wilt, P. H. Masson, Front. Plant Sci.
5, 148 (2014).
35. J. C. Sedbrook, R. Chen, P. H. Masson, Proc. Natl. Acad. Sci. U.S.A.
96, 1140–1145 (1999).
36. K. Boonsirichai, J. C. Sedbrook, R. Chen, S. Gilroy,
P. H. Masson, Plant Cell 15, 2612–2625 (2003).
37. B. R. Harrison, P. H. Masson, Plant J. 53, 380–392 (2008).
38. P. Chotewutmontri, K. Holbrook, B. D. Bruce, Int. Rev. Cell
Mol. Biol. 330, 227–294 (2017).
39. M. Nakamura, M. Toyota, M. Tasaka, M. T. Morita, Methods
Mol. Biol. 1309, 57–69 (2015).
40. J. Z. Kiss, R. Hertel, F. D. Sack, Planta 177, 198–206 (1989).
41. T. Caspar, B. G. Pickard, Planta 177, 185–197 (1989).
42. N. Saether, T.-H. Iversen, Planta 184, 491–497 (1991).
43. C.-Y. Liao et al., Nat. Methods 12, 207–210, 2, 210 (2015).
44. P. Grones et al., Sci. Rep. 8, 10279 (2018).
45. M. L. A. Simon et al., Plant J. 77, 322–337 (2014).
46. Y. Abe et al., Plant Biotechnol. (Tokyo) 37, 405–415 (2020).
47. G. Leitz, B.-H. Kang, M. E. A. Schoenwaelder, L. A. Staehelin,
Plant Cell 21, 843–860 (2009).
48. H. Chauvet, O. Pouliquen, Y. Forterre, V. Legué, B. Moulia, Sci. Rep.
6, 35431 (2016).
49. A. Bérut et al., Proc. Natl. Acad. Sci. U.S.A. 115, 5123–5128 (2018).
AC KNOWLED GME NTS
We are grateful to N. Kawamoto (National Institute for Basic
Biology), M. Furutani (Kumamoto University), K. Toyooka (RIKEN),
M. Sato (RIKEN), Y. Kodama (Utsunomiya University), and
H. Takeuchi (Nagoya University) for helpful discussions and to
T. Uemura (Ochanomizu University) for providing “Deep” cells. arg1-3
was a kind gift from P. Masson (University of Wisconsin–Madison).
mRFP1-Spo20p-PABD was a kind gift from M. Potocký (Czech
Academy of Sciences). The support of plant cultivation rooms was
provided by the Model Plant Research Facility of the National
Institute for Basic Biology. We also thank W. Takase, H. Motomura,
K. Miyoshi, M. Hamada, Y. Yamada, and Y. Soma for technical
assistance; the Salk Institute Genomic Analysis Laboratory for
p
providing the sequence-indexed Arabidopsis transfer DNA (T-DNA)
insertion mutants; and the Arabidopsis Biological Resource Center
and GABI-Kat for providing seeds of the A. thaliana T-DNA insertion
mutants and lipid biosensor lines (P5R and P24Y). Funding: This
work was supported by Japan Society for the Promotion of Science
(JSPS) KAKENHI, Grant-in-Aid for Scientific Research 19H03254
(M.T.M.); JSPS KAKENHI, Grant-in-Aid for Scientific Research on
Innovative Areas 18H05488 (M.T.M.), 18H05491 (M.T.), and
g
18H05492 (T.H.); Japan Science and Technology Agency (JST),
Core Research for Evolutionary Science and Technology (CREST)
JPMJCR14M5 (M.T.M.); Takeda Science Foundation (M.T.M.);
Naito Foundation (M.T.M.); JSPS KAKENHI, Grant-in-Aid for Scientific
Research 22H00302 (H.Y.Y.); JSPS KAKENHI, Grant-in-Aid for
Challenging Exploratory Research 20K21117 (H.Y.Y.); JSPS KAKENHI,
y
Grant-in-Aid for Scientific Research 20H03289 (T.H.); JST, CREST
JPMJCR2121 (T.H.); JSPS KAKENHI, Grant-in-Aid for Early-Career
Scientist 18K14731 (M.N.); and JSPS KAKENHI, Grant-in-Aid
for Early-Career Scientist 20K15826 (H.S.). Author contributions:
Conceptualization: M.T.M., T.N., and H.S. Design of experiments:
M.T.M., T.N., H.S., and M.T. Methodology: H.Y.Y. and M.T. Investigation:
T.N., H.S., S.M., M.N., Y.A., T.H., Y.H., and T.H. Funding acquisition:
M.T.M., M.N., M.T., T.H., H.Y.Y., and H.S. Project administration: M.T.M.
Supervision: M.T.M. Writing: M.T.M., T.N., H.S., M.T., and T.H.
Competing interests: The authors declare no competing financial
interests. Data and materials availability: All materials are available
upon request subject to a material transfer agreement. All data are
available in the main text or the supplementary materials. License
y g
information: Copyright © 2023 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
p
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.adh9978
Materials and Methods
Figs. S1 to S25
,
Table S1
References (50–70)
MDAR Reproducibility Checklist
Movies S1 to S9
Submitted 11 April 2023; accepted 2 August 2023
Published online 10 August 2023
10.1126/science.adh9978
5 of 5
RES EARCH

