12 JANUARY 2024, VOL 383, ISSUE 6679 Science

Citation preview

EDI TO R I A L

Rooting out scientific misconduct

S

Barbara Redman is a courtesy adjunct professor at New York University’s School of Nursing and an associate of New York University Langone’s Division of Medical Ethics, New York, NY, USA. [email protected]

y g y , 10.1126/science.adn9352

SCIENCE science.org

y

“…ORI…lacks the resources and authority needed to make a difference.”

Ivan Oransky is a cofounder of Retraction Watch, New York, NY, USA; distinguished journalist in residence at New York University, New York, NY, USA; and editor-in-chief of The Transmitter, The Simons Foundation, New York, NY, USA. ivan@ retractionwatch.com

g

or distort the actual findings.” Still, ORI has missed an opportunity to hold institutions accountable. Although the agency suggests that it is an institution’s responsibility to foster an environment that promotes integrity, how should this be measured and judged? The revision still exclusively addresses misconduct by individuals. It would be best if an institution could be held responsible for a toxic, unsupportive research environment. Even if its recommendations are further adjusted, ORI lacks the personnel and budget to address the potential scope of alleged misconduct. The office is largely limited to supervising university investigations instead of carrying them out itself, which would avoid the obvious institutional conflict of interest. ORI also lacks subpoena power to compel witness testimony. This point may help explain why the National Science Foundation’s Office of the Inspector General, which has subpoena power, tends to make far more findings of misconduct than ORI each year. There is good news, though. Some publishers have become more willing to correct the scientific record. This led to more than 10,000 retractions in 2023—reflecting about 0.2% of the literature across all fields, as indicated in a recent analysis. According to the study, this is a 10-fold increase compared with two decades ago. Not all of these were because of misconduct, but studies have consistently found that two-thirds of retractions are. However, it is not clear whether the incidence of misconduct has increased over time. There is no question that the work of today’s sleuths, who often use software not available 20 years ago, has pushed these numbers higher. And the fraudulent activity of research paper mills that produce fake manuscripts is likely also a factor. On a larger scale, universities are starting to take a harder look at the suitability of perverse “publish or perish” incentives for faculty promotion and tenure. Thirty years ago, ORI was created in response to a series of scandals at prominent institutions, some involving faked data, that caught the attention of Congress. Congress should strengthen what it set out to do— address misconduct in science by giving ORI the teeth it needs to sink into the problem. –Ivan Oransky and Barbara Redman

p

cientific misconduct is an issue rife with controversy, from its forms and definitions to the policies that guide how allegations are handled. A survey published nearly 15 years ago reported that 2% of researchers said they had fabricated or falsified data in their published work. This is not just an academic issue. Fake data promote ineffective or even dangerous treatments, for example, and thwart the discovery of real solutions for society. In the United States, the Office of Research Integrity (ORI) is tasked with rooting out misconduct in research funded by the National Institutes of Health (NIH). Last October, ORI proposed changes to how it functions. The agency’s recommendations—the first since 2005— have evoked mixed reactions, but the real problem is that ORI is underfunded and lacks the resources and authority needed to make a difference. Unless its charter is revised by Congress, the ORI can sadly do little more than tinker at the edges of scientific fraud. It is a wonder that the ORI accomplishes much at all. Its current budget is $12 million per year to oversee work funded by NIH, a $48 billion agency. Add to that the frequent internal strife over ORI’s proper role and a directorship that has often been vacant, and one can see how its ability to be effective does not meet the expectations for upholding the integrity of research activities. The regulations proposed by ORI’s new director, Sheila Garrity, include fine-tuning of definitions and processes that is long overdue. For example, they clarify the term “reckless,” which was used more often recently by the ORI to prosecute fraud cases. The term emphasizes indifference to or disregard for the truth of the matter being asserted, according to a recent article. But what should happen if someone has supervised, but not performed, the research at issue, as in the case of former Stanford University president Marc TessierLavigne, who failed to correct problems in work by his trainees? As the article asks, what is reasonable supervision and when is it so lacking that it becomes reckless? The agency seems much more open to disclosing the results of university investigations, a transparency that has predictably been met with criticism from academic institutions that claim it might “violate privacy laws

12 JANUARY 2024 • VOL 383 ISSUE 6679

131

NEWS IN BRIEF

“

We’ll never be a science superpower behind a visa paywall. Former U.K. science minister George Freeman, in Times Higher Education, on rising immigration fees and the government’s plan to raise the minimum salary needed to get a skilled worker visa to £38,700—above the typical starting salary of a U.K. postdoctoral researcher.

”

Edited by Jeffrey Brainard

Billionaire vows plagiarism checks | Hedge fund manager and billionaire Bill Ackman says he plans to look for signs of plagiarism in work by all of the faculty and board members of the Massachusetts Institute of Technology (MIT), including its president. His campaign, announced last week, comes after Ackman was angered by two Business Insider exposés this month reporting that a 2010 doctoral dissertation and other academic papers by Neri Oxman, his wife, included multiple passages substantially lifted from other sources without proper attribution. Without offering evidence, Ackman alleged that MIT leaders likely prompted the articles. He has pushed to oust several university presidents including MIT’s head, cell biologist Sally Kornbluth, claiming they did not take antisemitism on campus seriously enough. (Kornbluth has held on, but political scientist Claudine Gay stepped down last week from Harvard University’s helm after examinations of her scholarly work revealed instances of plagiarism, which Ackman and others had highlighted publicly.) Plagiarism and publishing experts told Science that identifying the hundreds of thousands of research papers and other writings of nearly 1100 MIT faculty members, obtaining the full text, and running the documents through plagiarism-checking software could take months, even with an unlimited budget. RESEARCH INTEGRITY

science.org SCIENCE

,

12 JANUARY 2024 • VOL 383 ISSUE 6679

P O L I C Y | Spain’s new approach to evaluating scientists at public universities is drawing praise for emphasizing quality over quantity of research publications. Spanish researchers face evaluations every 6 years to gain promotions and salary increases, and in the past, evaluations hinged on just one criterion: whether scientists published at least five papers in high-impact journals in that period. Critics complained the approach incentivized publishing mediocre papers and unethical ways to boost authorship

y

132

Spain reforms research reviews

y g

H

ours after the 8 January launch of the Peregrine lander from Cape Canaveral, Florida, a catastrophic leak crippled its propulsion system, ending hopes it would become the first commercial mission to land on the Moon. Developed by Astrobotic Technology, a Pittsburgh startup, Peregrine was the first mission in NASA’s Commercial Lunar Payload Services program, which has sought to spur lunar-landing capabilities by financing private missions. NASA has repeatedly warned that it expects some of these low-cost probes to fail; a second mission, led by Houston’s Intuitive Machines, will launch next month. The mishap raises questions for Astrobotic, which has a contract to deliver NASA’s $500 million Volatiles Investigating Polar Exploration Rover to the Moon later this year.

y

Failure spoils U.S. return to Moon

g

PLANETARY SCIENCE

p

The ill-fated Peregrine Lunar Lander is shown nestled within a Vulcan rocket before its launch.

ANIMAL PHYSIOLOGY

Snake teeth predict strike tactics

T

SCIENCE science.org

Mauna Kea slim-down begins | Workers last month completed dismantling and removing the Caltech Submillimeter Observatory (CSO) atop Hawaii’s Mauna Kea, a first step toward fulfilling a deal that was hoped to soften opposition to construction there of a proposed observatory, the Thirty Meter Telescope (TMT)—the largest in the Northern Hemisphere. Native Hawaiians, who consider Mauna Kea sacred, have

ASTRONOMY

| Scientists this week proposed a cause of odd radio circles (ORCs), the mysterious, giant rings that pop up in sky surveys at radio wavelengths: They form when distant galaxies expel gas after intense star formation. Fewer than a dozen ORCs have been spotted since the first, in 2019, and astronomers did not even know whether they originated in or near our own galaxy or in distant ones. New clues emerged after the team discovered that a distant galaxy, which appeared to be enveloped in an ORC, produced a bright fluorescent glow from excited oxygen atoms. The astronomers developed a computer model suggesting that during an intense period of star formation there 1 billion years earlier, many stars quickly went supernova, blasting gas out of the galaxy. That outflow hit cooler surrounding gas, generating the radio ring; the fluorescent glow resulted as some gas fell back into the galaxy. The team describes its findings, linking the ORC to that galaxy, this week in Nature. ASTRONOMY

12 JANUARY 2024 • VOL 383 ISSUE 6679

133

,

grown organoids from fetal brain tissue, providing a new means of studying brain cancers and developmental disorders. In previous studies, researchers created brainlike organoids using stem cells that can turn into many cell types. In this week’s issue of Cell, a research team in Europe reports using fetal brain tissue to grow 3D structures that resemble those that naturally occur in the cortex, forebrain, or spinal cord. The new organoids may offer a realistic new model of how individual brain regions grow and could also enable better testing of drugs for brain cancer. Pregnant people donated the tissue from fetuses aborted between 12 and 15 weeks postconception. The team worked with bioethicists to ensure that the rice grain–size organoids could not feel pain or become conscious. But in the United States, the study of fetal tissue is politically fraught. In 2021, President Joe Biden’s administration reversed a ban on

Radio circles linked to gas blasts

y

PHOTO: LEONARDO MERCON/VWPICS/SCIENCE SOURCE

| Researchers say they have

| A planned demolition of part of Marie Curie’s historic laboratory in Paris has been suspended after an outcry from preservationists and an intervention from France’s culture minister. The targeted building, the Pavillon des Sources, was one of three constructed when the Radium Institute, now known as the Curie Institute, was established in 1909. Curie used the building for some of her pioneering work on radioactivity and later became the first woman to win a Nobel Prize, for discovering polonium and radium. The institute, which still manages the site, proposed to build a new building there to expand its cancer research, saying Curie used the existing one mainly to store radioactive materials. Last week, the institute said it would now consider alternatives to demolishing the building, which remains contaminated.

H I S T O RY

y g

BIOMEDICINE

Curie lab demolition postponed

opposed the TMT and long campaigned against overdevelopment of the summit, which is home to 13 observatories. The CSO is the first to be taken down after a 2015 pledge by Hawaii’s then-Governor David Ige (D) to reduce the number. From 1987 to 2015, the CSO pioneered studies in the wavelength range between infrared and radio. Its components will be shipped to Chile and reassembled in the high Atacama Desert, where it will be reborn as the Leighton Chajnantor Telescope.

y

Fetal tissue yields brain organoids

such research, but some federal and state lawmakers have introduced new proposals to prohibit it.

g

credits. Under the new method, which took effect on 1 January, the National Evaluation and Accreditation Agency will also consider a range of research outputs besides papers—for example, patents and exhibitions—along with the work’s social impact. The reforms are also crafted to reduce plagiarism and deceitful multiauthorship, and to reward scholars for publishing in open-access journals. Spain is one of the first countries to adopt such a policy; in several others, only individual institutions have.

p

he shape and position of a snake’s teeth say a lot about how it hunts, a study has found. Using 3D x-ray scans of their mouths, herpetologist William Ryerson of the Cornell University College of Veterinary Medicine identified two distinct groups, each correlated with different methods of striking prey. “Strikers,” such as boa constrictors (right), quickly attack from above, immediately grabbing the prey with their tall lower teeth to enable the snake to coil around the prey and strangle it. “Lungers,” such as king snakes, attack an animal straight on and less quickly, piercing it with both their top and bottom teeth at the same time. Their gape is smaller than that of strikers. Ryerson, who presented the findings last week at the Society for Integrative and Comparative Biology’s annual meeting, studied about 70 individuals from 13 snake species, representing four of the five main groups. He filmed them as they attacked a dead rodent that he had heated to attract them. He hopes a better understanding of snake teeth will inspire practical engineering applications.

NE WS

IN DEP TH

A church in Sancourt, France, is pictured at sunset during a July 2023 heat wave.

p

CLIMATE CHANGE

g

The hottest year was even hotter than expected Greenhouse gases, El Niño, and cleaner air fueled record heat in 2023

I

science.org SCIENCE

,

12 JANUARY 2024 • VOL 383 ISSUE 6679

y

134

time Organization added to the effect when ships began to cut sulfur pollution—and inadvertently curbed the light-reflecting clouds that the sulfur particles help create (Science, 4 August 2023, p. 467). A preprint on Research Square suggests the loss of these clouds alone can explain half of the increase in the warming rate seen so far this decade, says Yuan, who led the work. “[It] would not account for all the warming we see this year, but it would represent a significant additional warming.” In a November 2023 paper, famed climate scientist James Hansen suggested curbing pollution has accelerated warming to 0.27°C per decade, up from the 0.18°C per decade rate experienced from 1970 to 2010. But the acceleration has yet to show up in records of heat in the ocean depths, which resist the short-term fluctuations of the atmosphere and offer a truer sense of long-term trends. The mystery of the past year leaves projections for this year less certain than usual. El Niño may inflate temperatures further, pushing the world briefly past the arbitrary 1.5°C “limit” settled on by policymakers in 2015’s Paris agreement to protect small island nations from extreme sea-level rise. But extreme heat will again have to develop over the northern oceans for the world to breach the threshold—hardly a sure bet. Regardless, the long-term warming pattern is certain to continue, as it has for decades—until fossil fuel burning ends. j

y g

t comes as no surprise to anyone who sweated through it: 2023 was the hottest year in human history. Average surface temperatures rose nearly 0.2°C above the previous record, set in 2016, to 1.48°C over preindustrial levels, the European Union’s Copernicus Climate Change Service reported this week. Only Australia was spared record-setting heat. The extreme conditions are a “dramatic testimony of how far we now are from the climate in which our civilization developed,” said Carlo Buontempo, Copernicus’s climate director, in a statement. Yet 2023’s record temperatures— confirmed days later by analyses from NASA, the U.S. National Oceanic and Atmospheric Administration, the United Kingdom’s Met Office, and Berkeley Earth— come with a mystery. Humanity’s unabated burning of fossil fuels is the dominant driver of the long-term trend, but it is insufficient to explain 2023’s sudden spike, says Michael Diamond, an atmospheric scientist at Florida State University. One exacerbating factor was the end of a La Niña climate pattern, which from 2020 to 2022 stirred up an increased amount of deep cold water in the eastern Pacific Ocean that absorbed heat and suppressed global temperatures. In 2023, the pattern flipped into an El Niño event, which blanketed the

equatorial Pacific with warm waters and began to boost global temperatures. But the flip is not enough to explain 2023’s record, Gavin Schmidt, director of NASA’s Goddard Institute for Space Studies, wrote in a blog post last week. Typically, El Niño plays a larger role in global temperatures the year after it starts—in this case, this year. And in 2023 heat surged far from El Niño’s influence, above the northern Atlantic and Pacific oceans, Schmidt noted. The 2022 eruption of Hunga Tonga– Hunga Ha‘apai, a volcano in the south Pacific, had been a suspect in the global temperature jump because of the vast amounts of climate-warming water vapor it injected into the stratosphere. But early studies neglected the sulfate particles it also sent into the upper atmosphere, which reflected light and canceled out the water vapor’s warming effect, says Mark Schoeberl, an atmospheric scientist at the Science and Technology Corporation. “For 2022, it was a nonevent. I have continued my computations into 2023—still a nonevent.” Perhaps the best explanation for the extra warming is the continued drop in light-blocking pollution as society shifts to cleaner sources of energy, says Tianle Yuan, an atmospheric physicist at NASA’s Goddard Space Flight Center. In 2022, satellites began to detect this decrease from space (Science, 22 July 2022, p. 353). In 2020, new regulations from the International Mari-

y

By Paul Voosen

N E WS

EARTH SCIENCE

U.S. undersea mapping is a boon for science New maps of continental shelf could vastly expand U.S. territory and resource claims By David Malakoff

y ,

GRAPHIC: M. HERSHER/SCIENCE

y g

135

y

12 JANUARY 2024 • VOL 383 ISSUE 6679

g

SCIENCE science.org

p

tive and aren’t likely to draw fierce objections, says David Mosher, an emeritus geoscientist at Natural Resources Canada. He helped with he United States has unveiled the the U.S. mapping effort and is also a member results of a monumental undersea of the U.N. commission that evaluates ECS mapping effort that could add 1 milclaims. For now, however, the U.S. isn’t sublion square kilometers of sea floor— mitting its maps to the commission, because twice the area of California—to its the Senate still hasn’t ratified the Law of the territory. In addition to enabling the Sea and some nations are likely to argue that U.S. to claim valuable geological and biothe U.S. is not entitled to have the commislogical resources, particularly in the Arctic, sion vet its claims. Mosher adds that even the project has produced a wealth of seaif the commission agreed to review the U.S. floor data that are fueling a wide range of maps, it could take decades to evaluate them. scientific advances. Meanwhile researchers are feasting “It’s just a treasure trove of informaon the data. For his 2020 doctoral distion,” says marine mapping specialist sertation, for example, Sowers used an Derek Sowers of the Ocean Exploration Pushing the boundaries algorithm to sort seafloor features along Trust, who has used the data to identify Newly charted extensions to the U.S. continental shelf hold the U.S. Atlantic coast into habitat types, deep-sea habitats that could be rich in resources and scientific riches. including mounds and seamounts that biodiversity. The maps have also shed may host deep water sponges and corlight on the geologic evolution of ocean Current maritime boundaries Added shelf territory als. Such analyses, he notes, could help basins, identified areas at risk of proArctic tic Ocean U.S. officials identify rich seafloor ecoducing submarine mudslides that could systems that need protection. trigger tsunamis, and pointed to potenOther Atlantic shelf research retial seafloor mineral deposits. “There vealed a surprising number of submacertainly are a lot of other potential rine landslides. It even provided the applications,” says geoscientist Larry first complete picture of a massive Mayer of the University of New Hamp375-kilometer-long slide off North Carshire, one leader of the effort. Bering Sea olina’s Cape Fear, which tumbled down The impetus to make the maps, the continental slope 10,000 to 27,000 which were released last month by the years ago. U.S. Department of State, came from Deborah Hutchinson, a marine geothe United Nations Convention on the physicist at USGS who helped lead Law of the Sea. One provision of that United States several cruises, says the mapping off 1982 pact gives a coastal nation the Atlantic Alaska is helping settle debates over right to claim sea floor that sits outOcean O Pacific Ocean how shifting tectonic plates formed side its exclusive economic zone, which stretches 200 nautical miles offshore, if parts of the Arctic Ocean. “It was a huge it can demonstrate that the territory is breakthrough to get ships into the Arc0 2000 a “natural prolongation” of its continentic,” she says, noting that U.S. funding tal shelf. Defining this extended contifor such cruises has been scarce. And in Gulf of Mexico km nental shelf (ECS) involves measuring the Bering Sea, seismic and other data the thickness of marine sediments, confirmed that parts of the sea floor which are typically thicker on the continental After analyzing terabytes of data, researchare composed of oceanic and not continental slope, where the shelf descends to the abyss, ers identified ECS additions in seven areas crust. That “helped eliminate some hypotheand identifying the “foot” of the slope. (see map, above). By far the largest is in the ses about its formation,” marine geophysicist Two decades ago, many nations launched Arctic, where the U.S. is claiming 520,400 Gail Christeson, who is a program director seafloor mapping efforts in order to submit square kilometers, about the area of Spain. In for marine geology and geophysics at the Naan ECS claim to the U.N. (Science, 6 Decemthe Atlantic, the claim totals 239,100 square tional Science Foundation. ber 2002, p. 1877). In 2003, the U.S. joined the kilometers. The third largest chunk—176,300 Scientists from many fields also “piggyrush even though the Senate had not ratisquare kilometers—is in the Bering Sea. In backed” on the mapping voyages, obtaining evfied the pact, which Republicans opposed on places, the U.S. claims overlap those made by erything from seafloor rock and mud samples grounds of national security. Mayer and other Canada, Japan, and the Bahamas; diplomatic to one of the first bits of icelike gas hydrate— scientists argued that emerging technologies, negotiations will resolve those boundaries a potential energy source—recovered from especially multibeam sonars that produce de(Science, 21 June 2019, p. 1120). the Arctic. “The U.S. investment in ECS maptailed 3D images, offered an unprecedented Though vast, the U.S. claims are fairly conping,” Hutchinson says, “produced much opportunity to document little known sea servative from a technical and legal perspecmore than just maps.” j

T

floors off the U.S. “We have much better maps of Mars than we do of most of our continental margin,” Mayer told Science at the time. Federal agencies, including the National Oceanic and Atmospheric Administration and the U.S. Geological Survey (USGS), ultimately spent tens of millions of dollars on nearly 50 mapping cruises, some held jointly with Canada. The ships completed multibeam surveys of 3 million square kilometers of sea floor and conducted acoustic surveys, which use sound to map sediments, along nearly 30,000 linear kilometers.

NE WS | I N D E P T H

p g

By sampling helium gases in hundreds of springs across southern Tibet, researchers identified places where mantle rocks were closer to the surface.

y

GEOPHYSICS

Tectonic plate under Tibet may be splitting in two Peeled-apart Indian Plate could be affecting earthquake hazards By Maya Wei-Haas

science.org SCIENCE

,

12 JANUARY 2024 • VOL 383 ISSUE 6679

y

136

uncertainties remain about the process, and the data are limited. “It’s just a snapshot,” he says. But the work is an important step toward understanding how our modern landscape came to be, he says. “It’s definitely the type of work that we need to move [forward].” Scientists have long suggested tectonic plates could unzip like this. The plates are a layered combination of buoyant crust and more dense upper mantle rock. When squeezed and thickened, a plate might split along the weak interface between the layers. But the process was mostly studied in the interiors of thick continental plates and simulated in computer models. “This is the first time that … it’s been caught in the act in a downgoing plate,” van Hinsbergen says. The Himalayan collision is a promising place to look for a plate being torn apart, says Peter DeCelles, a geologist at the University of Arizona. Before the smashup began, the Indian Plate varied in thickness and composition, which helps explain the crescent shape of the 2500-kilometer-long Himalayan front. He compares the ancient

y g

T

he towering peaks of the Himalayas are a geologic battleground—a slowmotion collision of the Indian and Eurasian tectonic plates. The crunch began some 60 million years ago as India, then an island, plowed into Eurasia, buckling the surface and forming the highest mountains on Earth. But the peaks are just the noise and smoke of the battle; tectonic maneuvers tens of kilometers below them drive the clash—and hold mysteries. Continental tectonic plates, unlike their dense oceanic cousins, are thick and buoyant, so they don’t easily sink, or subduct, into the mantle during collisions. Some scientists believe the Indian Plate resists plunging into the mantle and continues to slide horizontally under Tibet. Others suggest the most buoyant part of the Indian Plate rumples like a rug along the front edge of the collision, making it easier for the lower half of the plate to sink into the depths.

Now, a new analysis of earthquake waves traveling beneath Tibet and telltale gases rising to the surface points to yet another possibility, one that in effect splits the difference between the two scenarios. Part of the Indian Plate appears to be “delaminating” as it slides under the Eurasian Plate, with the dense bottom part peeling away from the top. The study also finds evidence for a vertical fracture, or tear, at the boundary between the peeled-apart section of the slab and its intact neighbor. “We didn’t know continents could behave this way and that is, for solid earth science, pretty fundamental,” says Douwe van Hinsbergen, a geodynamicist at Utrecht University. The work, presented in December 2023 at the American Geophysical Union conference and in a preprint posted online, could help scientists better understand the formation of the mighty Himalayas and perhaps even earthquake hazards in the region. Fabio Capitanio, a geodynamicist at Monash University, cautions that plenty of

S

cientists who study aging are howling about the possible demise of one of the field’s biggest studies, the Dog Aging Project. The effort has been probing cognitive and physical aspects of aging in about 50,000 dogs and is running a clinical trial to test a drug that may boost their longevity. But organizers say the project will probably lose funding this year from the National Institute on Aging (NIA), which has furnished at least 90% of its annual budget, now about $7 million. “It is a big loss if this project in dogs does not continue,” says gerontologist João Pedro de Magalhães of the University of Birmingham, who notes that large, long-lived animal models promise valuable insights into human aging. “It was going to be the most informative study of aging that was not done in humans,” says biogerontologist Steve Austad of the University of Alabama at Birmingham. (Neither scientist has a role in the research, but Austad’s 2-year-old dachshund is a participant.) Organizers are pessimistic about continued funding because the project received marginal scores on its grant renewal application late last year. They are striving to raise money from other sources and have launched a petition drive to convince the director of the National Institutes of Health to intervene to restore funding. “I’m doing everything possible to keep [the project] going in its current form,” says co-director Daniel Promislow, an evolutionary geneticist at the University of Washington (UW). Although longevity labs teem with rodents and other small animals, “Dogs are probably the most powerful model for studying the biology of aging,” says project co-director Matt Kaeberlein, a former UW geroscientist who is now CEO of the Seattlebased biotech Optispan. Thanks to decades 12 JANUARY 2024 • VOL 383 ISSUE 6679

137

,

GRAPHIC: N. BURGESS/SCIENCE

By Mitch Leslie

y

SCIENCE science.org

Maya Wei-Haas is a science journalist in Washington, D.C.

Organizers hope to save long-running project on canine aging and longevity

y g

Subduction

Massive study of dog aging likely to lose funding

y

Delamination

ANIMAL RESEARCH

g

ing under Tibet before plunging into the mantle. But in one area south of the line, near the eastern border of Bhutan, a trio of springs also contained mantle signatures— a hint that one section of the Indian Plate might be peeling apart, with hot mantle rock flowing into the space in between. Support for this picture came from an analysis of earthquake waves rippling across the boundary between crustal and mantle rock. Waves recorded at hundreds of seismic stations enabled the researchers to construct images of structures underground. Two blobs seen in one image seem to hint at a lower slab of the Indian Plate detaching from its top. A more recent analysis based on a different set of earthquake waves points to a tear on the western edge of the delaminated slab. West of the proposed break, the bottom of the Indian Plate appears to be some 200 kilometers deep, suggesting it is still intact; to the east, where the slab splits in two, mantle rock is flowing in around a depth of 100 kilometers. Almost every landmass on Earth was built from a series of collisions like the Himalayas, says Anne Meltzer, a seismologist at Lehigh University. So understanding how continents collide sheds light not only on our modern landscape, but also the hazards posed by earthquakes that can occur along the ancient scars of continental crashes. Klemperer points out the newly proposed tear may also be influencing earthquake hazards in Tibet today. A deep fracture in the Tibetan Plateau known as the ConaSangri rift overlies the tear—a tantalizing hint that the tumult in the Indian Plate’s underbelly might somehow ripple to A geological battleground the surface. Although the diThe continental collision of the Indian and Eurasian tectonic plates rect link to quakes remains has created the Himalayas. New evidence suggests part of the Indian uncertain, van Hinsbergen Plate may be splitting away and plunging into the mantle. notes that tears and peelHimalayas ing of tectonic plates could affect how and where stress builds, and thus the potential for quakes. Continents are complex to study, Klemperer says, as the ancient collisions that built the modern landscape Eurasian Plate left behind a network of Indian Plate overlapping scars. But that’s what makes the work excitUnderplating ing, he says. “They have this palimpsest of a billion years Mantle of history.” He and other sciMantle flow entists are gradually learning to read it. j

p

plate to a manta ray: a pair of thin wings of oceanic crust flanking a thick middle of continental crust. The thin oceanic slabs readily plunged below the Eurasian Plate while the thick continental crust plowed headlong into Eurasia like a battering ram. The difference in subduction speed likely tugged and tore the Indian Plate in multiple directions. Scientists have proposed so many tears in recent years, “it’s become a cottage industry,” quips study author Simon Klemperer, a geophysicist at Stanford University. Klemperer was intrigued by a zone in northeastern India, near Bhutan, where the subduction zone curves, making it a prime locale for potential tears. “That’s where things get ugly,” he says. The suspicion drove him on a multiyear quest for clues to the subsurface violence. One line of evidence came from isotope measurements of helium that burbles up in Tibetan springs. Klemperer and his colleagues, including collaborators in China, traversed dirt roads and forded streams to collect samples from some 200 natural springs across nearly 1000 kilometers of southern Tibet. Gas samples rich in helium-3, a light isotope trapped within Earth at the time of its birth, signal mantle rocks underfoot, whereas those depleted in helium-3 likely rise from buried crust. A stark pattern emerged when they mapped the springs. To the south of a line lay springs with crustal signatures, and to the north lay springs with mantle fingerprints. The team interpreted the line as the farthest edge of the intact Indian Plate slid-

NE WS | I N D E P T H

By Andrew Curry

L

science.org SCIENCE

,

ike many diseases, multiple sclerosis hits some populations harder than others. For example, Scandinavians are an estimated 17 times more likely than people from sub-Saharan Africa to develop the devastating chronic disease, in which the body’s immune system attacks nerves. “It’s very puzzling,” says University of Copenhagen paleogeneticist Eske Willerslev. Some have speculated that “Viking genes” or some aspect of Northern Europe’s diet or environment might boost risk. New data from ancient skeletons show part of the answer arose in the Bronze Age. About 5000 years ago, people from the steppes near the Black Sea, whom archaeologists call the Yamnaya, moved west across central and Northern Europe. In Scandinavia and parts of Northern Europe, these mobile cattle herders took just a few centuries to largely replace the sedentary farmers who had tilled the soil with stone tools for millennia. The Yamnaya brought with them not just a different way of life, but also genes linked to a higher risk of multiple sclerosis (MS), DNA from the bones shows. “With MS, you can see it’s Yamnaya ancestry, and the Yamnaya are basically Danes,” says Willerslev, who led the research, published this week in Nature. “Everything fits beautifully.” It’s not just MS. Willerslev, a pioneer in the study of DNA from ancient remains, and co-authors have completed a wide-ranging genetic survey of early Europeans, from the earliest modern humans on the continent 45,000 years ago to the later waves of farmers and herders who swept in from the east. In multiple papers, the team reports analyzing the whole genomes of more than 1000 ancient individuals and comparing them

y

12 JANUARY 2024 • VOL 383 ISSUE 6679

Among Europeans, risk of multiple sclerosis rises with genes from Bronze Age Yamnaya herders

y g

138

Ancient DNA ties modern diseases to ancestry

y

The Dog Aging Project’s chief veterinary officer, Kate Creevy (foreground), examines a dog.

PALEOGENOMICS

g

also helps dogs live longer. “The project is really just beginning to hit its stride,” says its chief veterinary officer, Kate Creevy of Texas A&M University. Kaeberlein says that, despite the disruptions from the COVID-19 pandemic, the project’s organizers thought they had achieved enough for NIA to approve a renewal of their grant. But, “The reviewer score wasn’t as positive as we had hoped for,” Promislow says, almost certainly putting it outside the cutoff for funding. Although the agency hasn’t announced the grant recipients for this cycle, “It’s very unlikely that NIA will be able to fund us this cycle,” he says. An NIA press representative said the agency does not comment on grant deliberations. Dog geneticist Heather Huson of Cornell University, who isn’t connected to the study, says such projects need long-term support. “It takes you 5 years to start accumulating data.” She knows the challenges firsthand because she took part in the Vaika Project before it shut down last year. The project was a casualty of the war between Russia and Ukraine, because philanthropies in Russia provided much of the money, Huson says. To avoid that fate, the leaders of the Dog Aging Project are trying to raise money from other sources to tide the project over for the next year, and they plan to reapply for NIA funding in 2025. Promislow and the other organizers are also creating a charitable foundation and hope to raise $40 million to $50 million for an endowment that would ensure continued funding. The Vaika Project also tried to coax donors to open their wallets for dog aging research but failed, Huson cautions. Yet she and others are hoping the larger project— which has suddenly become an underdog— can be rescued. j

p

of veterinary research, scientists know more about how health changes over time in dogs than they do in rodents. And unlike lab-dwelling rodents, dogs are exposed to the same environments as we are, and they develop many of the same age-related ailments, including heart disease and dementia. “It’s remarkable how much dog aging is teaching us about human aging,” says biochemist and geneticist Laura Niedernhofer of the University of Minnesota, who isn’t connected to the project. Other researchers and companies have joined the pack. The Vaika Project tracked the health of a group of 103 retired sled dogs for 5 years until it shut down, and the Golden Retriever Lifetime Study is still following more than 3000 members of that breed. Several companies are also developing therapies to slow aging in dogs, and one, San Francisco–based Loyal, has received a preliminary Food and Drug Administration endorsement for its drug— although the company has not revealed what’s in the treatment. The Dog Aging Project began in 2014, but research didn’t start in earnest until 2018, when the effort received a 5-year grant for nearly $29 million from NIA. Dog owners fill out annual questionnaires on their pets’ health to chart the animals’ physical deterioration. Some dogs get a closer look, furnishing DNA and other samples and going through tests of mobility and cognition. So far, scientists have sequenced the genomes of 1000 dogs and cached 14,000 tissue samples. Published studies have tracked dogs’ cognitive decline, gauged their susceptibility to tumors, and investigated how their eating schedules affected their health, among other topics. A clinical trial is also underway to test whether the drug rapamycin, which extends life in rodents,

y ,

PHOTO: DANISH NATIONAL MUSEUM

y g

139

y

12 JANUARY 2024 • VOL 383 ISSUE 6679

g

SCIENCE science.org

p

farmers arrived, in people with lots of Yamleast in part, the result of past selection fawith the UK Biobank, a massive data set of naya ancestry. “That’s really surprising,” voring survival against infectious diseases,” health histories and DNA from more than says University of Kiel bioarchaeologist Quintana-Murci says. 430,000 modern British people. The comBen Krause-Kyora. “This classical arms race Comparing ancient samples with the UK parisons revealed how the proportion of anbetween the human genome and bacteria Biobank’s database, the team was able to see cestry from ancient populations predicted and viruses isn’t happening until the end of when other traits appeared in ancient popudifferences in disease risk and physical atthe Neolithic.” lations and begin to think about why. “You tributes such as height and body mass in A recent preprint by members of the can see the impact of selection in real time,” present-day Europeans. “There’s a striking same team suggests a reason: Many dissays Pontus Skoglund, a paleogeneticist at difference in a range of traits, whether skin eases themselves were latecomers. Analysis the Francis Crick Institute. For example, a and eye color or dietary and mental health,” of pathogen DNA in the ancient bones suggene variant that today brings a higher risk says co-author Evan Irving-Pease, a compugests the incidence of diseases passing from of diabetes and high cholesterol appears tational geneticist at Copenhagen. “Subtle animals to humans, such as the bubonic more frequently in samples from about differences in ancestry have a distinct implague and leptospirosis, did not rise until 12,000 years ago and may have helped ice pact. … The legacy of these groups is still the time of the first Yamnaya migrations, age hunter-gatherers weather famine. Later, very present in modern people.” even though animals had been domestiancient farmers contributed other variants, Because they analyzed ancient samples, cated thousands of years earlier. The steppe linked to the ability to thrive on a more vegmostly from previously excavated skeletal herders may have kept animals at higher etarian diet, to the genetic makeup of modremains, the researchers could show when ern Europeans. and where key genetic variSuch knowledge has “impliants first appeared and infer cations for modern precision how shifts in ancestry and medicine,” Krause-Kyora says. selective pressures such as “The legacy of ancient populadisease could explain them. tions could explain why peo“Disentangling the main ple react differently to chronic ancestries helped us to deor infectious disease.” tect whether selection proThe studies also bring ancesses happened, and when,” cient population shifts into says co-author Alba Refoyo sharper focus. In two of the Martínez, a computational Nature papers, researchers geneticist at Copenhagen. use ancient DNA to track The studies are “an exthe movement of the first emplary tour de force,” says farmers from Anatolia north population geneticist Lluís into Europe, where hunterQuintana-Murci of the Pasgatherers were already living. teur Institute. “These papers Researchers had wondered illustrate how the intricate whether those encounters, interplay between ancient sewhich happened over millenlection and admixture events nia, were peaceful. But the has profoundly shaped … new data show that genes present-day Eurasians.” from the farmers increase Willerslev and his colabruptly, in stuttering surges. leagues started with a narIn many places, including rower goal: to understand Denmark about 5900 years what happened when hunterThe DNA of Denmark’s Porsmose Man, pierced by arrows around 2600 B.C.E., gatherer groups that had ago, they completely replace came mostly from early farmers, not the herders arriving from the east. survived the most recent ice those of earlier hunterage gave way to farmers, in gatherer populations. The a 6000-year period known as the Neolithic densities, or early Neolithic communities suddenness and thoroughness of the change that began around 9000 B.C.E. in Europe. may have been scattered widely, slowing the “lets us seriously question whether the meetThat was when humans first began to settle spread of disease. ing of populations was one of love and peace in densely populated villages and live close The timing could help explain the link and admixture,” Willerslev says. “At least to domestic animals. Researchers expected between Yamnaya ancestry and higher in Denmark, it’s a replacement, and a very those conditions would spread diseases, inrates of autoimmune diseases such as MS. abrupt replacement—reminiscent of Europecluding those originating in animals, and The same genes that bolster the immune ans entering the Americas, with disease and therefore would have favored genetic adapresponse to pathogens are thought to raise violence and devastating effects for the pretations against infection. “It’s a reasonable the risk that a person’s immune system will vious population.” theory,” Irving-Pease says. turn against their own cells. Those genes The clutch of new papers is centered on But, he says, “It’s not borne out by the were probably beneficial in the past, but Europe. But growing numbers of ancient data.” The team did find what they suspect became problematic in the past century DNA samples from Africa and Asia could are anti-infection adaptations, many in when the advent of vaccines and sanitamake it possible to apply similar approaches a region of the genome called the human tion lowered exposure to dangerous pathoelsewhere. “It would be good to expand to leukocyte antigen complex, which regulates gens. “It’s a compelling illustration of the other regions, and go further back in time,” immune response. But those variants prohypothesis that the contemporary high Refoyo Martínez says. “There’s still a lot to liferated thousands of years after the first prevalence of autoimmune diseases is, at look at and study.” j

NE WS | I N D E P T H

SEISMOLOGY

Mysterious seismic swarm led up to Japan quake Researchers study connection of rising fluids and quake flurry to magnitude 7.5 monster By Dennis Normile

B

,

science.org SCIENCE

y

12 JANUARY 2024 • VOL 383 ISSUE 6679

y g

140

Peninsula, which lies about 320 kilometers northwest of Tokyo. And Kato says a 2014 modeling study concluded a magnitude 7.6 event was possible. But seismologists did not expect an earthquake swarm there because the region has no volcanic or geothermal activity to generate high-pressure fluids. Nonetheless, thousands of quakes each year have rattled a small part of Noto’s northeastern tip since November 2020, in an unusually long-lived swarm. Studies suggest fluid is rising from the upper mantle through faults. A 2023 study by Nishimura and colleagues found the land above the swarm had risen by 70 millimeters, suggesting upwelling fluid was swelling the crust at a depth of about 16 kilometers.

y

activity. But the swarms almost always taper off and end with a whimper. The Noto Peninsula swarm produced a bang. “I can’t think of another earthquake swarm globally that preceded such a large event,” says Zachary Ross, a geophysicist at the California Institute of Technology. Scientists are now puzzling over the details of that process and how the swarm may have led to the 1 January shock. “There are many questions to be resolved,” says Kyoto University geodesist Takuya Nishimura. Earthquake swarms typically occur when heat from deep underground produces highpressure fluid—water, gas, or magma—that permeates fault systems, says University of Tokyo seismologist Aitaro Kato. The fluid lubricates faults, resulting in slow slips that

g

A 1 January earthquake triggered the collapse of many traditional wooden buildings in Japan.

The overwhelming majority of the earthquakes in the swarm have been weaker than magnitude 4. But magnitudes have gradually increased, Kato says. A magnitude 5.4 earthquake occurred in June 2022 and a magnitude 6.5 event in May 2023 killed one person and caused widespread damage. Nishimura says the deep fault movements associated with the earthquake swarm probably added strain and stress to faults higher up in the crust until they finally ruptured in the larger earthquakes. Then came 1 January. It was the strongest temblor to hit the Sea of Japan coast since 1993. In addition to the deaths, scores of people remain missing, and upward of 30,000 people are in emergency shelters. Landslides blocked highways connecting the peninsula to Japan’s main island, hampering rescue and relief efforts. And ongoing aftershocks threaten further destruction. How the earthquake might be connected to the swarm is unclear, Nishimura says. Rising fluids might have greased the fault that ruptured, for example, or slip from the swarm could have loaded the fault with stress. The earthquake epicenter is within the swarm region, and researchers are now plotting the distribution of the aftershocks to determine whether the quake occurred on faults involved in the swarm. Even if it began there, the rupture extended far beyond the swarm region, both farther along the Noto Peninsula and undersea toward Sado Island. “It is important to learn how to distinguish an earthquake swarm that might lead to a large earthquake from those that do not pose such a threat,” Nishimura says. Some scientists think that’s ambitious. “I think we should take a wait-and-see attitude,” says Robert Geller, a seismologist and professor emeritus at the University of Tokyo. The earthquake highlighted the vulnerability of Japan’s traditional wooden buildings. Engineers have long warned that their heavy tile roofs, which stand up well to Japan’s frequent typhoons, are prone to collapse during earthquakes. (Authorities say collapsing houses caused most of the casualties.) And they are flammable. A fire sparked by the quake in the city of Wajima on Noto’s northern coast spread through a historic market area in the center of town, burning some 200 wooden buildings to the ground. j

p

ig earthquakes sometimes bring big surprises. The magnitude 7.5 earthquake that struck Japan on New Year’s Day, killing more than 160 people, is no exception. Over the past 3 years, tens of thousands of small to moderate earthquakes, most barely noticeable, had rattled the Noto Peninsula, a finger of land that juts 100 kilometers into the Sea of Japan from the west coast of Honshu, Japan’s main island. Such earthquake swarms occur throughout the world, typically in areas of volcanic or geothermal

produce small earthquakes. Unlike the brief flurry of small quakes that often precedes a major temblor, swarms continue rumbling for days or years without generating an identifiable main shock. The ongoing collision of tectonic plates off Japan’s east coast produces the country’s most damaging earthquakes, including the 2011 magnitude 9 Tohoku earthquake that also triggered a mammoth tsunami (Science, 18 March 2011, p. 1375). But Japan’s west coast is also seismically active, with an intensely fractured and permeable fault zone. Over the past 20 years, several magnitude 6 earthquakes occurred near the Noto

N E WS

FEATURES

p g

Four largely ignored coronaviruses circulate in humans without causing great harm and may portend the future for SARS-CoV-2 By Jon Cohen

y ,

PHOTOS: (TOP TO BOTTOM) MATT ELLENBOGEN/LA JOLLA INSTITUTE FOR IMMUNOLOGY; SCOTT CAMAZINE/SCIENCE SOURCE

y g

COVID’S COLD COUSINS

y

In the wake of COVID-19, labs have ramped up studies of four widespread human coronaviruses that cause common colds.

O

ver a few weeks in November 1889, a respiratory disease attacked half the residents of St. Petersburg, Russia, and it soon began to race through Europe and the rest of the world. Two years later, in a spectacularly detailed book, a British medical officer, H. Franklin Parsons, described what was dubbed the “Russian influenza” epidemic, which raged until 1894. People seemed to spread the disease before developing symptoms, the young did not suffer as much as the old, a dry cough was common among the ill, some had a “perversion of taste and smell,” and deaths SCIENCE science.org

The human coronavirus dubbed OC43.

rose. Suspicions ran high that a pathogen had jumped from an animal into humans. Sound like COVID-19? In 2005, scientists in Belgium proposed that the earlier pandemic’s cause was not

an influenza virus, but rather a coronavirus. Three years before their theory was published, a coronavirus had passed from an animal to humans, touching off a highly lethal outbreak of what was called severe acute respiratory syndrome (SARS). The disease spread from China and brought new attention to these once-obscure viruses. The Belgian team wondered whether something similar happened in Russia more than a century ago. Based on molecular clues, they suggested that the once-deadly virus is still circulating today, as a coronavirus known as OC43 that in most people causes nothing worse than a cold. So far there’s no direct evidence to back the group’s theory, but two 12 JANUARY 2024 • VOL 383 ISSUE 6679

141

NE WS | F E AT U R E S

Family history

observations. “Referees said the images she produced were just bad pictures of influenza virus particles,” wrote Tyrrell in a book he co-authored, Cold Wars: The Fight Against the Common Cold. Her new images of B814 made a convincing case that COLD CORONAS the various viruses were a reNatural reservoir host Intermediate host Human virus lated, unrecognized group. Discovered | genus “So what should we call them? ‘Influenza-like’ seemed a bit 229E feeble, somewhat vague, and 1962 | alpha probably misleading,” Tyrrell recalled. But he and Almeida noticed “a kind of halo surrounding them … and so the OC43 name coronavirus was born.” 1966 | beta Around the same time, infectious disease specialists Dorothy Hamre and John Procknow at the University of Chicago were NL63 ? conducting their own hunt for 2003 | alpha new cold viruses in medical students there. In 1966, they reported having grown a virus, designated 229E, from a particiHKU1 ? 2004 | beta pant who had a “minor upper respiratory illness.” They gave samples to Tyrrell, whose team intentionally infected people KILLER COUSINS with it and showed, again by handkerchief counts, that 229E caused a mild cold, à la B814. SARS-CoV-1 The two viruses looked identical 2003 | beta under the microscope, but researchers could adapt only 229E to a cell line—and B814 was lost to history before any genetic MERS-CoV comparison could take place. 2012 | beta The researchers behind a long-running cold study at the U.S. National Institutes of Health (NIH) in 1967 reported what ? SARS-CoV-2 would prove to be a clearly dis2020 | beta tinct second coronavirus, OC43. “We advertised to the employees at the NIH to come by Building THE FIRST HUMAN coronavirus was isofrom the trachea—the natural habitat for 7, third floor if you had a cold, and we would lated 6 decades ago from the runny respiratory viruses—extracted from aborted be very delighted to wash out your nasal pasnoses of English school boys. In the winfetuses. One sample, dubbed B814, yielded a sages and collect the fluids,” remembers Ken ter of 1960–61, virologist David Tyrrell, new virus. “After considerable initial doubts McIntosh, then a young medical doctor who who ran the Common Cold Unit in the we now believe that the B814 strain is a viran the project in the lab of Robert Chanock. United Kingdom, and co-workers looked rus virtually unrelated to any other known Once again, electron microscopy showed for viruses in the boys’ handkerchief goop. virus of the human respiratory tract,” a virus similar in shape to the one that When they couldn’t identify any known cold Tyrrell and colleagues reported in 1965. causes avian infectious bronchitis. (Initially, virus, they inoculated adult volunteers with The next year Tyrrell sent samples of B814 McIntosh could only grow it in the organ culextracts from the nasal washings to confirm to June Almeida, a talented electron microture medium Tyrrell had used—hence the OC that something in the samples caused colds. scopist who did not have a university degree, in the isolate’s name—but it, too, was eventuYet nothing from those disease-bearing at St. Thomas’ Hospital in London. She really adapted to a cell line.) samples would grow in standard culture ported back that she had seen similar viral Yet research on the new viruses languished. media. So they turned to an odd culture particles in samples from chickens with in“Working with them was so awkward and difsystem recently developed for certain influfectious bronchitis and mice with hepatitis, ficult that nobody wanted to do it,” McIntosh enza and adenoviruses: cilia-bearing cells although she had been unable to publish her says. By January 2003, only a few hundred

y g y ,

science.org SCIENCE

y

12 JANUARY 2024 • VOL 383 ISSUE 6679

g

142

Coronaviruses have repeatedly spilled over from bats or rodents— the natural reservoir for many of them—to other animals such as camels, cows, or civets before jumping into humans. With SARS-CoV-2, raccoon dogs and other mammals sold at a Wuhan, China, market could have been intermediate hosts, although that remains contentious.

p

other teams soon hope to look at tissue samples from the late 19th century to see whether they can spot when the virus first became a human pathogen. This upcoming search for OC43’s roots is part of a flurry of research, since COVID-19 erupted globally 4 years ago this month, on it and the three other coronaviruses that cause common colds. Long ignored except by a tiny scientific community, these pathogens with clunky, alphanumeric names— NL63, 229E, and HKU1 are the other three—are now getting their due. Some groups are reexamining how the viruses leapt from animals to people, in part to understand how SARS-CoV-2, COVID-19’s cause, may have emerged. Studying the four may also illuminate whether other coronaviruses discovered in wild and domesticated animals pose a threat to humanity. And some scientists are exploring how immune responses to these four overlap and interact with the response to SARS-CoV-2. The four viruses currently show up each fall and winter, accounting for up to 30% of the colds we endure. But all may once have caused more serious disease, suggesting to some virologists that they offer a hopeful glimpse of COVID-19’s future. “Those four are the model system of what’s ahead for us,” predicts Lia van der Hoek, a virologist at the Amsterdam University Medical Centers who in 2003 discovered NL63. “SARS-CoV-2 is going to become a common cold. At least that’s what we want.”

studies had appeared about human coronaviruses, and most who did coronavirus research were interested in those that sickened animals. “[Coronavirus] people who studied human medicine were rare,” says Leiden University virologist Eric Snijder, who recalls struggling that January to draw scientists to a meeting he co-organized on nidoviruses, the order that includes coronaviruses. Then, in April 2003, researchers reported that the deadly, atypical pneumonia spreading through China, soon to be called SARS, was caused by a coronavirus. As the disease began to sicken people elsewhere and triggered international alarm, last-minute registration for the May meeting jumped

man coronavirus, HKU1, in a 71-year-old man who had an unexplained pneumonia. Both van der Hoek and Woo, now at National Chung Hsing University, doubt that there are more human coronaviruses circulating widely that researchers have yet to detect. “For years and years and years, people have screened respiratory samples … and no other [common cold] coronavirus has been identified,” van der Hoek says. “I’m convinced that there are only those four.” But some veteran coronavirologists are more circumspect. “How could there only be four?” asks Susan Weiss at the University of Pennsylvania, who has studied coronaviruses for 40 years. “It doesn’t make sense

143

PHOTO: PA IMAGES/ALAMY

,

12 JANUARY 2024 • VOL 383 ISSUE 6679

y

coronavirus that caused SARS, scientists had mapped out a convincing origin scenario. A virus in civets and raccoon dogs sold in marketplaces in southern China matched the one that sickened humans, and a virus later found in bats looked like its ancestor. This triggered an international

y g

WITHIN A FEW YEARS of the discovery of the

y

SCIENCE science.org

to me.” The University of Iowa’s Stanley Perlman, another coronavirus old-timer, says it’s important to keep looking for new human ones. “In 2002, we felt we were finished when we had 229E and OC43,” Perlman says. “We always get deceived when we think we’re finished.”

g

from 130 to 170, and SARS was added to the schedule. A human coronavirus had finally caught the wider scientific community’s attention, and two more were soon uncovered. Van der Hoek found the one she called NL63 in a nasal sample from a 7-monthold girl in the Netherlands who recently had fever, pink eye, and a runny nose. Ron Fouchier’s lab at nearby Erasmus Medical Center simultaneously uncovered what seemed to be the same virus, and both team’s findings appeared online within a few weeks of each other in the early spring of 2004. Before the year was out, a team led by clinical microbiologist Patrick Woo at the University of Hong Kong discovered another hu-

p

David Tyrrell in 1966 searched for common cold viruses by squirting nasal washings into the noses of volunteers.

push to sample bats and other animals for coronaviruses that might pose threats to humans, leading to thousands of viral sequences being cataloged. Although most of these coronaviruses have been identified only by sequencing fragments of their genome—getting intact viruses that grow in culture is often difficult—the viral family is clearly abundant in many species. And other mammals appear to be the source of all known cold-causing coronaviruses. The Belgian researchers studying the 1890s pandemic, for example, sequenced the genome of OC43 and found “remarkable” genetic similarities to a coronavirus found in cows. Using the estimated mutation rates of the bovine virus and OC43, they created a molecular clock and calculated that the two viruses shared a common ancestor somewhere around 1890. (The range went from 1815 to 1918.) The timing led the scientists to wonder whether the bovine cousin hopped into humans as a much more lethal pathogen and over time became the relatively mild OC43 seen today. “It seemed like an interesting coincidence that when we estimated the divergence time of the bovine virus and human OC43 it was basically spot on the date you would expect with the Russian flu epidemic,” says Philippe Lemey of KU Leuven, a co-author of the study, published in the Journal of Virology. He and his colleagues pointed out that between 1870 and 1890 an epidemic of pneumonia in cows led to “massive culling” of the animals in industrialized countries. This provided “ample opportunity for the culling personnel to come into contact with bovine respiratory secretions” that could have contained OC43’s precursor, they wrote. In 2022, a French team published a study in Microbial Biotechnology reporting “very preliminary” biological evidence that supports the OC43 hypothesis: They found antibodies to the virus in the dental pulp of World War I soldiers who were alive at the time of the Russian flu and died in battle in 1914. None of the analyses that link OC43 to the Russian flu persuade Michael Worobey, a University of Arizona evolutionary biologist who has collaborated with Lemey on high-profile studies about the origin of SARS-CoV-2. “I see it as extremely unlikely,” Worobey says. As he argued in a 2014 paper in the Proceedings of the National Academy of Sciences, “compelling evidence” ties the global outbreak to a specific influenza viral variant—including a study of stored samples from people born as far back as 1876 that found antibodies to a novel flu virus from the time of the pandemic. Worobey is now hoping to resolve the debate by obtaining archived tissue from

NE WS | F E AT U R E S

y g y ,

science.org SCIENCE

y

12 JANUARY 2024 • VOL 383 ISSUE 6679

g

144

p

people seen at a London hospital virus that infected a few Malaysians, around 1890 and looking for lingerhumanity is under constant, low-level ing genetic sequences of influenza or siege from the viruses. “I think there coronaviruses. A research group from are certainly other animal coronaSpain has identified “suitable samviruses circulating that are challengples” from that time period as well, ing human immune systems.” at the Basque Museum of the History of Medicine and Science. It plans to WHEN SARS-COV-2 BEGAN to gallop probe them soon. around the world, researchers wonThe other common cold coronadered whether our immune memories viruses are also thought to have leaped of its four milder relatives could blunt from animals. Hipposideros bats in the impact of the ferocious new virus. Ghana harbor a relative of 229E, a team All coronaviruses share the same baled by virologist Christian Drosten, sic repertoire of proteins, suggesting now at the Charité University Hospiimmune responses built up over retal of Berlin, reported in 2009. The peated exposure to colds might ease researchers estimated that the bat COVID-19. The evidence is mixed. virus and 229E have a common ancesFor one thing, the surface protor that dates back to between 1659 tein of SARS-CoV-2, called spike, and 1803, suggesting that’s the period differs markedly from the ones when it found a way to humans. that stud its cold-causing cousins. Like the SARS virus, it may have As a result, antibodies to the cold come via an intermediate species. coronaviruses don’t prevent infecStudies, some done by Drosten’s tions with SARS-CoV-2 or blunt the team, found 229E relatives in healthy symptoms it causes. A report in the dromedary camels in the Arabian 6 September 2023 issue of Science Peninsula and Africa, firming up the Translational Medicine even sugtheory. Drosten’s team also charted a gests that previous exposure to OC43 bat-to-camel-to-human pathway for might leave people with antibodies to the highly lethal coronavirus that its spike that can interfere with the causes Middle East respiratory synimmune system’s attempt to make drome, which was first recognized antibodies against the SARS-CoV-2 in 2012. A similar scenario also looks surface protein, increasing the risk likely for SARS-CoV-2, which some of developing the lasting, debilitatevidence suggests may have passed ing symptoms known as Long Covid. from bats to people via an animal Yet a bevy of studies early in the host such as raccoon dogs or other pandemic showed that other immune susceptible species known to have memories of the common cold coronabeen sold in a Wuhan, China, food viruses did help. “It has been well market that had the earliest cluster established that prepandemic, some of COVID-19 cases. people had preexisting immune reThe other two cold coronaviruses activity to SARS-CoV-2, and it was have less certain origins. NL63 has an consequential,” says immunologist ancestor found in tricolored bats in Alessandro Sette. His group at the Maryland. A genetic comparison with La Jolla Institute for Immunology is Electron microscopist June Almeida (top, in 1963) discovered that the bat virus suggests NL63 crossed among several to have reported that, the human virus dubbed B814 (bottom) resembled certain chicken into humans 563 to 822 years ago, in test tube experiments, T cells from and mouse viruses in its shape and surrounding “halo.” according to a 2012 estimate in the people who had never been infected Journal of Virology. HKU1 has the murkishe’s leading an effort to find novel coronaby SARS-CoV-2 could sometimes recognize est evolutionary history, but its genetic viruses in farmers who handle livestock. and destroy other cells infected by the visequence clusters close to the murine hepa“There probably is fairly frequent transrus. “We and others have shown that, at titis virus, suggesting it has a rodent origin. fer of zoonotic coronaviruses into the huleast in some cases, this could be mapped to In a chapter about human coronaviruses man population,” says J. Glenn Morris, an similarities between common cold [coronathat Drosten and co-authors wrote for Adepidemiologist who heads the Emerging virus] sequences and SARS-CoV-2 sevances in Virus Research in 2018, they Pathogens Institute at the University of quences,” Sette says. noted it was “peculiar” that no great apes Florida. But many then fail to spread furAnother study found that health care other than humans have their own coronather, he suspects. Indeed, over the years workers who had T cell responses to certain viruses. “This absence provides further Morris, Vlasova, and others have identified coronavirus proteins, besides spike, that are support to the suspicion that contact with coronaviruses from cows, dogs, cats, and similar in the cold viruses and SARS-CoV-2 domestic animals may have been essential pigs that appear to have infected people, appeared to abort infections with the latter. in human acquisition of most or all endemic then petered out. Other research documented that household CoV,” they concluded. Anastasia Vlasova, a To Gregory Gray, an epidemiologist at contacts of people with SARS-CoV-2 had a virologist at Ohio State University, may soon the University of Texas Medical Branch who lower risk of becoming infected themselves have further evidence for that theory as helped Vlasova uncover a canine coronaif they had T cells that reacted to proteins in

g

lockdowns, all four of the common cold coronaviruses returned, according to her unpublished analysis. “I don’t think that SARS-CoV-2 has any effect on their circulation,” she says.

y

“other” coronaviruses is different: She thinks they foreshadow the likely future of SARS-CoV-2. She is struck by how sharply severe disease and death from SARS-CoV-2 has dropped over the past 4 years, shifting its status from a widely feared killer to yet another human coronavirus that, at least in people under age 65 who have no comorbidities, causes little acute harm. Indeed for many, Long Covid has become more of a worry than immediate hospitalization. The early ferocity of the virus has much to do with the fact that aside from some possible modest protection from previous colds, the world population in January 2020 was immunologically blindsided by the new infection. But van der Hoek suspects that an evolutionary “trade-off” has also defanged SARS-CoV-2: As the virus has spread to billions of people, it may have become less virulent so it can spread more readily. “When viruses jump species, they are not adapted to their hosts at all, and they don’t take into account that the host must survive for them to survive,” she says. Each of the four common cold coronaviruses, she contends, probably came in lethally hot and then cooled down. “This must have happened with all four of them,

12 JANUARY 2024 • VOL 383 ISSUE 6679

145

,

SCIENCE science.org

TO VAN DER HOEK, the significance of the

and this is just number five,” she says. “Coughing in your own bed is bad for transmission of an acute respiratory virus. As soon as they start to adapt to their hosts, they let those infected people walk on the streets and go shopping.” But evolutionary biologist Jemma Geoghegan at the University of Otago is skeptical. Geoghegan co-authored a December 2018 article in Nature Reviews Genetics that calls into question the entrenched idea that emerging viruses become less virulent to persist. “I think the classic view is wrong,” says Geoghegan, whose article offers several examples of viruses—including HIV—that did not weaken over time. She notes that SARS-CoV-2 begins to spread before people develop symptoms and often doesn’t even sicken the immunologically naïve—which means there’s little evolutionary pressure for it to become less virulent. “There’s no selection for this reduced virulence/transmission trade-off.” The procession of SARS-CoV-2 variants contributes to Geoghegan’s skepticism. Delta was more virulent than the original virus that emerged in Wuhan. Omicron, the next to emerge, took over because it spreads more quickly, not because it’s milder. There’s no sign of the supposed trade-off, she says. So put an asterisk on the notion that SARS-CoV-2 is heading down an evolutionary path to becoming as docile as OC43 and the other cold coronaviruses. “Omicron still hospitalizes and kills lots of people,” Geoghegan says. “It’s not there yet.” j

y

the viral capsule of OC43 and HKU1. Immunity to the common cold cousins also seems to lead to less severe COVID-19, and Sette’s group showed that it improved responses to COVID-19 vaccines. Early in the pandemic, this preexisting coronavirus immunity might have significantly reduced the toll of SARS-CoV-2. But it may have little importance today, Sette says, because “the vast majority of the planet has been exposed to SARS-CoV-2 and vaccinated against SARS-CoV-2.” Infectious disease specialist Manish Sagar of Boston University and co-workers have flipped this issue on its head, asking whether immunity to SARS-CoV-2 protects against the common cold. They looked for the cold-causing coronaviruses in nasal swabs of nearly 5000 people who came to the Boston Medical Center between November 2020 and October 2021. People who had prior SARS-CoV-2 infections were 50% less likely to have symptomatic disease from one of the four, they reported in a bioRxiv preprint on 24 October 2023. T cells that targeted two of the internal proteins of OC43, the most frequently found cold coronavirus in their study, likely explained the benefit. But van der Hoek has also examined the cross-immunity question and come to a different conclusion. In the fall of 2021 her team began to test respiratory samples to see whether SARS-CoV-2 affected the presence of common cold coronaviruses. After the Netherlands ended its COVID-19

y g

ILLUSTRATION: SCIENCE HISTORY IMAGES/ALAMY

p Was the “Russian flu” pandemic from 1889 to 1894, depicted here in the United Kingdom’s Illustrated Police News, actually caused by a coronavirus now linked to mild colds?

INSIGHTS PERSPECTIVES

IMMUNOLOGY

p

Decoding the language of immunity g

Optimized transfer RNA (tRNA) codon use can speed up antibody generation

y y g

146

12 JANUARY 2024 • VOL 383 ISSUE 6679

ent factors, codon usage bias is particularly prominent in highly expressed genes and appears to correlate with the copy number or abundance of cognate tRNAs. The number of tRNAs not only differs from species to species but also can vary even across tissues and developmental stages in multicellular organisms. In humans, there are several hundred tRNA genes with varying copy numbers spread across the genome. Furthermore, posttranslational modifications of tRNAs, including so-called “wobble modifications,” increase the range of syn1Department of Medicine, Icahn School of Medicine

at Mount Sinai, New York, NY, USA. 2Ichor Biologics, LLC, New York, NY, USA. 3Centre for Immunobiology, Blizard Institute, Queen Mary University of London, London, UK. Email: [email protected]

science.org SCIENCE

,

A

ntibodies are critical for human health, providing long-lived and exquisitely specific immunity to pathogens. Plasma cells secrete tens of thousands of antibody molecules every second and sustain this output continuously for several decades. Plasmablasts and plasma cells [collectively, antibodysecreting cells (ASCs)] are generated following the activation and differentiation of B cells, which involves complete reprogramming of the transcriptional machinery and remodeling of the endoplasmic reticulum and Golgi apparatus. ASCs have co-opted the unfolded protein response (UPR), a

stress response, to accommodate the exceptional rate of protein synthesis and secretion required. On page 205 of this issue, Giguère et al. (1) reveal that mRNA encoding antibody genes is enriched with codons recognized by modified transfer RNA (tRNA). This codon optimization is accompanied by an increase in tRNAs with complementary modifications, which serves to enhance antibody biosynthesis by ASCs. This has wider implications for therapeutic protein production, vaccine design, and beyond. Eighteen of the 20 amino acids are encoded by two or more synonymous codons. Organisms display preferential biases in the selection of synonymous codons, which is thought to play a role in efficient translation. Although dependent on several differ-

y

By Raymond A. Alvarez1,2 and Louisa K. James3

In transfer RNAs (tRNAs), deamination of adenosine-34 (A34) to generate inosine-34 (I34) by the enzyme adenosine deaminase acting on RNA 1 (ADAR1) extends the range of pairing from U-ended codons to include A- and C-ended codons on eight different tRNAs. Anti-codon

tRNA

C G A G C U

ADAR1

Codon tRNA

G C U

tRNA A34 mRNA

C G I

G C C

I34

G C A

Optimizing translation I34-dependent codons and inosinecontaining tRNAs are enriched in antibody-secreting cells (ASCs), resulting in enhanced antibody secretion. l34 dependence

Memory B cell

ASC

Differentiation

y ,

served in aggressive cancers and may also provide insights into their initiation and evolution, offering new avenues for understanding and therapeutic development targeting aberrant translational processes. Giguère et al. revealed that codon usage plays a role in the selection of B cells into the memory compartment. Following activation, B cells undergo affinity-based selection to expand rare clones with the strongest antigen-binding variable regions. By examining the human postvaccination antibody repertoire, Giguère et al. identified an enrichment in I34-dependent codons in the mRNA encoding the antigen binding domains (variable regions) of plasma cells and memory B cells relative to the naïve B cell repertoire. Notably, these insights hold

y g

Naïve B cell

Antibodies

y

SCIENCE science.org

Boosting antibody generation

promise for optimizing protein expression in the mammalian cell lines used for manufacturing recombinant proteins, such as monoclonal antibodies for clinical use. Strategies such as media formulation, selective pressure, and cell and vector engineering have all been used to increase expression and improve the stability of these cell factories (10, 11). Although these strategies have enhanced stability and yields, a greater understanding of the natural process that governs antibody secretion from ASCs may yield considerable advances. For instance, designing constructs that tailor codon usage to cell-specific tRNA expression may revolutionize recombinant protein production, especially the synthesis of antibodies. Alternatively, modifications to the expression profiles of tRNA in producer cell lines could also be engineered to complement the requirements of particular monoclonal antibody sequences, enhancing the efficacy and yields of therapeutic protein production. Similarly, understanding how specific codons interact with the tRNA pool opens avenues for more efficient design of mRNAbased therapeutics with enhanced protein expression and potentially vaccine efficacy and duration of immunity. Indeed, tailoring mRNA sequences to align with tRNA availability may enable optimal vaccine designs and fine-tune protein expression systems. Deciphering the intricate network of signaling pathways and regulatory factors driving tRNA remodeling in ASCs is crucial; understanding how tRNA modifications affect translational fidelity, efficiency, and protein folding during antibody synthesis warrants further investigation. Whether altered tRNA pools influence broader aspects of cellular fitness beyond antibody production remains unknown. The study by Giguère et al. provides a foundational understanding of tRNA remodeling and codon usage in humoral immunity, representing a paradigm shift in our understanding of translation dynamics. j

g

onymous codons that can be translated, for example, inosine-34 (see the figure). Giguère et al. found that, relative to other genes, antibody mRNA sequences from both human and mouse ASCs display a conserved bias toward codons recognized by inosine-34– modified tRNA and that ASCs are enriched with tRNA with inosine modifications. In a mouse model, reducing the frequency of inosine-34–dependent codons resulted in decreased antibody production, confirming that coordination of inosine-34–dependent codon bias and inosine tRNA abundance is a mechanism to enhance antibody biosynthesis. Understanding the regulatory mechanisms governing tRNA modification requires further research. This coordination may extend beyond antibody production, potentially offering insight into fundamental regulatory mechanisms governing cellular fitness and function. Adenosine deaminase acting on RNA (ADAR) enzymes, which are responsible for catalyzing RNA adenosine-to-inosine posttranscriptional modifications, have been associated with a wide range of human diseases, including cancer and neurological, metabolic, and autoimmune diseases (2, 3). Humans express three ADAR proteins that are highly conserved across vertebrates. ADAR1 is essential for mammalian development (4). Ablation of Adar1 in mice leads to lethal defects during embryonic development, notably with severe defects in hematopoiesis (5). In human embryonic stem cells (hESCs), inhibiting ADAR1 expression reduces their ability to differentiate into neurons (6). Analogous to ASCs, during hematopoiesis, stem cells rapidly synthesize high levels of proteins to undergo cellular differentiation and (re)generate new or damaged tissues and maintain tissue homeostasis. Recent studies have suggested that RNA editing is critical for regulating stem cell fate and function (7), but how dynamic regulation of tRNA profiles dictates cell fate decisions, influencing differentiation versus proliferation, is unknown. Indeed, the correlations drawn from antibody production may be relevant to understanding how tRNA modifications may orchestrate the delicate balance between cell differentiation and maintenance of pluripotency, offering new perspectives on the regulatory networks governing cellular fate determination. In the case of cancer, the metabolic demands are somewhat analogous to those observed in ASCs. Increased ADAR1 expression and A-to-I editing activity have been as-

sociated with progression in several cancers, including leukemias and solid tumors (8, 9). Although increases in ADAR-mediated inosine modifications have been identified, alterations in tRNA pool availability and their impact on cellular translational dynamics remain undefined. The parallels drawn from antibody production may reveal the potential relevance of tRNA adaptations in elevated metabolism and proliferation ob-

p

GRAPHIC: N. BURGESS/SCIENCE

Specific codon usage in mRNAs (blue) and modification of transfer RNAs (tRNAs; center) enhance antibody production.

RE FE REN C ES AN D N OT ES

1. S. Giguère et al., Science 383, 205 (2024). 2. W. Slotkin, K. Nishikura, Genome Med. 5, 105 (2013). 3. B. Song, Y. Shiromoto, M. Minakuchi, K. Nishikura, Wiley Interdiscip. Rev. RNA 13, 1665 (2022). 4. J. C. Hartner et al., J. Biol. Chem. 279, 4894 (2004). 5. Q. Wang, J. Khillan, P. Gadue, K. Nishikura, Science 290, 1765 (2000). 6. T. Chen et al., Cell Res. 25, 459 (2015). 7. D. Lu, J. Lu, Q. Liu, Q. Zhang, Biomark. Res. 11, 61 (2023). 8. K. Fritzell, L. D. Xu, J. Lagergren, M. Öhman, Semin. Cell Dev. Biol. 79, 123 (2018). 9. X. Xu, Y. Wang, H. Liang, Curr. Opin. Genet. Dev. 48, 51 (2018). 10. B. Bachhav, J. de Rossi, C. D. Llanos, L. Segatori, Biotechnol. Bioeng. 120, 2441 (2023). 11. S. Xiao, J. Shiloach, M. J. Betenbaugh, Curr. Opin. Struct. Biol. 26, 32 (2014).

10.1126/science.adn1067 12 JANUARY 2024 • VOL 383 ISSUE 6679

147

I NS I GHTS | P E R S P E C T I V E S

PHYSICS

Electrons catch light pulses on the fly Energy exchange between electrons and photons enables ultrafast probing of materials By Albert Polman1 and F. Javier García de Abajo2,3

y g y ,

science.org SCIENCE

y

12 JANUARY 2024 • VOL 383 ISSUE 6679

g

148

p

ity, thereby increasing the light intensity to a high value and enabling PINEM to be performed with a low-power continuous-wave lthough free electrons are widely used laser light source (7). This capitalizes on the as probes to analyze the morphology, ability of integrated optical circuits to comatomic structure, and optical propbine waveguides, multiplexers, and other operties of nanoscale materials using tical components to process information entransmission or scanning electron tirely using light, rather than using electrons microscopy, they can also interact as in common electronic integrated circuits. strongly with light and, thus, sample the Yang et al. exploited a fascinating property distribution of optical fields in and around of some optical materials such as silicon nithose materials with high spatial resolution tride—a relatively strong nonlinear optical (1–3). Specifically, in photon-induced nearresponse. The speed of light in silicon nitride field electron microscopy (PINEM), intense depends on light intensity even for moderate light fields bring an electron laser powers. Above a certain into multiple energy states sipower threshold, such nonlinemultaneously—a quantum-mearity causes light to propagate Probing electron–light interaction chanical superposition state— in complex ways. This includes A high-energy beam aligned with a microring cavity can probe the ultrafast evolution which are then measured with the formation of Kerr solitons of optical soliton pulses. Strong electron–photon interaction tailors the production an electron spectrometer. On (8), light pulses generated from of quantum-mechanical electron wave packets in space and time. page 168 of this issue, Yang et al. a continuous laser that circulate Soliton pulse Laser light (4) report the use of PINEM to in the microring cavity. examine the creation of special Yang et al. brought together Photonic chip light pulses known as solitons free electrons and nonlinear Continuous-wave in an optical integrated circuit. optics by using PINEM with pump direction The electron–soliton interac200-keV electrons to probe Kerr tion shapes the electron’s probsolitons in a silicon nitride miability distribution in space croring of hundreds of microme– Circulating light and time, thereby enabling new eters in diameter. This estabways to probe ultrafast dynamlished a disruptive approach to Electron wave e– packet ics in matter with an electron explore electron–light interacmicroscope. tions and to shape the free-elecElectron beam Microring In quantum mechanics, a tron probability density. The free-space electron is described structure is pumped with a as a wave packet characterized by a spatiwhich is a few femtoseconds in the visible continuous-wave laser at a wavelength comotemporal distribution of probability denregime. PINEM thus pushes the combined monly used for on-chip telecommunication sity—the likelihood of an electron being presspatiotemporal resolution of electron microapplications (1.55 µm). The authors aligned ent at a given position in space as a function scopes to the subfemtosecond and nanomthe electron beam (which travels in vacuum of time. A distinct feature of PINEM is that, eter (sub-fs/nm) regime. Besides reshaping along a straight path) and the ring such that by tailoring the light field, electron–light the electron probability density, the sidethe electron interacts twice along the cavity interaction enables control over this distriband amplitudes provide a direct measure of circumference, thereby probing the optical bution of the electron probability density. the intensity associated with the optical near field at two different moments in time (see Starting with an incident electron of energy field traversed by the electron. Yang et al. the figure). The delay between the two interE0 and a smooth spatial density distribution leveraged this effect to measure the intensity actions is controlled by the electron beam curve, its interaction with light expands the associated with the generation of solitons. distance from the cavity center (i.e., zero deenergy spectrum, creating energy states The creation of strong electron–light inlay when the beam is tangent to the ring, and (sidebands) that are separated from E0 by teractions in PINEM requires intense optia maximum delay determined by the ring positive and negative multiples of the phocal fields, which are commonly achieved by diameter when the beam intersects the ring ton energy "v. After the initial observation using ultrashort intense laser pulses under center). This provides a means to examine of PINEM and associated electron sideband far-field illumination conditions (5). Intense the temporal propagation and phase of light fields are also achieved by using optical mias it circulates the ring cavity. In particular, 1Center for Nanophotonics, NWO Institute AMOLF, croring waveguides made of highly transparthe phase is proportional to the delay, thus Amsterdam, Netherlands. 2ICFO-Institut de Ciencies ent silicon nitride and fabricated on a silicon allowing Yang et al. to observe characteristic Fotoniques, The Barcelona Institute of Science and chip through lithographic techniques (6). Ramsey interference patterns of light as the Technology, Castelldefels, Spain. 3ICREA-Institució Catalana When light is injected into a microring, it cirdistance between the two points of light– de Recerca i Estudis Avançats, Barcelona, Spain. Email: [email protected] culates hundreds of times inside the ring cavelectron interaction is gradually changed.

A

spectra (1), a well-defined coherence relation was established between the electron waves at different sideband energies for every electron (2, 3). This coherence has important implications for propagation of the electron because sidebands with different energies have different electron velocities, and therefore, the electron wave packet that is formed by the superposition of all sidebands undergoes a spatiotemporal reshaping as the electron propagates (3). Specifically, the probability density distribution of each electron takes the shape of a train of pulses of short duration compared with the period of the light,

Practical challenges for precision medicine The prediction of individual treatment responses with machine learning faces hurdles By Frederike H. Petzschner

P

B. Barwick et al., Nature 462, 902 (2009). F. J. García de Abajo et al., Nano Lett. 10, 1859 (2010). A. Feist et al., Nature 521, 200 (2015). Y. Yang et al., Science 383, 168 (2024). M. Liebtrau et al., Light Sci. Appl. 10, 82 (2021). Y. Liu et al., Science 376, 1309 (2022). J.-W. Henke et al., Nature 600, 653 (2021). T. Herr et al., Nat. Photonics 8, 145 (2014). L. A. Lugiato, R. Lefever, Phys. Rev. Lett. 58, 2209 (1987).

AC KN OWL EDG M ENTS

The authors receive support from European Research Council (ERC) 789104/eNANO; ERC 01019932/QEWS; and European Commission FET-Proactive 101017720/EBEAM.

Training Dataset

Training

Test

Training data from single patients

Classification “Responder”

“Nonresponder”

Cross-validation

Independent dataset

Validation “Responder?”

Out-of-sample validation

10.1126/science.adn1876 SCIENCE science.org

12 JANUARY 2024 • VOL 383 ISSUE 6679

149

,

1. 2. 3. 4. 5. 6. 7. 8. 9.

y

Supervised machine learning for individual treatment prediction is based on the development of classifiers. To prevent overfitting, model validation is crucial, typically achieved through cross-validation or out-of-sample validation. Out-of-sample validation requires a completely independent dataset and is more resource-intensive, but this approach is less susceptible to overfitting and can provide more generalizable results.