SILICONE CHEMISTRY [5 equivalents (equiv.)] relative to tBuOK/cryptand

initiator (0.5 mol % versus D4) (21, 22) at 80°C.
Ring-opening polymerization of cyclic oligosiloxanes Careful monitoring of the ROP demonstrated
that the polymer ratio to cyclic oligomers (D4-7)
without producing cyclic oligomers reached 95% after 6 hours and then gradually
decreased to 89% in 12 hours. This result sug-
Limiao Shi1, Aurélie Boulègue-Mondière2, Delphine Blanc2, Antoine Baceiredo1, gests that there is a certain effect preventing
Vicenç Branchadell3, Tsuyoshi Kato1* the backbiting reactions that gradually dis-
appears over time. This disappearance can be
A long-standing problem associated with silicone synthesis is contamination of the polymer products with 10 to attributed to the loss of coordinating alcohols
15% cyclic oligosiloxanes that results from backbiting reactions at the polymer chain ends. This process, in at the anionic chain ends either by chain trans-
competition with chain propagation through ring-opening polymerization (ROP) of cyclic monomers, was fer or by dissociation at the last stage of poly-
thought to be unavoidable and routinely leads to a thermodynamically controlled reaction mixture (polymer/ merization, leading to the conventional situation
cyclic oligosiloxanes = 85/15). Here, we report that simple alcohol coordination to the anionic chain ends with the propagation/backbiting equilibrium.
prevents the backbiting process and that a well-designed phosphonium cation acts as a self-quenching system Probably for this reason, the effect of alcohol
in response to loss of coordinating alcohols to stop the reaction before the backbiting process begins. The coordination has been ignored previously.
combination of both effects allows a thermodynamically controlled ROP of the eight-membered siloxane A possible solution is an automatic quench
ring D4 without producing undesirable cyclic oligosiloxanes. system responsive to the loss of coordinating
alcohols. For this purpose, we considered the

S
use of (i-propyl)tris(dialkylamino)-phosphonium

ince the discovery of polysiloxanes, com-
monly called silicones, in the first at-
tempt to synthesize a stable silanone
by F. Kipping in 1901 (1) and their first
industrial production by Dow-Corning
Corporation in 1943 (2), these organosilicon
methylcyclotetrasiloxane D4) (12). In this case,
with proper choice of initiator/catalyst and
conditions, the ROP of D3 is kinetically con-
trolled and achieves complete monomer con-
version (Fig. 1C, dashed line) (15–21), whereas
the anionic ROP of other monomers (D4-7) is
p
alkoxides 1-xROH that are stable only with the
alkoxide anion stabilized by coordination of al-
cohols through weak hydrogen-bonding interac-
tions (Fig. 2A) (23). Without a sufficient number
of stabilizing alcohols, the alkoxide readily at-
tacks the P atom of the phosphonium cation,

materials have found extremely wide-ranging
applications in different forms (oil, rubber, gel,
and resin) because of their particular chemical
and physical properties (3). Despite important
progress in their synthesis (4–6), a long-known
under thermodynamic control, affording the
conventional polymer/cyclic oligomer mixture
(Fig. 1C, solid line) (12). Thus, the achievement
of the ROP of D4 (most thermodynamically fa-
vored) without producing cyclic oligomers by
g
which, through a pentacoordinate intermediate
3, affords either amine 4 and phosphine oxide
5 (x = 0) or ether 6, amine 7, and phosphine
oxide 5 (x = 1) (Fig. 2B) (24). The use of such
phosphonium alkoxides 1-xROH as the initia-