RE F ER E NC ES AND NOTES

y g

Individual treatment prediction using machine learning

y

recision medicine promises treatments tailored to individual patient profiles. Machine learning models have been heralded as the tools to accelerate precision medicine by sifting through large amounts of complex data to pinpoint the genetic, sociodemographic, or biological markers that predict the right treatment for the right person at the right time. However, the initial enthusiasm for these advanced predictive tools is now facing a sobering reality check. On page 164 of this issue, Chekroud et al. (1) show that machine learning models that predict treatment response to antipsychotic medication among individuals with schizophrenia in one clinical trial failed to generalize to data from new, unseen clinical trials. The findings not only highlight the necessity for more stringent methodological standards for machine learning approaches but also require reexamination of the practical challenges that precision medicine is facing. What predicts whether a patient will benefit from a particular treatment? The answer may lie in their genetics, biology, sociodemographic background, social environment, past experiences, or a myriad of other potential factors. Machine learning techniques

have the capacity to analyze large datasets and identify the most effective combination of features that accurately predicts a variable of interest. They thus offer a promising avenue for discovering relevant features or biomarkers that predict individual treatment responses. Typically, this involves training the model on a dataset for which the outcome, such as the response to a given treatment, is already known. This is known as supervised learning. One common pitfall of this method is overfitting. Overfitting occurs when a model is too flexible relative to the data it is trained on, which limits its generalizability. A sign of overfitting is when the model accurately predicts outcomes on the data it was trained on but performs poorly on new, unseen data. To address the issue of overfitting, it is essential to validate models on unseen data. Cross-validation is a widely used technique for this purpose. It involves repeatedly dividing the data into subsets, training the model on one subset, and then evaluating its prediction accuracy on the remaining “heldout” data (see the figure). However, cross-validation is not infallible. Chekroud et al. revealed that models trained to predict responses to antipsychotic medication in schizophrenia within a specific clinical trial using cross-validation failed to predict treatment responses in other independent

g

GRAPHIC: F. PETZSCHNER, ADAPTED BY K. HOLOSKI/SCIENCE

MEDICINE

p

These patterns evolve between constructive and destructive interference as the delay-induced phase varies from 0 to p. The soliton signature emerges as a broad background signal superimposed on the Ramsey pattern and involves higher light intensities (optical fields accumulated in the pulse-shaped soliton) that, consequently, produce higher-order sidebands. Nonlinear optical behavior can also be observed through the formation of chaotic light fields, trains of solitons, or individual solitons, depending on how the input light wavelength and its power are tuned. These phenomena are driven by the silicon nitride nonlinear optical response (5, 8). As the laser wavelength and power were varied, Yang et al. could probe the transitions between nonlinear optical regimes because of the strong sensitivity of PINEM to the light field intensity and its distribution around the ring. The data of Yang et al. could be explained through a combination of well-established theoretical models for both the nonlinear optical response of microcavities (9) and the electron– photon interaction (2, 3). Simulations based on these models were in good quantitative agreement with the authors’ experiments. The interaction of electrons with enhanced light fields enabled by the microring geometry creates opportunities to obtain previously inaccessible information on light propagation inside integrated optical circuits. This includes the degree and spatial distribution of coherence associated with nonlinearly generated light pulses such as solitons. Furthermore, electron–soliton interaction enables a disruptive approach to shaping the electron probability density in space and time. Also, solitons circulating the microring cavity can interact with a continuous electron beam at a high repetition rate approaching the terahertz, which is inaccessible with current ultrafast optics technology. The resulting electron modulations could be synchronized with the optical excitation, presenting new ways to perform electron microscopy and probe ultrafast dynamics in material systems with sub-fs/nm spatiotemporal resolution. j

I NS I GHTS | P E R S P E C T I V E S

By Viacheslav Slesarenko1,2 and Lars Pastewka1,2

F

1Department of Microsystems Engineering, University of

ACK NOWL EDG M E N TS

F.H.P. was supported by the Brainstorm Program at the Robert J. and Nancy D. Carney Institute for Brain Science. 10.1126/science.adm9218

Freiburg, 79110 Freiburg, Germany. 2Cluster of Excellence livMatS, Freiburg Center for Interactive Materials and Bioinspired Technologies, University of Freiburg, 79110 Freiburg, Germany. Email: viacheslav.slesarenko@livmats. uni-freiburg.de; [email protected] science.org SCIENCE

,

riction controls daily life, often without being noticed. It allows walking without slipping, holds sandcastles together, and determines the perceived cleanliness of hair. Little resistance is desired when pedaling bikes, yet the expectation of pulling the brakes is to stop moving. Overall, machines use 20% of the world’s energy production to overcome frictional resistance (1). Present-day strategies to tune friction, derived from more than a century of engineering insights, often involve the lubrication of interfaces with oils or greases. On page 200 of this issue, Aymard et al. (2) report an alternative strategy of rationally designing the frictional properties of interfaces. Their approach to friction control may lead to the development of surfaces that adapt to the environment in real time. Aymard et al. show that small bumps of identical radii (3) constitute simple building blocks that can be combined into a frictional metainterface. By using many such bumps on a surface and adjusting their height distribution, the authors could prescribe a desired, even nonlinear, dependence of the frictional force that resists sliding motion on the external load that pushes the sliding interfaces together. The effect of surface topography on friction has long been known. CharlesAugustin Coulomb, one of the founders of tribology (the science of friction), wrote in 1779 about the interlocking of asperities (4), the name given to “bumps” on rough surfaces. Surface topography determines the amount of actual contact that two bodies make. Thus, two bodies typically

y

A. M. Chekroud et al., Science 383, 164 (2024). T. Wolfers et al., Neurosci. Biobehav. Rev. 57, 328 (2015). G. Varoquaux et al., Neuroimage 145 (Pt B), 166 (2017). K. E. Stephan et al., Lancet Psychiatry 3, 77 (2016). M. P. Paulus, W. K. Thompson, Psychopharmacology 238, 1231 (2021). 6. Q. J. M. Huys et al., Nat. Neurosci. 19, 404 (2016). 7. B. Kirkpatrick et al., Schizophr. Bull. 32, 214 (2006).

The frictional properties of material interfaces can be rationally designed

y g

12 JANUARY 2024 • VOL 383 ISSUE 6679

1. 2. 3. 4. 5.

The bumpy road to friction control

y

150

R EFER ENC ES A ND N OT ES

MATERIALS SCIENCE

g

Robert J. and Nancy D. Carney Institute for Brain Science, Brown University, Providence, RI, USA. Email: [email protected]

driven computational models that aim to describe underlying disease mechanisms, a method gaining traction in the field of computational psychiatry. These models are increasingly being used alongside data-driven machine learning techniques, forming powerful tools to tackle the issue of heterogeneity in patient populations (5, 6). Another form of heterogeneity may stem from systematic differences across studies, locations, or time points. As a result, predictions of machine learning models trained on data from a specific context—a population, country, setting, or time period—might rely on features that are associated but not causally related with a clinical outcome in a given study but are not predictive in other contexts. One way to address this heterogeneity is to pool data across multiple studies and sites. Unreliable predictions may also be the result of outdated outcome measures. Many existing symptom scores are based on questionnaires that may no longer align with understanding of the disease and potentially lead to inaccurate assessments of treatment response. For example, the positive and negative syndrome scale (PANSS) used in the clinical trials from Chekroud et al. is gradually being supplanted by more contemporary assessment tools, specifically in the context of negative symptoms in schizophrenia (7). If a questionnaire fails to fully capture the true disease burden, it might not accurately detect genuine improvements resulting from treatment. This discrepancy can lead to misclassification of who has or has not benefited from the treatment, which hinders the accurate training of the machine learning model. Similar to the heterogeneity within diagnostic categories, outcome measures will become more accurate with increasing insight into the underlying disease mechanism. The challenges of using machine learning to predict individual treatment response in medicine, specifically in the context of psychiatry, stem from a complex interplay of issues related to model validation standards, diagnostic heterogeneity, and the relevance of outcome measures used. Addressing these challenges is essential for impactful clinical research and to enable progression toward effective precision medicine. j

p

clinical trials. One reason that cross-validation can inadvertently result in overfitting the held-out data is that the modeler, through iterative model adjustments, may eventually use all the available data. The issue is likely more widespread than typically acknowledged. For example, a comprehensive review of 116 studies across various psychiatric diagnoses found signs of overfitting specifically in studies with small sample sizes (40% (3, 4). Perovskite solar cells (PSCs) are promising for such tandem integration owing to their tunable bandgap (which is needed to maximize the spectral efficiency) (5) combined with their potential for high performance (small-area, single-junction devices have reached PCEs of >26%) and their potential for low-cost manufacturing (2). Silicon-based tandems are predicted by the International Technology Roadmap for Photovoltaic (ITRPV) to enter the market in the next few years (6). However, mainstream c-Si module technology already delivers reliable products at a cost of less than $0.25 per watt with a scaled and rapidly expanding manufacturing capacity that exceeded 600 GW per year at the end of 2022 (6). To be market competitive, perovskite/silicon tandems must demonstrate clear benefits in terms of cost and LCOE. Despite the advantages of higher-performance modules toward a lowered cost per nameplate watt of PV systems, uncertainties over module durability and energy yield may compromise the bankability of perovskite/silicon tandems. Hence, accelerated degradation tests, targeted specifically to perovskite technology, and out-

Most experimental efforts have focused on monolithic tandems in the two-terminal (2T) configuration with a perovskite cell fabricated on top of a c-Si solar cell, series-connected by an internal junction (Fig. 1A) (7). Although monolithic modules require fewer transparent electrodes and BOM and BOS components compared with other tandem configurations, their optimal performance requires current matching, which demands the tuning of the top cell bandgap through compositional engineering of the perovskite. Eliminating the need for bandgap tuning is possible when incorporating a third terminal, for instance by using c-Si bottom cells with interdigitated back contacts. However, this strategy introduces increased complexity in both processing and interconnections. (8, 9). Tandems can also be constructed by the mechanical stacking of the subcells. In this case, the subcells are manufactured independently, which offers greater freedom in processing but adds costs. More transparent electrodes are required [usually in the form of transparent conductive oxides (TCOs)], which may also increase parasitic absorption losses (10). Moreover, the large perovskite submodule requires several film patterning steps (Fig. 1B), usually by laser scribing. Finally, the perovskite and c-Si subcells need to be electrically isolated through appropriately chosen module lamination materials (Fig. 1C). Mechanically stacked tandems come in principle in a four-terminal (4T) configuration. This allows the submodules to operate independently but requires two dc-ac inverters, which, along with extra cabling, adds further to the total system cost. Yet, a 2T module configuration is also possible for mechanically stacked tandems by connecting the two submodules either in series or in parallel. This requires current matching or voltage matching of the constituent submodules, respectively; in either case, a single inverter is sufficient. One advantage of voltage matching is the less stringent demand on bandgap engineering and better performance stability against spectral and operating temperature changes (11). Overall, for mechanically stacked tandems, perovskite submodules may be integrated with minimal adjustments to commercial c-Si module manufacturing, which could enable faster entry into the mainstream PV market. Nevertheless, their monolithic counterparts may offer decisive benefits in terms of device performance, reliability, and cost.

g

T

he large-scale deployment of photovoltaics (PV) witnessed in recent years has been driven by a sharp decline in PV module prices (dropping >85% since 2010), increased power conversion efficiencies (PCEs) and system lifetimes, as well as inherently low variable running costs (1). All of these factors help to lower the levelized cost of electricity (LCOE) from PV, which is a cost metric that accounts for the overall system costs and energy generated over a PV system’s lifetime. Yet, PV technologies based on crystalline silicon (c-Si), which currently dominate the market, are approaching their theoretical and technical PCE limits. Today, the record PCE for c-Si PV has reached 26.8% for cells (the theoretical maximum is ~29.4%) and >24% for modules, and mass-produced commercial modules are now sold with PCEs in the range of 22 to 24% (2). Advanced cell interconnection and module integration schemes can help to further reduce cell-to-module losses, but ultimately c-Si PV technologies are nearing their practical PCE limits. One way to further decrease the LCOE of PV is through improvements in PCEs beyond what can be achieved with c-Si technology alone.

Tandem module configurations

p

Perovskite/silicon tandem solar cells offer a promising route to increase the power conversion efficiency of crystalline silicon (c-Si) solar cells beyond the theoretical single-junction limitations at an affordable cost. In the past decade, progress has been made toward the fabrication of highly efficient laboratory-scale tandems through a range of vacuum- and solution-based perovskite processing technologies onto various types of c-Si bottom cells. However, to become a commercial reality, the transition from laboratory to industrial fabrication will require appropriate, scalable input materials and manufacturing processes. In addition, perovskite/silicon tandem research needs to increasingly focus on stability, reliability, throughput of cell production and characterization, cell-tomodule integration, and accurate field-performance prediction and evaluation. This Review discusses these aspects in view of contemporary solar cell manufacturing, offers insights into the possible pathways toward commercial perovskite/silicon tandem photovoltaics, and highlights research opportunities to realize this goal.

door testing coupled with accurate energy yield forecasting are critically needed to reduce such uncertainties to translate performance gains at the module level to a lower LCOE and to thereby advance the market entry of this technology.

RES EARCH | R E V I E W

p g y

Device scaling Silicon bottom-cell choices

Silicon heterojunction (SHJ) solar cells in a both-sides contacted layout (where electrons and holes are collected at opposite sides of the cell) have been the preferred choice to construct monolithic perovskite/silicon tandems, including the recent record tandems to date. The attractiveness of SHJ technology lies in its high operating voltages that result from passivating contacts, efficient light-trapping features (10, 16), inherently bifacial nature, and fabrication process that can easily enable p-i-n or n-i-p tandem configurations (17). Moreover, in a 2T tandem, the top TCO of the SHJ

12 January 2024

cell can be used as a recombination junction layer with only minor modifications (e.g., by keeping its thickness 100 cm2). The impact of localized shunts may be minimized by engineering the contact resistance at the recombination junction/perovskite

y

Reducing electrical shunts in perovskite subcells

interface, as well as the use of recombination junctions with high lateral resistance, such as those based on doped nanocrystalline silicon (nc-Si) films (Fig. 4A) (46) or thinner and therefore more resistive TCOs (53). In the case of mechanically stacked tandems, the perovskite top cell will cover dimensions as large as the module glass onto which it is coated (>1 m2). To pattern the perovskite top cell and reduce resistive losses at the module level, several laser scribing steps (usually three, called P1, P2, and P3 in specific processing orders) are necessary. These steps divide the cell into smaller units and interconnect the adjacent cells electrically in series, as shown in Fig. 4B. Such compartmentalizing also reduces the impact of local shunts on the overall module performance. An additional fourth scribe (P4) to subdivide the cells can isolate underperforming areas. The benefit of this approach is seen through an overall improved electroluminescence (EL) response from a module, evidenced by a decrease in dark zones (implying low EL emission and thus low PV performance) and an increase in pink zones (implying high EL emission and high PV performance) (Fig. 4B). An alternative strategy to minimize the impact of imperfections may lie in the fabrication of smaller-sized perovskite subcell coupons, integrated into tandems at the module lamination level. Despite these mitigation strategies, developing large-scale, current leakage–free perovskite films is essential for reproducible large-scale perovskite manufacturing.

g

Aydin et al., Science 383, eadh3849 (2024)

control (39, 44). However, the integration of additive engineering strategies, often used in solution-processed perovskites to passivate defects and fully unlock the optoelectronic properties of the bulk material, remains difficult in this approach. Hybrid sequential deposition, such as through vacuum deposition of a porous inorganic template, followed by solution- (e.g., slot-die or spray coating) or gas-based (chemical vapor transport) conversion into the desired perovskite phase combines the advantages of vacuum- and solution-based techniques in providing conformal coverage on textured surfaces (Fig. 3G), thickness control, and facile compositional tuning (45–47). Moreover, additive engineering approaches can be introduced during the solution processing step, which facilitates defect passivation and improved perovskite crystallization (47–50). Guaranteeing full infiltration of the solution into the mesoscopic template to yield full conversion into the perovskite phase remains the biggest challenge for this approach. Overall, linear evaporation, slot-die coating, or their combination in a hybrid approach may be the most promising candidates for scaled perovskite deposition for monolithic tandems. Yet, in the absence of a well-established scaling platform, alternative techniques merit deeper exploration too, such as pulsed laser deposition (PLD) or even magnetron sputtering— both of which enable perovskite deposition with stoichiometric material transfer from target to substrate (51). Finally, we also note the importance of process robustness when scaling technologies, with consistent batch-to-batch reproducibility for depositing the perovskite absorber layer, passivation methods, and depositing contacting materials. Ultimately, the most adequate processing technique should yield a combination of high device performance, throughput, and reproducibility.

p

Spin coating of perovskite-precursor inks, combined with antisolvent treatments for film crystallization and additive engineering for defect passivation, has so far led to the highestperforming perovskite single-junction and perovskite/silicon tandem solar cells (28, 29). However, materials waste and limited scalability and throughput undermine the commercial prospects of this fabrication route. These limitations may be overcome by slotdie coating, for which the suitability for topcell fabrication, even on textured surfaces, has been reported for laboratory-scale tandem devices. Yet, several challenges arise related to slot-die coating, which likely explains the lack of large-scale demonstrations to date (24, 30). First, deposition of the perovskite through slot-die coating is incompatible with antisolvent treatments, so film crystallization has been accomplished by temperature control (31) or gas quenching (32). Second, dedicated ink compositions are needed to obtain uniform coatings and morphologies over larger areas, where surface terminations of the substrate must strike a balance between wettability to form closed films and hydrophobicity to enhance crystal growth (33, 34). Third, the leading and trailing edges of linear-printed coatings often have uncontrolled thicknesses, morphologies, and electrical properties, for which mitigation strategies still need to be formulated (35). This is of particular concern when coating on device-active silicon bottom cells and may compromise tandem performance. Finally, given the modulated textured surface of silicon wafers or surface imperfections of tempered glass (e.g., flatness), ensuring that the perovskite consistently covers all of the surface features may require thicknesses of multiple micrometers (Fig. 3F), which could limit effective charge-carrier extraction (12, 36, 37). Overcoming this challenge requires effective defect-passivation strategies combined with control of the substrate morphology. Physical vapor deposition techniques, such as thermal evaporation, can yield directional material transfer, which is attractive for scalable, conformal coatings on textured surfaces with accurate film-thickness control (38, 39). Yet, fully evaporated perovskites have rarely been reported even for single-junction PSCs (40–42), which is likely in part because of the high capital expenditure (CapEx) and low throughput of laboratory-scale tools. For halide perovskites, one-step vacuum deposition requires coevaporation of multiple precursors (such as PbI2, MAI, FAI, and MABr). Greater insight into the compositional control of deposited films would be a first critical step toward closing the large performance gap between evaporated and solution-processed PSCs (43). Sequential deposition of the perovskite constituents in a linear evaporator, followed by thermal annealing, could simplify the compositional

RES EARCH | R E V I E W

A

B

No P4

With P4

P4 scribing lines

Transparent electrode Electron/hole selective contact

p

Perovskite absorber Electron/hole selective contact Transparent electrode Glass or flexible transparent substrate

g

Aydin et al., Science 383, eadh3849 (2024)

12 January 2024

Module reliability and lifetime predictions

To determine the durability of perovskite/silicon tandem modules intended to last for more than 25 years, accelerated tests that can give predictions within a substantially shorter time frame are required. c-Si PV manufacturers have established protocols that effectively anticipate potential early failures that may arise during field operations. For instance, the IEC 61215 protocol consists of a range of controlled-laboratory accelerated degradation tests, such as damp-heat, thermal cycling, humidity-freeze, ultraviolet (UV) light exposure, and potential induced degradation (PID) testing, among others. Confidence in this protocol is based on statistical feedback from reported long-term outdoor test data, which remain scarce for perovskite/ silicon tandems (Fig. 5C).

,

To be commercially competitive, the warranty of perovskite/silicon tandem modules should be on par with that of mainstream c-Si–based modules, which at present is typically about 25 years, although some manufacturers already offer up to 40 years (Fig. 5A) (59). A simple LCOE calculation considering 25 years of service gives insight into acceptable annual degradation rates for perovskite/silicon tandems. As a reference, we used a 22%-PCE c-Si module with a 0.4% relative annual degradation rate for our LCOE calculations, which is representative of the commercial state of the art (representing >90% relative PCE retention over 25 years). We then considered higher annual degradation rates (2 and

decomposition of the perovskite as the primary issues that require attention. Substantial progress has been made in addressing critical stability issues [e.g., damp-heat and maximum power point tracking (MPPT) at elevated temperature] for low-bandgap (~1.55 eV) cells based on FAPbI3 (63, 64). However, the use of wide-bandgap perovskites (~1.67 to 1.70 eV) in perovskite/silicon tandems with mixed halide (iodide, bromide, and chloride) compositions remains a more complex challenge.

y

LCOE versus stability

5% relative, which means less PCE retention after 25 years) in combination with different variable module costs (0, 10, and 30%) to compare the LCOE of tandems with c-Si modules (Fig. 5B) (60). If a similar annual degradation rate can be guaranteed (0.4% relative), and assuming only a 30% additional cost resulting from the perovskite subcell processing, then the tandem module PCE should be >24% to become competitive with mainstream c-Si PV. Under similar cost assumptions, if the annual degradation rate is 2% relative, a >32% tandem module PCE will be needed to be competitive. Given unavoidable cell-to-module losses, the required tandem cell PCEs will need to be even higher (61). To date, the longest reported annual degradation rate of small-area (1 cm2, 21.4% initial PCE for encapsulated cell) perovskite/silicon tandems based on outdoor data is >17% relative. This large value underlines the urgency of improving the stability of perovskite/silicon tandem solar cells rather than merely enhancing their PCEs (62). Outdoor testing of perovskite/silicon tandem cells provides valuable insights into actual stability concerns that may arise during field operations, covering synchronously all aspects of degradation such as heat, light, humidity, temperature cycles, and voltage biasing. Initial tests have identified ion migration and phase

y g

to be smaller than 1.68 eV for monofacial monolithic tandems under such conditions. This could offer increased resilience against lightinduced phase segregation because less bromide is needed to obtain such smaller bandgaps (11). Therefore, the intended geographical location of operation may become an additional parameter to consider when designing commercial tandems. Whether such market-specific bandgap tailoring will truly be needed must be determined through accurate energy yield and cost calculations, based on realistic climate and device data, and validated by extensive field testing.

scribing (horizontal lines) process aims to isolate the underperforming areas into smaller cells (image is used with the permission of CubicPV). The magnified area (left) shows the sketch of the P1, P2, and P3 scribing steps to build a series-connected string of thin-film cells [used with permission from (109)].

y

Fig. 4. Electrical shunt in tandem solar cells. (A) Schematic illustrations of the shunting paths in TCO-based recombination junctions, and the tunneling junctions show localizing the shunts as a result of the low lateral conductivity. (B) Photographs (top) and photoluminescence images (bottom) of perovskite minimodules aimed at mechanically stacked tandem utilization. The P4

6 of 13

RES EARCH | R E V I E W

Today's champion tandem cell PCE

1.5

10 20

1

10

0.1

0 1970

1980

1990

2000

2010

2020

LCOEtandem / LCOEc-Si

Assuming 30% module PCE

30

Module cost (US$ per W)

Module warranty (Years)

C

B 100

5 % / year

1

0.8

2030

2 % / year Today's c-Si PV panels

22

Years

D

0% Extra Cost 10% Extra Cost 30% Extra Cost

0.4 % / year

30 PCE (%)

Module outdoor test periods (Years)

A

Goal: 30 years stability with >80-92% PCE retention

30 Aydin et al. 2020 (79% retention, first outdoor data)

28 De Bastiani et al. 2021 (51% retention)

26 Liu et al. 2021 (95% retention)

2 Babics et al. 2023 (83.4% retention)

0

2020

33.9

2022

2024

2026

2028

2030

2032

Years

F

E

p g

Module packaging

12 January 2024

7 of 13

,

Mainstream c-Si module encapsulation consists of vacuum lamination of strings of cells between a glass sheet at the front and a polymeric back sheet at the rear, with two layers of encapsulant. The porous nature of typical polymeric back sheets is not suitable for perovskitebased PV because of moisture ingress—water rapidly degrades perovskite devices. Instead, the polymeric back sheet needs to be replaced with a second glass sheet, and module edges are to be sealed with rubber (Fig. 5D) (68). However, in this case, initial degradation products of perovskite absorbers, such as iodine gas, iodomethane, and hydroiodic acid, may remain trapped inside the module. These degradation products can then further accelerate module degradation, particularly in the interconnections between different cells (69). One mitigation

pathway may be through additive engineering during perovskite processing that targets the removal of initial degradation products, which, however, will require new materialsscience insights. Furthermore, typical mainstream PV encapsulants, such as ethylene vinyl acetate (EVA), polyolefin elastomer (POE), coextruded EVA/ POE/EVA, and thermoplastic polyurethane (TPU), require lamination at 120° to 140°C for 15 min under a mild vacuum (68). However, a temperature of >110°C is usually sufficient to induce degradation of the perovskite, the organic interlayers, and their interfaces (70), which has spurred the search for alternative encapsulants or more thermally stable perovskite cells. Moreover, the forces that occur during lamination and temperature cycles (Fig. 5E), which can be driven by a mismatch in coefficients of thermal expansion, present another challenge. Specifically, the ubiquitous fullerene (C60) used as the electron transport layer (ETL) in the perovskite top contact is only weakly bonded to the adjacent layers and can easily be fractured and peeled off. Figure 5F shows an example of delamination of perovskite top contact in tandems at their C60/SnO2 interface, where SnO2 is widely used as a buffer layer against the

y

Aydin et al., Science 383, eadh3849 (2024)

can stay operational under field conditions. Accurate degradation rates are crucial to determine warranty costs for the modules. To achieve this, longer outdoor tests are necessary, covering realistic device areas (larger than 1 cm2 and with maximized active area/wafer size ratios) at both the module and system levels (consisting of strings of modules).

y g

Additionally, some perovskite-specific degradation modes, although known, have yet to be included in mainstream industrial testing protocols, and others doubtlessly remain to be discovered. For example, simultaneous exposure to light and elevated temperature is a known critical cause of failure of PSCs, but this mechanism is not thoroughly probed by the IEC 61215 protocol. To gauge its impact, the US initiative Perovskite PV Accelerator for Commercializing Technologies (PACT) recommends four sequences of light exposure at elevated temperatures (1-sun, 75°C for 1000 hours) with module loss analysis after every 250 hours (65). This is similar to the so-called light and elevated temperature–induced degradation (LeTID) test applied to c-Si PV modules defined by IEC TS63342. Determining the activation energy of degradation of the perovskite cells through MPPT tests at elevated temperatures can also help to predict the operational lifetimes of the tandem cells (66, 67). Table 1 provides an overview of the aging protocols that have been proposed to date for perovskite modules. Correlation analysis between these laboratory tests and field performance can be used to assess how long these devices—specifically the perovskite subcells—

perovskite/silicon tandem modules outdoor data. The PCE retention goal is based on the longest product warranties in the market. (D) Sketch of state-of-the-art encapsulation for perovskite/silicon tandem solar cell modules. (E) Photograph of failed encapsulation for 156 mm by 156 mm perovskite/silicon tandem cells because of contact delamination. (F) Detailed view of contact delamination at the C60/SnO2 interface, part of the perovskite top contact, resulting from module lamination at 120°C [adapted with permission from (71), copyright 2022 American Chemical Society].

y

Fig. 5. Reliability and LCOE of perovskite/silicon tandem modules. (A) Market status of the c-Si PV industry in terms of cost and module warranties. The previous years’ data are retrieved from Jordan et al. (110), and future predictions are adapted from ITRPV 2022. The PV prices are retrieved from (111). From the 1970s, the market status changed from high-cost and short warranties to low-cost, long warranties. (B) Relative LCOEs of perovskite/silicon tandems versus silicon cells, based on module degradation rates and additional perovskite submodule costs. (C) Reported

RES EARCH | R E V I E W

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

Critical tests

Reports for monolithic perovskite/ silicon tandems

Conditions

Reports for mechanically stacked perovskite/ silicon tandems or single-junction perovskite modules

Research pathway

Passing damp heat test Passed one cycle (112) one time might not and three cycles (113) for be sufficient to ensure perovskite single-junction >30 years of field stability. More module, and one cycle for correlation analysis is needed. small perovskite pixel (64). ............................................................................................................................................................................................................................................................................................................................................ Passed for perovskite Interfacial toughening Thermal cycling 95% PCE retention after 98.8% of its initial single-junction methods or new materials −40° to +85°C, 200 cycles PCE after 200 thermal module (112). should be developed that can cycles (tabbing was resist expansion and shrinking copper tape) (114) during thermal cycles (115). ............................................................................................................................................................................................................................................................................................................................................ Moisture-resistant and Humidity freeze 95% PCE retention after No test data reported No test data reported perovskite-compatible −40° to +85°C and 85% for tandem modules. Passed encapsulation materials relative humidity, 50 cycles for small single-junction should be developed for perovskite pixels (116). freezing-heating cycles. ............................................................................................................................................................................................................................................................................................................................................ Longer (>>1 year) outdoor 40 days, 90% No test data reported for Outdoor stability 95% PCE retention tests at different climates, PCE retention (117) tandem modules. For after 60 kWh/m2 total realistic module sizes, and the single-junction solar irradiance multiple number of module pixel, 1-year data (118). tests are required to correlate them with laboratory tests. ............................................................................................................................................................................................................................................................................................................................................ This is not a well-defined standard MPPT at elevated 95% PCE retention after No test data reported No test data reported for protocol but is critical for temperature 65° to 85°C, MPPT, tandem modules. understanding the halide >1000 hours 97%-PCE retention at 65°C segregation issue. The after 1200 hours for ideal testing temperature small perovskite pixel (119). should be chosen according to the defect activation energies of the perovskite layers. ............................................................................................................................................................................................................................................................................................................................................ Perovskite film quality PID 60°C, 85% humidity, ~50% loss after No test data reported for should be improved, and new 1000-V load for a 24 hours (75) modules. For 1-cm2 cell, encapsulants and ionic period of 96 hours no degradation observed barrier layers should at ±500 V (120). be developed (120–122). ............................................................................................................................................................................................................................................................................................................................................ For monolithic tandems, Reverse bias degradation Hotspot test, partial Negligible degradation No loss for perovskite contact quality should be or full shading after biasing at minimodule with carbonimproved. For thin film modules, of subcells −2.0 V for 12 hours (123) based electrodes bypass diodes should be (area of 56.8 cm2) (124). integrated to the film stacks.

Damp heat

95% PCE retention after testing at 85°C and 85% relative humidity for 1000 hours

Passed one cycle (93)

8 of 13

,

12 January 2024

y

Bifacial solar cells have transparent rear contacts to exploit the diffuse and reflected light from the environment and the ground (referred to as albedo) to enhance their current output by 10 to 30% relative, depending on the specific

y g

Emerging concepts Bifacial tandems

deployment. Bifacial c-Si modules are rapidly becoming the dominant utility-scale technology (76). Calculations show that adopting the bifacial concept also enhances the energy yield of perovskite/silicon tandems (77). Such calculations require accurate knowledge of the front and rear irradiance, as well as potential (nonuniform and temporal) shading from the surroundings (78). Bifacial tandems may also offer superior stability because they require the use of lower-bandgap (50% and thus decrease the LCOE. In Fig. 6D, conceptual processing lines for mechanically stacked and monolithic tandem cells are depicted. Mechanically stacked tandems require extra processing steps for the perovskite submodules, including multiple laser scribing, and they demand intricate module assembly lines, which leads to increased overall CapEx. On the other hand, monolithic cells Aydin et al., Science 383, eadh3849 (2024)

have packaging procedures similar to those of c-Si single-junction modules and only necessitate wafer-scale processing steps for the perovskite subcells. A recent techno-economic analysis corroborates that the production cost of mechanically stacked tandems might be higher compared with that of their monolithic counterparts (105). Additional challenges arise in characterizing industrially produced perovskite/silicon tandem technology. Current-voltage (I–V) analysis within light exposure at the millisecond timescale is commonly used for industrial characterization of c-Si cells. New methods for fast and reliable measurements of perovskite/ silicon tandems must be developed, which so

12 January 2024

far typically require several seconds and power stabilization. REFERENCES AND NOTES

1.

2. 3.

4.

N. M. Haegel et al., Terawatt-scale photovoltaics: Transform global energy. Science 364, 836–838 (2019). doi: 10.1126/ science.aaw1845; pmid: 31147512 National Renewable Energy Laboratory, “Best Research-Cell Efficiency Chart”; https://www.nrel.gov/pv/cell-efficiency.html. B. A. Veith-Wolf, S. Schäfer, R. Brendel, J. Schmidt, Reassessment of intrinsic lifetime limit in n-type crystalline silicon and implication on maximum solar cell efficiency. Sol. Energy Mater. Sol. Cells 186, 194–199 (2018). doi: 10.1016/j.solmat.2018.06.029 S. Schäfer, R. Brendel, Accurate Calculation of the Absorptance Enhances Efficiency Limit of Crystalline Silicon Solar Cells With Lambertian Light Trapping. IEEE J. Photovolt. 8, 1156–1158 (2018). doi: 10.1109/ JPHOTOV.2018.2824024

10 of 13

RES EARCH | R E V I E W

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

Country

Company or consortia

Europe

PEPPERONI Project

Budget

Duration

€18.8 million

2022–2026

Remarks

12 January 2024

9.

10.

P. Wagner, P. Tockhorn, S. Hall, S. Albrecht, L. Korte, Performance of Monolithic Two- and Three-Terminal Perovskite/Silicon Tandem Solar Cells Under Varying Illumination Conditions. Sol. RRL 7, 2200954 (2023). doi: 10.1002/solr.202200954 Z. C. Holman et al., Infrared light management in high-efficiency silicon heterojunction and rear-passivated solar cells. J. Appl. Phys. 113, 013107 (2013). doi: 10.1063/1.4772975

11 of 13

,

8.

M. De Bastiani et al., Recombination junctions for efficient monolithic perovskite-based tandem solar cells: Physical principles, properties, processing and prospects. Mater. Horiz. 7, 2791–2809 (2020). doi: 10.1039/D0MH00990C E. L. Warren et al., A Taxonomy for Three-Terminal Tandem Solar Cells. ACS Energy Lett. 5, 1233–1242 (2020). doi: 10.1021/acsenergylett.0c00068

y

Aydin et al., Science 383, eadh3849 (2024)

7.

y g

6.

Z. Yu, M. Leilaeioun, Z. Holman, Selecting tandem partners for silicon solar cells. Nat. Energy 1, 16137 (2016). doi: 10.1038/ nenergy.2016.137 VDMA Photovoltaic Equipment, “International Technology Roadmap for Photovoltaic (ITRPV)” (2023); https://www.vdma.org/international-technology-roadmapphotovoltaic.

y

5.

g

............................................................................................................................................................................................................................................................................................................................................

p

17 partners from 12 countries aim to bring the TRL level of 2T tandems from 4 to 5 up to TRL 7, with a long-term goal of reaching gigawatt-scale manufacturing in Europe. ............................................................................................................................................................................................................................................................................................................................................ Europe NEXUS Project €3.6 million 2022–2025 12 partners from nine countries aim to reach >30% tandem module PCE in 2T configuration by using dry perovskite processing. ............................................................................................................................................................................................................................................................................................................................................ Europe PrEsto Project €4.25 million 2021–2024 Nine partners from three countries aim to identify suitable production processes for scaled 2T tandems and perform techno-economic evaluations. ............................................................................................................................................................................................................................................................................................................................................ UK Oxford PV $142 million Since 2010 Oxford PV is the first perovskite-based tandem PV company and received investment from multiple sources. The company currently holds the highest certified PCE (28.6%) for industry-size 2T tandems (274 cm2). The budget data are acquired from www.crunchbase.com. ............................................................................................................................................................................................................................................................................................................................................ USA TEAMUP Project $12.2 million Since 2023 12 partners from academic, industrial, and federal laboratories in the US aim to enhance the stability of both 2T and 4T tandems in real-world conditions. ............................................................................................................................................................................................................................................................................................................................................ USA ADDEPT Project $11.25 million Since 2023 Eight partners from academic institutes and companies from the US aim to enhance the performance and durability of both 2T and 4T tandem cells through machine learning and robot screening. ............................................................................................................................................................................................................................................................................................................................................ USA PV PACT Project $9 million Since 2021 12 partners including national laboratories, universities, and companies from the US aim to develop performance and reliability test protocols for industrial applications. ............................................................................................................................................................................................................................................................................................................................................ China Auner >$14 million Since 2017 Auner is a start-up company and claims a 32.44% PCE for 1 cm2 and 30.83% certified PCE on 25 cm2 aperture area. The company reports that an industrial-size pilot line is under construction. ............................................................................................................................................................................................................................................................................................................................................ China Jinko Solar >$6.6 million Since 2021 Jinko ranks first in cumulative PV module shipment in the world. The company reported 32.33% PCE for perovskite/TOPCon 2T tandems, with limited information about other tandem research activities. ............................................................................................................................................................................................................................................................................................................................................ China Trina Solar >$17 million Since 2021 Trina claims >31% laboratory-scale PCE, and the company is one of the leading enterprises (top three) in silicon solar cell and PV modules. The given budget is for the initial investment. ............................................................................................................................................................................................................................................................................................................................................ China Longi No information No information Longi is the highest–market value PV company in China, holding the record for the PCE achieved in SHJ cells (usually used as bottom cells). The company reported 33.9% PCE for ~1-cm2 2T tandems, and there is no other public release. ............................................................................................................................................................................................................................................................................................................................................ China Tongwei No information Since 2022 Tongwei stands as the foremost company in the shipment of polysilicon and c-Si cells, asserting an impressive 31.1% PCE for 2T tandems, albeit with only limited public disclosure. ............................................................................................................................................................................................................................................................................................................................................ Japan Kaneka No information No information Kaneka is an expert in amorphous silicon and has held the world record for interdigitated back contact (IBC) cells for several years. Also, the company has background with a-Si thin-film tandem cells. Kaneka released 29.2% PCE (~1 cm2) for 2T tandems with no other public release about R&D activities. ............................................................................................................................................................................................................................................................................................................................................ Korea Hanwha Q Cells $100 million 2023–2024 This Korea-based company, with its research hub in Germany, has disclosed 29.9% PCE achievement for 2T tandem solar cells, covering an area of 1 cm2. They have also earmarked $100 million for a pilot production line in Jincheon, Korea, and they are one of the partners of the PEPPERONI project. ............................................................................................................................................................................................................................................................................................................................................ Singapore SERIS and REC Solar $57 million 2023–2025 SERIS and Singapore-based panel manufacturer REC aim to develop a >30%-PCE 2T tandem module in the REC@NUS corporate R&D laboratory.

RES EARCH | R E V I E W

11.

36.

37.

38.

39.

40.

41.

42.

43.

44.

45.

47.

48.

49.

51.

54.

55.

56.

57.

58.

12 January 2024

62.

63.

64.

65.

66.

67.

68.

69.

70.

71.

72.

73.

74.

75.

76.

77.

78.

79.

80.

81.

82.

12 of 13

,

53.

61.

y

52.

60.

J. F. Weaver, “How the numbers may work for 50-year solar” (PVMagazine, 2021); https://www.pv-magazine.com/2021/ 11/05/how-the-numbers-may-work-for-50-year-solar/. S. J. Andrews, B. Smith, M. G. Deceglie, K. A. Horowitz, T. J. Silverman, “NREL Comparative PV LCOE Calculator” (2021); https://www.nrel.gov/pv/lcoe-calculator/. J. Qian, M. Ernst, N. Wu, A. Blakers, Impact of perovskite solar cell degradation on the lifetime energy yield and economic viability of perovskite/silicon tandem modules. Sustain. Energy Fuels 3, 1439–1447 (2019). doi: 10.1039/ C9SE00143C M. Babics et al., One year outdoor operation of monolithic perovskite/silicon tandem solar cell. Cell Rep. Phys. Sci. 4, 101280 (2023). doi: 10.1016/j.xcrp.2023.101280 S. M. Park et al., Engineering ligand reactivity enables high-temperature operation of stable perovskite solar cells. Science 381, 209–215 (2023). doi: 10.1126/science.adi4107; pmid: 37440655 R. Azmi et al., Damp heat-stable perovskite solar cells with tailored-dimensionality 2D/3D heterojunctions. Science 376, 73–77 (2022). doi: 10.1126/science.abm5784; pmid: 35175829 L. Schellhaas, J. Stein, PACT Perovskite PV Module Stress Testing Protocol Version 0.0. (Perovskite PV Accelerator for Commercializing Technologies, 2022); https://pvpact.sandia. gov/publications-and-protocols/. X. Zhao et al., Accelerated aging of all-inorganic, interfacestabilized perovskite solar cells. Science 377, 307–310 (2022). doi: 10.1126/science.abn5679; pmid: 35709247 Q. Jiang et al., Towards linking lab and field lifetimes of perovskite solar cells. Nature 623, 313–318 (2023). doi: 10.1038/s41586-023-06610-7; pmid: 37696288 M. De Bastiani et al., All Set for Efficient and Reliable Perovskite/Silicon Tandem Photovoltaic Modules? Sol. RRL 6, 2100493 (2022). doi: 10.1002/solr.202100493 Y.-H. Lin et al., A piperidinium salt stabilizes efficient metalhalide perovskite solar cells. Science 369, 96–102 (2020). doi: 10.1126/science.aba1628; pmid: 32631893 G. Divitini et al., In situ observation of heat-induced degradation of perovskite solar cells. Nat. Energy 1, 15012 (2016). doi: 10.1038/nenergy.2015.12 M. De Bastiani et al., Mechanical Reliability of Fullerene/Tin Oxide Interfaces in Monolithic Perovskite/Silicon Tandem Cells. ACS Energy Lett. 7, 827–833 (2022). doi: 10.1021/ acsenergylett.1c02148 R. Cheacharoen et al., Design and understanding of encapsulated perovskite solar cells to withstand temperature cycling. Energy Environ. Sci. 11, 144–150 (2018). doi: 10.1039/C7EE02564E M. D. Kempe et al., Acetic acid production and glass transition concerns with ethylene-vinyl acetate used in photovoltaic devices. Sol. Energy Mater. Sol. Cells 91, 315–329 (2007). doi: 10.1016/j.solmat.2006.10.009 R. Cheacharoen et al., Encapsulating perovskite solar cells to withstand damp heat and thermal cycling. Sustain. Energy Fuels 2, 2398–2406 (2018). doi: 10.1039/C8SE00250A L. Xu et al., Potential-induced degradation in perovskite/ silicon tandem photovoltaic modules. Cell Rep. Phys. Sci. 3, 101026 (2022). doi: 10.1016/j.xcrp.2022.101026 R. Kopecek, J. Libal, Bifacial Photovoltaics 2021: Status, Opportunities and Challenges. Energies 14, 2076 (2021). doi: 10.3390/en14082076 J. Lehr et al., Energy yield of bifacial textured perovskite/ silicon tandem photovoltaic modules. Sol. Energy Mater. Sol. Cells 208, 110367 (2020). doi: 10.1016/j.solmat.2019.110367 A. Onno et al., Predicted Power Output of Silicon-Based Bifacial Tandem Photovoltaic Systems. Joule 4, 580–596 (2020). doi: 10.1016/j.joule.2019.12.017 M. De Bastiani et al., Efficient bifacial monolithic perovskite/silicon tandem solar cells via bandgap engineering. Nat. Energy 6, 167–175 (2021). doi: 10.1038/ s41560-020-00756-8 E. Ugur et al., Front-contact passivation through 2D/3D perovskite heterojunctions enables efficient bifacial perovskite/silicon tandem solar cells. Matter 6, 2919–2934 (2023). doi: 10.1016/j.matt.2023.05.028 M. T. Patel, R. Asadpour, J. Bin Jahangir, M. Ryyan Khan, M. A. Alam, Current-matching erases the anticipated performance gain of next-generation two-terminal Perovskite-Si tandem solar farms. Appl. Energy 329, 120175 (2023). doi: 10.1016/j.apenergy.2022.120175 M. Babics et al., Unleashing the Full Power of Perovskite/ Silicon Tandem Modules with Solar Trackers. ACS Energy Lett. 7, 1604–1610 (2022). doi: 10.1021/ acsenergylett.2c00442

y g

50.

59.

y

46.

A. Verma et al., Towards industrialization of perovskite solar cells using slot die coating. J. Mater. Chem. C 8, 6124–6135 (2020). doi: 10.1039/D0TC00327A M. Rai, L. H. Wong, L. Etgar, Effect of Perovskite Thickness on Electroluminescence and Solar Cell Conversion Efficiency. J. Phys. Chem. Lett. 11, 8189–8194 (2020). doi: 10.1021/ acs.jpclett.0c02363; pmid: 32893636 A. Farag et al., Efficient Light Harvesting in Thick Perovskite Solar Cells Processed on Industry-Applicable Random Pyramidal Textures. ACS Appl. Energy Mater. 5, 6700–6708 (2022). doi: 10.1021/acsaem.1c04028 M. Roß et al., Co-Evaporated Formamidinium Lead Iodide Based Perovskites with 1000 h Constant Stability for Fully Textured Monolithic Perovskite/Silicon Tandem Solar Cells. Adv. Energy Mater. 11, 2101460 (2021). doi: 10.1002/aenm.202101460 H. Li et al., Sequential vacuum-evaporated perovskite solar cells with more than 24% efficiency. Sci. Adv. 8, eabo7422 (2022). doi: 10.1126/sciadv.abo7422; pmid: 35857518 J. Li et al., Highly Efficient Thermally Co-evaporated Perovskite Solar Cells and Mini-modules. Joule 4, 1035–1053 (2020). doi: 10.1016/j.joule.2020.03.005 C. Momblona et al., Efficient vacuum deposited p-i-n and n-i-p perovskite solar cells employing doped charge transport layers. Energy Environ. Sci. 9, 3456–3463 (2016). doi: 10.1039/C6EE02100J Y.-H. Chiang et al., Vacuum-Deposited Wide-Bandgap Perovskite for All-Perovskite Tandem Solar Cells. ACS Energy Lett. 8, 2728–2737 (2023). doi: 10.1021/acsenergylett.3c00564; pmid: 37324541 M. Kroll et al., Insights into the evaporation behaviour of FAI: Material degradation and consequences for perovskite solar cells. Sustain. Energy Fuels 6, 3230–3239 (2022). doi: 10.1039/D2SE00373B J. Feng et al., High-throughput large-area vacuum deposition for high-performance formamidine-based perovskite solar cells. Energy Environ. Sci. 14, 3035–3043 (2021). doi: 10.1039/ D1EE00634G S. Siegrist et al., Triple-cation perovskite solar cells fabricated by a hybrid PVD/blade coating process using green solvents. J. Mater. Chem. A 9, 26680–26687 (2021). doi: 10.1039/D1TA07579A; pmid: 34976398 F. Sahli et al., Fully textured monolithic perovskite/silicon tandem solar cells with 25.2% power conversion efficiency. Nat. Mater. 17, 820–826 (2018). doi: 10.1038/s41563-018-0115-4; pmid: 29891887 X. Luo et al., Efficient Perovskite/Silicon Tandem Solar Cells on Industrially Compatible Textured Silicon. Adv. Mater. 35, e2207883 (2023). doi: 10.1002/adma.202207883; pmid: 36599055 E. Aydin, M. De Bastiani, S. De Wolf, Defect and Contact Passivation for Perovskite Solar Cells. Adv. Mater. 31, e1900428 (2019). doi: 10.1002/adma.201900428; pmid: 31062907 F. Sahli et al., Vapor Transport Deposition of Methylammonium Iodide for Perovskite Solar Cells. ACS Appl. Energy Mater. 4, 4333–4343 (2021). doi: 10.1021/acsaem.0c02999 L. Mao et al., Fully Textured, Production-Line Compatible Monolithic Perovskite/Silicon Tandem Solar Cells Approaching 29% Efficiency. Adv. Mater. 34, e2206193 (2022). doi: 10.1002/adma.202206193; pmid: 35985840 N. Rodkey et al., Pulsed Laser Deposition of Cs2AgBiBr6: From Mechanochemically Synthesized Powders to Dry, Single-Step Deposition. Chem. Mater. 33, 7417–7422 (2021). doi: 10.1021/acs.chemmater.1c02054; pmid: 34602745 Y. Rong et al., Challenges for commercializing perovskite solar cells. Science 361, eaat8235 (2018). doi: 10.1126/ science.aat8235; pmid: 30237326 E. Aydin et al., Enhanced optoelectronic coupling for perovskite/silicon tandem solar cells. Nature 623, 732–738 (2023). doi: 10.1038/s41586-023-06667-4; pmid: 37769785 J. P. Mailoa et al., A 2-terminal perovskite/silicon multijunction solar cell enabled by a silicon tunnel junction. Appl. Phys. Lett. 106, 121105 (2015). doi: 10.1063/1.4914179 S. Albrecht et al., Monolithic perovskite/silicon-heterojunction tandem solar cells processed at low temperature. Energy Environ. Sci. 9, 81–88 (2016). doi: 10.1039/C5EE02965A G. E. Eperon, M. T. Hörantner, H. J. Snaith, Metal halide perovskite tandem and multiple-junction photovoltaics. Nat. Rev. Chem. 1, 0095 (2017). doi: 10.1038/s41570-017-0095 J. Xu et al., Triple-halide wide-band gap perovskites with suppressed phase segregation for efficient tandems. Science 367, 1097–1104 (2020). doi: 10.1126/science. aaz5074; pmid: 32139537 L. Xu et al., Heat generation and mitigation in silicon solar cells and modules. Joule 5, 631–645 (2021). doi: 10.1016/ j.joule.2021.01.012

g

Aydin et al., Science 383, eadh3849 (2024)

35.

p

E. Aydin et al., Interplay between temperature and bandgap energies on the outdoor performance of perovskite/silicon tandem solar cells. Nat. Energy 5, 851–859 (2020). doi: 10.1038/s41560-020-00687-4 12. B. Chen et al., Enhanced optical path and electron diffusion length enable high-efficiency perovskite tandems. Nat. Commun. 11, 1257 (2020). doi: 10.1038/s41467-020-15077-3; pmid: 32152324 13. S. Kim et al., Over 30% efficiency bifacial 4-terminal perovskite-heterojunction silicon tandem solar cells with spectral albedo. Sci. Rep. 11, 15524 (2021). doi: 10.1038/ s41598-021-94848-4; pmid: 34330966 14. G. Coletti et al., Bifacial Four-Terminal Perovskite/Silicon Tandem Solar Cells and Modules. ACS Energy Lett. 5, 1676–1680 (2020). doi: 10.1021/acsenergylett.0c00682 15. M. A. Green et al., Solar cell efficiency tables (version 62). Prog. Photovolt. Res. Appl. 31, 651–663 (2023). doi: 10.1002/ pip.3726 16. T. G. Allen, J. Bullock, X. Yang, A. Javey, S. De Wolf, Passivating contacts for crystalline silicon solar cells. Nat. Energy 4, 914–928 (2019). doi: 10.1038/s41560-019-0463-6 17. S. De Wolf, A. Descoeudres, Z. C. Holman, C. Ballif, Highefficiency Silicon Heterojunction Solar Cells: A Review. Green 2, 7–24 (2012). doi: 10.1515/green-2011-0018 18. J. Zheng et al., Efficient monolithic perovskite–Si tandem solar cells enabled by an ultra-thin indium tin oxide interlayer. Energy Environ. Sci. 16, 1223–1233 (2023). doi: 10.1039/ D2EE04007G 19. A. Razzaq, T. G. Allen, W. Liu, Z. Liu, S. De Wolf, Silicon heterojunction solar cells: Techno-economic assessment and opportunities. Joule 6, 514–542 (2022). doi: 10.1016/ j.joule.2022.02.009 20. N.-G. Park, K. Zhu, Scalable fabrication and coating methods for perovskite solar cells and solar modules. Nat. Rev. Mater. 5, 333–350 (2020). doi: 10.1038/s41578-019-0176-2 21. Z. Li et al., Scalable fabrication of perovskite solar cells. Nat. Rev. Mater. 3, 18017 (2018). doi: 10.1038/natrevmats.2018.17 22. Y. Hou et al., Efficient tandem solar cells with solutionprocessed perovskite on textured crystalline silicon. Science 367, 1135–1140 (2020). doi: 10.1126/science.aaz3691; pmid: 32139544 23. B. Chen et al., Blade-Coated Perovskites on Textured Silicon for 26%-Efficient Monolithic Perovskite/Silicon Tandem Solar Cells. Joule 4, 850–864 (2020). doi: 10.1016/ j.joule.2020.01.008 24. A. S. Subbiah et al., High-Performance Perovskite SingleJunction and Textured Perovskite/Silicon Tandem Solar Cells via Slot-Die-Coating. ACS Energy Lett. 5, 3034–3040 (2020). doi: 10.1021/acsenergylett.0c01297 25. F. Fu et al., Monolithic Perovskite-Silicon Tandem Solar Cells: From the Lab to Fab? Adv. Mater. 34, e2106540 (2022). doi: 10.1002/adma.202106540; pmid: 35060205 26. K. Jäger, J. Sutter, M. Hammerschmidt, P.-I. Schneider, C. Becker, Prospects of light management in perovskite/silicon tandem solar cells. Nanophotonics 10, 1991–2000 (2020). doi: 10.1515/nanoph-2020-0674 27. E. Köhnen et al., 27.9% Efficient Monolithic Perovskite/ Silicon Tandem Solar Cells on Industry Compatible Bottom Cells. Sol. RRL 5, 2100244 (2021). doi: 10.1002/ solr.202100244 28. J. Park et al., Controlled growth of perovskite layers with volatile alkylammonium chlorides. Nature 616, 724–730 (2023). doi: 10.1038/s41586-023-05825-y; pmid: 36796426 29. P. Tockhorn et al., Nano-optical designs for high-efficiency monolithic perovskite-silicon tandem solar cells. Nat. Nanotechnol. 17, 1214–1221 (2022). doi: 10.1038/ s41565-022-01228-8; pmid: 36280763 30. K. Xu et al., Slot-Die Coated Triple-Halide Perovskites for Efficient and Scalable Perovskite/Silicon Tandem Solar Cells. ACS Energy Lett. 7, 3600–3611 (2022). doi: 10.1021/ acsenergylett.2c01506; pmid: 36277135 31. Y. Deng et al., Surfactant-controlled ink drying enables high-speed deposition of perovskite films for efficient photovoltaic modules. Nat. Energy 3, 560–566 (2018). doi: 10.1038/s41560-018-0153-9 32. Y. Yu et al., A Review on Gas-Quenching Technique for Efficient Perovskite Solar Cells. Sol. RRL 5, 2100386 (2021). doi: 10.1002/solr.202100386 33. W.-Q. Wu et al., Bilateral alkylamine for suppressing charge recombination and improving stability in blade-coated perovskite solar cells. Sci. Adv. 5, eaav8925 (2019). doi: 10.1126/sciadv.aav8925; pmid: 30873433 34. K. Liu et al., Zwitterionic-Surfactant-Assisted Room-Temperature Coating of Efficient Perovskite Solar Cells. Joule 4, 2404–2425 (2020). doi: 10.1016/j.joule.2020.09.011

RES EARCH | R E V I E W

83.

84.

85.

86.

87.

88.

89.

90.

92.

95.

96.

97.

102.

103.

104.

105.

106.

107.

108.

109.

110.

111. 112.

113.

114.

AC KNOWLED GME NTS

Competing interests: The authors declare no competing interests. License information: Copyright © 2024 the authors, some rights reserved; exclusive licensee American Association for the Advancement of Science. No claim to original US government works. https://www.science.org/about/science-licenses-journalarticle-reuse

y g

98.

101.

y

94.

100.

115. Z. Dai et al., Interfacial toughening with self-assembled monolayers enhances perovskite solar cell reliability. Science 372, 618–622 (2021). doi: 10.1126/science.abf5602; pmid: 33958474 116. L. Shi et al., Gas chromatography-mass spectrometry analyses of encapsulated stable perovskite solar cells. Science 368, eaba2412 (2020). doi: 10.1126/science. aba2412; pmid: 32439657 117. J. Liu et al., 28.2%-efficient, outdoor-stable perovskite/silicon tandem solar cell. Joule 5, 3169–3186 (2021). doi: 10.1016/ j.joule.2021.11.003 118. Q. Emery et al., Encapsulation and Outdoor Testing of Perovskite Solar Cells: Comparing Industrially Relevant Process with a Simplified Lab Procedure. ACS Appl. Mater. Interfaces 14, 5159–5167 (2022). doi: 10.1021/ acsami.1c14720; pmid: 35108814 119. S. Yang et al., Stabilizing halide perovskite surfaces for solar cell operation with wide-bandgap lead oxysalts. Science 365, 473–478 (2019). doi: 10.1126/science.aax3294; pmid: 31371610 120. K. Brecl et al., Are Perovskite Solar Cell Potential-Induced Degradation Proof? Sol. RRL 6, 2100815 (2022). doi: 10.1002/solr.202100815 121. Z. Purohit et al., Impact of Potential-Induced Degradation on Different Architecture-Based Perovskite Solar Cells. Sol. RRL 5, 2100349 (2021). doi: 10.1002/solr.202100349 122. J. Carolus et al., Potential-Induced Degradation and Recovery of Perovskite Solar Cells. Sol. RRL 3, 1900226 (2019). doi: 10.1002/solr.201900226 123. Z. Xu et al., Reverse-bias resilience of monolithic perovskite/ silicon tandem solar cells. Joule 7, 1992–2002 (2023). doi: 10.1016/j.joule.2023.07.017 124. D. Bogachuk et al., Perovskite Photovoltaic Devices with Carbon-Based Electrodes Withstanding Reverse-Bias Voltages up to −9 V and Surpassing IEC 61215:2016 International Standard. Sol. RRL 6, 2100527 (2022). doi: 10.1002/solr.202100527 125. I. J. Park et al., A Three-Terminal Monolithic Perovskite/Si Tandem Solar Cell Characterization Platform. Joule 3, 807–818 (2019). doi: 10.1016/j.joule.2018.11.017 126. E. Bellini, “Solliance hits 29.2% efficiency on perovskite/ silicon tandem solar cell” (PV Magazine, 2021); https://www. pv-magazine.com/2021/11/01/solliance-hits-29-2-efficiencyon-perovskite-silicon-tandem-solar-cell/. 127. L.-L. Senaud et al., “High Efficiency 3-Terminal Perovskite/ Silicon Tandems Based on Tunnel Junction IBC” paper presented at The 8th World Conference on Photovoltaic Energy Conversion, Milan, Italy, 26 to 30 September 2022.

g

93.

99.

and opportunities. Prog. Photovolt. Res. Appl. 31, 429–442 (2023). doi: 10.1002/pip.3499 G. S. Han et al., Cutting-Edge Studies Toward Commercialization of Large Area Solution-Processed Perovskite Solar Cells. Adv. Mater. Technol. 8, 2201387 (2023). doi: 10.1002/admt.202201387 F. Schmidt, L. Ledermann, A. Schäffer, H. J. Snaith, M. Lenz, Rapid sequestration of perovskite solar cell-derived lead in soil. J. Hazard. Mater. 436, 128995 (2022). doi: 10.1016/ j.jhazmat.2022.128995; pmid: 35525217 S. A. Khalifa et al., Fate and Exposure Assessment of Pb Leachate from Hypothetical Breakage Events of Perovskite Photovoltaic Modules. Environ. Sci. Technol. 57, 5190–5202 (2023). doi: 10.1021/acs.est.2c05815; pmid: 36966415 R. Vidal et al., Assessing health and environmental impacts of solvents for producing perovskite solar cells. Nat. Sustain. 4, 277–285 (2021). doi: 10.1038/s41893-020-00645-8 T. D. Siegler et al., The Path to Perovskite Commercialization: A Perspective from the United States Solar Energy Technologies Office. ACS Energy Lett. 7, 1728–1734 (2022). doi: 10.1021/acsenergylett.2c00698 L. Duan et al., Stability challenges for the commercialization of perovskite–silicon tandem solar cells. Nat. Rev. Mater. 8, 261–281 (2023). doi: 10.1038/s41578-022-00521-1 M. Hull, J. Rousset, V. S. Nguyen, P.-P. Grand, L. Oberbeck, Prospective Techno-Economic Analysis of 4T and 2T Perovskite on Silicon Tandem Photovoltaic Modules at GW-Scale Production. Sol. RRL 7, 2300503 (2023). doi: 10.1002/solr.202300503 M. De Bastiani et al., Monolithic Perovskite/Silicon Tandems with >28% Efficiency: Role of Silicon-Surface Texture on Perovskite Properties. Adv. Funct. Mater. 33, 2205557 (2023). doi: 10.1002/adfm.202205557 E. Aydin et al., Ligand-bridged charge extraction and enhanced quantum efficiency enable efficient n–i–p perovskite/silicon tandem solar cells. Energy Environ. Sci. 14, 4377–4390 (2021). doi: 10.1039/D1EE01206A X. Y. Chin et al., Interface passivation for 31.25%-efficient perovskite/silicon tandem solar cells. Science 381, 59–63 (2023). doi: 10.1126/science.adg0091; pmid: 37410835 C. Altinkaya et al., Tin Oxide Electron-Selective Layers for Efficient, Stable, and Scalable Perovskite Solar Cells. Adv. Mater. 33, 2005504 (2021). doi: 10.1002/ adma.202005504; pmid: 33660306 D. C. Jordan, N. Haegel, T. M. Barnes, Photovoltaics module reliability for the terawatt age. Prog. Energy 4, 022002 (2022). doi: 10.1088/2516-1083/ac6111 Our World in Data, “Solar (photovoltaic) panel prices”; https://ourworldindata.org/grapher/solar-pv-prices. Solliance, “Scaled perovskite solar modules pass three critical stability tests” (2020); https://www.solliance.eu/2020/scaledperovskite-solar-modules-pass-three-critical-stability-tests/. Microquanta, “Microquanta demonstrates high stability beyond IEC standards” (2020); https://www.microquanta. com/en/newsinfo/112E5B4ABC180F9D/. G. Wang et al., Molecular engineering of hole-selective layer for high band gap perovskites for highly efficient and stable perovskite-silicon tandem solar cells. Joule 7, 2583–2594 (2023). doi: 10.1016/j.joule.2023.09.007

p

91.

C. D. Rodríguez-Gallegos et al., Global Techno-Economic Performance of Bifacial and Tracking Photovoltaic Systems. Joule 4, 1514–1541 (2020). doi: 10.1016/j.joule.2020.05.005 T. Leijtens, K. A. Bush, R. Prasanna, M. D. McGehee, Opportunities and challenges for tandem solar cells using metal halide perovskite semiconductors. Nat. Energy 3, 828–838 (2018). doi: 10.1038/s41560-018-0190-4 J. Werner et al., Perovskite/Perovskite/Silicon Monolithic Triple-Junction Solar Cells with a Fully Textured Design. ACS Energy Lett. 3, 2052–2058 (2018). doi: 10.1021/ acsenergylett.8b01165 I. Almansouri, A. Ho-Baillie, S. P. Bremner, M. A. Green, Supercharging Silicon Solar Cell Performance by Means of Multijunction Concept. IEEE J. Photovolt. 5, 968–976 (2015). doi: 10.1109/JPHOTOV.2015.2395140 E. T. Hoke et al., Reversible photo-induced trap formation in mixed-halide hybrid perovskites for photovoltaics. Chem. Sci. 6, 613–617 (2015). doi: 10.1039/C4SC03141E; pmid: 28706629 Z. Wang et al., Suppressed phase segregation for triplejunction perovskite solar cells. Nature 618, 74–79 (2023). doi: 10.1038/s41586-023-06006-7; pmid: 36977463 M. T. Hörantner et al., The Potential of Multijunction Perovskite Solar Cells. ACS Energy Lett. 2, 2506–2513 (2017). doi: 10.1021/acsenergylett.7b00647 M. Morales-Masis, S. De Wolf, R. Woods-Robinson, J. W. Ager, C. Ballif, Transparent Electrodes for Efficient Optoelectronics. Adv. Electron. Mater. 3, 1600529 (2017). doi: 10.1002/ aelm.201600529 B. A. McNulty, S. M. Jowitt, Barriers to and uncertainties in understanding and quantifying global critical mineral and element supply. iScience 24, 102809 (2021). doi: 10.1016/ j.isci.2021.102809; pmid: 34308296 Y. Zhang, M. Kim, L. Wang, P. Verlinden, B. Hallam, Design considerations for multi-terawatt scale manufacturing of existing and future photovoltaic technologies: Challenges and opportunities related to silver, indium and bismuth consumption. Energy Environ. Sci. 14, 5587–5610 (2021). doi: 10.1039/D1EE01814K J. Liu et al., Efficient and stable perovskite-silicon tandem solar cells through contact displacement by MgFx. Science 377, 302–306 (2022). doi: 10.1126/science.abn8910; pmid: 35737811 J. Warby et al., Understanding Performance Limiting Interfacial Recombination in pin Perovskite Solar Cells. Adv. Energy Mater. 12, 2103567 (2022). doi: 10.1002/aenm.202103567 E. Aydin et al., Sputtered transparent electrodes for optoelectronic devices: Induced damage and mitigation strategies. Matter 4, 3549–3584 (2021). doi: 10.1016/j.matt.2021.09.021 A. F. Palmstrom et al., Interfacial Effects of Tin Oxide Atomic Layer Deposition in Metal Halide Perovskite Photovoltaics. Adv. Energy Mater. 8, 1800591 (2018). doi: 10.1002/ aenm.201800591 N. Arora et al., Perovskite solar cells with CuSCN hole extraction layers yield stabilized efficiencies greater than 20. Science 358, 768–771 (2017). doi: 10.1126/science. aam5655; pmid: 28971968 A. u. Rehman et al., Electrode metallization for scaled perovskite/silicon tandem solar cells: Challenges

Submitted 28 February 2023; accepted 1 December 2023 10.1126/science.adh3849

y ,

Aydin et al., Science 383, eadh3849 (2024)

12 January 2024

13 of 13

RES EARCH

RESEARCH ARTICLE SUMMARY

◥

CANCER IMMUNOLOGY

Deterministic reprogramming of neutrophils within tumors Melissa S. F. Ng†*, Immanuel Kwok†, Leonard Tan†, Changming Shi†, Daniela Cerezo-Wallis, Yingrou Tan, Keith Leong, Gabriel F. Calvo, Katharine Yang, Yuning Zhang, Jingsi Jin, Ka Hang Liong, Dandan Wu, Rui He, Dehua Liu, Ye Chean Teh, Camille Bleriot, Nicoletta Caronni, Zhaoyuan Liu, Kaibo Duan, Vipin Narang, Iván Ballesteros, Federica Moalli, Mengwei Li, Jinmiao Chen, Yao Liu, Lianxin Liu, Jingjing Qi, Yingbin Liu, Lingxi Jiang, Baiyong Shen, Hui Cheng, Tao Cheng, Veronique Angeli, Ankur Sharma, Yuin-han Loh, Hong Liang Tey, Shu Zhen Chong, Matteo Iannacone, Renato Ostuni, Andrés Hidalgo, Florent Ginhoux, Lai Guan Ng*

INTRODUCTION: Neutrophils are the first re-

Mature Growth Angiogenesis

Immature

VEGFα

Stromal

Hypoxic-glycolytic niche

Tumor

Tumor-infiltrating neutrophils undergo convergent reprogramming into pro-angiogenic neutrophils that support tumor growth. In cancer, both immature and mature neutrophils infiltrate the tumor. After entering the tumor microenvironment, these neutrophils undergo differentiation, leading to the formation of transitional populations. Through reprogramming, these populations ultimately converge into a terminal neutrophil state. Reprogrammed neutrophils strongly express VEGFa and localize to a unique hypoxic-glycolytic niche near the tumor core. This places them in an optimal position to exert their pro-angiogenic function within hypoxic and nutrientpoor tumor regions, thereby promoting tumor growth. The emergence of tumor reprogramming reflects the adaptability of neutrophils to environmental cues, allowing them to consolidate their protumoral responses. Ng et al., Science 383, 163 (2024)

12 January 2024

▪

The list of author affiliations is available in the full article online. *Corresponding author. Email: [email protected] (L.G.N.); [email protected] (M.S.F.N.) †These authors contributed equally to this work. Cite this article as M. S. F. Ng et al., Science 383, eadf6493 (2024). DOI: 10.1126/science.adf6493

READ THE FULL ARTICLE AT https://doi.org/10.1126/science.adf6493 1 of 1

,

Reprogrammed neutrophils

y

Transitional states

y g

Tumor reprogramming

Developmental trajectory

CONCLUSION: By examining neutrophils in the context of their ontogeny, we uncovered their intrinsic flexibility in adapting to environmental signals regardless of their initial maturation stage. This implies that neutrophils infiltrating a tissue niche follow a common path, merging their different functional states into a single terminal phenotype as guided by the tissue. Within the tumor, this deterministic program likely ensures a continual supply of pro-angiogenic T3 neutrophils that fuel tumor growth. Our findings thus demonstrate how short-lived effector cells such as neutrophils effectively tailor their functions to accommodate tissue requirements, highlighting the untapped possibilities of targeting the local neutrophil response as immunotherapy.

y

RESULTS: We identified three distinct neutrophil states within the tumor microenvironment, T1, T2, and T3, which were epigenetically and transcriptionally distinct from neutrophils in the bone marrow, spleen, and blood. By assessing nuclear morphology and maturation status, we determined that immature and mature neutrophils infiltrating the tumor differentiated into transitional T1 and T2 populations, respectively. T1 and T2 neutrophils underwent further

g

RATIONALE: To identify the mechanisms by which disparate neutrophil states are coordinated into a concerted protumoral response, we used single-cell RNA sequencing and ATACseq (assay for transposable chromatin sequencing) on neutrophils from various organs and tu-

mors in a murine orthotopic model of pancreatic cancer. Tumor neutrophil states identified from these analyses were validated by multiparametric flow cytometry, and spatial mapping at the RNA and protein levels were performed to reveal their localization within the pancreatic tumors. In vitro and in vivo approaches were then used to examine how the tumor environment shapes neutrophil phenotype, lifespan, and protumoral functions.

p

sponders to infection and injury and are rapidly recruited to affected tissues in large numbers to enact their protective function. As such, neutrophils were historically perceived as a homogeneous and transient population. Recently, however, a diverse array of neutrophil states has been reported in cancer, varying in their maturation, surface marker expression, and transcript profiles. The relationship between these neutrophil states and their organization into a unified protumoral response have yet to be elucidated, limiting the therapeutic targeting of neutrophils in cancer.

reprogramming to converge into the T3 neutrophil state, which was terminally differentiated and expressed the surface marker dcTRAIL-R1. dcTRAIL-R1 up-regulation in tumor-naïve neutrophils could be induced by exposure to tumorconditioned medium in vitro or entry into the tumor in vivo, and was accompanied by the expression of T3-specific genes. More importantly, this phenomenon was independent of their initial maturation phenotype. These findings thus underscore the capability of neutrophils to adopt a new functional phenotype, overlaying it onto their existing differentiation stage. The T3 phenotype was strongly correlated with a prolonged lifespan, with dcTRAIL-R1+ neutrophils persisting for more than 5 days within the tumor. Furthermore, T3 neutrophils were mainly localized to a unique hypoxicglycolytic niche within the tumor, where they optimally exerted their pro-angiogenic function. This indicated that neutrophil reprogramming plays a critical role in enabling their survival under hypoxia, oxidative stress, and metabolic perturbations within the tumor microenvironment. Specifically, T3 neutrophils expressed high levels of vascular endothelial growth factor alpha (VEGFa) and substantially enhanced blood vessel formation within the tumor core, and only coinjection of T3 neutrophils with tumor cells accelerated tumor growth. Therefore, the ablation of either T3 neutrophils or VEGFa inhibits this growth enhancement. Finally, all three tumor neutrophil states were observed across mouse models and in multiple human cancers, with the T3 signature predicting poorer patient outcomes in two independent human pancreatic cancer cohorts and other solid tumor types.

RES EARCH

RESEARCH ARTICLE

◥

CANCER IMMUNOLOGY

Deterministic reprogramming of neutrophils within tumors Melissa S. F. Ng1*†, Immanuel Kwok1†, Leonard Tan1†, Changming Shi2†, Daniela Cerezo-Wallis3,4, Yingrou Tan1,5, Keith Leong1, Gabriel F. Calvo6, Katharine Yang1, Yuning Zhang7,8, Jingsi Jin2, Ka Hang Liong1, Dandan Wu9, Rui He2, Dehua Liu1, Ye Chean Teh1, Camille Bleriot10,11, Nicoletta Caronni12, Zhaoyuan Liu9, Kaibo Duan1, Vipin Narang1, Iván Ballesteros3, Federica Moalli13,14, Mengwei Li1, Jinmiao Chen1, Yao Liu15, Lianxin Liu15, Jingjing Qi16,17, Yingbin Liu16,17, Lingxi Jiang18,19,20, Baiyong Shen18,19,20, Hui Cheng21, Tao Cheng21, Veronique Angeli7,8, Ankur Sharma22,23,24, Yuin-han Loh25, Hong Liang Tey5,26,27, Shu Zhen Chong1,28, Matteo Iannacone13,14,29, Renato Ostuni12,29, Andrés Hidalgo3,4, Florent Ginhoux1,9,10,30, Lai Guan Ng2,1,28*

y y g

trum of neutrophil states in cancer, with differences in density (4, 5), surface markers (6–10), and transcript expression (11–13). Neutrophil heterogeneity also exists in the bone marrow (BM) in the form of various maturation stages. Numerous studies have shown that granulocytic progenitor cells differentiate sequentially into precursor, immature, and finally mature neutrophils, and each subset of cells has distinct functional capabilities (14–19). Tumor-induced chronic inflammation triggers the premature egress of these precursors into the circulation and subsequently into the tumor (4, 14, 17, 19, 20), and extramedullary sites such as the spleen (13, 14), have been proposed to be priming sites of protumoral neutrophils

To investigate neutrophil heterogeneity in cancer, we used an orthotopic mouse model of pancreatic ductal adenocarcinoma (PDAC). Using a pancreatic cancer cell line previously established from a tumor in the Pdx1Cre; KrasG12D/+; Trp53R172H/+ (KPC) genetically modified mouse model (23), cultured cells were orthotopically transplanted into the pancreas, which grew in situ to form a tumor characterized by copious neutrophil infiltration (14). Single-cell RNA sequencing (scRNAseq) was performed on total CD11b+CD115–Ly6G+ cells from the BM, spleen, blood, and pancreatic tumors to characterize neutrophil heterogeneity in the context of their ontogenetic order and origin (n = 2; Fig. 1A and fig. S1, A and B). Uniform manifold approximation and projection (UMAP) analysis revealed that neutrophils from the BM, spleen, and blood clustered according to their developmental states, from pre-neutrophils (preNeus) to immature neutrophils (IMM 1 and 2) and mature neutrophils (MAT 1 to 5) (Fig. 1B and fig. S1C). Integration with a previous dataset containing healthy mouse neutrophils from matching tissues (19) showed that despite

g

N

eutrophils play a substantial role in the immune response to infection and injury. As one of the first cells to enter a damaged site from the circulation, the rapid recruitment of large numbers of neutrophils into the tissue is key for their protective function (1). This process is co-opted in pathological settings such as cancer, in which persistent neutrophil infiltration into the tumor has been consistently associated with poorer patient outcomes (2). Many studies have associated protumoral functions to neutrophils, which in this context have been referred to as granulocytic myeloid-derived suppressor cells (3). This uniform view has since been disrupted by the identification of a wide spec-

Intratumoral neutrophils converge into a distinct transcriptional state

p

Neutrophils are increasingly recognized as key players in the tumor immune response and are associated with poor clinical outcomes. Despite recent advances characterizing the diversity of neutrophil states in cancer, common trajectories and mechanisms governing the ontogeny and relationship between these neutrophil states remain undefined. Here, we demonstrate that immature and mature neutrophils that enter tumors undergo irreversible epigenetic, transcriptional, and proteomic modifications to converge into a distinct, terminally differentiated dcTRAIL-R1+ state. Reprogrammed dcTRAIL-R1+ neutrophils predominantly localize to a glycolytic and hypoxic niche at the tumor core and exert pro-angiogenic function that favors tumor growth. We found similar trajectories in neutrophils across multiple tumor types and in humans, suggesting that targeting this program may provide a means of enhancing certain cancer immunotherapies.

(12, 21, 22). Collectively, this spectrum of neutrophil states, encompassing phenotypic and maturation differences, is proposed to make up the functional diversity of neutrophils in cancer (4–15). Here, we used multi-omics approaches in pancreatic tumors at single-cell transcriptional and spatial resolution to examine neutrophil functional diversity in the context of their ontogenetic order, origin, and influence from tissue signals. We unexpectedly found that although various populations of neutrophils entered the tumor, it was only inside the tumor that neutrophils entered a convergent trajectory that directed them toward a specific protumoral state. Our findings challenge current models and suggest the potential of targeting neutrophils for cancer immunotherapy.

y ,

1

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

Ng et al., Science 383, eadf6493 (2024)

12 January 2024

1 of 16

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

g

Diffusion component 2

p

Diffusion component 2

UMAP2

Fig. 1. Neutrophils infiltrate scRNAseq workflow A pancreatic tumors and undergo Harvest selected tissues Sort CD11b+Ly6G+ neutrophils 10X V3 3' scRNAseq Generate tumor further differentiation and Orthotopic PDAC Demultiplex, Hashtag Bone marrow tumor cell injection downstream label Blood converge upon a transcriptionally analysis individual Spleen tissues distinct T3 neutrophil state. Tumor (A) Schematic showing scRNAseq Bone marrow (BM) Spleen Blood Tumor B workflow. CD11b+Ly6G+ neutrophils T3 were sorted from the BM, spleen, preNeu T1 blood, and pancreatic tumor of T2 MAT 5 MAT 3 tumor-bearing mice 6 weeks after MAT 1 IMM 1 MAT 4 orthotopic injection (n = 2 biological IMM 2 MAT 2 replicates). Each sample was individually tagged with cell-hashing UMAP1 antibodies before being pooled for analysis with 10X V3 3′ scRNAseq D E Maturation signature Diffusion map RNA velocity C (see also fig. S1A and the materials and methods). (B) UMAP projection preNeu preNeu of total neutrophils in the BM, or spleen, blood, and tumor show strong urs rec P enrichment of three clusters (T1, preNeu Immature IMM 1 re Immature T2, and T3) in the tumor. Louvain atu IMM 2 m Im MAT 1(BM) clustering was performed, and colors T1 e MAT 2 T1 correspond to the clusters identified. tur a MAT 3 Mature M Mature MAT 4 (C) The neutrophil maturation MAT 5 score can be used to identify Louvain T1 T2 or T2 m clusters along their differentiation T2 T3 Tu T3 T3 trajectory. Histograms show the 1 0 Maturation score Diffusion component 1 Diffusion component 1 module score of the maturation gene signature for each cluster identified Immature Open chromatin regions (OCRs) at T3 genes ATACseq in (B), with scores closest to 1 being F BM,BM,Spleen, G Blood Immature WT Tumor the most mature. Dotted line in Tumor Mature Spleen BM Tumor Blood BM BM,BM, gray defines the cutoff for mature 0 10 Spleen, IMM MAT IMM MAT IMM MAT IMM MAT IMM MAT Blood 0 neutrophils and is set at the lower Tumor 20 0 bound of the Mature 2 cluster. mouse: 10 0 -10 0 Mature BM 0 0 -20 Tumor -1 Spleen (D) Low dimensional embedding of 0 0 0 -2 Blood 20 0 00 7 0 Tumor 1 all neutrophils using a diffusion map 30 0 0 -4 Hypoxia: WT 00 00 -1 approach reveals a branch point mouse: Alkbh5, Bnip3, 00 BM -2 Bnip3l,Cited2, between mature and tumor neutrophils. 6 Ddit4, Egln3, Scatterplot shows diffusion components T3 enriched TF regulons Hilpda, Hmox1, H from scRNAseq Itpr2, Npepps, 1 and 2. Light gray arrows are T3 T2 T1 MAT IMM pre 5 P4ha1, Rbpj, used to indicate the trajectory from Scaled Rgcc, Vegfa Sap30 (514) AUC Ddit3 (31) precursor, immature, and mature score Atf4 (867) 3 Glycolysis: Mafk (86) 4 neutrophil states. Black arrows denote 2 Bhlhe41 (195) Ddit4, Usf2 (1632) 1 the branching into T3 neutrophils Eno1, Hk2, Atf5 (436) 0 Hif1a (248) Ier3, Pfkp -1 from T1 and T2 states. (E) RNA Zmiz1 (861) -2 3 Atf3 (1326) Jun (1664) velocity suggests the convergent Angiogenesis: Tgif1 (476) Hk2, Hmox1, Maff (463) differentiation of T1 and T2 into T3 Bhlhe40 (1971) Itgb1, Lgals3, 2 Rest (138) neutrophils. RNA velocity vectors Rhob, Thbs1, Egr1 (2600) Nfil3 (1096) Vegfa were projected on the diffusion map Atf1 (1412) Creb5 (143) embedding with velocity vectors Fos (800) 1 Junb (1510) Cebpb (1454) terminating in the tumor and mature Nfe2l2 (49) Fosb (238) neutrophils. (F) Principal component Runx1 (66) 0 -2kb 0 +2kb genomic dist.(bp) analysis of ATACseq indicated that changes in chromatin accessibility established in tumor neutrophils clustered them away from other neutrophil indicated neutrophil subsets; histograms indicate average mapping intensity. T3 genes linked to hypoxia, glycolysis, and angiogenesis are annotated (see also subsets. ATACseq was performed for immature (circles) and mature (squares) neutrophils sorted from the BM (gray and dark blue, n = 3 each), spleen table S2). (H) Heatmap showing scaled AUC scores of the top 25 transcription factor regulons enriched in T3 neutrophils computed by PySCENIC (see also (light blue, n = 3 each), blood (red, n = 3 mature, n = 2 immature), and tumor (orange, n = 3 each) in wild-type (WT) or tumor-bearing (tumor) mice. Numbers table S2). Numbers in brackets denote the number of genes assigned to the transcription factor regulon. Transcription factors that also had increased motif denote the number of biological replicates used. (G) OCRs matched to differentially expressed T3 genes have increased accessibility only in immature and mature enrichment in tumor-associated immature and mature bulk neutrophil tumor neutrophils. Heatmap shows intensity of fragment mapping across populations are shown in bold.

y

2:

PC

PC

1:

46

.6

7%

3%

: 7.9

PC3

8%

.5

19

,

2 of 16

y

12 January 2024

y g

Ng et al., Science 383, eadf6493 (2024)

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

,

3 of 16

y

12 January 2024

y g

Ng et al., Science 383, eadf6493 (2024)