but still unsolved drawback is their systematic
contamination with cyclic oligosiloxanes (~10 to
15%) regardless of the methods used (3, 4). Con-
sidering the potential toxicity of low-molecular-
weight cyclic siloxanes (7) capable of overcoming
the human skin barrier (8), it is critical to ad-
dress this contamination problem (9). Currently,
nearly 50% of new skin care products contain
at least one type of silicone (10).
The problem arises from backbiting reac-
somehow preventing the backbiting process
and altering the thermodynamic polymer/cyclic
oligomer ratio (from 85/15 to 100/0) (Fig. 1E)
would be a fundamental solution for this issue.
Here, we report that simple coordination of al-
cohols to the active silanolate chain ends pre-
vents the backbiting process (Fig. 1D), and that
well-designed phosphonium counter-cations act
as a smart self-quenching system in response
to the loss of coordinating alcohols to stop the
y
tor of ROP leads to the formation of silicones
with alcohol-stabilized anionic silanolate chain
ends that should behave in the same way. In
other words, once coordinating alcohols are lost,
the ionic pair (silanolate anion/phosphonium cat-
ion) at the chain end start reacting to gener-
ate a neutral chain end (Si-NR2 or Si-OR), thus
quenching the polymerization process.
Alcohol-free and -stabilized phosphonium
alkoxides 1-xBnOH were synthesized by the

tions at the polymer chain end to produce cy-
clic oligosiloxanes of different sizes [four- to
seven-membered rings (D4-7); Fig. 1A] (11, 12).
The rate of backbiting is competitive with the
reaction before the backbiting process starts.
The combination of these two effects allows the
thermodynamically controlled anionic ROP of
D4 without producing undesirable cyclic oligo-
y g
addition of benzyl alcohol (BnOH) to phos-
phonium ylides 2a-c with different dialkyla-
mino substituents (Fig. 2, A and C) (24). The
number of coordinating alcohols (x) can be

propagation through ring-opening of cyclic
monomers, leading to a backbiting/propaga-
tion equilibrium and therefore to a polymer/
cyclic oligomer mixture in a thermodynami-
cally controlled ratio (~85/15; Fig. 1B). This
means that in the polymer chemistry of sili-
cones, most of the reactions catalyzed by either
acid or base end with a ring-opening polymer-
ization (ROP) process as the final stage, leading
to the same contamination (4). An exception
is the ROP of hexamethylcyclotrisiloxane (D3)
(13, 14), which allows a substantially faster pro-
cess (100 to 1000 times faster than that of octa-
1
Laboratoire Hétérochimie Fondamentale et Appliquée (UMR
5069), Université de Toulouse, CNRS, F-31062 Toulouse,
France. 2Elkem Silicones, ATRiON, 69190 Saint-Fons, France.
3
Departament de Química, Universitat Autònoma de
Barcelona, 08193 Bellaterra, Spain.
*Corresponding author. Email: tsuyoshi.kato@univ-tlse3.fr
siloxane contaminants (Fig. 1E).
As mentioned previously, the origin of the
problem, incomplete ROP of cyclic oligosilox-
anes, is the competitive rate of backbiting rela-
tive to propagation. Thus, we hypothesized that
flexible coordination between weakly acidic
alcohols and the anionic silanolate chain end
should increase steric hindrance around the
nucleophilic site and modulate the relative
rates without excessively reducing the propa-
gation rate (Fig. 1D). Nucleophilic attack of the
anionic chain end on silicon centers of the
linear siloxane chains (i.e., the backbiting pro-
cess) could be significantly more hindered by
this effect than attack on the sterically more
accessible silicon atom of a rigid cyclic mono-
mer for propagation process. On the basis of
this hypothesis, we performed the ROP of D4
in the presence of an excess of benzyl alcohol
p
precisely controlled by the quantity of BnOH
(x + 1 equiv.) added to the ylide 2. As ex-
pected, they were thermally labile, and the
least-hindered variant, 1a-4BnOH, fully de-
,
composed within 1 hour (t1/2 = 15 min) at 80°C
in C6D6 (M = 0.07 mol/liter) to generate the
corresponding dibenzylether 6, dimethylamine
7, and phosphine oxide 5. This result indicates
that the decomposition proceeds with the in-
tervention of an alcohol [Fig. 2B, (x = 1)] and
thus starts before complete loss of stabilizing
alcohols. The decay gradually slowed down as
free benzyl alcohol in the medium increased
each time 1a-4BnOH decomposed. The ef-
fects of the phosphonium cation structure and
the quantity of coordinating alcohol on their
stability (at 80°C in C6D6) were then assessed
to regulate the stability/decomposition balance
(Fig. 2C).