Having shown that immature T1 and mature T2 neutrophils in the tumor undergo transcriptional and epigenetic reprogramming toward the T3 neutrophils, we next assessed whether we could distinguish these subsets by protein expression. This was particularly important because increases in RNA expression in neutrophils typically precedes protein expression, especially for genes tightly regulated in neu-

y

To confirm that T3 neutrophils can be reprogrammed from both immature (T1) and mature (T2) tumor neutrophils, we performed

CD101 and dcTRAIL-R1 discriminate tumor neutrophil states

trophil development, such as those for granule proteins (14). T1, T2, and T3 neutrophils had differential expression of surface marker genes, suggesting that a combination of surface marker protein staining could be used to identify them (fig. S3A). Using a multiparametric flow cytometry approach, live CD45+ cells from the PDAC tumor were screened for 249 cell surface markers. Then, all captured cells were clustered with InfinityFlow (28, 29), and the coexpression patterns of all tested markers were evaluated (fig. S3B), which revealed two distinct neutrophil clusters (Fig. 2A). High levels of immuno-modulatory or suppressive markers such as dcTRAIL-R1, PD-L1 (CD274) (6), CD14 (13), CD371 (30), VISTA (31), and CD39 (32), distinguished Cluster 2 neutrophils from those of Cluster 1 (Fig. 2B). These markers correlated with surface marker genes enriched in T3 neutrophils (fig. S3C), suggesting that they may serve to identify the T3 state. Indeed, we found that dcTRAIL-R1 was expressed mainly by a population of tumor neutrophils (Fig. 2C) while having minimal expression in tumor-infiltrating monocytes and macrophages and in neutrophils in the BM, spleen, and blood (Fig. 2C and fig. S3D). By contrast, staining for other markers such as VISTA and CD14 showed expression across multiple neutrophil subsets and was less restricted to tumor neutrophils (Fig. 2D and fig. S3D). dcTRAIL-R1 expression in tumor-infiltrating neutrophils was also conserved across several other tumor types, including orthotopic breast cancer [median: 20.4%; interquartile range (IQR): 18.0 to 23.1%] and orthotopic lung cancer (median: 12.1%; IQR: 4.54 to 19.5%) mouse models (fig. S3E). Across all experimental tumor models examined, dcTRAIL-R1 expression was similarly limited to a subset of tumor neutrophils, suggesting a possible candidate marker for T3 neutrophils. Consistent with its protein expression pattern, gene expression and RNA velocity (indicative of active gene transcription) for the gene encoding dcTRAIL-R1 (Tnfrsf23) were highest in the T3 neutrophil cluster (fig. S4A). Therefore, we attributed the T3 population to the dcTRAILR1+ cluster (cluster 2), whereas the T1 and T2 populations were likely contained within the dcTRAIL-R1– cluster (cluster 1) identified in our InfinityFlow analysis (Fig. 2A). Because T1 and T2 neutrophils were transcriptionally identified as immature and mature, respectively, in our scRNAseq data (Fig. 1D), we next evaluated CD101 expression, which separates immature neutrophils from mature neutrophils (14). Cluster 1 exhibited a continuum of CD101 expression (fig. S4B), suggesting that CD101 can be used to distinguish T1 and T2 neutrophils within the dcTRAIL-R1– population. Using CD101 and dcTRAIL-R1 to identify the different neutrophil types by flow cytometry (Fig. 2E), we isolated putative T1, T2, and T3 neutrophils and discovered that putative T1 neutrophils had

g

Epigenetic reprogramming of tumor-infiltrating neutrophils

ATACseq (assay for transposase-accessible chromatin followed by sequencing) on sorted CD101– (immature) and CD101+ (mature) neutrophils (14) from various tissue compartments of control and tumor-bearing mice. Principal component analysis revealed that both immature and mature tumor neutrophils clustered separately from corresponding subsets of neutrophils from other tissues (Fig. 1F), indicating that signals from the tumor microenviroment imprint specific chromatin accessibility changes within infiltrating neutrophils. We then identified all open chromatin regions (OCRs) that were differentially accessible in immature and mature tumor neutrophils relative to the nontumor neutrophil subsets (table S2). Increased chromatin accessibility for genes up-regulated in T3 neutrophils (T3 genes) in both immature and mature tumor neutrophils (Fig. 1G) included genes involved with hypoxia, glycolysis, and angiogenesis, such as Vegfa and Hk2 (fig. S2A). These changes in chromatin were undetectable in neutrophils from other tissues, which still shared similar accessibility for canonical neutrophil genes such as Cebpa and Gfi1 (fig. S2B). To elucidate putative upstream regulators of the T3 program, we used pySCENIC (27) to identify transcription factor regulons predicted in T3 neutrophils that had minimal presence in the other neutrophil subsets (Fig. 1H). This was then correlated to motifs found in the bulk ATACseq analysis (fig. S2C), which revealed T3specific transcription factors with enriched motifs in immature and mature tumor neutrophil differentially accessible OCRs (fig. S2, D and E). Among the T3-specific transcription factors, Mafk, Nfe2l2, and Atf3 were implicated in regulating metabolic and oxidative stress typical of the tumor microenvironment (fig. S2, D and E). Correlation of the underlying chromatin accessibility with transcription factors predicted to govern the T3 state suggested that these epigenetic changes initiated within T1 and T2 neutrophils are reinforced by transcription factor activity as they transition toward a T3 state. These findings suggest that tumorinfiltrating neutrophils can be reprogrammed by the tumor microenvironment regardless of their stage of maturity by converging transcriptional and chromatin trajectories toward a distinct T3 state.

p

potential perturbations from the tumor, no new developmental clusters or trajectories emerged in neutrophils from tissues of the tumor-bearing mice (fig. S1, D and E). This suggests a robust adhesion of neutrophils to a single developmental trajectory within the tissues where they develop and circulate. By contrast, neutrophils acquired specific transcriptional profiles in the tumor (Fig. 1B and fig. S1C) and formed three distinct clusters (T1, T2, and T3) separate from the other compartments. Because immature and mature neutrophils can infiltrate tumoral tissue (14, 15), we investigated the relationship between the maturation and functional states of the T1 to T3 neutrophil populations using a neutrophil maturation gene signature (table S1) curated by Xie et al. (19); each cluster in our dataset was graded by a maturation score. T1 neutrophils had a maturation score comparable to immature neutrophils (i.e., the IMM 2 cluster), whereas T2 neutrophils had a maturation score similar to mature neutrophils (i.e., the MAT 1 to 5 clusters) (Fig. 1C). Next, we sought to uncover more potential links between neutrophil clusters in the tumor and the other tissue compartments by using a diffusion map approach that orders cells on the basis of transitional probabilities and better preserves differentiation trajectories (24, 25). We found that tumor neutrophils deviated from the steady-state neutrophil developmental trajectory (Fig. 1D). Specifically, T1 and T2 neutrophils progressed along the tumor-specific branch and converged at the T3 state, suggesting that T1 and T2 may both give rise to the T3 population, which is predicted to be the most terminally differentiated subset. RNA velocity (26) indicated a pseudotime progression from immature neutrophils to T1 and mature neutrophils to T2, suggesting that T1 and T2 neutrophils are transitory states from immature and mature neutrophils that have migrated into the tumor (Fig. 1E). By contrast, T3 neutrophils exhibited a maturation score that was in between that of the T1 and T2 neutrophils (Fig. 1C), representing a potential admixture of T1 and T2 cells. Accordingly, we observed that velocity vectors originating from both T1 and T2 populations terminated in the T3 cluster (Fig. 1E). These findings thus corroborate earlier studies demonstrating that both mature and immature neutrophils infiltrate tumor tissue (4) and become T1 and T2 neutrophils, respectively. Further, our data predict that these stages are transitional, because both T1 and T2 neutrophils remained amendable for further differentiation into T3 neutrophils.

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

UMAP 2

UMAP 2

Fig. 2. Combinatorial Gated on live CD45+ cells in the tumor, A B dcTRAIL-R1 PD-L1 CD14 249 predicted markers by InfinityFlow: Z score Z score Z score CD101 and dcTRAIL-R1 8 0 24 0 0 4 1.4 -2.1 expression identifies Eosinophils Cluster 2 Ly6G intensity (log2MFI) 1000 T1, T2, and T3 neutrophils. (A) UMAP Macrophages/ NK DCs 800 projection of live CD45+ Neutrophil Cluster 2 Monocytes immune cells within the Cluster 1 600 PDAC tumor reveals CD8 T CD371 (MICL) VISTA CD39 two clusters of neutroZ score Z score Z score 0 8 1000 0 4 0 5 400 phils. High-parameter 750 flow cytometry was 500 performed on single-cell 200 B CD4 T cells suspensions of pooled 250 Neutrophil Cluster 1 tumors (n = 10 biological 0 0 1000 0 800 200 600 400 replicates) using the 0 250 500 750 1000 UMAP 1 UMAP 1 LEGENDScreen panel (BioLegend). Live CD45+ Peripheral tissue neutrophils F C Gene signature expression well-compensated FCS Spleen BM Blood were first analyzed and 0.96% 0.44% 0.98% 10 10 10 T3 T2 T1 then exported out for 10 10 10 dcTRAIL-R1dcTRAIL-R1dcTRAIL-R1+ analysis using the Immature Immature/mature Mature 10 10 10 InfinityFlow package in 0 0 0 -10 -10 -10 R. UMAP shows Ly6G Atf3 -10 0 10 10 10 -10 0 10 10 10 -10 0 10 10 10 Ccl3 expression intensity Tumor Ccl4 imputed by InfinityFlow Neutrophils Monocytes Macrophages Cd274 Cstb 10 10 10 8.94% 37.7% 9.85% from high (red) to low Cxcl3 10 10 10 (blue). Clusters are Hcar2 Hilpda annotated by canonical 10 10 10 Hk2 0 0 0 Hmox1 surface marker expres-10 -10 -10 Ier3 sion (see also fig. S3B). -10 0 10 10 -10 0 10 10 -10 0 10 10 Jun dcTRAIL-R1 Plin2 (B) Neutrophil cluster 2 Spp1 has increased expression Tgif1 D Validated markers Tnfrsf23 of immunosuppressive Vegfa dcTRAIL−R1 Zeb2 and/or immunomodulaScaled CD371 Ldha MFI tory surface markers Mif 1.00 CD14 0.75 compared with cluster 1. Cd300ld CD39 0.50 Surface marker expresCxcr2 VISTA 0.25 Dusp1 sion intensities are 0.00 Gbp2 PD-L1 Ifitm1 shown for curated BM Spleen Blood Il1b surface markers (clockIsg15 Jaml wise): dcTRAIL-R1, Junb Tumor Peripheral tissues PD-L1, CD14, CD371, Msrb1 Osm VISTA, and CD39. Data S100a6 Sorting for transcriptomic validation E Selplg are represented as a Slpi Z score based on Total live,CD45+ Ltc4s predicted log2 mean neutrophils in tumor Mmp8 Mmp9 fluorescence intensity dcTRAIL-R1dcTRAIL-R1+ Ppia (MFI) from high (red) to Prr13 Ptma Immature Mature low (blue) (see also Retnlg fig. S3C). (C) dcTRAIL-R1 Scaled log counts Putative T3 Putative T2 Putative T1 expression marks and +/+ CD101 CD101 CD101 is restricted to a sep-3 -2 -1 0 1 2 dcTRAIL-R1+ dcTRAIL-R1- dcTRAIL-R1arate population of tumor-infiltrating neu(E) Proposed gating strategy to isolate T1, T2, and T3 neutrophils by dcTRAIL-R1 and trophils. Representative contour flow cytometry plots (top) show dcTRAIL-R1 CD101 expression from the tumor. (F) Sorted dcTRAIL-R1+ tumor neutrophils have the expression against Ly6G expression in the tumor for indicated cell populations. highest expression of the T3 transcriptional signature. Heatmap shows scaled Nanostring (D) Heatmap shows scaled MFI for markers in (B), scaled between 0 and gene counts (normalized against internal positive controls and housekeeping genes) 1 across all populations, with 0 being the lowest MFI. Populations were analyzed for T1 (n = 4), T2 (n = 4). and T3 (n = 4) neutrophils sorted according to (D). Numbers by flow cytometry, in which total neutrophils were gated as CD11b+CD115–Ly6G+ represent the biological replicates across two independent experiments. Genes belonging in the BM, spleen, blood, and tumor. Tumor macrophages (CD11b+Gr-1–F4/80hiMHCIIhi) to either the T1, T2, or T3 transcriptional signature are indicated (see also fig. S4D). and monocytes (CD11b+Ly6G–Ly6Chi) were gated accordingly (see also fig. S3D). 5

5

4

4

4

3

3

3

4

5

3

3

3

4

5

3

5

5

4

4

4

3

3

3

3

3

4

5

3

3

3

4

3

3

4

3

4

4 of 16

,

12 January 2024

y

Ng et al., Science 383, eadf6493 (2024)

y g

Neu.

Mono.

Macs.

WT

Tum.

WT

Tum.

WT

Tum.

T1 genes

3

y

5

3

g

Ly6G

3

T3 genes

3

3

T2 genes

3

p

5

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

,

5 of 16

y

12 January 2024

Because T3, but not T1 or T2, neutrophils had transcriptional and epigenetic profiles that were enriched for the hypoxia, glycolysis, and angiogenesis pathways, we evaluated whether this was reflected in their localization to hypoxic, glycolytic, or angiogenic regions within the tumor, respectively. All spots within the tumor region were first scored for GO pathways corresponding to glycolysis (fig. S6C), hypoxia (fig. S6D), and angiogenesis (fig. S6E), with a 50th percentile cutoff used to determine low and high regions. We then determined the frequency of neutrophil-containing spots falling into low and high regions for each tumor neutrophil subset. T3 neutrophils were found at greater frequencies in high scoring regions for glycolysis (68.8 ± 19.7%), hypoxia (53.3 ± 4.65%), and angiogenesis (62.2 ± 14.2%) (Fig. 3H). By contrast, most T2 (fig. S6F) and T1 (fig. S6G) neutrophil-containing spots fell into regions with low glycolysis, hypoxia, and angiogenesis scores, indicating that partitioning of the neutrophil subsets potentially results from T3 neutrophils occupying a specific regional hypoxic and glycolytic tumor niche. We next set out to validate these observations at single-cell resolution because the current Visium spatial transcriptomic resolution is limited to 55-mm spots. We used MACSima imaging cyclic staining (MICS) (Fig. 4A), in which iterative immunofluorescence staining and photobleaching allows for the evaluation of large numbers of antibody targets within the same tissue slice (39). Optimization was first performed to determine working markers within orthotopic pancreatic tumor tissues (fig. S7, A and B), given that MICS staining times were relatively short at 10 min. Testing revealed that CD29, Galectin-3, and CD44, which typically mark fibroblasts (40–42), displayed the greatest staining at peripheral stromal and desmoplastic (panCK–/lowDAPIhi) regions (fig. S7C), whereas CD31 and CD105 marked vessels within the stromal and tumor region (panCK+). CD105 was able to better resolve intratumoral vessels (fig. S7C), consistent with their role in marking tumor-associated endothelium (43). To identify hypoxic regions within the tumor, Hif1a, CD39, and CD73 staining was evaluated. CD73, an adenosine monophosphate ectoenzyme, has been found to be directly induced by hypoxia and Hif1a in cancer (44–47). Immunofluorescence analysis revealed diverse staining patterns of CD73 in both stromal and desmoplastic regions and tumor (panCK+) regions (fig. S7C), resembling the spatial distribution of hypoxia gene signature scores observed in

y g

Ng et al., Science 383, eadf6493 (2024)

Reprogrammed neutrophils occupy a hypoxic and glycolytic tumor niche

y

Spatial mapping of immune microenvironments in tumors has been highly informative by providing insights into their diversity and functionality on the basis of their localization within the tumor (33). To gain a better understanding of how tumor neutrophils are affected by their localization in the tumor, we first investigated possible functional differences between the three tumor neutrophil subsets by performing gene ontology (GO) analysis on differentially expressed T1, T2, and T3 genes (table S4). T1 neutrophils were enriched for pathways relating to transcription and translation (including genes governing ribosomal biogenesis such as Npm1) and oxidative phosphorylation (proton membrane transport) (Fig. 3A), which is consistent with the immature phenotype of T1 neutrophils (14). T2 neutrophils were enriched in pathways related to transcriptional regulation, amide and reactive oxygen species metabolism, and immune responses, as well as type I interferon genes, including Ifit1, Ifit2, Ifit3, and Isg15 (Fig. 3B), which likely reflects neutrophil activation upon tumor infiltration. Finally, T3 neutrophils featured pathways of cell stress and survival, including response to hypoxia, oxidative stress, and glycolysis, (Fig. 3C), suggesting adaptation to the tumor environment. T3 neutrophils were also enriched for angiogenic genes, including Vegfa, Thbs1 (34), and Lgals3 (35), which, when combined, is suggestive of a strong protumoral role. We next characterized how neutrophils with distinct transcriptional states would interact

suggest that neutrophils are spatially organized within the tumor, and T3 neutrophils occupy a tumor niche that is distinct from that of T1 and T2 neutrophils.

g

Spatial compartmentalization of neutrophils in tumors

with the neighboring tumor microenvironment. Using 4′,6-diamidino-2-phenylindole (DAPI) and pan-cytokeratin (panCK) costaining, we noted that our orthotopic tumors were mainly composed of tumor cell regions (PanCKhighDAPIhigh), fibrotic and/or necrotic regions (panCKhighDAPI–/low), and sparse stromal regions along the edges (panCK–/lowDAPIhigh) (fig. S5, A and B). We performed spatial transcriptomics on four cryosections of three different PDAC tumor samples (Fig. 3D). As expected, necrotic regions with low numbers of DAPI+ cells contained low-quality counts, which were manually annotated and filtered to mitigate RNA contamination from nearby spots and to improve downstream analyses (fig. S5, C and D). We next assessed the tumor architecture by using BayesSpace (36) to identify transcriptionally similar neighborhoods within the tumor (fig. S5E). Clustering and UMAP embedding revealed that 10 clusters were shared across all four sections (fig. S5E). Indicative of the complexity of the tumor environment, the distributions of spatial clusters identified were different in the two cryosections [regions of interest 1 and 2 (ROI1 and ROI2)] taken from different regions of the same tumor (fig. S5E). Each spatial cluster had specific functional pathways associated with their distribution (fig. S5F). For example, regions enriched for pathways related to epithelial-to-mesenchymal transition (EMT, clusters 1 and 10), typically associated with the invasive tumor front (37), were present closer to the tumor periphery, whereas regions associated with hypoxia (cluster 6) and cell cycle progression (G2-M checkpoint, cluster 2) were located close to the core of the tumor (fig. S5, E and F). Using Cell2location (38) with a pre-annotated scRNAseq dataset (fig. S6A) allowed us to map the full transcriptional signature of each cell cluster within the tumor and estimate cell type abundances within each spot. To ensure the accuracy of neutrophil assigned spots, we ascertained the presence of Ly6G immunofluorescence staining falling within the spot as a selection cutoff (fig. S6B). Cell2location revealed that T1, T2, and T3 neutrophils mapped to different regions within the tumor (Fig. 3E), with visualization of T3 enriched spots on the UMAP embedding indicating that T3 neutrophils had the highest enrichment in spatial cluster 6, which was associated with hypoxia (Fig. 3F). T3 neutrophils were also observed to be mostly distributed adjacent to PanCKhighDAPI–/low necrotic zones (Fig. 3G), consistent with their localization to cluster 6 (fig. S5E). By contrast, T1 neutrophils were highly enriched in the cluster 1 (EMT) and clusters 7 and 8 (IL-2 and STAT5 signaling, respectively) regions, whereas T2 neutrophils were mostly found in cluster 3 associated with p53 signaling (Fig. 3F). Both T1 and T2 annotated spots were generally localized closer to the tumor periphery compared with T3 neutrophils (Fig. 3G). Our data

p

a toroidal nuclear morphology resembling immature neutrophils from the BM (fig. S4C). Putative T2 neutrophils had hypersegmented nuclei similar to those of mature BM neutrophils (fig. S4C), whereas putative T3 neutrophils contained cells with hypersegmented or toroidal nuclear morphology (fig. S4C), congruent with our findings that T3 neutrophils are reprogrammed from both immature and mature neutrophils. Next, we found that dcTRAIL-R1+ neutrophils showed the highest expression of T3 genes (table S3), whereas the dcTRAIL-R1–CD101– and dcTRAIL-R1–CD101+ populations were strongly enriched for genes associated with T1 and T2 neutrophils, respectively (Fig. 2F and fig. S4D). Finally, all three neutrophil populations were identified in all experimental tumor models examined (fig. S4E), and the frequencies of the T1 to T3 populations in PDAC as determined by scRNAseq were comparable to the median frequencies of putative T1 to T3 populations isolated by flow cytometry (fig. S4, E and F). Thus, dcTRAIL-R1 and CD101 expression phenotypically divides tumor neutrophils into three distinct populations and successfully recapitulates the T1 to T3 populations defined in our transcriptomic and epigenetic approaches.

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

B

T1 enriched GO pathways

Regulation of ROS metabolic process

Proton transmembrane transport

rRNA processing

C

T2 enriched GO pathways Immune response

Regulation of cell proliferation

Protein folding

Mapped neutrophils

T2

T1

H

T3

Pe he

re Co

ROI 4

Periphery

*

Hypoxiahigh Hypoxialow

55

Hypoxia

UMAP2

20

ry

ry

50 45

g

Periphery

ry he

rip

Pe

P

ROI 3

Glycolysishigh Glycolysislow

40

rip

he

ry

e ph eri

*

60

re

rip

UMAP1

Etf1, Eif4a1,Eif3a, Rps2,Eif5, Denr, Zfp36l2,Npc1 Acod1,Rpl21, Mdm2,Ppp1r15a, Dnajb9

T3 localization 80

Co

re

Spatial clustering 1: EMT 2: G2-M checkpoint 3: p53 signaling 4: Angiogenesis 5: Myc targets 6: Hypoxia 7: IL-2/STAT5 8: signaling 9: IFN response 10: EMT 2

Alkbh5,Bnip3, Bnip3l,Cited2, Egln3, Egr1 Fam162a, Higd1a

ROI 2 (Tumor 1)

ROI 1 (Tumor 1)

re

Cell type deconvolution (Cell2location)

1.76

Ddit3,Bax, Atf4,Atf3, Rgcc,Rbpj

1.00

Co

10x Visium Spatial Gene Expression

Data analysis

2.14

Co

F

Hsp90aa1, Hsp90b1, Hspa1a, Hspa1b, Hspa5,P4hb, Ppia,St13

2.52

1.38

UMAP1

Freeze and cryosection

Avg. log2FC 2.90

T1 mapped

*

Angiogenesishigh Angiogenesislow

80

T2 mapped

Core 1 2 3 4 5 6 7 8 9 10

Core

40

Core

T3 mapped

60

20

Core

Low

12 January 2024

ically evaluated staining for the tumor vasculature (CD105), hypoxia (CD73), and glycolysis (GLUT1) across all five ROIs (Fig. 4C and fig. S8, A and B). CD105+ blood vessels were present as a diffuse network scattered throughout the tumor (Fig. 4C and fig. S8, A and B). A gradient of CD73 staining was apparent throughout the tumor, whereas GLUT1 staining was more tightly localized; however, GLUT1hi areas strongly colocalized with CD73hi regions marking a distinct hypoxic and glycolytic niche (Fig. 6 of 16

,

flow cytometric strategy (Fig. 2D), and were successfully annotated in both confocal (fig. S7D) and MICS (fig. S7E). We performed MICS for five different ROIs across two different orthotopic pancreatic tumors at 4 to 6 weeks after injection. ROIs were selected to unequivocally cover an entire region from the left to right margin encompassing the core of the tumor (fig. S8, A and B), and we observed staining across the full panel for all markers surveyed (fig. S9). We next specif-

y

neutrophils localize to different spatial clusters. Projection of T1, T2, and T3 enriched spots identified by Cell2location on merged UMAP derived from BayeSpace enhanced clustering analysis of tumor sections (n = 4 biological replicates). (F) Merged UMAP representation of spots of tumor sections were analyzed and color-coded according to BayesSpace-identified clusters (top). Violin plots show frequency of T1, T2, and T3 neutrophils enriched spots that map to each cluster (bottom). (G) Spatial mapping of T1, T2, and T3 neutrophils across tumor sections (n = 4) by Cell2location. Black lines denote the outline of the section; gray colored areas indicated the excluded DAPI–panCKhi regions annotated to be fibrotic and or necrotic. Spots are filtered based on Ly6G+ staining (see also fig. S6B). (H) Quantification of percentages of deconvoluted T3 neutrophil-enriched spots falling into high- or low-scoring spots for GO ontology pathways: glycolysis (GO: 0061621), hypoxia (GO: 001666), and angiogenesis (GO: 0045766). Center line of boxplots show median, box hinges represent 25th and 75th percentiles, and whiskers extend to minimum and maximum values. *P < 0.05 by one-tailed t test.

y g

Fig. 3. Spatial compartmentalization of T1, T2, and T3 neutrophils in the pancreatic tumor. (A) Chord diagram showing differentially expressed genes (DEGs) in T1 neutrophils that are enriched for GO pathways linked to transcription and oxidative phosphorylation. (B) Chord diagram showing DEGs in T2 neutrophils that are enriched for GO pathways linked to metabolism and immune response. (C) Chord diagram showing DEGs in T3 neutrophils that are enriched for GO pathways linked to survival and angiogenesis. In (A) to (C), bars associated with each gene are colored by strength of fold change of differential expression and are sized based on the number of pathways with which it interacts (see also table S4 for DEG lists). (D) Spatial transcriptomic analysis workflow. PDAC tumors were isolated and cut into quarters, where the sharp edges denote the core facing regions, before flash-freezing. Fresh frozen PDAC tumors were sectioned and placed on 10X Visium slides containing spatially barcoded capture spots. After processing and sequencing, the data were clustered spatially (BayesSpace) and cell type deconvolution was performed (Cell2location). Gene signatures of various biological processes were then probed and mapped with the UCell package. (E) Tumor

Ng et al., Science 383, eadf6493 (2024)

High

UCell score

Cluster

our 10X Visium analysis (fig. S6B). By contrast, Hif1a did not stain within the MICS time frame, whereas CD39 mostly identified vessels (fig. S7C). Similarly, GLUT1 (glucose transporter 1) staining showed regional restriction (fig. S7C) similar to glycolysis signature scores (fig. S6B). Thus, CD73 and GLUT1 representative staining was used to define the hypoxic and glycolytic tumor niche. T1, T2, and T3 neutrophils were identified by staining for Ly6G, CD101, and dcTRAIL-R1, respectively, based on our

y

Gene signature mapping (UCell)

Angiogenesis

Spatial clustering and enhancement (BayesSpace)

1

Rps28, Rps26, Rplp1, Rpl8,Rpl38, Rpl36, Rpl35a,Rpl30,Rpl26, Rpl22,Rpl19,Rpl18

egfa, Thbs1,V aja1, Canx,Dn bp1a Fk

Pe

Necrotic core

Eif1 Eef1a

l1a, Hmox1,Ials3, Itgb1,Lg,Rhob Ptgs2

p

UMAP2

Periphery

G

Rplp1, Rpl35 Rpl32, Rpl14 Rpl12, Rpl10, Fau

Gapdh,Ier3, Mif,Pgam1, Pgk1,Tpi1

Glycolysis

Mapped neutrophils T1 T2 T3

Cc Ccl3 r Cd1,C , Cc 27 crl2 l4, 4, ,C Cc C d l6 Clelec4 14, , c4 e, n Csf1, Cxcl1 Ctss, Cxcl30,Cxcl2, ,Gbp Gbp5, 2, Gbp7, Ifi204

A Rp tp5 Rp l10 l l1 Rp 1 l1 2 Rp l3 Rpl58 Rpl6 Rplp0 Rps10 Rps14

Cell type deconvolution

Rps8,Rps5,Rps3a1, Rps3,Rps20,Rps19

Npepps Itpr2

E

Isg20,Isg15, Irgm1,Il1a, Ifit3,Ifit2,Ifit1

Harvest orthotopic PDAC tumors 6 weeks post injection

3.40 2.92 2.44 1.96 1.48 1.00

Tnf 1 Xbp 6 3 Zfp gcc t1 R ta 1 S bs Th rn, d Il1 1 1, l2a am c Ic 1, B od c A

Egln Zfp 3 Rh 36l1 oh U N sp1 Tribfkbia8, P Ptg 1, , Ly lac8 s2 Tes 6a, , , 1, sp 9, Du d 1, 1a,C 1, s1 Fthdkn ,Zbp oc C tg1 1,S , B rem ad2 T ,Rs sn1 tafr, Samb20,Pm, id Ra p,Os fkb Pn biz,N Nfk

D

Avg log2FC

1 s2 6 R p s1 Rp psa R s5 Rp s28 Rp 27l Rps 27 Rps Rps25 Rps2 Rps19 Rps15

2.10 1.88 1.66 1.44 1.22 1.00

Glycolytic process Hk2,Pfk Pfkp,Aldo l, a, Ddit4,Eno1

Nr4a1 Marcksl1 Junb, Cox17, Pim1, Cish, Socs3, Il1r2,Igtp, Ifi47

Id1 Il1b Jun Nfil3,Plk3, Ppp1r15a,il Rbpj,Sk

Trmt112 Rps7 Rps24 Rps17 Rpl35 Rpl26 a Rpl1 No 4 Sn p10 Rp u13 R l35 N pl1 R pm 0a ps 1 8

Positive regulation of angiogenesis

Translation

3b, Rtp4, Ifits, Crsnp1, Ier2, Fo Tgm2

Atf3,Bhlhe40, Btgs2,Egr1

Hk2 Ier3 Mif Nme1 Nme2 1 Pgam i1 Tp p Pn 5e Atp g1 5 2 Atp p5g 5j Avg. t p log2FC A At tp5j2 A

Response to hypoxia

Response to oxidative stress

Neg. regulation of translation by RNA pol. II

Regulation of cellular amide metabolic process

Nucleoside disphophate phosphorylation

T3 enriched GO pathways

Frequency (%)

A

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

A

B

Tumor E dg e

or

Tu m o

e

rE

dge

CD73 & GLUT1 F

D

C

p

2.08e−17 1.56e−17 1.04e−17 5.20e−18 5.93e−24

E g y

Normalised Mixing Score

1.5

3000

2000

1000

network of CD105+ vessels (Fig. 4E and fig. S8, A and B). By contrast, T3 (Ly6G+CD101–/+dcTRAILR1+) neutrophils across all ROIs were more likely to cluster together at the CD73hiGLUT1hi hypoxic and glycolytic niche (Fig. 4E), thus giving rise to a clear spatial segregation of the three neutrophil populations, as suggested by our 10X Visium analysis (Fig. 4F and fig. S8, A and B).

12 January 2024

Co-localisation with Hypoxichigh Glycolytichigh Region T1

T2

T3

1.0

,

Stromal

Nearest Distance to Hypoxichigh Glycolytichigh Region

y

Tumor

Hypoxichigh Glycolytichigh

H

Mean Min Distance (Pixels)

Gated on CD45negCD105neg segmented cells

0.5

0.0

0 T1

T2

T3

y g

G

Hypoxichigh Glycolytichigh

100um

4D and fig. S8, A and B). Annotation of T1 (Ly6G+CD101–dcTRAIL-R1–) and T2 (Ly6G+ CD101+dcTRAIL-R1–) neutrophils (fig. S8C) revealed that they were distributed throughout the tumor parenchyma in two different patterns reflective of their likely routes into the tumor: bordering the tumor edges in CD73– stromal regions or diffused throughout GLUT1–CD73+ regions, both of which contained a substantial Ng et al., Science 383, eadf6493 (2024)

ROI 1

C

Fig. 4. T3 neutrophils occupy a hypoxicglycolytic niche in pancreatic tumors. (A) Workflow diagram of multiplexed imaging using MICS technology. Cryosections of halved PDAC tumors were placed on slides, fixed, and permeabilized before ROI selection with DAPI staining. Sections were stained for 10 min per cycle containing antibodies in FITC, PE, and APC. After scanning, sections were photobleached and scanned to subtract background signals. After imaging the desired cycles, images were registered, segmented, and exported for conventional flow cytometric annotation of cell types, which were further used for spatial statistical analysis with SPIAT. (B) Tumor pictograph showing ROI1 selection area. (C) Immunofluorescent image of ROI1 with indicated stain markers of respective tumor regions. Scale bar, 100 mm. (D) Left, Expression marker intensity map of CD73 and GLUT1. Right, Coexpression plot of CD73 and GLUT1 marker intensities. (E) Spatial mapping of each annotated tumor neutrophil subsets in ROI1. (F) Comapping of neutrophil subsets in ROI1. (G) Left, Gating strategy of CD45– CD105– tumor regions demarcated by CD73 and GLUT1. Right, Mapped gated regions on segmented data of ROI1. (H) Spatial statistical analysis of segmented tumor neutrophils (n = 5 biological replicates). Left, Average minimum distance of each segmented neutrophil subset to hypoxichighglycolytichigh regions. Center line of boxplots show median, box hinges represent 25th and 75th percentiles, and whiskers extend to minimum and maximum values. *P < 0.05, **P < 0.01 by Mann-Whitney U test. Right, Colocalization of tumor neutrophils with hypoxichighglycolytichigh regions measured using a normalized mixing score for each radius of interaction. Dots represent mean across five replicates. Error bars indicate SEM. ***P < 0.001 by MannWhitney test, corrected for multiple comparisons with a false discovery rate (FDR) of 1% using the two-stage step up (Benjamini, Krieger, and Yekutieli) method.

50 100 150 200 250 300 350 400 450 500

Radius of interaction

Next, we agnostically quantified the three neutrophil subgroups’ relative locations within the tumor. Using marker intensities on segmented cells, we excluded CD45+ immune cells and CD105+ endothelial cells and subdivided the remaining cells into three distinct regions: the stromal niche (CD73–GLUT1–), the tumor parenchyma niche (CD73+GLUT1–), and the hypoxichighglycolytichigh tumor niche 7 of 16

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

,

8 of 16

y

12 January 2024

To better understand the temporal regulation of neutrophil differentiation into the T3 state within the tumor microenvironment in vivo, we administered 5-bromo-2′-deoxyuridine (BrdU) intravenously at selected time points before harvest, pulse-labeling all proliferating neutrophil precursors (Fig. 5G). This approach defines a time stamp at the point of labeling, allowing us to evaluate all time points simultaneously at the time of harvest, thereby minimizing batch effects. Modeling the disappearance rate from the peak of BrdU+ signal (fig. S12, A and B) and incorporating the entire temporal interval measured allowed the prediction of neutrophil half-life and lifespans (fig. S12C), reflecting their dwell time within each tissue compartment (Fig. 5G). The earliest peak of BrdU+ neutrophils was observed in the BM and spleen, because of active granulopoiesis within these organs in the tumor-bearing state, whereas recruitment of BrdU-labeled neutrophils into the blood and subsequently the tumor placed their peak of recruitment at ~4 to 5 days after labeling (Fig. 5G). Because the dwell times of neutrophils in the BM and spleen are complicated by continual neutrophil production, we focused our comparison between the blood and tumor neutrophils. Blood neutrophils had a half-life of 31.4 hours, reflecting an increased transit time in circulation compared with previously predicted times for WT blood neutrophils (49). Even so, tumor neutrophils had an even longer half-life [41.8 hours; 95% confidence interval (CI) = 39.6 to 46.6] and a predicted life span of 135 hours (up to 5.625 days; 95% CI = 126.5 to 142.4) (Fig. 5G and fig. S12, C and D), representing a marked extension of neutrophil life span within the tumor microenvironment. Compared with the unmarked BrdU– fraction, the proportion of BrdU+ dcTRAILR1+ neutrophils increased over time, surpassing baseline levels at day 6 after labeling (fig. S12, E and F). Newly recruited BrdU+ tumor neutrophils already expressed dcTRAIL-R1 at 1 day after labeling (fig. S12E), suggesting that T3 reprogramming is initiated upon tumor entry. dcTRAIL-R1 expression steadily increased and was the highest in neutrophils at 15 days after labeling (Fig. 5, H and I), confirming that acquisition of the T3 phenotype was associated with their extended lifespans in vivo. Our results show that a notable increase in the duration of neutrophil half-life and residence in the tumor compared with those in nontumor tissues (fig. S12D), indicating that neutrophils can survive long enough within the tumor to undergo reprogramming and sustained persistence

y g

Ng et al., Science 383, eadf6493 (2024)

T3 neutrophils are long-lived, terminal effectors within the tumor microenvironment

y

Because both mature and immature neutrophils were predicted to give rise to T3 neutrophils inside the tumor, we examined the impact of the maturation state on the acquisition of the T3 signature. We first isolated immature and mature neutrophils from WT mice and cultured them with tumor conditioned medium (TCM), which allowed us to mimic the infiltration of immature and mature neutrophils into the tumor microenvironment (Fig. 5A). We then measured up-regulation of dcTRAIL-R1 on cultured neutrophils across multiple time points as a proxy to estimate the acquisition of the T3 profile (Fig. 5A). After 3 days, a substantial proportion of BM immature and mature neutrophils cultured in TCM were dcTRAIL-R1+, but not those cultured in control media [i.e., complete Dulbecco’s modified Eagle’s medium (cDMEM)] (Fig. 5A). Specifically, exposure to TCM could trigger dcTRAIL-R1 expression 24 hours into culture and prolonged survival up to 3 days compared with culture in cDMEM

within neutrophils that enables them to acquire the T3 phenotype independent of their maturation stage.

g

Deterministic reprogramming of neutrophils in the tumor

(fig. S10, A and B). We obtained similar results regardless of tissue origin or maturation state, such that neutrophils isolated from the BM, spleen, and circulation cultured in TCM all exhibited dcTRAIL-R1 up-regulation and concurrent increased survival (Fig. 5B and fig. S10, A and B). Furthermore, exposure to TCM induced the up-regulation of the T3 gene signature in immature and mature neutrophils (Fig. 5C). We then assessed whether culture in hypoxic conditions would enhance neutrophil survival or dcTRAIL-R1 up-regulation of WT neutrophils. Culture in hypoxic conditions slightly augmented neutrophil survival in vitro, especially for mature neutrophils (fig. S10C), but did not induce dcTRAIL-R1 expression (fig. S10D), indicating that hypoxic conditions are not the main driver of T3 reprogramming. Adoptive transfer of CD45.1+ immature and mature neutrophils into tumor-bearing mice (Fig. 5D) further confirmed that exposure to the tumor microenvironment was necessary for the up-regulation of dcTRAIL-R1 (Fig. 5E), and dcTRAIL-R1 up-regulation in vivo followed the same kinetics as in vitro (Fig. 5F and fig. S11A). Therefore, exposure to tumor-derived stimuli supports extended survival and the acquisition of the T3 phenotype in WT neutrophils. Both TCM-cultured immature and mature neutrophils up-regulated dcTRAIL-R1 expression and the T3 gene signature to an equal degree, indicating that immature neutrophils could be reprogrammed directly into T3 neutrophils (Fig. 5, A and C). RNA velocity analysis suggested that this occurs in a stepwise fashion as immature neutrophils transit into T1 neutrophils and subsequently differentiate into T3 neutrophils (Fig. 1E). As expected, transfer of CD45.1+ immature neutrophils resulted in the appearance of dcTRAIL-R1–CD101– T1 neutrophils within the tumor 1 day after transfer (fig. S11B). By contrast, only dcTRAIL-R1–CD101+ T2 neutrophils, not T1 neutrophils, were observed in the tumor with CD45.1+ mature neutrophil adoptive transfer (fig. S11B). Both T1 and T2 populations derived from transferred cells were undetected after 3 days within the tumor because most of these transferred cells had completed their reprogramming into dcTRAILR1+ T3 neutrophils (fig. S11C). We further corroborated the transitory nature of T1 and T2 subsets by in vitro culture (fig. S11, D and E). Isolated T1 and T2 neutrophils up-regulated dcTRAIL-R1 expression after 24 hours in culture. Unlike tumor-naïve neutrophils, dcTRAILR1 up-regulation was independent of the culture medium used for both T1 (fig. S11F) and T2 (fig. S11G) neutrophils. Therefore, T1 and T2 neutrophils reflect immature and mature neutrophil populations captured in the process of T3 differentiation and thus do not require any further input from tumor soluble factors. Our findings suggest a deterministic program

p

(CD73hiGLUT1hi) (Fig. 4G and fig. S8C). In all five ROIs, T3 neutrophils had the shortest average minimum distance to the hypoxichighglycolytichigh niche compared with T1 and T2 neutrophils (Fig. 4H). We then assessed normalized mixing scores, which quantify the proportion of direct touches between the target regions and reference neutrophil cells (target-reference) against reference-reference interactions (48). T3 neutrophils had the greatest mixing scores within hypoxichighglycolytichigh tumor regions, which was conserved with each radius of area investigated (Fig. 4H). Scores for the stromal region showed a trend for highest scores in T1, followed by T2, with T3 being minimally detected within the stroma (fig. S8D). Finally, T2 neutrophils trended toward higher proportions of interspersion within the tumor parenchyma compared with T1 and T3 neutrophils (fig. S8D). These observations suggest that T1 and T2 neutrophils are mostly located away from the hypoxichighglycolytichigh regions, in contrast to T3 neutrophils. Therefore, our spatial data imply that T3 neutrophils likely migrate into and occupy the specialized hypoxicglycolytic tumor niche. This is consistent with our finding that transcription factor regulons enriched in T3 neutrophils (Fig. 1H) include those up-regulated in response to hypoxia (e.g., Hif1a and Bhlhe40) and to metabolic and ER stress (e.g., Atf4, Ddit3, Atf3, and Nfe2l2). Thus, epigenetic and transcriptional up-regulation of these pathways in T3 neutrophils could confer an added survival advantage within these tumor microenvironments. Collectively, our data provide a spatial representation by which tumor neutrophils converge upon the occupancy of a specialized tumor niche upon reprogramming.