Shi et al., Science 381, 1011–1014 (2023) 1 September 2023 1 of 4

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

p
g
y
y g
Fig. 1. Chain propagation/backbiting equilibrium. (A) Cylic oligosiloxanes of (D3); dashed line indicates D4 (12). (D) Strongly favored propagation process due
different sizes (D3-7). (B) Reversible processes: Propagation through ROP of cyclic to the hampered backbiting by the alcohol coordination on the silanolate end.
oligosiloxanes and backbiting of silanolate chain end. (C) Typical kinetic curves (E) Kinetic curves for the ROPs of cyclic oligosiloxanes (D3 and D4) improved by the
for the kinetically and thermodynamically controlled ROPs of cyclic oligosiloxanes effect of alcohols and of self-degradable phosphonium counter cation.

p
As mentioned above, the decomposition of posed at almost the same rate (t1/2(80°C) < 5 min) absence of stabilizing alcohols (1c-0BnOH,
phosphonium alkoxides 1-xBnOH starts with but with a poor selectivity due to the excessive 60% of decomposition in 24 hours), and thus

the nucleophilic attack of alkoxide on the cen-
tral P atom of phosphonium cation (Fig. 2B).
Therefore, the stability should be strongly re-
lated to the steric hindrance around the P atom.
Indeed, an enhanced persistence of BnOH-
stabilized phosphonium alkoxides 1a,b,c-
4BnOH with bulkier amino groups on the
P atom (1a < 1b < 1c) was clearly shown by
gradual slowdown of decomposition (Fig. 2C).
A strong acceleration of decomposition was
observed for monocoordinated phosphonium
alkoxides 1a,b-1BnOH (t1/2(80°C) < 5 min),
which is critical for using them as smart self-
quenching systems automatically triggered by
the loss of coordinating alcohols. Alcohol-free
phosphonium alkoxides 1a,b-0ROH decom-
reactivity of nonstabilized alkoxide ion. In this
case, in addition to the phosphine oxide 5, the
formation of phosphines was also observed.
The decomposition of phosphonium alkoxides
through b elimination triggered by deproto-
nation at the b position of the substituent by the
alkoxide to give a phosphine and an alcohol is
known (23, 24). High stability was observed for
the most-hindered variant, 1c-4BnOH, with
piperidino substituents, and it was not possible
to estimate its half-life because of the exceed-
ingly slow decomposition rate (4% of decom-
position at 80°C in 24 hours). Phosphonium
alkoxide 1c remained persistent, with a reduced
number of coordinating alcohols (1c-1BnOH,
28% of decomposition in 24 hours), even in the
,
it was not suitable for our purpose.
The ROP of D4 at 80°C in the presence of the
BnOH-stabilized phosphonium benzyloxide 1b-
4BnOH (0.5 mol % versus D4) as the initiator led
to complete conversion in 18 hours to produce a
clean silicone polymer (99.6%) with only a small
amount of D4 (0.4%) (Fig. 3A), as determined by
1
H- and 29Si-nuclear magnetic resonance (NMR)
analysis of the resulting mixture. Despite the un-
usually high conversion, the polymerization was
thermodynamically controlled, as suggested by
the large polydispersity index of the resulting poly-
mers (Ip = 1.54) (25, 26). The NMR analysis
demonstrated that the polymer chains are
end capped with the neutral Si-OBn group,
which would be expected to form through the

Shi et al., Science 381, 1011–1014 (2023) 1 September 2023 2 of 4

RES EARCH | R E S E A R C H A R T I C L E
Fig. 2. Synthesis and stability assessment of
initiators. (A) Synthesis of alcohol-stabilized
(i-propyl)(tris-amino)phosphonium alkoxides 1-xROH.
(B) Proposed mechanism for the decomposition of
phosphonium alkoxides 1-xROH (x = 0, 1). (C) Effects
of phosphonium cation (size) and of coordinating
alcohols on the stability of phosphonium benzyloxides
1-xBnOH. Half-life time of 1-xBnOH (x = 0, 1, or 4)
in C6D6 (M = 0.07 mol/liter) at 80°C. Half-life of
1-xROH was determined by 31P-NMR.
decomposition of ionic chain ends (x = 1; Fig. 2B).
After 18 hours, the system was totally de-
activated, as evidenced by unchanged poly-
mer size even after heating at 80°C for a
prolonged time (36 hours), as well as by the
absence of any further polymerizations upon
adding another equivalent of D4 to the mixture.
As expected, the stability of phosphonium
alkoxides substantially influenced the ROP of
D4 (Fig. 3A). Indeed, the least-hindered cation,
1a-4BnOH, was not persistent enough to com-
plete the ROP (10% silicone and 90% unreacted
D4). The ROP with the excessively stable initia-
tor 1c-4BnOH (persistent even in the absence
of stabilizing alcohols) led to a conventional
polymer/cyclic oligomer mixture (polymer/D4-7 =
83/17). These results clearly confirm the impor-
tance of an automatic quench system triggered,
Shi et al., Science 381, 1011–1014 (2023) 1 September 2023
before backbiting begins, by the loss of alcohols
to benefit from the effect of alcohol coordination
to prevent the cyclic oligomer formation.
The size of the resulting polymer can be
roughly regulated by the amount of alcohol
added (~1/5 of added ROH can effectively
give rise to polymer chains) (Fig. 3A) (27).
Therefore, the loss of coordinating alcohols
should result mainly from dissociation rather
than from chain transfer reactions. According to
density functional theory calculations on the mod-
el reaction [Me3Si-(OSiMe2)4-O–(BnOH)x1a+→
Me3Si-(OSiMe2)4-O–(BnOH)x-11a+ + BnOH, where
x = 1 to 4], the Gibbs free energy for the dis-
sociation of benzyl alcohol (BnOH) from the sila-
nolate chain end is small for x = 4 (1.6 kcal/mol)
and x = 3 (5.0 kcal/mol), whereas it is larger for
x = 2 (11.8 kcal/mol) and x = 1 (15.2 kcal/mol).
p
g
y
y g
p
More precise polymer size control can be achieved
using a disiloxane end-capping reagent, 1,3-
divinyl(tetramethyl)siloxane (M2Vi), without
,
losing the full conversion character of ROP.
Indeed, varying the amounts of M2Vi (2.5 to
10 equiv.) relative to 1b-4BnOH, the end-capped
silicone polymers with a Mn close to the expected
value (MnTheo) have been obtained with an
excellent yield (>99%) (Fig. 3B).
Polydimethylsiloxane (PDMS)–based initia-
tors (28, 29) are common tools for increasing
initiator solubility and minimizing impurities
in silicone synthesis. Therefore, we consid-
ered preparing a PDMS-based initiator using
short a,w-dihydroxy-polydimethylsiloxanes
[PDMSOH = HO(Me2SiO)8H] instead of al-
cohols. The addition of PDMSOH (5 equiv.) to
phosphonium ylide 2b affords the PDMS-based
3 of 4
RES EARCH | R E S E A R C H A R T I C L E
Fig. 3. Fully converted anionic ROP of D4. (A) Study of the effects of phosphonium cations on the ROP of
D4 with different initiators 1-xROH. (B) Study of polymer size control by the end-capping agent M2Vi in the
ROP of D4 with the phosphonium benzyloxide initiator b-4BnOH. (C) Synthesis of the PDMS-based initiator
1b-4PDMSOH and its tests for the full ROPs of D4 in the presence of M2Vi.
initiator 1b-4PDMSOH, which was stable at
room temperature and decomposed at 80°C in
2 hours (t1/2 = 15 min) (Fig. 3C). The ROP of D4
with this initiator (1.0 mol %) in the presence of
M2Vi (2.5 or 5 equiv. versus 1b-4PDMSOH) af-
forded silicones with yields >99% and with the
expected polymer sizes (Fig. 3C). After the reac-
Shi et al., Science 381, 1011–1014 (2023) 1 September 2023
tion, no traces of short oligomer HO(Me2SiO)8H
could be detected by NMR and gel-permeation
chromatography analysis, probably because of