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

105

104

104

103

103

0

0

-10

-10

3

3

104

105

105

105

104

104

103

0

104

105

0

104

105

103

0

0

-10

-10

3

3

104

0

105

105

105

+ 104

104

+

103

103

0

0

-103

3

-10

0 103

104

105

Ly6G

104

105

-103 -103 0 103

104

105

10

10

104

104

103

103

0

0

-103

3

0 10

3

4

10

10

5

Fold vs baseline (dcTRAIL-R1 gMFI)

Ly6G

Log normalized Counts

0

0

10-3

10-3

10

Ng et al., Science 383, eadf6493 (2024)

% of total neutrophils

103

103

12 January 2024

-3

0

10

3

4

10

10

5

10-3

0

103

104

105

9 of 16

,

104

5

y

105

104

4

y g

106

105

3

y

3

g

%BrdU+ (normalized to max)

-10

106

p

0 103

5

* *

5

-103 -103

* *

% dcTRAIL−R1+

Ly6G

0

Log normalized Counts

105

% dcTRAIL-R1+ of live cells

Sorted Fig. 5. Reprogramming within the dcTRAIL-R1 expression A immature/ B C T3 gene signature Culture in cDMEM Flow vs TCM (tumor cytometry BM Immature BM Mature tumor environment results in mature *** *** conditioned media) at D1, D3 cDMEM neutrophils TCM long-lived, terminally differentiated 80 *** *** cDMEM TCM 60 T3 neutrophils. (A) Experimental D3 7.5 40 *** setup of in vitro culture of sorted *** Immature 20 (CD101-) neutrophils from WT mice in cDMEM 0 versus TCM. Representative flow 5.0 77.0% 6.45% Blood Mature Spleen Mature cytometry plots show dcTRAIL-R1 80 D3 * *** expression increases on sorted Mature 60 2.5 + (CD101 ) immature and mature WT neutrophils 40 * from the BM after 3 days of culture in 20 75.1% D0 D3 D0 D3 0 TCM but not in cDMEM. (B) NeutroD1 D3 D1 D3 WT BM IMM WT BM MAT dcTRAIL-R1 phils cultured in TCM up-regulate dcTRAIL-R1 over time. Line plots show the percentage of dcTRAIL-R1+ cells Blood Tumor Immature Mature CD45.1+ CD45.1+ D Sort CD45.1+ Immature E F transferred transferred D3 Immature transferred Mature gated as in (B), dots indicate the BM neutrophils transferred transferred Tumor Gated on median, and error bars indicate Q1 and Spleen live, Ly6G 79.6% 6.06% Blood 100 100 CD45.1 Q3 intervals for neutrophil subsets neutrophils 80 80 Inject i.v. into cultured in cDMEM (dotted line) and 60 60 PDAC tumor D3 Mature transferred TCM (solid line) over 1 and 3 days. bearing mice 40 40 Each group contains the following 20 20 80.4% number of samples: day 1 cDMEM: BM 0 0 Flow cytometry Blood Spleen Tumor D1 D3 D1 D3 at D1, D3 immature (n = 8), BM mature (n = 8), dcTRAIL-R1 Days post transfer spleen mature (n = 8), blood mature (n = 3); day 1 TCM: BM immature Harvest all dcTRAIL-R1 gMFI on live, Ly6G+ Orthotopic BrdU label G H Gated I (n = 8), BM mature (n = 8), spleen timepoints tumor neutrophils PDAC injection for timepoints + Fold vs = BrdU gMFI + mature (n = 8), blood mature (n = 5); BrdU ( BrdU- gMFI ( 0 6 weeks baseline D15 BrdUday 3 cDMEM: BM immature (n = 10), D15 D12 D8 D6 D4 D3 D2 D1 BM mature (n = 10), spleen mature 2.5 * BM Spleen D12 (n = 10), and blood mature (n = 4); 100 100 * D8 day 3 TCM: BM immature (n = 10), BM 2.0 50 50 mature (n = 10), spleen mature (n = 10), ** D6 blood mature (n = 5) performed 0 0 0 5 10 15 0 5 10 15 1.5 across seven independent experiments. D4 Blood Tumor * *P < 0.05, **P < 0.01, ***P < 0.001 by 100 t½= 41.8h t½= 31.4h 100 D3 Mann-Whitney U test. (C) Neutrophils t5%= 91.5h t5%= 135.0h 1.0 50 50 cultured in TCM up-regulate the D2 T3 gene signature. Scatter dot plots 0 0 0 5 10 15 0 5 10 15 -10 0 10 10 10 2 3 4 6 8 12 15 for T3 gene signature expression in Days post BrdU labelling dcTRAIL-R1 Days post BrdU labelling BM immature (WT BM IMM) and mature (WT BM MAT) neutrophils that Sorted T3 neutrophils were freshly sorted (D0) or cultured dcTRAIL-R1 expression T3 gene signature J K L n.s. for 3 days (D3) (n = 3 for all samples). D0 D1, D3 10 n.s. n.s. n.s. n.s. n.s. 100 culture Each dot denotes a single gene, lines 8 TCM Sorted cDMEM denote the mean, and error bars 75 indicate the SEM. ***P < 0.001 by 6 Loss of T3 phenotype? Wilcoxon signed-rank test with 50 4 Bonferonni’s correction, with compar25 isons indicated on the graph. 2 (D) Experimental setup for transfer of 0 0 CD45.1+ neutrophils into pancreatic 91.7% 94.3% d d M M EM TCM MEM TCM rte rte M ME TC o tumor-bearing mice. Sorted WT S So cD cD cD dcTRAIL-R1 D0 D1 D1 D0 D3 BM immature and mature neutrophils were intravenously injected into WT (G) Experimental setup for BrdU pulse labeling in tumor-bearing mice. WT mice PDAC tumor-bearing mice. At 1 and 3 days after transfer, CD45.1+ neutrophils received orthotopic injection of the PDAC cells, and the tumor was allowed were evaluated within the blood, spleen, and tumor for dcTRAIL-R1 expression by to grow. At days 15 (n = 4), 12 (n = 4), 8 (n = 5), 6 (n = 3), 4 (n = 4), 3 (n = 4), flow cytometry. (E) Up-regulation of dcTRAIL-R1 expression was restricted 2 (n = 3), and 1 (n = 4) before the harvest, mice were injected with BrdU, thus to the tumor. Representative flow plots show dcTRAIL-R1 expression present on labeling proliferating neutrophil precursors within the BM and spleen. Data were transferred CD45.1+ immature and mature neutrophils present in the blood or collected across three independent experiments. At day 42 (6 weeks) after tumor at 3 days after transfer. (F) Line plots show proportion of WT BM CD45.1+ injection, mice were sacrificed and BrdU+ neutrophils within the BM, spleen, blood, immature (n = 4 at day 1 and n = 3 at day 3) or mature neutrophils (n = 3 and tumor were quantified. BrdU percentages at all time points were then normalized at both time points) expressing dcTRAIL-R1 performed across two independent to the maximal BrdU+ percentage value for each tissue, which was set to experiments. Each dot denotes a single gene, lines denote the mean, and error 100%. Dots represent mean expression, with error bars denoting 95% CIs. bars indicate SEM. *P < 0.05 by Kruskal-Wallis test with Dunn’s post test.

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

A mathematical model capturing the full temporal window was fitted to estimate half-life (t1/2) and life span (t5%) for each organ, denoted on each plot in hours (see also fig. S12C and the materials and methods). (H) BrdU labeled neutrophils up-regulate dcTRAIL-R1 over time. Histograms show geometric MFI (gMFI) of dcTRAIL-R1 within BrdU+ (orange) and BrdU– (gray) neutrophils at days 2, 3, 4, 6, 8, 12, and 15 after BrdU labeling. (I) Quantification of gMFI in (H). Line plots show fold change of dcTRAIL-R1 gMFI of BrdU+ against BrdU– neutrophils. BrdU– neutrophils served as a measure of baseline dcTRAIL-R1 gMFI within the tumor. Each dot denotes the mean, with error bars indicating SEM. *P < 0.05, **P < 0.01 by Mann-Whitney U test, one-tailed, alternative = “greater.” (J) Experimental setup of in vitro culture of sorted T3 neutrophils from PDAC mice in cDMEM or TCM at 1 or 3 days. Representative flow cytometry plots show

10 of 16

,

Phenotypic differences in tumor-infiltrating neutrophils have been observed across mouse and human cancers, and multiple subsets have been characterized by differential surface marker or transcriptome expression (6–13). Whether such neutrophil profiles are conserved across tumor types and species remains unclear. We therefore assessed whether T1 to T3 neutrophil states can be detected in previously published mouse cancer scRNAseq datasets containing annotated neutrophils (11, 13). We projected these datasets onto a reference UMAP embedding, mapping each cell on the same UMAP space as our dataset (fig. S14A). As in the PDAC model, neutrophils within Lewis lung carcinoma tumors (13) could be assigned to the T1 to T3 states, whereas spleens from WT or tumorbearing mice mostly contained nontumor neutrophil clusters (fig. S14B). An intermediate neutrophil subset identified in cancer-bearing individuals [PMN2 in (13)] was enriched in the preNeu and IMM 1 and 2 clusters (fig. S14C), whereas T1 to T3 clusters scored highly for the tumor-specific neutrophils (PMN3; fig. S14D). Similarly, reanalysis of a different scRNAseq dataset showed that neutrophils from their KP1.9 tumor-bearing lung model (11) mapped as T1 to T3 clusters, whereas the normal lung

y

12 January 2024

Evidence for conserved neutrophil reprogramming across tumor types and human PDAC

y g

Ng et al., Science 383, eadf6493 (2024)

y

Prior studies have identified key roles for neutrophils in tumor progression, especially through the promotion of tumor angiogenesis (50–52). We therefore sought to determine the functional specialization of T1 to T3 neutrophils in promoting tumor growth. Given their phenotypic stability, distribution near angiogenic regions, and the expression of a pro-angiogenic transcriptional signature (Fig. 3C), we speculated that T3 neutrophils promoted angiogenesis from their hypoxic-glycolytic niche to support continual tumor growth. By contrast, we anticipated that the transient T1 and T2 neutrophil states would not yet feature proangiogenic ability. T3 neutrophils had the highest transcript (Fig. 6A) and protein expression (Fig. 6, B and C) of Vegfa (vascular endothelial growth factor a) compared with the other neutrophil subsets in the tumor and periphery. We then evaluated whether T3 neutrophils had a greater capacity to induce blood vessel formation in vivo using a modified Matrigel plug assay and measured angiogenesis by the rate of vascular flow measured by Doppler imaging (53) (fig. S13A). Matrigel plugs co-injected with T3 neutrophils showed the greatest flux intensity compared with control WT BM

trols showed greater distribution along the tumor edges (Fig. 6H and movie S2). Normalization for tumor size revealed that both T3 and WTBM MAT–co-injected tumors had similar densities of CD31+ blood vessels (fig. S13F) but substantially differed in their distribution (Fig. 6I). Although T3 neutrophils were the highest expressors of Vegfa in the pancreatic tumor, Vegfa expression was also detected in macrophage populations (fig. S13G), indicating that neutrophils may not be the sole proangiogenic contributor for tumor growth support. Nonetheless, blockade of T3 neutrophils with an anti–dcTRAIL-R1 antibody decreased tumor growth in T3-co-injected tumors compared with isotype controls (Fig. 6J), indicating that T3 neutrophils are the predominant cells responsible for increasing the tumor growth rate, most likely through vascular remodeling.

g

T3 neutrophils are pro-angiogenic and promote tumor growth

neutrophil plugs within the same mouse (fig. S13B), and this increase in vascularization, although modest, was consistent across all mice observed in the assay (fig. S13C). By contrast, co-injection of T1 and T2 neutrophils did not enhance vascular flow (fig. S13, B and C), indicating that the T3 population contains the most potent angiogenic ability (Fig. 3C). The tumor triggers angiogenesis to maintain the supply of oxygen and other nutrients to the core of the tumor to sustain continual growth (54). To test whether T3 neutrophils residing in their hypoxic-glycolytic niche at the tumor core support this angiogenic switch, we modified our PDAC model by implanting PDAC tumors subcutaneously to enable longitudinal tumor size measurements. We then assessed whether injection of neutrophils within the tumor affected their subsequent growth (Fig. 6D). PDAC cells co-injected with T3 neutrophils formed tumors that grew rapidly (Fig. 6D), whereas co-injection with other neutrophil types had little influence on tumor growth. T3co-injected tumors had a 100% tumor growth rate (Fig. 6E, as evaluated in fig. S13D) and had the greatest mass at the end point (Fig. 6F). T2co-injected tumors showed the lowest tumor incidence rates (5 of 12) (Fig. 6F), suggestive of increased tumor rejection and consistent with the expression of proinflammatory genes in T2 neutrophils detected in our scRNAseq dataset (Fig. 3B). T3-co-injected tumors continued to grow rapidly for up to 3 weeks after the last injection of neutrophils, suggesting that even after their disappearance, transferred T3 neutrophils generated long-lasting effects that sustained this accelerated growth rate. Neutralization of VEGFa in our model resulted in the reduction of growth in T3-co-injected tumors, but had no impact on the growth of WT BM mature (WTBM MAT)–co-injected tumors (Fig. 6G). To determine whether this growth enhancement was due to increased angiogenesis driven by T3 neutrophils, we optically cleared T3- and WTBM MAT–co-injected tumors (fig. S13E) and visualized the intratumoral threedimensional (3D) vessel network through CD31 staining (Fig. 6H). T3-co-injected tumors had greater CD31 staining density toward the tumor core (Fig. 6H and movie S1), whereas the vessel network in WTBM MAT–co-injected con-

p

within the tumor as long-lived dcTRAIL-R1+ T3 neutrophils. To determine whether the T3 state is transitory or is stably maintained once it has been acquired, we isolated and cultured T3 neutrophils overnight and up to 3 days in the presence or absence of TCM (Fig. 5J). Survival of T3 neutrophils in culture was not affected by the medium type (fig. S12G), and dcTRAIL-R1 expression was maintained despite the absence of tumor-derived factors (Fig. 5K), with expression levels comparable to freshly isolated T3 neutrophils at both time points (fig. S12H). Additionally, T3 neutrophils did not down-regulate their gene signature in culture (Fig. 5L). This in vitro evidence indicates that once neutrophils have been reprogrammed into T3 neutrophils, they do not revert their phenotype in the absence of supportive tumor factors and they represent the terminally differentiated neutrophil population within the tumor.

that dcTRAIL-R1 expression is retained on T3 neutrophils after 1 day of culture in both cDMEM and TCM. (K) Boxplots show quantification of frequency of dcTRAIL-R1+ neutrophils (n = 5 each performed across three independent experiments) in (I). Each dot represents one biological replicate, center line of boxplots show median, box hinges represent 25th and 75th percentiles, and whiskers extend to minimum and maximum values. P = n.s. (not significant) by Kruskal-Wallis followed by many-to-one U test comparing against the D0 time point. (L) T3 neutrophils cultured overnight do not down-regulate the T3 gene signature. Scatter dot plots for T3 gene signatures in sorted, cDMEM-, or TCM-cultured neutrophils (n = 3 each). Each dot denotes a single gene, lines denote the mean, and error bars indicate SEM. P = n.s. by Kruskal-Wallis followed by Dunn’s post test.

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

Volume (mm3)

UMAP2

% VEGF

+

0. 0 0. 00 0 0. 25 0 0. 50 07 5

Fig. 6. Protumoral T3 neutrophils * ** Exp. density A C 100 D Vegfa * promote tumor growth and associate Co-inject tumor 80 Tumor volume ± neus.(s.c.) with poorer patient outcomes. (A) T3 60 D0 Co-injected: PBS control T1 * 40 neutrophils have the highest expression WTBM IMM T2 D3 Boost 20 WTBM MAT T3 1000 neus. of Vegfa transcripts. UMAP projection 0 of total neutrophils in BM, spleen, blood, Spleen T1 T2 T3 750 D7 ** UMAP1 Tumor Measure and tumor show expression density of tumor Gated on live, Ly6G neutrophils 500 D14 volume B Vegfa. (B) T3 neutrophils have the Spleen T1 T2 T3 * 250 highest expression of VEGFa. Repre* D21 Endpoint sentative flow plots show intracellular volume/ 14 21 28 weight D28 8.12% 10.2% 66.4% 88.5% VEGFa protein staining in neutrophils Days post tumor injection from the spleen (n = 3) and T1, T2, and VEGF T3 neutrophils (n = 5 biological replicates) in the pancreatic tumor across Tumor incidence E G VEGF inhibition H 3D vessel quantification Tumor No tumor two independent experiments. (C) BoxCo-inject Boost Measure WTBM WTBM tumor + neus. tumor vol. T3 co-injected tumor PBS IMM MAT 87.5% T3/WTBM + ab. to endpoint plots quantify VEGFa expression as 0.63X 2X 70% 83.3% MAT neus. CD31 Core T1 T2 T3 ± ab. Left margin shown in (B). Each dot represents one 85.7% 58.3% 100% D0 D3 D14 D7 D21 biological replicate, center line of boxplots Co-injected antibody: Endpoint weight show median, box hinges represent F Isotype WTBM Isotype T3: 25th and 75th percentiles, and whiskers α -VEGFα α -VEGFα MAT: 2.5 * 1500 µM 1500 µM * ** extend to minimum and maximum 1000 2.0 values. *P < 0.05, **P < 0.01 by KruskalWTBM MAT co-injected tumor 750 1.5 Left margin Core CD31 n.s. Wallis followed by many-to-one U test 500 n.s. n.s. n.s. n.s. n.s. 1.0 comparing against T3 neutrophils. (D) T3 250 0.5 neutrophils promote rapid tumor growth. 0 Schematic of experimental setup to 0 7 14 7 14 PBS IMM MAT T1 T2 T3 1500 µM 1500 µM determine the ability of neutrophils to Days post last ab. injection WT BM Tumor promote tumor growth in vivo. Equal numbers of neutrophils and PDAC cells Core Left margin Right margin dcTRAIL-R1 inhibition J I were mixed in Matrigel before subCo-injected antibody: Co-inject tumor + T3 neus. Isotype α -dcTRAIL-R1 Centre Co-injected with cutaneous injection into the right flank. + isotype/ * T3 α-dcTRAIL-R1 * Neutrophils in the tumor was boosted 800 WTBM MAT D0 0.20 on day 3 (D3) and D7 by direct 600 D3 0.15 Boost * * subcutaneous injection into the visible T3 neus. 400 + antibody 0.10 matrigel plug–tumor. Tumor growth D7 was measured by calipers weekly until 200 0.05 D14 Measure the day 28 end point. Line plots show tumor 0.00 0 volume to 0 0 0 0 0 0 0 0 0 0 10 20 30 40 50 60 70 80 90 00 volume of measured subcutaneous 7 14 D21 10 9 8 7 6 5 4 3 2 1 1 endpoint Quantiles from centre Days post last ab. injection tumors co-injected with PBS (n = 8), WT BM immature (WTBM IMM, n = 10), Overall survival TCGA Pan-cancer dataset L K TCGA (PAAD) dataset Australia (PACA) dataset (Liu et al., 2018) WT BM mature (WTBM MAT, n = 12), and 1.0 T3 sig. T3 sig. (427 days) (652 days) Scored for Scored for Scored for T1 (n = 7), T2 (n = 12), and T3 (n = 7) 0.8 T3 sig. T3 sig. T3 signature T2 signature T1 signature (709 days) (1332 days) 0.6 neutrophils. Data were collected across 0.4 BRCA five independent experiments. Dots show SKCM 0.2 means, with error bars indicating SEM p.val = 0.045* COAD p.val = 0.026* 0.0 KIRC over all three measurement time points. Disease-free survival ACC *P < 0.05, **P < 0.01 by Kruskal-Wallis 1.0 SARC T3 sig. T3 sig. (293 days) T3 sig. 0.8 LUAD test followed by many-to-one U test T3 sig. (476 days) HNSC 0.6 comparing all other conditions to PBS PAAD 0.4 LGG co-injection. (E) Quantification of PDAC 0.2 ESCA subcutaneous tumor incidence after p.val = 0.022* p.val = 0.055 CESC 0.0 0 0 00 00 00 00 500 00 0 neutrophil co-injection (see fig. S13D for 1 2 3 1 2 3 1 2 3 00 500 000 5 10 15 20 2 30 50 2 1 10 Days Hazard ratio Days quantification method). Piecharts show frequency of mice with tumors (light gray) or no tumor growth (dark gray) at day 28 end point in (D). (F) Tumors n = 5). Data were collected across three independent experiments. Neutrophils were further injected on D3 and D7 with the corresponding antibody added, co-injected with T3 neutrophils had the biggest weights compared other conditions. Each dot represents one biological replicate, with boxplots showing and tumor growth was measured at D14 and D21, 7 and 14 days after tumor injection. Boxplots show median tumor volume, where each dot represents one median with IQR and whiskers indicating lowest and highest measurement. P < 0.05* by Kruskal-Wallis test followed by many-to-one U test comparing all biological replicate, box hinges represent the interquartile range, and whiskers extend to the minimum and maximum values. P < 0.05* by by Mann-Whitney other conditions with PBS co-injection. (G) Antibody neutralization of VEGFa reduces T3-mediated acceleration of tumor growth. Schematic of experimental U test. (H) Visualization of CD31 vessels within T3 and WT BM MAT co-injected setup with sorted T3 neutrophils are co-injected with PDAC tumor with antisubcutaneous tumors. Subcutaneous tumors from (A) for T3 (n = 3) and VEGFa targeting antibody (n = 10) or isotype control (n = 7). WT BM mature WTBM MAT (n = 3 biological replicates) were dissected, optically cleared, and neutrophils were used as controls (anti-VEGFa targeting antibody, n = 4, isotype, permeabilized, stained with anti-CD31 antibody, and imaged in 3D at 0.65× 106

106

106

105

105

105

105

104

104

104

104

103 0

103 0

103 0

-103 0 103

104

105

103 0

-103

-103

-103

-103 0 103

104

105

-103

-103 0 103

104

105

-103 0 103

104

105

Left margin

Volume (mm3)

Volume (mm3)

CD31+ volume (normalized to quantile vol.)

y

high

high

low

low

y

high

y g

high

low

low

11 of 16

,

Cumulative survival

g

12 January 2024

p

Ng et al., Science 383, eadf6493 (2024)

Right margin

Weight (g)

Left margin

Ly6G

+

106

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

(100% of total volume) and 2× (35% of total volume starting from midpoint) resolution. Data were collected across three independent experiments. Representative 3D immunofluorescence images show T3 (top, quarter) and WTBM MAT (bottom, whole) co-injected tumors, with tumor margins marked out in white dotted lines and indicated. (I) Subcutaneous PDAC tumors co-injected with T3 neutrophils have greater CD31 vessel density within the tumor core. Quantification of CD31 staining intensity at 2× resolution as in (E). CD31 staining was surfaced with a seedpoint of 16.2, and binned in 10% quantiles according to the distance from left or right margins toward the tumor core. To account for differences in tumor sizes, CD31 staining intensity was further normalized over total slice volume for each quantile. Line plots represent volume-normalized staining intensity, with dots representing the mean and errors bars indicating SEM for T3 (orange)– or WTBM MAT (gray)–co-injected tumors. *P < 0.05 by Mann-Whitney U test, one-tailed, alternative = “greater.” (J) Antibodymediated blockade of T3 neutrophils reduces their ability to promote rapid tumor growth. Schematic of experimental setup. Sorted T3 neutrophils were co-injected with PDAC tumor with anti–dcTRAIL-R1 targeting antibody (n = 8) or isotype control (n = 8). T3 neutrophils were further injected on D3 and D7 with the corresponding antibody added, and tumor growth was measured at D14

y

Discussion

,

12 of 16

y

Environmental cues can fine-tune immune responses by inducing cell recruitment and expansion of immune cells, generating productive responses with adequate numbers and specialized phenotypes. Unable to further proliferate, neutrophils rely on their ability to swiftly mobilize into tissues to perform their functions effectively, and in diseases such as cancer, neutrophils at various maturation stages, tissue origins, and phenotypes are recruited into the tumor in large numbers (12–14, 17, 19). Given the likely coexistence of multiple tumor neutrophil states with different functional phenotypes, it is thus unclear how neutrophils exert a focused local effect in the tumor. Here, we examined how neutrophils decouple their initial maturation phenotype from their eventual protumoral function by undergoing convergent reprograming within the tumor. Our study demonstrates that tumor-infiltrating immature and mature neutrophils acquire distinct epigenetic, transcriptomic, and proteomic phenotypes while retaining features of their initial maturation status. We found that both immature (T1) and mature (T2) tumor neutrophils within the tumor converged toward a third population (T3) expressing dcTRAIl-R1. T3 neutrophils thus show intermediate maturation scores when quantified at the population level, and have

y g

12 January 2024

cancers, including pancreatic cancer [from the The Cancer Genomics Atlas Pancreatic Adenocarcinoma (TCGA-PAAD); see Fig. 6L and table S7 for P values and confidence intervals]. By contrast, T1 and T2 signatures were both protective against (lower hazard ratios, HRs) or associated with (higher HRs) patient death depending on the type of solid tumor (Fig. 6L and table S7), which is consistent with their nature as transitional subsets in the process of reprogramming. These data support a model in which tumoreducated T3 neutrophils drive tumor progression in both mouse and human cancers.

g

Ng et al., Science 383, eadf6493 (2024)

tissue (fig. S15B). Our findings suggest that tumor-induced reprogramming of neutrophils is conserved in humans. Because T3 neutrophils promote the growth of pancreatic tumors, we hypothesized that the genetic signature associated with the T3 state might be predictive of pancreatic cancer outcomes in human patients. To test this, we performed survival analysis on two independent pancreatic cancer cohorts from The Cancer Genome Atlas (TCGA) and International Cancer Genome Consortium Pancreatic Cancer-Australia (PACA-AU) by scoring patients on the basis of high and low expression of each signature (see the materials and methods for scoring criteria). Patients with high expression of the T3 neutrophil signature had poorer overall survival (OS) across both datasets, with a median of 652 (TCGA) and 427 (PACA-AU) days, respectively (Fig. 6K), and this was independent of potential confounders such as patient gender, age, and tumor stage (table S5). Similarly, when disease-free survival (DFS, assessed as time to an adverse event from the initial treatment) was evaluated, patients with high T3 neutrophil signature expression had reduced DFS (Fig. 6M), with equal distribution of potential confounders across the two groups (table S6). When T1 and T2 signatures were considered, high T2 signature expression correlated with worse OS only in the TCGA dataset (fig. S15, C and D), but this was confounded by gender (table S5), whereas high T1 signature expression was associated with lower DFS only in the PACA-AU dataset (fig. S15, E and F). Therefore, only the T3 signature was consistently associated with poorer OS and DFS across both datasets. We next considered whether the T3 neutrophil signature was also associated with poorer OS in other solid tumors. Within the TCGA pancancer database (57), higher expression of the T3 signature was associated with a significantly higher risk of death across a subset of solid

p

tissue was mostly enriched for mature neutrophils (fig. S14E). Likewise, neutrophil types identified as tumor-specific neutrophil clusters in this model [mN3-6 in (11)] corresponded to T1 to T3 clusters (fig. S14F), confirming that the tumor microenvironment induces a prototypical transcriptional trajectory that is deterministic in nature. By contrast, the signature of neutrophils from healthy lungs (49) was largely independent of tumor-induced transcriptional changes (fig. S14G). These data suggest that reprogramming of tumor neutrophils is conserved across different tumor types, and that T3 neutrophils represent terminal differentiated tumor neutrophils. To examine cross-species conservation of the neutrophil tumoral trajectory found in mice, we investigated whether our neutrophil classification could account for the heterogeneity present in human tumors. We examined independent scRNAseq datasets (55, 56) from two human PDAC cohorts (fig. S14, H and I) and mapped them to a simplified reference UMAP embedding through label transfer (fig. S14J). T2 and T3 neutrophils were amply labeled in both datasets, and a smaller cluster of T1 neutrophils was also identified (fig. S14, K and L). These tumor neutrophil subsets were predominantly enriched in the pancreatic tumor and not in adjacent normal pancreatic tissue (fig. S14K) or in the peripheral blood (fig. S14L). A pro-angiogenic neutrophil type expressing genes linked to hypoxia and glycolysis in one of these studies [referred to as TAN-1 (56)] strongly matched the T3 state in the murine tumor, indicative that our T3 classification can capture subsets independently identified to be protumoral (fig. S15A). Similarly, tumor neutrophil gene signatures derived from our mouse dataset (fig. S15B) were conserved in their ability to distinguish between human tumor neutrophils and showed strong specificity in identifying T1, T2, and T3 neutrophils within the tumor and not in the adjacent normal

and D21, 7 and 14 days after tumor injection. Data were collected across four independent experiments. Boxplots show median tumor volume, with each point representing one biological replicate, box hinges represent the interquartile range with whiskers extending to minimum and maximum values. *P < 0.05 by Mann-Whitney U test. (K) The T3 neutrophil gene signature is associated with poorer patient OS and DFS in pancreatic cancer. Kaplan-Meier plots show OS (top) and DFS (bottom) for patients in the TCGA-PAAD and PACA-AU datasets. Patients were split into high and low expression of the T3 curated signature. Median OS and DFS survival are represented on the graph when available. Events are represented by vertical lines and were defined from days to death (from initial pathological diagnosis) or days to first event (from initial treatment, for TCGA, and from clinical disease-free diagnosis for PACA-AU). *P < 0.05 as calculated by log-rank test. (L) The T3 neutrophil gene signature is associated with poorer patient OS in a subset of solid human cancers. Each data set within the curated TCGA Pan Cancer dataset was scored for T1, T2, and T3 signatures (see also table S4). Forest plots show hazard ratio (HR) scores associated with patient OS that were significant (P < 0.05) by Cox proportional hazards test across all signatures. Dots indicate the calculated HR and whiskers indicate 95% CIs.

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

y

Methods summary scRNAseq

,

13 of 16

y

Sorted total neutrophils (Lin–CD45+CD115– Ly6ClowSiglec-F–Gr1+CD11b+Ly6G+) from the BM, spleen, blood, and tumor of mice bearing orthotopic pancreatic tumors [generated as described in (66)] were incubated with 0.5 mg of Totalseq-A anti-mouse Hashtag antibodies (BioLegend) per 100,000 cells for 30 min at 4°C. Cells were washed with fluorescence-activated cell sorting (FACS) buffer, spun down, and resuspended in PBS with 1% bovine serum albumin. Cells were pooled accordingly for 10X Genomics 3’ (v3) sequencing on NovaSeq (Illumina) following the manufacturer’s protocol. Sequencing reads were evaluated by FastQC and MultiQC. High-quality reads were aligned to the GRCm38 mm10-2020-A genome assembly and quantified using CellRanger (version 2.2.0, 10X Genomics). The gene expression matrix was analyzed in R using the Seurat package (4.0.5) (68). Hashtags were demultiplexed using CITE-seq-Count. Doublets and multiplets were filtered out, as well as unique molecular identifiers (UMIs) with two or more hashtags associated with them. UMIs with excess mitochondrial reads (>5%), number of features 4200 (outliers) were also removed. Normalization, scaling, and clustering were performed with the default Seurat pipeline. The destiny package (3.4.0.) was used for diffusion mapping (24). Velocity analysis was performed with scVelo (2.1.0) (26) using the default stochastic model

y g

12 January 2024

states coexist within the tumor-bearing mouse, the neutrophil maturation trajectory remains unchanged, consistent with findings in other proinflammatory settings (19, 20, 64, 65). Therefore, whereas our findings do not fully exclude the possibility of neutrophil reprogramming outside of the tumor, they do suggest that it is unlikely that upstream changes in neutrophil progenitors specifically drive the formation of a protumoral neutrophil population. Instead, it is the intrinsic capacity of recruited neutrophils to adapt in response to the tumor environment regardless of their initial phenotype that allows them to adopt convergent trajectories to settle upon a final, protumoral state. This feed-forward loop thus ensures the continued supply and differentiation of long-lived, pro-angiogenic neutrophils that support tumor growth. Expanding from the results of our study, we propose that it is advantageous for tissues to induce functional homogeneity in neutrophils at the local scale to support tissue growth and function. This process is then hijacked by the tumor to favor a functional neutrophil state that promotes aberrant tumor growth. Our findings thus suggest a general mechanism by which short-lived effector cells such as neutrophils efficiently adjust their functions to meet the demands of a tissue, and that local neutrophil responses can be therapeutically targeted.

g

Ng et al., Science 383, eadf6493 (2024)

tive stress are switched on in T3 neutrophils compared with T1 or T2 populations. Deletion or inhibiting these transcription factors to trace the timeline of T3 reprogramming thus represent important goals for future studies within the field. T3 neutrophils accelerated early tumor growth experimentally, and although we still observed tumor growth with other co-injected neutrophil subsets, this occurred at a slower rate, likely due to recruitment and reprogramming of endogenous T3 neutrophils at the later time points. T1 and T2 neutrophils did not show statistically significant enhancement or inhibition of tumor growth compared with other control neutrophil populations or phosphate-buffered saline (PBS), reflecting their transitional nature as opposed to a fully anti- or protumoral population. By contrast, the sustained growth advantage conferred by T3 neutrophils stemmed, at least in part, from T3-dependent remodeling of the vasculature toward the core of the tumor. Given their localization within the glycolytic-hypoxic niche, an intriguing scenario is that T3 neutrophils serve as possible guide rails to direct angiogenesis to relieve hypoxic and nutrient stress in areas that would most require it. Consistent with this, resistance toward anti-angiogenic therapies in human cancer has been associated with neutrophil infiltration (61), and neutrophil depletion reduces tumor vascularization and growth (51, 52, 62, 63). Studies facilitating further understanding of T3-mediated vascular remodeling would reveal new therapeutic targets against pathological angiogenesis within the tumor. A particularly interesting finding is the conservation of this differentiation program in tumor-infiltrating neutrophils across tumor type and species. When scRNAseq datasets from other tumor models were examined, T3 annotation could identify protumoral clusters found both in mouse [i.e., mN5 (11)] and in human [i.e., TAN-1 (56)]. T3 neutrophils promoted PDAC tumor growth in mice, whereas ablation of T3 neutrophils or their proangiogenic function removed this growth advantage. In parallel, the T3 signature consistently predicted poorer patient outcomes in two independent human PDAC cohorts, as well as in a subset of other solid tumors. Thus, different studies converge upon our identification of a terminally differentiated neutrophil state, the signature of which might be used to better understand neutrophil function in cancer and possibly to predict tumor progression. We propose that the reported heterogeneity of neutrophils across tumors more likely reflects transitional states derived from populations at different stages of maturation and/or reprogramming. Collectively, by ordering neutrophils through the lens of their ontogeny, we have assessed global neutrophil phenotypic heterogeneity through maturation stages, extramedullary sources, and the specialization of each neutrophil to each tissue (49). Although all of these

p

both toroidal (immature) and hypersegmented (mature) neutrophil nuclei, representative of their being an admixture of reprogrammed neutrophils of T1 and T2 origin. Our data reveal the capacity of neutrophils to simultaneously integrate different types of signals from the environment, allowing them to layer a new functional phenotype onto their pre-existing differentiation stage. We demonstrated this plasticity with immature and mature neutrophils from tumor-naïve mice, which are both capable of acquiring phenotypic (dcTRAIL-R1 expression) and transcriptional traits of the T3 state when cultured in tumor-conditioned medium or upon entering the tumor in vivo. By circumventing neutrophil maturation as a ratelimiting step, this adaptability allows immature neutrophils to be equally mobilized and reprogrammed within the tumor in a shorter time frame. Subsequently, this intrinsic ability embedded in neutrophils consolidates the various functional neutrophil states into one terminal neutrophil phenotype as directed by the tissue, in this case, a tumor. Circulating neutrophils have a predicted halflife of ~10 hours (58), whereas tissue-resident neutrophils persist for up to 1 day (49). Although studies have hinted that neutrophils persist far longer within the tumor (10, 59, 60), whether this coincides with terminal differentiation in the tumor remains an open question. Using a BrdU pulse-labeling approach, we found that up to 5% of the originally labeled neutrophils could remain within the tumors for as long as 5.625 days upon entry. In addition, whereas our curve fit was accurate for the first labeling time points, there was a noticeable underfitting at 12 and 15 days after labeling, representing a possible underestimation of the full tumor neutrophil life span and indicating that a small population of neutrophils could persist even longer within the tumor. Long-lived neutrophils were predominantly of the dcTRAIL-R1hi T3 phenotype, which indicates a correlation between their ability to survive within challenging hypoxic and glycolytic environments and their continued persistence within the tumor. Spatial mapping at the transcriptome and protein level placed T3 neutrophils predominantly within a hypoxic-glycolytic niche nearer to the tumor core. By contrast, T1 and T2 neutrophils were positioned at the stromal and tumor parenchyma, where a large vessel network exists, supporting the notion that these cell states are in the process of reprogramming after tumor entry. Given that hypoxia is not required to trigger T3 reprogramming, which can occur in normoxic conditions, we propose that migration toward the hypoxicglycolytic niche occurs after acquisition of T3 epigenetic and transcriptional programs is fully complete to ensure neutrophil survival. Consistent with this possibility, upstream transcription factors regulating metabolic and oxida-

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

and velocity vectors were projected onto the diffusion map embedding. Regulatory network analysis was carried out with PySCENIC (0.10.4) (27) in Python and exported for visualization in R. High-parameter flow cytometry staining and analysis

A total of 1 × 105 tumor cells were resuspended with or without with 1 × 105 neutrophils (specified in text) in 1× PBS, mixed in a 1:3 ratio of Matrigel, and injected subcutaneously into the right flanks of mice using a 30-gauge insulin needle. At 3 and 7 days after the initial injection, 1 × 105 neutrophils were resuspended in 30 mL of PBS and injected directly into the observable Matrigel-tumor plug. If antibodies were co-injected alongside neutrophils, 10 mg

12 January 2024

REFERENCES AND NOTES

1. L. G. Ng, R. Ostuni, A. Hidalgo, Heterogeneity of neutrophils. Nat. Rev. Immunol. 19, 255–265 (2019). doi: 10.1038/ s41577-019-0141-8; pmid: 30816340 2. A. J. Gentles et al., The prognostic landscape of genes and infiltrating immune cells across human cancers. Nat. Med. 21, 938–945 (2015). doi: 10.1038/nm.3909; pmid: 26193342 3. V. Bronte et al., Recommendations for myeloid-derived suppressor cell nomenclature and characterization standards. Nat. Commun. 7, 12150 (2016). doi: 10.1038/ncomms12150; pmid: 27381735 4. J. Y. Sagiv et al., Phenotypic diversity and plasticity in circulating neutrophil subpopulations in cancer. Cell Rep. 10, 562–573 (2015). doi: 10.1016/j.celrep.2014.12.039; pmid: 25620698 5. B. E. Hsu et al., Immature low-density neutrophils exhibit metabolic flexibility that facilitates breast cancer liver metastasis. Cell Rep. 27, 3902–3915.e6 (2019). doi: 10.1016/ j.celrep.2019.05.091; pmid: 31242422 6. J.-I. Youn, S. Nagaraj, M. Collazo, D. I. Gabrilovich, Subsets of myeloid-derived suppressor cells in tumor-bearing mice. J. Immunol. 181, 5791–5802 (2008). doi: 10.4049/ jimmunol.181.8.5791; pmid: 18832739

14 of 16

,

Ng et al., Science 383, eadf6493 (2024)

Subcutaneous pancreatic tumor model

y

Tumor cryosections were fixed in 4% paraformaldehyde for 10 min, washed in PBS, and

Mice were injected intraperitoneally with 2 mg of BrdU (Sigma-Aldrich) at the indicated time points. To detect BrdU incorporation, cells from the BM, spleen, blood, and tumor in pancreatic tumor-bearing mice were stained with fixable vitality dye (LIVE/DEAD Fixable Blue Dead Cell Stain Kit, Invitrogen) and surface makers. Cells were then fixed, permeabilized, and finally stained intracellularly with FITC-conjugated anti-BrdU antibody according to the manufacturer’s protocol before analysis with flow cytometry.

y g

MICS sample preparation and data analysis

BrdU pulse-chase assay

y

Frozen tumor samples were quartered, embedded in optimal cutting temperature filled molds, and sectioned to a thickness of 10 mm at −20 °C. To ensure capture of tumor-infiltrating neutrophils, cryosections were screened for Ly6G immunofluorescence before mounting on Visium Spatial Gene Expression Slides (10X Genomics). Tissue sections were fixed with methanol; stained for cytokeratin, Ly6G, and DAPI; and imaged (EVOS Auto FL2, Thermo Scientific). After imaging, Visium Spatial Gene Expression libraries were prepared according to the manufacturer’s protocol. Libraries were sequenced using HiSeq X or NovaSeq 6000 S4 (Illumina) at PE150 with 50,000 read pairs per tissue-covered spot. Fastq data were processed with SpaceRanger (version 1.2.2, 10X Genomics) and mapped to the GRCm38 mm10-2020-A genome assembly. Spots were filtered to remove DAPI-low/negative necrotic areas with Loupe browser 6 (10X Genomics). Downstream analysis was performed with Seurat (v3.2.3) with default parameters, and joint clustering was performed with BayesSpace (36). The cell2location package (38) was used for deconvolution. The UCell package (80) was used to score spots independently for enrichment in various GO processes.

TCM was collected from the culture of FC1242L PDAC cell lines, spun down at 1000g for 10 min at 4°C to pellet down dead cells and debris, after which the supernatant was aliquoted and stored at –20°C. A total of 2.5 × 105 sorted neutrophils per well were cultured in cDMEM (Gibco) or TCM in normoxia or in a hypoxia incubator at 5% O2. After 1 and 3 days, dcTRAIL-R1 upregulation was analyzed by flow cytometry or wells were pooled, counted, and lysed for analysis of RNA expression.

g

Spatial transcriptomics for orthotopic pancreatic tumors

In vitro cell culture

of antibody was added to 1 × 105 neutrophils and incubated for 10 to 20 min on ice before the addition of 1 × 105 tumor cells. From day 14 after the injection of the tumor, tumors were measured weekly with Vernier calipers and tumor volumes were calculated using the following formula: 0.5 × length × width2. Mice were euthanized at day 28 after injection and tumor presence and weights were recorded. Tumors were marked as rejected if there was complete absence of tumor formation or only a clear Matrigel plug left. To evaluate vessel network formation in coinjected tumors, tumors were excised, fixed with 4% paraformaldehyde overnight at 4°C, and then washed with PBS. Samples were permeabilized with the SHANEL method, blocked, and finally stained for CD31 and propidium iodide (PI). Samples were then subjected to solvent-based optical clearing and transferred into ethyl cinnamate for imaging. All tumors were fully imaged from one lateral edge to the other at 2.0× magnification. Images were acquired using a lightsheet LaVision Ultramicroscope II and captured using the PCO edge 4.2 sCMOS camera and then analyzed using Imaris 9.5.0 (Bitplane). Total tumor volume imaged was first determined by surfacing positive PI signal using the surface function. To standardize the tumor volume evaluated, all tumor images were processed to obtain the same percentage volume (~35%) from the tumor midpoint, which was then used for the downstream analysis. The same image thresholds were applied consistently across tumors within the same experiment. Surfaces based on CD31+ vessel staining were then created with a standardized seed point value of 16.2, and their distribution from tumor edge to the tumor midpoint was calculated using the shortest distance function. Finally, to quantify the vessels present, the total tumor length was then divided into percentile bins (10%), after which CD31+ surface intensity was quantified and normalized against the total tumor volume intensity for each bin.

p

Pancreatic tumor single-cell suspensions pooled from five tumor-bearing mice were stained with a backbone panel of fluorophore-conjugated antibodies. The stained cell suspension was then distributed equally across all antibody wells in the LEGENDScreenTM kit (BioLegend). Staining and washing were performed according to the manufacturer’s instructions. Cells were finally stained with 1 mM DAPI, and 1 million events per well were acquired. Flow cytometry standard (.fcs) files corresponding to each unique PE marker were compensated and analyzed using FlowJo software (BD Biosciences). Compensated live, CD45+ singlets were analyzed using the InfinityFlow (1.4.0) in R (28, 29). The analysis was transferred back to FlowJo, where clusters were manually annotated and the imputed PE intensity values for each cluster were exported. The geometric mean was calculated for each cluster for heatmap plotting and comparison.

permeabilized for 20 min in blocking buffer (5% donkey serum, 0.3% Triton X-100, and DAPI). Sections were iteratively stained with fluorescein isothiocyanate (FITC)-, PE- and allophycocyanin (APC)-conjugated anti-mouse antibodies, after which image acquisition and processing was performed on the MACSima instrument as described previously (39). Images were preprocessed with MACS iQView software. Stitched and registered TIFF files of individual channels were then imported into Imaris (Bitplane). Staining artifacts were masked using the surface/ spot tool, and the masked channel was exported back to MACS iQView. Image segmentation based on DAPI-stained nuclei was performed using the Advanced Tissue Morphology method with the donut algorithm and confirmed by visual inspection. Identified cells and related features containing marker intensities for all channels were imported into FlowJo to gate relevant populations. Seurat V4 was used for further visualization and cell-type annotations, and SPIAT (48) was used for distance-based and colocalization metrics determining neutrophil subset localization with the tumor.

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

y y g y

AC KNOWLED GME NTS

,

12 January 2024

51. H. Nozawa, C. Chiu, D. Hanahan, Infiltrating neutrophils mediate the initial angiogenic switch in a mouse model of multistage carcinogenesis. Proc. Natl. Acad. Sci. U.S.A. 103, 12493–12498 (2006). doi: 10.1073/pnas.0601807103; pmid: 16891410 52. F. Shojaei et al., Bv8 regulates myeloid-cell-dependent tumour angiogenesis. Nature 450, 825–831 (2007). doi: 10.1038/ nature06348; pmid: 18064003 53. P. Kastana et al., “Matrigel plug assay for in vivo evaluation of angiogenesis,” in The Extracellular Matrix, D. Vigetti, A. D. Theocharis, Eds. (Springer, 2019), vol. 1952 of Methods in Molecular Biology, pp. 219–232; http://link.springer.com/ 10.1007/978-1-4939-9133-4_18. 54. V. Baeriswyl, G. Christofori, The angiogenic switch in carcinogenesis. Semin. Cancer Biol. 19, 329–337 (2009). doi: 10.1016/j.semcancer.2009.05.003; pmid: 19482086 55. N. G. Steele et al., Multimodal mapping of the tumor and peripheral blood immune landscape in human pancreatic cancer. Nat. Cancer 1, 1097–1112 (2020). doi: 10.1038/ s43018-020-00121-4; pmid: 34296197 56. L. Wang et al., Single-cell RNA-seq analysis reveals BHLHE40driven pro-tumour neutrophils with hyperactivated glycolysis in pancreatic tumour microenvironment. Gut 72, 958–971 (2023). doi: 10.1136/gutjnl-2021-326070; pmid: 35688610 57. J. Liu et al., An integrated TCGA pan-cancer clinical data resource to drive high-quality survival outcome analytics. Cell 173, 400–416 (2018). 58. J. Pillay et al., In vivo labeling with 2H2O reveals a human neutrophil lifespan of 5.4 days. Blood 116, 625–627 (2010). doi: 10.1182/blood-2010-01-259028; pmid: 20410504 59. Y. Sawanobori et al., Chemokine-mediated rapid turnover of myeloid-derived suppressor cells in tumor-bearing mice. Blood 111, 5457–5466 (2008). doi: 10.1182/blood-2008-01-136895; pmid: 18375791 60. P.-B. Ancey et al., GLUT1 expression in tumor-associated neutrophils promotes lung cancer growth and resistance to radiotherapy. Cancer Res. 81, 2345–2357 (2021). doi: 10.1158/ 0008-5472.CAN-20-2870; pmid: 33753374 61. L. M. Schiffmann et al., Tumour-infiltrating neutrophils counteract anti-VEGF therapy in metastatic colorectal cancer. Br. J. Cancer 120, 69–78 (2019). doi: 10.1038/ s41416-018-0198-3; pmid: 30377339 62. J. Jablonska, S. Leschner, K. Westphal, S. Lienenklaus, S. Weiss, Neutrophils responsive to endogenous IFN-b regulate tumor angiogenesis and growth in a mouse tumor model. J. Clin. Invest. 120, 1151–1164 (2010). doi: 10.1172/JCI37223; pmid: 20237412 63. Y. Itatani et al., Suppressing neutrophil-dependent angiogenesis abrogates resistance to anti-VEGF antibody in a genetic model of colorectal cancer. Proc. Natl. Acad. Sci. U.S.A. 117, 21598–21608 (2020). doi: 10.1073/ pnas.2008112117; pmid: 32817421 64. R. Grieshaber-Bouyer et al., The neutrotime transcriptional signature defines a single continuum of neutrophils across biological compartments. Nat. Commun. 12, 2856 (2021). doi: 10.1038/s41467-021-22973-9; pmid: 34001893 65. E. Montaldo et al., Cellular and transcriptional dynamics of human neutrophils at steady state and upon stress. Nat. Immunol. 23, 1470–1483 (2022). doi: 10.1038/ s41590-022-01311-1; pmid: 36138183

g

Ng et al., Science 383, eadf6493 (2024)

29. E. Becht et al., High-throughput single-cell quantification of hundreds of proteins using conventional flow cytometry and machine learning. Sci. Adv. 7, eabg0505 (2021). doi: 10.1126/ sciadv.abg0505; pmid: 34550730 30. K. Neumann et al., Clec12a is an inhibitory receptor for uric acid crystals that regulates inflammation in response to cell death. Immunity 40, 389–399 (2014). doi: 10.1016/ j.immuni.2013.12.015; pmid: 24631154 31. W. Xu et al., Immune-checkpoint protein VISTA regulates antitumor immunity by controlling myeloid cell–mediated inflammation and immunosuppression. Cancer Immunol. Res. 7, 1497–1510 (2019). doi: 10.1158/2326-6066.CIR-18-0489; pmid: 31340983 32. S. V. Ryzhov et al., Role of TGF-b signaling in generation of CD39+CD73+ myeloid cells in tumors. J. Immunol. 193, 3155–3164 (2014). doi: 10.4049/jimmunol.1400578; pmid: 25127858 33. M. Casanova-Acebes et al., Tissue-resident macrophages provide a pro-tumorigenic niche to early NSCLC cells. Nature 595, 578–584 (2021). doi: 10.1038/s41586-021-03651-8; pmid: 34135508 34. S. Kaur et al., Functions of thrombospondin-1 in the tumor microenvironment. Int. J. Mol. Sci. 22, 4570 (2021). doi: 10.3390/ijms22094570; pmid: 33925464 35. A. I. Markowska, F.-T. Liu, N. Panjwani, Galectin-3 is an important mediator of VEGF- and bFGF-mediated angiogenic response. J. Exp. Med. 207, 1981–1993 (2010). doi: 10.1084/ jem.20090121; pmid: 20713592 36. E. Zhao et al., Spatial transcriptomics at subspot resolution with BayesSpace. Nat. Biotechnol. 39, 1375–1384 (2021). doi: 10.1038/s41587-021-00935-2; pmid: 34083791 37. W. Lu, Y. Kang, Epithelial-mesenchymal plasticity in cancer progression and metastasis. Dev. Cell 49, 361–374 (2019). doi: 10.1016/j.devcel.2019.04.010; pmid: 31063755 38. V. Kleshchevnikov et al., Cell2location maps fine-grained cell types in spatial transcriptomics. Nat. Biotechnol. 40, 661–671 (2022). doi: 10.1038/s41587-021-01139-4; pmid: 35027729 39. A. Kinkhabwala et al., MACSima imaging cyclic staining (MICS) technology reveals combinatorial target pairs for CAR T cell treatment of solid tumors. Sci. Rep. 12, 1911 (2022). doi: 10.1038/s41598-022-05841-4; pmid: 35115587 40. Y. Kinugasa, T. Matsui, N. Takakura, CD44 expressed on cancer-associated fibroblasts is a functional molecule supporting the stemness and drug resistance of malignant cancer cells in the tumor microenvironment. Stem Cells 32, 145–156 (2014). doi: 10.1002/stem.1556; pmid: 24395741 41. A. Costa et al., Fibroblast heterogeneity and immunosuppressive environment in human breast cancer. Cancer Cell 33, 463–479.e10 (2018). doi: 10.1016/j. ccell.2018.01.011; pmid: 29455927 42. R. J. Slack, R. Mills, A. C. Mackinnon, The therapeutic potential of galectin-3 inhibition in fibrotic disease. Int. J. Biochem. Cell Biol. 130, 105881 (2021). doi: 10.1016/ j.biocel.2020.105881; pmid: 33181315 43. E. Fonsatti, M. Altomonte, M. R. Nicotra, P. G. Natali, M. Maio, Endoglin (CD105): A powerful therapeutic target on tumor-associated angiogenetic blood vessels. Oncogene 22, 6557–6563 (2003). doi: 10.1038/sj.onc.1206813; pmid: 14528280 44. A. Giatromanolaki et al., Ectonucleotidase CD73 and CD39 expression in non-small cell lung cancer relates to hypoxia and immunosuppressive pathways. Life Sci. 259, 118389 (2020). doi: 10.1016/j.lfs.2020.118389; pmid: 32898522 45. K. Losenkova et al., Compartmentalization of adenosine metabolism in cancer cells and its modulation during acute hypoxia. J. Cell Sci. 133, jcs241463 (2020). doi: 10.1242/ jcs.241463; pmid: 32317394 46. N. Petruk et al., CD73 facilitates EMT progression and promotes lung metastases in triple-negative breast cancer. Sci. Rep. 11, 6035 (2021). doi: 10.1038/s41598-021-85379-z; pmid: 33727591 47. Z. Fu et al., Proteolytic regulation of CD73 by TRIM21 orchestrates tumor immunogenicity. Sci. Adv. 9, eadd6626 (2023). doi: 10.1126/sciadv.add6626; pmid: 36608132 48. Y. Feng et al., Spatial analysis with SPIAT and spaSim to characterize and simulate tissue microenvironments. Nat. Commun. 14, 2697 (2023). doi: 10.1038/s41467-023-37822-0; pmid: 37188662 49. I. Ballesteros et al., Cooption of neutrophil fates by tissue environments. Cell 183, 1282–1297.e18 (2020). doi: 10.1016/ j.cell.2020.10.003; pmid: 33098771 50. L. Yang et al., Expansion of myeloid immune suppressor Gr+CD11b+ cells in tumor-bearing host directly promotes tumor angiogenesis. Cancer Cell 6, 13 (2004). pmid: 15488763

p

7. T. Condamine et al., Lectin-type oxidized LDL receptor1 distinguishes population of human polymorphonuclear myeloid-derived suppressor cells in cancer patients. Sci. Immunol. 1, aaf8943 (2016). doi: 10.1126/sciimmunol. aaf8943; pmid: 28417112 8. C. Engblom et al., Osteoblasts remotely supply lung tumors with cancer-promoting SiglecFhigh neutrophils. Science 358, eaal5081 (2017). doi: 10.1126/science.aal5081; pmid: 29191879 9. F. Veglia et al., Fatty acid transport protein 2 reprograms neutrophils in cancer. Nature 569, 73–78 (2019). doi: 10.1038/ s41586-019-1118-2; pmid: 30996346 10. C. Pfirschke et al., Tumor-promoting Ly-6G+ SiglecFhigh cells are mature and long-lived neutrophils. Cell Rep. 32, 108164 (2020). doi: 10.1016/j.celrep.2020.108164; pmid: 32966785 11. R. Zilionis et al., Single-cell transcriptomics of human and mouse lung cancers reveals conserved myeloid populations across individuals and species. Immunity 50, 1317–1334.e10 (2019). doi: 10.1016/j.immuni.2019.03.009; pmid: 30979687 12. H. Alshetaiwi et al., Defining the emergence of myeloidderived suppressor cells in breast cancer using single-cell transcriptomics. Sci. Immunol. 5, eaay6017 (2020). doi: 10.1126/sciimmunol.aay6017; pmid: 32086381 13. F. Veglia et al., Analysis of classical neutrophils and polymorphonuclear myeloid-derived suppressor cells in cancer patients and tumor-bearing mice. J. Exp. Med. 218, e20201803 (2021). doi: 10.1084/jem.20201803; pmid: 33566112 14. M. Evrard et al., Developmental analysis of bone marrow neutrophils reveals populations specialized in expansion, trafficking, and effector functions. Immunity 48, 364–379.e8 (2018). doi: 10.1016/j.immuni.2018.02.002; pmid: 29466759 15. Y. P. Zhu et al., Identification of an early unipotent neutrophil progenitor with pro-tumoral activity in mouse and human bone marrow. Cell Rep. 24, 2329–2341.e8 (2018). doi: 10.1016/ j.celrep.2018.07.097; pmid: 30157427 16. I. Kwok et al., Combinatorial single-cell analyses of granulocytemonocyte progenitor heterogeneity reveals an early uni-potent neutrophil progenitor. Immunity 53, 303–318.e5 (2020). doi: 10.1016/j.immuni.2020.06.005; pmid: 32579887 17. H. Q. Dinh et al., Coexpression of CD71 and CD117 identifies an early unipotent neutrophil progenitor population in human bone marrow. Immunity 53, 319–334.e6 (2020). doi: 10.1016/ j.immuni.2020.07.017; pmid: 32814027 18. D. E. Muench et al., Mouse models of neutropenia reveal progenitor-stage-specific defects. Nature 582, 109–114 (2020). doi: 10.1038/s41586-020-2227-7; pmid: 32494068 19. X. Xie et al., Single-cell transcriptome profiling reveals neutrophil heterogeneity in homeostasis and infection. Nat. Immunol. 21, 1119–1133 (2020). doi: 10.1038/ s41590-020-0736-z; pmid: 32719519 20. T. E. Khoyratty et al., Distinct transcription factor networks control neutrophil-driven inflammation. Nat. Immunol. 22, 1093–1106 (2021). doi: 10.1038/s41590-021-00968-4; pmid: 34282331 21. C. Wu et al., Spleen mediates a distinct hematopoietic progenitor response supporting tumor-promoting myelopoiesis. J. Clin. Invest. 128, 3425–3438 (2018). doi: 10.1172/JCI97973; pmid: 29771686 22. M. Liu et al., PERK reprograms hematopoietic progenitor cells to direct tumor-promoting myelopoiesis in the spleen. J. Exp. Med. 219, e20211498 (2022). doi: 10.1084/jem.20211498; pmid: 35266960 23. S. R. Hingorani et al., Trp53R172H and KrasG12D cooperate to promote chromosomal instability and widely metastatic pancreatic ductal adenocarcinoma in mice. Cancer Cell 7, 469–483 (2005). doi: 10.1016/j.ccr.2005.04.023; pmid: 15894267 24. P. Angerer et al., destiny: Diffusion maps for large-scale single-cell data in R. Bioinformatics 32, 1241–1243 (2016). doi: 10.1093/bioinformatics/btv715; pmid: 26668002 25. L. Haghverdi, M. Büttner, F. A. Wolf, F. Buettner, F. J. Theis, Diffusion pseudotime robustly reconstructs lineage branching. Nat. Methods 13, 845–848 (2016). doi: 10.1038/nmeth.3971; pmid: 27571553 26. G. La Manno et al., RNA velocity of single cells. Nature 560, 494–498 (2018). doi: 10.1038/s41586-018-0414-6; pmid: 30089906 27. B. Van de Sande et al., A scalable SCENIC workflow for single-cell gene regulatory network analysis. Nat. Protoc. 15, 2247–2276 (2020). doi: 10.1038/s41596-020-0336-2; pmid: 32561888 28. C.-A. Dutertre et al., Single-cell analysis of human mononuclear phagocytes reveals subset-defining markers and identifies circulating inflammatory dendritic cells. Immunity 51, 573–589.e8 (2019). doi: 10.1016/j.immuni.2019.08.008; pmid: 31474513

We acknowledge and thank all members of the L.G.N. laboratory for helpful discussion and feedback on the manuscript, the SIgN Flow Cytometry team for sorting and flow cytometry assistance, the SIgN Immunogenomics team for assistance with generating and running the ATACseq and scRNAseq libraries, and the SIgN mouse core facility for technical help and support. Funding: This research was funded by Singapore Immunology Network (SIgN) core funding and A*STAR, Singapore. L.G.N. is supported by National Natural Science Foundation of China (grant 32270956) and Shanghai Science and Technology Commission (grant 20JC1410100). M.S.F. is supported by the A*STAR Career Development Fund (grant 202D8150). I.W.K. is supported by the A*STAR Career Development Fund (grant 202D8197). S.Z.C. is supported by the A*STAR Career Development Award (192D8043) and core funding from SIgN. Y.T. and H.L.T. are supported by Clinician Scientist Awards (NMRC/CSA-INV/0023/2017 and CSAINV20nov-0003) from the National Medical Research Council of Singapore. The SIgN Flow Cytometry facility is supported by National Research Foundation (NRF) Singapore under Shared Infrastructure Support (SIS) (NRF2017_SISFP09). For the use of the Ultramicroscope II, we would like to thank A*STAR core funds,

15 of 16

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

the HMBS IAF-PP grant (H1701a0004), and the National Research Foundation’s Shared Infrastructure Support grant for SingaScope, a Singapore-wide microscopy infrastructure network (NRF2017_SISFP10), for continued support of the A*STAR Microscopy Platform. D.C.-W. is supported by a CRI Irvington postdoctoral fellowship (CRI3511). G.F.C. is supported by grant PID2022-142341OB-I00 funded by the Spanish MCIN/AEI/ 10.13039/501100011033. I.B. is supported by a Ramon y Cajal fellowship (RYC2021-033511-I) from MICINN and by a Leonardo fellowship (LEO22-2-2596) from the BBVA Foundation. A.V. and Y.Z. are supported by NMRC OF-IRG grant to VA (OFIRG19nov-0112). M.I. is supported by the European Research Council (ERC) Consolidator Grant 725038, ERC Proof of Concept Grants 957502 and 101138728, Italian Association for Cancer Research (AIRC) Grants 19891 and 22737, Italian Ministry for University and Research Grants PE00000007 (INF-ACT) and PRIN (2022FMESXL), and by a Funded Research Agreement from Asher Biotherapeutics and VIR Biotechnology. A.H. is supported by grant R01AI165661 from NIH/NIAD, RTI2018-095497- B-I00 from MCIN, HR17_00527 from Fundación La Caixa, the Transatlantic Network of Excellence (TNE-18CVD04) from the Leducq Foundation, and FET-OPEN (grant 861878) from the European Commission. The CNIC is supported by the MCIN and the Pro CNIC Foundation and is a Severo Ochoa

Center of Excellence (CEX2020-001041-S). F.G. is supported by Singapore Immunology UIBR award from the Agency for Science, Technology and Research (A*STAR), a Singapore NRF Senior Investigatorship (NRFI2017-02), the Fondation Gustave Roussy, and the ARC Foundation. Author contributions: Conceptualization: L.G.N., M.S.F., L.T., I.W.K., A.H., F.G.; Formal analysis: K.B.D., V.N., J.M.C., M.W.L., L.X.J., B.Y.S., G.F.C.; Funding acquisition: L.G.N., M.S.F., I.W.K., Y.H.L.; Investigation: M.S.F., L.T., I.W.K., Y.R., C.M.S., K.L., Y.N.Z., J.J., K.H.L., D.H.L., Y.C.T., D.C.-W., K.Y., C.B., C.N.; Methodology: M.S.F., L.T., I.W.K., Y.R., C.M.S., D.C.-W., K.Y.; Project administration: L.G.N.; Supervision: L.G.N.; Visualization: M.S.F., L.T., I.W.K., Y.T., C.M.S., K.L., D.C.-W., K.Y.; Writing – original draft: L.G.N., M.S.F., L.T., I.W.K., A.H., F.G., D.C.-W.; Writing – review & editing: All authors Competing interests: M.S.F., L.T., F.G., and L.G.N. are inventors on patent WO2022132041 submitted by A*STAR that covers a method of characterizing an immunosuppressive neutrophils in cancer. M.I. participates in advisory boards and consults for Gilead Sciences, Third Rock Ventures, Antios Therapeutics, Asher Biotherapeutics, GentiBio, Clexio Biosciences, Sibylla Biotech, BlueJay Therapeutics, and Aligos Therapeutics. H.L.T. is a cofounder and medical adviser for RNAscence Pte. Ltd. F.G. participates in advisory boards for iTheos and Genoskin. The remaining authors declare no competing interests. Data and materials availability: All data

needed to evaluate the conclusions in the paper are present in the paper and/or in the supplementary materials. scRNAseq (GSE243466), ATACseq (GSE244531), and spatial transcriptomics data (GSE244534) have been deposited in the NCBI Gene Expression Omnibus. All three datasets have been linked under into one SuperSeries GSE244536 for ease of access. License information: Copyright © 2024 the authors, some rights reserved; exclusive licensee American Association for the Advancement of Science. No claim to original US government works. https://www. science.org/about/science-licenses-journal-article-reuse SUPPLEMENTARY MATERIALS

science.org/doi/10.1126/science.adf6493 Materials and Methods Figs. S1 to S16 Tables S1 to S7 References (66–81) MDAR Reproducibility Checklist Movies S1 and S2 Submitted 9 November 2022; resubmitted 23 August 2023 Accepted 27 November 2023 10.1126/science.adf6493

p g y y g y ,

Ng et al., Science 383, eadf6493 (2024)

12 January 2024

16 of 16

RES EARCH

RESEARCH ARTICLE

◥

NEUROSCIENCE

Illusory generalizability of clinical prediction models Adam M. Chekroud1,2*, Matt Hawrilenko1, Hieronimus Loho2, Julia Bondar1, Ralitza Gueorguieva3, Alkomiet Hasan4, Joseph Kambeitz5, Philip R. Corlett2, Nikolaos Koutsouleris6, Harlan M. Krumholz7, John H. Krystal2, Martin Paulus8 It is widely hoped that statistical models can improve decision-making related to medical treatments. Because of the cost and scarcity of medical outcomes data, this hope is typically based on investigators observing a model’s success in one or two datasets or clinical contexts. We scrutinized this optimism by examining how well a machine learning model performed across several independent clinical trials of antipsychotic medication for schizophrenia. Models predicted patient outcomes with high accuracy within the trial in which the model was developed but performed no better than chance when applied out-of-sample. Pooling data across trials to predict outcomes in the trial left out did not improve predictions. These results suggest that models predicting treatment outcomes in schizophrenia are highly context-dependent and may have limited generalizability.

the potential for statistical models to improve clinical practice (7–9). Open data opens possibilities

Data sources

g y

Table 1. Treatment outcomes across trials.

,

Outcome definition

Adults first Adults Adults Older episode Chronic #1 Chronic #2 adults Teens (n = 321) (n = 430) (n = 481) (n = 99) (n = 182)

Total (n = 1513)

264 208 266 32 47 816 (82.2%) (48.4%) (55.3%) (32.3%) (25.8%) (54.0%) ..................................................................................................................................................................................................................... 50% Reduction 119 85 82 7 12 306 PANSS (37.1%) (19.8%) (17.0%) (7.1%) (6.6%) (20.3%) ..................................................................................................................................................................................................................... RSWG remission 152 129 153 24 58 517 (34.2%) criteria (47.4%) (30.0%) (31.8%) (24.2%) (31.9%) ..................................................................................................................................................................................................................... Percentage change -44.1 -26.9 -28.4 -18.0 -13.7 -28.8 in PANSS total (23.1) (28.2) (25.3) (21.8) (21.5) (26.7) score (SD) ..................................................................................................................................................................................................................... Baseline total 103.0 92.4 92.9 91.1 90.0 94.4 PANSS (SD) (14.3) (13.0) (10.9) (8.8) (13.1) (13.2) 25% Reduction PANSS

1

2

Spring Health, New York City, NY 10010, USA. Department of Psychiatry, Yale University School of Medicine, New Haven, CT 06520, USA. 3Department of Biostatistics, Yale University, New Haven, CT 06520, USA. 4Department of Psychiatry, Psychotherapy and Psychosomatics, University Augsburg, 86159 Augsburg, Germany. 5Department of Psychiatry and Psychotherapy, University of Cologne, Faculty of Medicine and University Hospital of Cologne, Cologne, Germany. 6Department of Psychiatry and Psychotherapy, Ludwig-MaximiliansUniversity, Munich, Germany. 7Center for Outcomes Research and Evaluation, Yale New Haven Hospital, New Haven, CT 06520, USA. 8Laureate Institute for Brain Research, Tulsa, OK 74136, USA.

.....................................................................................................................................................................................................................

*Corresponding author. Email: [email protected]

Chekroud et al., Science 383, 164–167 (2024)

y g

We used treatment data from five international, multisite RCTs (NCT00518323, NCT00334126, NCT00085748, NCT00078039, and NCT00083668) obtained through the YODA Project (https:// yoda.yale.edu/). These trials were selected because of their comparability and consistency. All patients had a current DSM-IV diagnosis of schizophrenia at the start of the trial; all trials randomized patients to an antipsychotic medication or placebo; all trials used the same scale to measure treatment outcomes (the Positive and Negative Syndrome Scale, PANSS); all trials included a 4-week timepoint to measure outcomes; and all trials collected similar data about the patients at baseline. Combined, the trials also provide a heterogeneous patient

y

As efforts toward mandatory randomized controlled trial (RCT) data deposition, archival data sharing, and open science continue to advance, opportunities arise to more rigorously examine how well treatment prediction models will fare in different contexts. The Yale Open Data Access (YODA) Project is one such effort, which now includes a data archive of over 246 clinical trials from all medical fields. The YODA project included several RCTs evaluating the comparative efficacy of antipsychotic medications for treating schizophrenia. Predicting treatment outcomes in schizophrenia could be especially advantageous because the clinical response to pharmacological interventions is heterogeneous and depends on many

p

O

ne fundamental problem in medicine is that despite similar treatments some patients get better whereas others show no improvement. One goal of precision medicine is to use machine learning to find models that will help predict who will respond to what type of treatment (1). For precision medicine to affect clinical practice and improve outcomes, the models that we develop must robustly predict outcomes for unseen, future patients (2–5). However, models are not usually tested on new patients in a different context because data—especially data from controlled designs— are scarce and expensive (6). Instead, researchers typically split a study’s participants into two or more random groups, build a model using the data from one of the groups, and test its predictions on the other group (e.g., k-fold cross-validation) (3, 4). When we use this kind of approximation based on one data set or clinical sample, we have a fundamentally limited insight into the true potential for a model to improve outcomes in the future. Validating clinical prediction models in different clinical samples is an essential step in the model development process. It generally results in predictive performance measures that are lower but allows for a more realistic assessment of

environmental factors such as individual and family-related stress, drug abuse, homelessness, and social isolation. Depending on the clinical outcome definition, up to 20 to 30% of first-episode individuals (10) and more than 50% with a relapse do not respond sufficiently to antipsychotic medications (11). We examined the generalizability of clinical prediction models across multiple clinical trials using the case study of antipsychotic treatments for schizophrenia. Critically, this study directly evaluated the performance of a model on its initial training sample as well as how the same model performed on truly independent clinical trial samples. This allowed us to assess two key risks: First, models may “overfit” the data by fitting the random noise of one particular dataset rather than a true signal likely to generalize across samples, leading to good predictions in the training data that do not generalize to the testing data. The second key risk is poor model transportability. Models may lack external validity due to patients, providers, or implementation characteristics varying across trials (12).

12 January 2024

1 of 4

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

population, with patients recruited from 194 sites across 4 continents, a pediatric trial, an older adult trial, and a trial of individuals with a first episode (see SM for more details). The study design, outcome measure, and crossvalidation approach were preregistered on 2 August 2016 (YODA 2016-1005). Minor updates to the preregistration were submitted on 2 May 2023 (included in the SM). Patients and outcomes

Within-trial cross-validation

y

We applied machine learning methods using baseline data to predict whether a patient would achieve clinically significant improvements in symptoms over four weeks of antipsychotic treatment. We used the elastic net algorithm (21, 22), a penalized regression method that is appropriate when covariates are

g

Within-trial no validation

Machine learning approach

correlated with one another and predictors may only be sparsely endorsed. It has been successful in research predicting psychiatric treatment outcomes (5, 23–25). The elastic net model uses two penalty parameters, lambda and alpha, which balance stability with parsimony. We examined 400 combinations of alpha and lambda penalties (see supplement) and selected the optimal penalties using repeated 10-fold cross-validation. The cross-validation part of this procedure separates the data set into 10 random folds and uses 9 of the subsets for training, repeating the process such that each subset is left out once for testing. The repeated part of this procedure re-splits the data ten times to reduce the impact of the random data split; in aggregate, 100 total models were fit to the 10 folds by 10 repeats. Model performance was calculated by averaging the performance metric across all 100 models. This entire procedure was run for each of the 400 combinations of alpha and lambda values, and the final values were chosen as the combination of alpha and lambda values that optimized the model performance metric. We used the metrics of area under the receiver operating curve for binary outcomes and root mean square error for continuous outcomes. The final alpha and lambda values were applied to the aggregate sample to estimate the prediction

p

From 29 March 2004 to 30 March 2009, 1962 total patients aged 12 to 81 years were enrolled across five randomized controlled trials at 194 sites in North America, Asia, Europe, and Africa. We assessed symptomatic outcomes based on the PANSS (13) at week 4 for the 1513 participants with baseline and 4-week follow-up data. Different definitions of response, remission, and recovery are used in schizophrenia research, which makes comparing and applying results in clinical practice difficult (14–16). The primary outcome reported here is the Remission in Schizophrenia Working Group criteria (RSWG) (17). To ensure that our findings were not driven by idiosyncrasies in how we defined treatment response, we included three other definitions commonly used in the field, including percentage change with baseline correction (15, 16), and two binary definitions of 25 and 50% symptom re-

duction. Table 1 reports treatment outcomes for all definitions across the five trials. We extracted all information available at baseline across all trials and retained it as a predictor variable if it was available for more than 80% of patients. We also computed condition (control versus treatment) X predictor interaction terms. Drug dose was standardized to paliperidone dose equivalents using the defined daily dose method (18). Together, this yielded 217 predictor variables that included basic demographic features, psychiatric history (DSM-IV diagnosis category, age of diagnosis, psychiatric hospitalizations), clinical data (PANSS, Clinical Global Impression) (17), extrapyramidal symptom scales (Abnormal Involuntary Movement Scale) (19) and Simpson Angus Scale (20), biometric data (blood chemistry panel, hematology, urinalysis), and treatment randomization. The detailed list of predictors, selection criteria, and missing data approach is provided in the SM.

Leave-onetrial-out

Paired-trial

1.0

Balanced accuracy

0.9

y g

0.8

0.7

0.6

y ,

Ch ro

Ad

ult

sF

E→ nic Ad #1 ult Ch sF → ron E Ch ic # Ol ron de 2 ic # rA → 1 du Ch lts ron → ic Ol # 2 de rA Te du en l ts s→ Ad ult Te sF en Ch s E ron → Ad ic # ult 1→ Ch sF ron E Ch ic # Ol de 2 → ronic rA #1 du Ch lts ron → ic # Ol 2 de rA Te d en ult s→ s Ad ult Te sF en E→ s Ad ult Ch sF Ad E → ronic ult #1 C sF E → hron ic Ol #2 de Ad rA ult du Ch sF lts ron E→ ic # Te Ch 1→ en ron s Ad ic Ch #1 ult sF ron → E ic # Ch ron 1→ ic Ol #2 Ch de ron rA ic d Ch ult #1 ron s → ic # Te Ch 2→ en ron s A i c# du Ch 2→ lts ron FE ic # C 2 → hron ic # Ol 1 Ch de ron rA Ol ic # du de l ts rA 2→ du Ol Te lts de en → rA s A du du Ol lts lts de → FE rA C du hro lts nic → Ol #1 Ch de ron rA du ic #2 lts Te → en Te s→ en Te s Ad en ult s→ sF E Te Ch en s → ronic Te #1 Ch en ron s→ ic Old #2 All er Ad oth ult ers s All → oth Ad ers ult sF All → E oth Ch ron ers All ic # → oth 1 Ch ers ron → ic Ol #2 de All rA oth du ers lts → Te en s

0.5

Study pair Balanced accuracy is higher than chance

no

yes

FE = first episode

Fig. 1. Balanced accuracy for models predicting treatment outcome (Remission in Schizophrenia Working Group criteria) across all modeling scenarios. Gray intervals represent 95% confidence intervals, not adjusted for multiple comparisons. Red markers denote statistical significance after applying the BenjaminiHochberg adjustment with the false discovery rate set to 5%. Repeated 10-fold cross-validation; FE, first episode. Chekroud et al., Science 383, 164–167 (2024)

12 January 2024

2 of 4

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

model coefficients. To interpret differential performance across samples, we report a metric known as balanced accuracy [(sensitivity + specificity) / 2] whose null distribution is centered on 50% (26, 27). To determine whether balanced accuracy was statistically significantly above chance, we bootstrapped confidence intervals and adjusted for multiple comparisons across all 35 comparisons using the Benjamini-Hochberg adjustment with the false discovery rate set to 5% (28). All analysis was conducted using R version 4.1 (29), with machine learning models fit using the caret package (30). Exploring the generalizability of machine learning models

Paired-trial validation

Next, we directly assessed out-of-sample performance in the paired-trial validation (16). We applied the prediction models developed using within-trial models across each of the other trials, for a total of 20 trial pairs. Model performance was low (mean across all trial pairs was 0.54, range 0.48 to 0.61) with only three trial pairs performing above chance. Leave-one-trial-out validation

12 January 2024

3 of 4

,

Machine learning prediction of treatment outcomes in medicine is exciting but challenging. Our modeling scenarios using antipsychotic treatment outcome prediction in schizophrenia suggest that predictive models are fragile and that excellent performance in one clinical context is not a strong indicator of performance on future patients. This is highly concerning as most predictive studies today rely on internal samples for testing and validation. When models were tested on the same sample on which they were developed, models routinely produced strong predictions. Cross-validation

y

To estimate more generalizable prediction accuracy, we employed within-trial cross-validation. Performance characteristics of the optimal alpha and lambda values were averaged across the 100 left out folds (10 folds * 10 repeats) from the repeated cross-validation procedure. Each trial’s data were divided into 10 subsets, with coefficients developed on 9 subsets and

Discussion

y g

Cross-validation

The pattern of results observed was not due to idiosyncrasies of how we measured treatment response. We found the same pattern of results when we reproduced all four modeling scenarios using other binary and continuous definitions of treatment response (see SM). This lack of model generalizability to unseen patients was also observed for another machine learning algorithm. When we applied random forest models, which can detect complex patterns of interactions amongst predictor variables, we observed the same pattern of results except that excessive overfitting occurred for no-validation conditions (see SM).