its consumption by polycondensation processes.
These results demonstrate that the silanol func-
tion (Si-OH) in PDMS can also act as a stabil-
izer of anionic chain ends to prevent the cyclic
oligomer formation. We conclude that the full
ROP of D4 is actually achievable.
REFERENCES AND NOTES
1. F. Kipping, L. L. Lloyd, J. Chem. Soc. Trans. 79, 449–459 (1901).
2. N. R. Thomas, Silicon 2, 187–193 (2010).
3. J. M. Zeigler, F. W. G. Fearon, Silicon-Based Polymer Science (ACS, 1990).
4. M. A. Brook, Silicon in Organic, Organometallic, and Polymer
Chemistry (Wiley, 1999).
5. I. E. Markó et al., Science 298, 204–206 (2002).
6. A. M. Tondreau et al., Science 335, 567–570 (2012).
7. L. Xu, Y. Shi, N. Liu, Y. Cai, Sci. Total Environ. 505, 454–463 (2015).
8. K. Mojsiewicz-Pienkowska, in Handbook of Polymers for
Pharmaceutical Technologies, Processing And Applications,
V. K. Thakur, M. K. Thakur, Eds. (Wiley, 2015), vol. 2, pp. 363–382.
9. K. Mojsiewicz-Pieńkowska, M. Jamrógiewicz, K. Szymkowska,
D. Krenczkowska, Front. Pharmacol. 7, 132 (2016).
10. https://cosmethicallyactive.com/silicones-in-cosmetics-and-
their-impact-on-the-environment/
11. D. W. Scott, J. Am. Chem. Soc. 68, 2294–2298 (1946).
12. J. Chojnowski, in Silicon Compounds: Silanes and Silicones,
B. Arkles, G. Larson, Eds. (Gelest, 2008) pp. 389–405.
13. D3 is more reactive than other larger cyclics (D4-7) because of
its more strained structure. The ring strain energy was
calculated to be 11 kcal/mol higher than that of D4.
p
14. J. D. Kress, P. C. Leung, G. J. Tawa, P. J. Hay, J. Am. Chem. Soc.
119, 1954–1960 (1997).
15. J. C. Saam, D. J. Gordon, S. Lindsey, Macromolecules 3, 1–4 (1970).
16. H. J. Hölle, B. R. Lehnen, Eur. Polym. J. 11, 663–667 (1975).
17. J. G. Zilliox, J. E. L. Roovers, S. Bywater, Macromolecules 8,
573–578 (1975).
18. P. Boehm, M. Mondeshki, H. Frey, Macromol. Rapid Commun.
33, 1861–1867 (2012).
g
19. B. G. G. Lohmeijer et al., Org. Lett. 8, 4683–4686 (2006).
20. K. Fuchise, M. Igarashi, K. Sato, S. Shimada, Chem. Sci. 9,
2879–2891 (2018).
21. M.-H. Yang, C.-K. Yang, C. Chou, J. Chin. Chem. Soc. (Taipei)
40, 557–561 (1993).
22. The reaction conditions modified by the procedure proposed
y
by S. Boileau has been used.
23. S. Boileau, Makromol. Chem. Macromol. Symp. 73, 177–181 (1993).
24. P. A. Byrne, D. G. Gilheany, Chemistry 22, 9140–9154 (2016).
25. A. Bessmertnykh, F. Ben, A. Baceiredo, G. Mignani, J. Organomet.
Chem. 686, 281–285 (2003).
26. Ip was calculated using the equation: Ip = Mw/Mn where Mw is the
mass-average molar mass and Mn is the number-average molar mass.
27. R. F. T. Stepto, Pure Appl. Chem. 81, 351–353 (2009).
28. The number of ROHs used to initiate the ROP relative to the
5 ROHs available with the 1b-4BnOH initiator. This was estimated
by the ratio of the experimental Mn to the theoretical Mn (Mntheo)
expected by the ROP using 1 ROH of 1b-4BnOH.
29. Mixtures obtained by the reaction of D4 with KOH are commonly
y g
used as PDMS-based initiators for ROP of cyclic oligosiloxanes.
AC KNOWLED GME NTS
We thank CNRS for financial support. Funding: This work was
supported by the CNRS, Agence National de Recherche (ANR)
p
grant SILISCY (D.B., T.K.), and Spanish AEI grant PID2020-
116861GB-I00 (V.B.). Author contributions: L.S. performed all
of the synthetic work. V.B., A.B., and T.K. conducted the
computational and supervised the experimental program. A.B.-M.
and D.B. managed the project (SILISCY). All authors jointly
,
wrote the manuscript. Competing interests: The authors declare
no competing interests. The findings have been patented by
Elkem Silicones with L.S., A.B.-M., D.B., A.B., and T.K. listed
as authors [French patent application no. FR2302798, (2023)].
Data and materials availability: All data are available in the main
text or the supplementary materials. License information:
Copyright © 2023 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.adi1342
Materials and Methods
Figs. S1 to S172
Tables S1 to S4
References (30–39)
Submitted 5 April 2023; accepted 31 July 2023
10.1126/science.adi1342
4 of 4
 WORKING LIFE
By Faye Harwell

The cost of being a student

I
n my first semester as a teaching assistant, a student came to me with a sensitive request. The
professor required students to use electronic clickers to answer multiple-choice questions during
the lecture; participation would count toward their grade. But this student confided that she could
not afford the $50 for a clicker and $30 to register it. She offered to write her answers on paper and
give them to me after every lecture, or just accept a 0% for attendance, knowing that she would
need to excel in other aspects of the class to make up for it. She was proud of her work ethic and
sure she belonged in academia. She reminded me of myself—and of our responsibility as educators to
support poor students without singling them out.

p
For my bachelor’s, I had chosen to When I learned Ph.D. students
attend the school where I received are typically paid a stipend, I was
the most financial aid, which was delighted. And once I started grad-
a private university. I felt privi- uate work, my financial struggles

g
leged to study there. But I was dissipated. Many students found
surrounded by people who could the stipend meager, but I was an
easily afford it, and I struggled with expert at making ends meet, and I
feelings of isolation and insecurity. felt my classmates and I were now

y
I felt I needed to hide my lower in the same boat. I finally felt I be-
socioeconomic status to fit in. I lied longed in higher education.
about the many hours I worked at When that student told me she
part-time jobs, why I couldn’t go couldn’t afford a clicker, though, I
to the movies, and why I couldn’t realized sharing my earlier strug-
ask my family for money. gles could help low-income stu-
Most professors didn’t help dents feel less alone. I told her I
much, often requiring expensive had once been in her shoes. And, of
textbooks and class materials with course, I let her submit her answers
seemingly little thought to the bur-
“I realized sharing my earlier on paper and manually entered her

y g
den this placed on some students. attendance into the gradebook.
Sometimes I could find the books
at the library or buy less expen-
struggles could help low-income In the years since then, I’ve
continued to talk about my expe-
sive used versions. And sometimes students feel less alone.” riences when they seem relevant.
I sought accommodations from I still feel vulnerable opening up

professors, even though it felt uncomfortable to share that I
needed them. But it didn’t always go well. When I asked my
organic chemistry professor, for example, whether I could buy
an older version of the textbook, he tersely said no; only the
new textbook had the access code for the homework software.
As he talked, I did the mental math of how many extra hours I
would need to work to pay for the $350 textbook.
There were rare exceptions, such as my calculus profes-
sor. On the same day as my conversation with the organic
chemistry professor, he gave me a much-appreciated boost
when he loaned the entire class his own textbooks for the
semester. Other students may have turned up their noses at
the books’ tattered condition, but I was incredibly grateful—
both to be saved the expense, and to not be set apart from
the rest of the students. His gesture, and similar efforts by a
handful of faculty and staff, helped me make it through my
undergraduate and master’s studies.
p
in classes where the majority of students comfortably af-
ford college, but it has been incredibly rewarding. Many
students have now shared their financial struggles with me.
,
This past semester, for example, one of my most diligent
students confided that he worked long hours in the univer-
sity’s cafeteria and needed to do the bulk of his coursework
late into the night. So, when he occasionally fell asleep in
class, I knew it reflected anything but lack of interest, com-
mitment, or ability.
All low-income students should have mentors who can help
them cope with the many expenses associated with college
and make them feel welcome. And I hope that more academ-
ics can be like my calculus professor, who with his simple ges-
ture made such a big difference for me. j
Faye Harwell recently received her Ph.D. from Boston University. Send
your career story to SciCareerEditor@aaas.org.

1018 1 SEP TEMBER 2023 • VOL 381 ISSUE 6661 science.org SCIENCE
 Produced by the Science/AAAS Custom Publishing Office
Flexible Positive Pressure Sensors
Tecan has announced the launch of the Resolvex
A200 Omics and Resolvex A200 24, the next
generation of Resolvex A200 positive pressure
processors. The Resolvex A200 is a standalone
system for automated positive pressure sample
preparation which is designed to address unmet
     ! #  $
% % &$#$$ %" 
#!'()) % %$$ *'#
and fully enclosed environment to reduce sample
contamination in highly sensitive mass spectrometry work. The
#!'())( %$ # #$#%$ #
including 3 and 6 ml columns and 24-well plates. The system is
$#$$ #$ # $%% $$$$$##
 %$ $ %# %$%  %$ 
methods. Dr. Klaus Lun, executive vice president and head of the
life sciences division at Tecan, commented: ìWe are excited to
#$% #!'()) %$($  
#!'())&$ # $ %  $ #  &
$  
& %$ $$#  %$#%
 '*$&$&  
  $%%$% $
throughput.î
Tecan
For info: +44 (0) 1480 405333
www.tecan.com/resolvex-a200-next-generation-positive-pressure-
processing
            
  
 $)      "%#&'
&  #  +& $&-
microphysiological systems that simulate the function of human
organs for medical research purposes. Designed to overcome
 * &-$, &,
 +  "   &)$&$$ ''
$&&$$ . $,&
$ -, &&,&*+ 
   &,' '$$&  +*$
& $ &
 &, -$& 
that include predicting drug-induced liver injury or determining
  $$&$+& &-, 
,&,    $ &&'$$  ( # !


For info: +44(0)7891477378
 
  $
     

  
        
and the rapid growth of the sector, these treatments remain
    

!  !     
  # !  
    
  
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
0901Product.indd 1015
LIFE SCIENCE TECHNOLOGIES
new produ cts: automation
    !    
therapy developers access to sterile manufacturing at scale.
   %$      
       
  ! % !
"
    
needs of each manufacturing facility and the therapy they produce.
      
including quality control systems and fully remote digital control and
analytics.
   

For info: +44 (0)7771 730919
https://cellularorigins.com/
     
 

!#$%#"2#&'(###"2#)!*+, # +!*#2+##
2-!#)")!#".#*"-'!#+"#*+)#-0!,!%+#11)+"%4#!5%!-#
,"'%-#+!#6%, #&'*1(#-'#)%%!#,!5!%+#-!7!, # +!*8#
+!#"2#&'(#)%#!#)"%35',!-#+"#,'%#"+#*%'#%-#
'+"*+!-#"2#)!*+, #,!)+"%4#!# +!*#* #!#)"%35',!-#
p
2+#% #)"*%+"%#".#'1#+"#."',#%-7-'#,!)+",#*"-'!4#+#
+!#)"*%!-#"+"(#%-#"+1(#!+!-#,!)+",#*"-'!#
"%5-!#+!#&",#6!,#&'#"2(#",#&",#6!,#&'#(#
), "5!%)#*"-'!8#%#"1!,+%5#,%5!#.,"*# # #+"# # ##
1"!4#!#--+"%#".#"2"%+,"##".+2,!#"2# "'#+"#
%+!5,+!# "',#"2#&'(#2+##.,)+"%#)"!)+",8#!%%5# "'#
+"#'+"*+!#,!)+"%#.,"*#!+!,#,!5!%+#"++!#",#'+"*+!-#
g
*1!#""1#.",#,!)+"%#"1+*+"%#+'-!4#!#)"*1'+!,8#
1'*18#%-#!+%5)""%5#*"-'!#%##"2#&'(#,!#
)"%%!)+!-#"7!,###'%5#%#/+!,%!+#'8#!%%5#+!# +!*#+"#
!#)"%+,"!-#,!*"+! # # 8#%-#"2%5#+!#)"%+,"#)"*1'+!,#
y
+"#!#"1!,+!-#"'+-!##.'*!#""-4


For info: +44-1223 942004
www.uniqsis.com
         
 
!,#-'+"%#,!#2-! #'!-#),"#*% #.!#)!%)!#-)1%!#
and performing them correctly underpins the success of later
+'-!4#/#6")!%)!#!!)+,"%)#11!++!8#'+"*+"%#
platforms, and peristaltic pumps are designed to ensure
)"%+!%+#'-#%-%5#%-#*%*!#'*%#!,,",8#+!,!.",!#
y g
%),!%5#+,"'51'+#%-#,!1,"-')+ #".#!,#-'+"%4#!!#
%+,'*!%+#,!#-1+!#.",##11)+"%#++#*1. ##,%5!#
".#'-#%-%5#2",
"2#%-#!%%)!##1,"-')+7+ 4#!!#
'-#%-%5#"'+"%#)%#!#'!-#+"#+%-,-!#%-#+,!*%!#
p
pipetting protocols in a variety of applications, including standard
curve preparation to quantify nucleic acid concentration and
7-+!#&# 8#*%*'*#%+", #)"%)!%+,+"%#+!+%5#%#
,
!+),!#%-#-,'5#-!7!"1*!%+8#%-#)+!,#!%'*!,+"%#.",#
epidemiological studies
  
For info: +44 (0)1480 405333
++1 2224%+!5, ")!%)!4)"*
1 SEP TEMBER 2023 • VOL 381 ISSUE 6661 1015
8/28/23 8:23 AM