There are three key reasons why predictive models might not generalize across trials. First, patient groups may be too different across trials. The umbrella category of schizophrenia is useful for clinical practice but also means that patients with different disease stages are coerced into the same diagnostic category in clinical trials. If key information that differentiates patients is not captured in the data or if the range of that information is more restricted in the dataset used to develop the model compared with the target trial, predictions may be inaccurate. Thus, patient populations may differ considerably between trials within the same diagnostic category. However, the current study found little evidence that results would generalize across even the most similar trials. The three cross-trial pairs with predictions slightly greater than chance were amongst the three studies of adults aged 18 and over but this pattern of results did not consistently replicate across other outcome definitions. Second, these trials may not have collected the type or volume of data needed to make good predictions. This study used clinical, sociodemographic, and simple biomarker databased on almost 2000 patients. However, additional data types may have been more relevant to treatment outcomes. Psychosocial information and social determinants of health were not included in this study but have previously been found to predict treatment outcomes in first episode psychosis (27). Preliminary research suggests that longitudinal patterns of symptom co-occurrence—either before or during treatment—can be specifically relevant to how a patient will respond to treatment although it may delay care to collect this data (31–34). Some have suggested the use of neuroimaging and genetic data but there is currently little evidence to suggest that such data would improve predictions; further, collecting these data would pose additional barriers for routine implementation (35–37). Finally, having data from more participants may allow for more nuanced modeling of individual differences. A third reason why predictive models may not generalize is that patient outcomes may be

y

In the scenario where we assessed within-trial performance without any external validation, the final prediction model created for a specific trial was applied to the entire sample from that same trial. The balanced accuracy was high and significantly above random chance for all models, with an average of 0.72 (range: 0.66 to 0.77) across all five prediction models. However, because the model was evaluated on the same sample used to develop it, there is a risk of overfitting, making these results less likely to generalize.

Sensitivity analyses

Why model generalizability is challenging

g

No validation

Given the availability of multiple archival trials for developing a prediction model, a natural extension of the paired-trial validation would be a leave-one-trial-out approach. This approach might enhance generalizability by allowing the algorithm to be exposed to more information through between-trial variability in baseline phenotypes. We aggregated data across four trials, leaving the fifth out for testing, and repeated the process 5 times so that each trial was left out once. Performance was once again poor with low balanced accuracy in all conditions (mean across all left out trials was 0.54 with range 0.50 to 0.58) and performance was significantly above chance in only two of the five testing sets.

tempered these performance estimates but even the models that performed well in crossvalidation were little better than chance when predicting outside of the sample in which they were developed—even when the unseen samples were well-phenotyped. In a world where we hope that predictive models might eventually improve clinical practice, the ability to generalize to other carefully controlled clinical contexts is only the first step to generalize to settings with more heterogeneity in patient presentations and methods of care delivery.

p

We evaluated the applicability of machine learning models across four distinct scenarios to gain insights into their generalizability: First, we assessed the predictive accuracy of the model within the trial, without any external validation beyond the training data. Second, we also focused on within-trial prediction accuracy but this time estimated using the data excluded from the training set in a repeated tenfold cross-validation process. Third, we conducted a paired-across-trial prediction accuracy assessment. In this case, models trained on one trial were applied to all other trials to evaluate their performance. Finally, in the fourth scenario, we implemented a leaveone-trial-out prediction accuracy assessment. Models were trained using data aggregated from four trials and their predictive accuracy was tested on the fifth trial (Fig. 1). Balanced accuracy for the RSWG criteria are shown in Fig. 1, and data for alternative outcome definitions and additional outcome metrics are shown in the supplement.

Chekroud et al., Science 383, 164–167 (2024)

then tested on the remaining subset. In this scenario, balanced accuracy was lower in each dataset, averaging 0.60 (range: 0.56 to 0.67) across all five prediction models. Only three out of five models performed above chance.

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

too context-dependent. Trials may have subtly important differences in recruiting procedures, inclusion criteria, or treatment protocols. Because these characteristics do not vary across patients within a trial, they cannot be modeled as predictors within a single trial. However, this study used multinational RCTs conducted by large pharmaceutical companies and contract research organizations, minimizing nonspecific concerns especially in comparison to the variability we would expect in real clinical practice going from one site or provider to the next. Of course, different antipsychotic drugs may differ from one another in ways that affect outcome prediction, and the D2 dopamine receptor blockade intended to correct overstimulation of D2 receptors by endogenous dopamine may be too far downstream from the primary pathology of schizophrenia or the symptom severity criteria used to measure it (38).

The present study offers an underwhelming but realistic picture of our current ability to develop truly useful predictive models for schizophrenia treatment outcomes. Models that performed with excellent accuracy in one sample routinely failed to generalize to unseen patients. These findings suggest that approximations based on a single data set are a fundamentally limited insight into future performance and represent a potential concern for prediction models throughout medicine. The field as a whole—present authors included— hope that machine learning approaches can eventually improve the allocation of treatments in medicine; however, we should a priori remain skeptical of any predictive model findings that lack an independent sample for validation.

Submitted 27 January 2023; resubmitted 20 July 2023 Accepted 10 November 2023 10.1126/science.adg8538

4 of 4

,

science.org/doi/10.1126/science.adg8538 Methods and Results Figs. S1 to S15 Tables S1 to S20 References (42) MDAR Reproducibility Checklist Data S1

y

SUPPLEMENTARY MATERIALS

y g

12 January 2024

Funding: No funding source had any role in the study design, data collection, data analysis, data interpretation, writing, or submission of this report. All trials were originally funded by Janssen Research and Development. The study design, outcome measure, and crossvalidation approach were preregistered on 1 Aug 2016 (YODA #20161005). The study was approved on 15 December 2016 (IRB/HSC# 1610018521) by the Yale University Institutional Review Board. Access was granted on 5 January 2017, and data were analyzed until October, 2023. Author contributions: Conceptualization: A.C. Methodology: A.C., M.H., R.G., H.L., J.B., A.H., N.K., J.H.K., H.K. Data Acquisition: H.K., on behalf of the Yale Open Data Access Initiative. Data Analysis: A.C., M.H., H.L., J.B. Visualization: M.H., H.L., R.G. Supervision: A.C., M.P., P.C., and J.H.K. Writing – original draft: A.C. Writing – substantial review and editing: A.C., M.H., A.H., N.K., P.C., H.K., J.H.K., and M.P. Competing interests: A.C. holds equity in Spring Care and is the lead inventor on 3 patent submissions relating to treatment for major depressive disorder (US Patent and Trademark Office number Y0087.70116US00 and provisional application numbers 62/491 660 and 62/629 041). M.H. and J.B. are employed by and hold equity in Spring Care. RG received royalties from the book Statistical Methods in Psychiatry and Related Fields published by CRC Press and is an inventor on US patent application 20200143922. A.H. was a member of advisory boards and received paid speakership by Boehringer-Ingelheim, Lundbeck, Otsuka, Rovi, and Recordati. He received paid speakership by AbbVie and Advanz. He is editor of the AWMF German guidelines for schizophrenia. J.H.K. has been a consultant and/or advisor to or has received honoraria from Janssen/ J&J, Lundbeck and Boehringer Ingelheim. P.C. is co-founder and Board Member of Tetricus Labs and reported holding stock and stock options in Tetricus Labs Inc. H.K. received funding from Johnson and Johnson through Yale University. J.Kr. reported holding patents licensed to Johnson and Johnson and Freedom Biosciences; cofounder of Freedom Biosciences, stock in Spring Health, Biohaven Pharmaceuticals, Neumora Pharmaceuticals; Consultant to Biogen, Bionomics Limited, Boehringer Ingelheim International, Cerevel Therapeutics, Jazz Pharmaceuticals, Otsuka American Pharmaceutical Inc., Perception Neuroscience, Sumitomo America, Taishisho, Takeda, BioXcel, Psychogenics Inc. Data and materials availability: The study design, outcome measure, and cross-validation approach was preregistered on 1 August 2016 (YODA #2016-1005). All data are available via the Yale University Open Data Access (YODA) platform (https://yoda.yale.edu/request). Data accession numbers for the five trials analyzed here are: R076477-PSZ-3001 (Teens trial), R076477-SCH-3015 (Adults First Episode), R076477-SCH-302 (Older Adults), R076477-SCH-305 (Adults Chronic #1), R076477SCH-303 (Adults Chronic #2). The code to reproduce all results in this manuscript and the supplement is available at Zenodo (42). This study, carried out under YODA Project #2016-1005, used data obtained from the Yale University Open Data Access Project, which has an agreement with JANSSEN RESEARCH & DEVELOPMENT, L.L.C. The interpretation and reporting of research using this data are solely the responsibility of the authors and does not necessarily represent the official views of the Yale University Open Data Access Project or JANSSEN RESEARCH & DEVELOPMENT, L.L.C. License information: Copyright © 2024 the authors, some rights reserved; exclusive licensee American Association for the Advancement of Science. No claim to original US government works. https://www. sciencemag.org/about/science-licenses-journal-article-reuse

y

1. F. S. Collins, H. Varmus, N. Engl. J. Med. 372, 793–795 (2015). 2. D. G. Altman, P. Royston, Stat. Med. 19, 453–473 (2000). 3. D. G. Altman, Y. Vergouwe, P. Royston, K. G. M. Moons, BMJ 338, 1432 (2009). 4. E. W. Steyerberg, in Clinical Prediction Models (Springer, 2009), pp. 11–31. 5. A. M. Chekroud et al., Lancet Psychiatry 3, 243–250 (2016). 6. G. C. M. Siontis, I. Tzoulaki, P. J. Castaldi, J. P. A. Ioannidis, J. Clin. Epidemiol. 68, 25–34 (2015). 7. E. W. Steyerberg, Y. Vergouwe, Eur. Heart J. 35, 1925–1931 (2014). 8. R. D. Riley et al., Stat. Med. 40, 4230–4251 (2021). 9. M. Pavlou et al., Stat. Methods Med. Res. 30, 2187–2206 (2021). 10. Y. Zhu et al., Eur. Neuropsychopharmacol. 27, 835–844 (2017). 11. S. Leucht et al., Am. J. Psychiatry 174, 927–942 (2017). 12. B. Li, C. Gatsonis, I. J. Dahabreh, J. A. Steingrimsson, Biometrics 79, 2382–2393 (2023). 13. S. R. Kay, A. Fiszbein, L. A. Opler, Schizophr. Bull. 13, 261–276 (1987). 14. M. Lambert, A. Karow, S. Leucht, B. G. Schimmelmann, D. Naber, Dialogues Clin. Neurosci. 12, 393–407 (2010). 15. S. Leucht, J. Clin. Psychiatry 75, 8–14 (2014). 16. M. Obermeier et al., BMC Psychiatry 11, 113 (2011). 17. J. Busner, S. D. Targum, Psychiatry 4, 28–37 (2007). 18. S. Leucht, M. Samara, S. Heres, J. M. Davis, Schizophr. Bull. 42, S90–S94 (2016). 19. W. Guy, Ecdeu Assessment Manual for Psychopharmacology (The George Washington University, 1976). 20. G. M. Simpson, J. W. Angus, Acta Psychiatr Scand Suppl. 45, 11–19 (1970). 21. H. Zou, T. Hastie, J. R. Stat. Soc. Series B Stat. Methodol. 67, 301–320 (2005). 22. J. Friedman, T. Hastie, R. Tibshirani, J. Stat. Softw. 33, 1–22 (2010). 23. A. M. Chekroud et al., Psychiatr. Serv. 69, 927–934 (2018). 24. A. M. Chekroud et al., JAMA Psychiatry 74, 370–378 (2017). 25. Z. D. Cohen, R. J. DeRubeis, Annu Rev Clin Psychol. 14, 209–236 (2018). 26. K. H. Brodersen, C. S. Ong, K. E. Stephan, J. M. Buhmann, in 2010 20th International Conference on Pattern Recognition (2010), pp. 3121–3124. 27. N. Koutsouleris et al., Lancet Psychiatry 3, 935–946 (2016). 28. Y. Benjamini, Y. Hochberg, J. R. Stat. Soc. B 57, 289–300 (1995). 29. R Core Team, R: A language and environment for statistical computing (R Foundation for Statistical Computing, 2021); www.R-project.org/. 30. M. Kuhn, J. Stat. Softw. 28, 1 (2008). 31. A. J. Fisher, J. D. Medaglia, B. F. Jeronimus, Proc. Natl. Acad. Sci. U.S.A. 115, E6106–E6115 (2018). 32. A. J. Fisher, P. Soyster, Generating Accurate Personalized Predictions of Future Behavior: A Smoking Exemplar. PsyArXiv e24v6 [Preprint] (2019); doi:10.31234/osf.io/e24v6 33. A. J. Fisher, J. F. Boswell, Assessment 23, 496–506 (2016). 34. A. J. Fisher et al., Behav. Res. Ther. 116, 69–79 (2019).

AC KNOWLED GME NTS

g

It is worth considering how we might improve the situation in the future. From a statistical modeling perspective, capturing important heterogeneity through phenotyping or stratification procedures might help improve the generalizability of models. Identifying triallevel characteristics that relate to patient outcomes may provide information to better equip prediction models to generalize across settings. Such trial-level variation can be studied using Bayesian approaches or recent techniques that incorporate replicability across contexts or populations into the algorithm training process (39). From a population perspective, there may be some patients for whom the choice of treatment has no impact on their clinical course, which represents an inherent limitation of predicting treatment outcomes. However, this could also be an opportunity for further improvement in identifying which patients have a wider range of potential outcomes and for whom selecting the optimal treatment would provide clinical benefit (40). Longitudinal validation methods, in which a validation sample is drawn from the same population at a later point in time, may provide a limited but pragmatic path to avoid generalizing from one clinical setting to another. The growth of large mental health care delivery systems provides the opportunity to collect large amounts of data and deploy prediction models in the same setting in which they were developed (41). This strategy can reduce challenges associated with patient heterogeneity and context-dependence, and also help identify temporal or geographic trends that affect a model’s predictions. However, when a model is trained and validated on samples from the same population, it may perform well in that specific context but fail when applied to a different population with different characteristics.

35. A. M. Chekroud, N. Koutsouleris, Mol. Psychiatry 23, 24–25 (2018). 36. A. M. Chekroud, JAMA Psychiatry 74, 1183–1184 (2017). 37. M. P. Paulus, JAMA Psychiatry 74, 1185–1186 (2017). 38. R. A. McCutcheon et al., Biol. Psychiatry 94, 561–568 (2023). 39. P. Patil, G. Parmigiani, Proc. Natl. Acad. Sci. U.S.A. 115, 2578–2583 (2018). 40. R. J. DeRubeis et al., PLOS ONE 9, e83875 (2014). 41. J. Bondar et al., JAMA Netw. Open 5, e2216349 (2022).

p

RE FERENCES AND NOTES

Improving model generalizability

Chekroud et al., Science 383, 164–167 (2024)

Conclusions

RES EARCH

OPTICS

Free-electron interaction with nonlinear optical states in microresonators Yujia Yang1,2*†, Jan-Wilke Henke3,4†, Arslan S. Raja1,2†, F. Jasmin Kappert3,4†, Guanhao Huang1,2, Germaine Arend3,4, Zheru Qiu1,2, Armin Feist3,4, Rui Ning Wang1,2, Aleksandr Tusnin1,2, Alexey Tikan1,2, Claus Ropers3,4*, Tobias J. Kippenberg1,2* The short de Broglie wavelength and strong interaction empower free electrons to probe structures and excitations in materials and biomolecules. Recently, electron-photon interactions have enabled new optical manipulation schemes for electron beams. In this work, we demonstrate the interaction of electrons with nonlinear optical states inside a photonic chip–based microresonator. Optical parametric processes give rise to spatiotemporal pattern formation corresponding to coherent or incoherent optical frequency combs. We couple such “microcombs” to electron beams, demonstrate their fingerprints in the electron spectra, and achieve ultrafast temporal gating of the electron beam. Our work demonstrates the ability to access solitons inside an electron microscope and extends the use of microcombs to spatiotemporal control of electrons for imaging and spectroscopy.

1 of 6

,

12 January 2024

y

Yang et al., Science 383, 168–173 (2024)

y g

*Corresponding author. Email: [email protected] (Y.Y.); [email protected] (C.R.); [email protected] (T.J.K.) †These authors contributed equally to this work.

We studied and harnessed nonlinear optical effects in a transmission electron microscope (TEM) using ultralow-loss Si3N4 photonic chip– based microresonators (24, 25). The fiberpackaged photonic chip was placed in the TEM with the quasi-monochromatic electron beam (e-beam) passing over the microresonator surface in an aloof geometry (Fig. 1A). The microresonator (Fig. 1B) features a high Q of ~3.0 × 106 and anomalous group velocity dispersion to resonantly enhance the intracavity field and facilitate nonlinear frequency mixing. When pumped by a continuous-wave (CW) laser above the parametric oscillation threshold (32), cascaded four-wave mixing (FWM) leads to the generation of incoherent or coherent optical frequency combs corresponding to various dissipative structures (4) attainable by scanning the pump laser frequency from the blue to the red side of the resonance—changing the frequency difference or detuning between the pump laser and the cavity resonance (Fig. 1C). In particular, DKSs exist in the bistability regime of the resonance tilt caused by Kerr nonlinearity. In the time domain, nonlinear dissipative structures

where z is the coordinate in the e-beam direction, (x, y) is the transverse coordinate, e is the electron charge, ℏ is Planck’s constant h divided by 2p, w0 is the angular frequency of the optical carrier wave, and v is the electron velocity. In a simplified picture, a continuous e-beam samples the electric field of the intracavity waveforms at random arrival times of the electrons within the beam. This results in a characteristic electron spectral shape for each nonlinear optical intracavity state (Fig. 1D). The experimental setup is illustrated in Fig. 2A. We transiently generated nonlinear optical states in the transverse-electric mode family by continuously scanning the pump laser frequency across a cavity resonance near 1551.5 nm and recorded electron energy spectra (200 keV center energy) in parallel. Several oscilloscope traces of the nonlinearly generated light and the detuning-dependent electron spectra are shown in Fig. 2B. The oscilloscope traces bear the characteristics of a detuning scan for a Kerr nonlinear cavity with anomalous dispersion, exhibiting a typical “soliton step” (5). We identified five regions corresponding to distinct intracavity states, agreeing with the simulated stability chart (Fig. 2C) in terms of the pump power and detuning: (i) At low power, nonlinear optical responses are negligible, leading to a monochromatic CW intracavity field; (ii) with an increased detuning, and thus a higher intracavity power, cascaded FWM generates new frequency components and an amplitude modulation of the intracavity field, known as a Turing pattern; (iii) when further red-detuning the pump, the intracavity state enters the chaotic MI regime, corresponding to a disordered, rapidly varying waveform and an incoherent Kerr frequency comb; (iv) when tuning the pump to the soliton step, localized structures are spontaneously formed, which at small detunings consist of breathing solitons with periodically oscillating temporal and spectral shapes; and (v) with an increased detuning on the soliton step, shape-invariant stable DKSs are generated, featuring one or multiple femtosecond temporal pulses and a coherent frequency comb.

y

Institute of Physics, Swiss Federal Institute of Technology Lausanne (EPFL), CH-1015 Lausanne, Switzerland. 2Center for Quantum Science and Engineering, EPFL, CH-1015 Lausanne, Switzerland. 3Department of Ultrafast Dynamics, Max Planck Institute for Multidisciplinary Sciences, D-37077 Göttingen, Germany. 4Georg-August-Universität Göttingen, D-37077 Göttingen, Germany.

Transient observation of electron spectral modulation by nonlinear optical states

w0 e þ∞ Ez ðx; y; z′Þei v z′ dz′ ð1Þ ∫ ∞ 2ℏw0

g

1

taneous parametric generation of electronphoton pairs (25). However, only the linear cavity response was exploited to enhance the intracavity field and the electron-light coupling, whereas nonlinear optical responses have been predicted to endow electron-light interaction with new capabilities (26, 27). Kerr microresonators support diverse intracavity nonlinear dissipative structures, including DKSs (5), Turing patterns (28), chaotic modulation instability (MI) (29), breathing solitons (30), and soliton crystals (31). We studied the in situ coupling of electron beams with such spatiotemporal patterns.

g ðx; yÞ ¼ 

p

N

onlinear optical phenomena are prevalent in science and technology, allowing broadband supercontinuum for optical frequency metrology (1), squeezed light for gravitational-wave astronomy (2), and entangled photons for quantum information science (3). Over the past decade, triggered by advances in ultralow-loss (ultrahigh quality factor Q) photonic microresonators, continuouswave driven microresonators with Kerr nonlinearity (c(3)) have given rise to a host of “dissipative structures”—namely, self-organizing spatiotemporal light patterns in nonlinear optical systems (4). In particular, dissipative Kerr solitons (DKSs) (5) lead to coherent optical frequency combs with widespread applications, from atomic clocks (6) and telecommunications (7) to photonic computing (8) and astrophysical spectroscopy (9). Photonic structures mediating electron-photon interactions have brought new optical manipulations to electron beams, enabling high space/ time/energy–resolution electron microscopy (10–12), quantum-coherent optical modulation and probing (13–17), longitudinal and transverse electron beam shaping (18–20), dielectric laser acceleration (21), and electron-driven light sources (22, 23). Recently, high-Q silicon nitride (Si3N4) photonic chip–based microresonators have been used to demonstrate continuousbeam electron phase modulation (24) and spon-

correspond to the spontaneous formation of diverse spatiotemporal patterns (Fig. 1D). Electrons traversing the optical evanescent near field of the air-cladded microresonator undergo inelastic electron-light scattering (IELS) by absorbing and emitting photons, leading to “photon sidebands” in the electron spectrum spaced by the photon energy (10). The sideband amplitude and the overall spectral width are determined by a coupling parameter g (33), which is sensitive to the electric field along the e-beam direction

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

C

Fiber Power

Photonic chip-based microresonator

Microresonator

Time

Light output

0.5 mm

in g tu n de bl ue

0.8

tiv e

e e- energy

0.6

Un sta ble

Electron gun -

Pump laser tuning direction 1

Ef fe c

B

Normalized intracavity power (arb. u.)

A

0.4 0.2

Effective red 0 -5

0

detuning 10

5

Power

Normalized detuning (arb. u.)

D

Continuous wave

Time

Dissipative Kerr soliton

Chaotic modulation instability

Light input

e- energy

w0

w0

 ein c Rf ei v z′ dz′

ð2Þ

Yang et al., Science 383, 168–173 (2024)

Fingerprints of dissipative structures in Ramsey-type interference

We gained further insights into the interaction of electrons with nonlinear dissipative structures by investigating the Ramsey-type interferences from two sequential interactions with the microresonator. The e-beam intersects the waveguide near field twice (Fig. 3A), result-

12 January 2024

ing in constructive or destructive interference owing to a phase delay between the two interactions (24, 35). We changed this phase delay by scanning the e-beam position along the chip surface and acquired electron spectra while maintaining a given intracavity state (Fig. 3). For dissipative structures, the evolution of both the optical carrier and envelope between the two interaction regions provides more complex free-electron modulations than the coherent phase modulation at a uniform field strength. For the monochromatic CW field (Fig. 3B), the electron spectral width oscillation along the radial position has the typical pattern of previously reported Ramsey-type interferences (Fig. 3D) (24). The oscillation nodes or destructive interference arises from a p-phase shift of the optical field at the two interactions. For a Turing pattern, the generated frequency components and the accompanying amplitude modulation (Fig. 3C) alter the electron spectrum. The electron spectrum (Fig. 3E) has a central, low-energy-change part of the spectrum that resembles the double-peak shape from a CW interaction but with a narrower spectral width caused by a reduced intracavity pump 2 of 6

,

JN is the Bessel function of the first kind. We obtained continuous e-beam spectra by means of an incoherent averaging over the angular offset f0 and the slow time t (34).

The simulated intracavity waveforms and optical spectra are shown in Fig. 2D, as well as the measured optical and electron spectra for four intracavity states. Electron spectra in Fig. 2, B and D, demonstrate disparate characteristic shapes for different states. For the CW state, the electron spectrum has the typical double-peak shape from coherent phase modulation. For the Turing pattern and chaotic MI, the electron spectrum broadens because of increasing intracavity power. A Gaussianshaped background appears, and the double peaks reduce in intensity. We attribute these observations to the intracavity field strength no longer being uniform. In the stable DKS state stochastically generated by the detuning scan, the electron spectrum features a strong, narrow central peak on a weak, broad plateau.

y

where A(f, t) is the intracavity waveform in the frame rotating at the optical group velocity with the microresonator angular coordinate f and the slow time t (4); D1/2p is the microresonator free spectral range (FSR); ef and er are the tangential and radial optical modal fields, respectively; n is the linear refractive index; c is the speed of light; R is the microresonator ring radius; and f0 is the angular offset of the rotating frame from the laboratory frame. The coupling parameter further determines the electron spectrum through the Nth photon sideh iN Þ band amplitude JN ½2jg ðx; yÞj jgg ððx;y x;yÞj , where

detuning showing a resonance tilt and bistability from the Kerr effect. (D) Illustration of intracavity waveforms and post-interaction electron spectra for CW, chaotic MI, and DKS states. The electron spectral broadening originates from incoherent summation of electrons that interact with the intracavity field at different times.

y g

For electrons interacting with these states, the coupling parameter (Eq. 1) becomes
 e þ∞ z′ ; t A f þ f þ D g ðx; yÞ ¼  1 0 2ℏw0 ∫∞ v    ef ðx; yÞcos f  er ðx; yÞsin f

Electron energy

y

Fig. 1. Free-electron interaction with nonlinear optical states. (A) Schematic of the experiment. Electrons in a TEM pass by a photonic chip–based nonlinear microresonator. Stimulated inelastic scattering between electrons and nonlinear optical states induces electron spectral broadening. (B) Photograph of the fiber-packaged Si3N4 photonic chip. (C) Intracavity power versus laser

Electron energy

Spectrum (log.)

Spectrum (log.)

Electron energy

g

Magnetic prism

Spectrum (log.)

p

Electron detector

Light output

ESA BS

PD FBG

OSC

PD

Si3N4

CIRC

FPC

AFG

EOM Trigger

Simulation 1

2

3

5

Resonator angular coordinate

1 0.5 1

0 -30

105

-20

104

-10 103

0 10

102

20

101

-100

2

3

4

5

0.1

CW Turing pattern Chaotic MI Breather Stable DKS

2.5

2

3.5

3

Simulation

Mode number

4

0.4

100

0.8

1.2

Detuning (GHz)

1490

Experiment

Wavelength (nm)

1620

-32

Experiment

Electron energy change (eV)

32

indicate regions corresponding to different intracavity states. (C) Simulated stability chart of the microresonator showing the two-dimensional parameter space of pump power and detuning. Colored regions indicate the existence ranges for the labeled states. The dashed arrow depicts the detuning scan in (B). (D) Simulated optical waveforms and spectra, as well as measured optical and electron spectra for (top to bottom, respectively) CW, Turing pattern, chaotic MI, and DKS states. The numbering and color-coding are equivalent to that in (B) and (C). Measured electron spectra are vertical slices of the electron spectrum in (B), indicated with the colored arrows.

,

3 of 6

y

where D denotes the pump detuning, k is the cavity decay rate, q0 is the soliton phase, and

D2 is the second-order dispersion (5). The observed corresponding electron spectra have a distinct shape featuring a strong, narrow, lowenergy-change region and a weak, broad plateau (Fig. 3I). The former exhibits the hallmark of a coherent phase modulation produced by the CW background. The latter has a low total spectral weight and a broad width originating from the interaction between electrons and the high-peak-power femtosecond soliton pulse. Because the DKS pulse duration (~100 fs) is much shorter than the round-trip time (~10 ps), only a small fraction of the electrons in the continuous e-beam interact with the pulse. Hence, the plateau is much lower than the central region. In addition, the plateau has a moderate, positiondependent spectral width oscillation, arising from the Ramsey interference of electrons interacting with the soliton pulse and the CW background one time each. This oscillation has the same period as that of the CW Ramsey pattern but with an offset that encodes the soliton phase (34). For all four intracavity states, experimental results were well reproduced with simulations

y g

and absence of nodes in the position-dependent scan (Fig. 3H), resulting from random intracavity field strengths that prohibit well-defined interference conditions. The e-beam modulated by the chaotic MI state possesses a Gaussian spectrum similar to that previously reported for thermal states (17). Rather than invoking photon statistics, our analysis attributes the spectral shape to the statistical fluctuations of the intracavity optical intensity, stochastically sampled by the electrons and averaged on the detector (34). In the stable DKS state, a single temporal soliton pulse circulates in the cavity on a weak CW background (Figs. 2D and 3A), forming a coherent, low-noise optical frequency comb with a sech2 spectral shape (Fig. 3G). The analytical waveform is rffiffiffiffiffiffi
rffiffiffiffiffiffi  4D iq0 2D e sech f ð3Þ Y ≃ Y0 þ k D2

12 January 2024

0

Time (µs)

y

Yang et al., Science 383, 168–173 (2024)

1

30

Fig. 2. Electron spectra imprinted by nonlinear dissipative structures. (A) Experimental setup. OSA, optical spectrum analyzer; ESA, electronic spectrum analyzer; OSC, oscilloscope; VNA, vector network analyzer; AFG, arbitrary waveform generator; PD, photodetector; EOM, electro-optic modulator; EDFA, erbium-doped fiber amplifier; FBG, fiber Bragg grating; CIRC, optical circulator; FPC, fiber polarization controller; and BS, beam splitter. (B) (Top) Oscilloscope traces of the generated light and (bottom) the electron spectrum when scanning the pump laser frequency across a resonance. Electron spectra are summed over multiple scans. Dashed lines and numbers

power due to FWM. Additionally, the electron spectrum has a broad shoulder that is absent in the CW case. This shoulder arises from the interaction of electrons with the intensity peaks of the intracavity waveform (Fig. 3E, inset). For the chosen e-beam position, the temporal periodicity of the Turing pattern amplitude modulation is approximately commensurate with the travel time difference of the electron and the optical envelope between the two interaction regions (DTtravel ≈ NTTuring, N ∈ ℤ) (34). Hence, the electron experiences nearly equal optical intensities in the two interactions because of the time-translation invariance of the periodic Turing pattern; this leads to similar modulation strengths just like the CW case. Therefore, the spectral widths of the low-energychange part and the broad shoulder jointly increase and decrease. For the chaotic MI (incoherent Kerr comb) (Fig. 3F), the characteristic double peaks from uniform phase modulation disappear, and the electron spectrum has a smooth Gaussian shape. We further found a reduced spatial dependence

0.2

g

-

soliton step

2

p

Simulated intracavity waveform (arb. units)

D

EDFA

0.3

5

1.5

Simulated optical spectrum (80 dB/div)

CW pump

Light input

C

4

e- counts (arb. u.)

VNA

e-

3

Measured electron spectrum (50 dB/div)

TEM

OSA

2

Pump power (W)

B

Measured optical spectrum (80 dB/div)

A

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

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

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

CW pump e -e

ne rgy sh ift

0

-40

-80 1500

sition

1550

1550

1600

Wavelength (nm)

Sim.

-15

0

15

30 -30

interaction

e-

-15

0

15

30

-40

1550

1600

Wavelength (nm)

G

1650

I

Sim.

-30

-15

0

30 -30

15

-15

0

15

30

Electron energy change (eV)

Single DKS

0

-40

sech2

-80 1450

1.38

1500

1550

1600

Wavelength (nm)

Exp.

1650

Sim.

-50

-30

-15

0

15

30 -30

Electron energy change (eV)

-15

0

15

30

Electron energy change (eV)

y g

0 0

e- spectrum (dB)

Radial position (µm)

Radial position (µm)

Exp.

-50

y

1500

e-

0 0

Electron energy change (eV)

Chaotic MI

0

-80 1450

second interaction

Electron energy change (eV) Optical spectrum (dBm)

-30

e- spectrum (dB)

Radial position (µm) -50

1650

g

Optical spectrum (dBm)

1500

p

e- spectrum (dB)

0 0

H 1.38

e- spectrum (dB)

-40

E 1.38 Exp.

Sim.

Electron energy change (eV)

F

Turing pattern

Radial position (µm)

Intensity (dB) -50 0

0

-80 1450

1600

Wavelength (nm)

Radial po

D 1.38 Exp.

C

CW

Optical spectrum (dBm)

B

e-

Optical spectrum (dBm)

A

0 0

-50

-30

-15

0

15

Electron energy change (eV)

30 -30

-15

0

15

30

Electron energy change (eV)

y

by using the reduced time-dependent Schrödinger equation for electrons and the Lugiato-Lefever equation for nonlinear optics (Fig. 3). We also observed additional electron spectral features caused by multisoliton states, incommensurate Turing patterns, soliton pulse edges, and breathing solitons (supplementary materials) (34). Probing soliton dynamics with free electrons

We used free electrons to probe basic properties of DKSs existing in the bistability regime Yang et al., Science 383, 168–173 (2024)

(Sim.) Ramsey interference patterns for these states. For each interference pattern, two line cuts (green and blue dashed lines) are plotted at bottom. The schematic in (E) depicts snapshots of the Turing pattern at the two interactions for one exemplary electron arrival time. The sech2 fitting of the optical spectrum in (G) corresponds to a DKS pulse duration of ~98.5 fs.

of the Kerr microresonator (Fig. 4A). Adjusting the detuning D alters the DKS pulse duration, peak field, and background CW amplitude. Shown in Fig. 4B is a DKS state characterized by scanning the electro-optic-modulation sidebands of the pump by means of a vector network analyzer (VNA) (Fig. 2A) for three different detunings (D1, D2, and D3). The VNA traces illustrate the soliton (S) and cavity (C) resonances, with the latter revealing the detuning (36). The measured optical spectra (Fig. 4C) demonstrate

12 January 2024

the detuning-dependent DKS frequency comb width, with increasing detuning leading to a broader spectrum—that is, a shorter DKS pulse pffiffi qffiffiffiffi 2  D2D2 . duration tFWHM ≃ 2 arcosh D1 When measuring the electron spectrum, an increasing detuning D leads to a wider plateau, induced by an increasing DKS peak field pffiffiffiffiffiffiffiffiffiffi ffi (∼ 4D=k ) (Fig. 4D). Meanwhile, the plateau height slightly decreases because the DKS pulse duration tFWHM decreases, and fewer electrons 4 of 6

,

Fig. 3. Ramsey-type interference of nonlinear intracavity states. (A) Illustration of the double interactions between the e-beam and the microresonator supporting an exemplary DKS state. Position-dependent electron spectra are shown at bottom. (B, C, F, and G) Measured optical spectra for (B) CW, (C) Turing pattern, (F) chaotic MI, (G) and single-DKS states. (D, E, H, and I) Measured (Exp.) and simulated

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

B

Soliton branch

CW-lower branch

-40

-60

λ

400

800

-30 -40

Δ1 Δ2 Δ3

-60 -70

1200

1500

1550

1600

1650

Wavelength (nm)

E

400

110

300

200

2

Δ1 Δ2 Δ3

fit

-50

Frequency (MHz) e- counts

e- spectrum (dB)

0

Δ1 Δ2 Δ3

0

Detuning

D

Power (dBm)

-20

0 -2 2 0 Electron energy change (eV)

-25

|g max |

Δ

C

100

200

90 100

Pulse width (fs)

Laser (pump)

Power (dB)

Intracavity power (a. u.)

up C pe W rb ran ch

A

80

-50 -30

-20

-10

0

10

20

30

400

600

700

800

Detuning (MHz)

p

Electron energy change (eV)

500

12 January 2024

AC KNOWLED GME NTS

Photonic chips were fabricated in the Center of MicroNanoTechnology (CMi) and the Institute of Physics cleanroom at EPFL. We thank J. Liu for helping with the fabrication. Funding: This material is based on work supported by the Air Force Office of Scientific Research under award FA9550-19-1-0250 and by the Swiss National Science Foundation under grant agreement 185870 (Ambizione). Y.Y. acknowledges support from the EU H2020 research and innovation program under the Marie Skłodowska-Curie IF grant agreement 101033593 (SEPhIM). The experiments were conducted at the Göttingen UTEM Lab, funded by the Deutsche Forschungsgemeinschaft (DFG; German Research Foundation) through grant 432680300/SFB 1456 (project C01) and the Gottfried Wilhelm Leibniz program, and the EU H2020 research and innovation program under grant agreement 101017720 (FET-Proactive EBEAM).

5 of 6

,

1. J. M. Dudley, G. Genty, S. Coen, Rev. Mod. Phys. 78, 1135–1184 (2006). 2. H. Grote et al., Phys. Rev. Lett. 110, 181101 (2013). 3. P. G. Kwiat et al., Phys. Rev. Lett. 75, 4337–4341 (1995). 4. T. J. Kippenberg, A. L. Gaeta, M. Lipson, M. L. Gorodetsky, Science 361, eaan8083 (2018). 5. T. Herr et al., Nat. Photonics 8, 145–152 (2014). 6. S. B. Papp et al., Optica 1, 10 (2014). 7. P. Marin-Palomo et al., Nature 546, 274–279 (2017). 8. J. Feldmann et al., Nature 589, 52–58 (2021). 9. E. Obrzud et al., Nat. Photonics 13, 31–35 (2019). 10. B. Barwick, D. J. Flannigan, A. H. Zewail, Nature 462, 902–906 (2009). 11. D. Nabben, J. Kuttruff, L. Stolz, A. Ryabov, P. Baum, Nature 619, 63–67 (2023). 12. A. Polman, M. Kociak, F. J. García de Abajo, Nat. Mater. 18, 1158–1171 (2019). 13. A. Feist et al., Nature 521, 200–203 (2015). 14. K. Wang et al., Nature 582, 50–54 (2020). 15. O. Kfir et al., Nature 582, 46–49 (2020). 16. V. Di Giulio, M. Kociak, F. J. G. de Abajo, Optica 6, 1524 (2019). 17. R. Dahan et al., Science 373, eabj7128 (2021). 18. K. E. Priebe et al., Nat. Photonics 11, 793–797 (2017). 19. G. M. Vanacore et al., Nat. Mater. 18, 573–579 (2019). 20. A. Konečná, F. J. G. de Abajo, Phys. Rev. Lett. 125, 030801 (2020).

y

Optical manipulation of free electrons is extended beyond the regimes of pulsed or continuouswave lasers. Bringing free electron–light interaction to the nonlinear optics regime opens up the potential to noninvasively probe nonlinear optical dynamics and devices with nanometerfemtosecond spatiotemporal resolution and direct access to the intracavity field. The integrated photonics toolbox provides diversity and flexibility for on-chip arbitrary optical waveform generation and frequency conversion, promising advanced electron control schemes with flexible and programmable sideband amplitudes. Furthermore, we achieved ultrafast electronlight interaction with chip-based femtosecond temporal solitons in the absence of pulsed lasers or electron sources. This approach facilitates ultrafast electron microscopy in a conventional TEM equipped with a photonic chip and a CW laser, using temporal photon-gating (37, 38)

21. R. J. England et al., Rev. Mod. Phys. 86, 1337–1389 (2014). 22. G. Adamo et al., Phys. Rev. Lett. 103, 113901 (2009). 23. M. Taleb, M. Hentschel, K. Rossnagel, H. Giessen, N. Talebi, Nat. Phys. 19, 869–876 (2023). 24. J.-W. Henke et al., Nature 600, 653–658 (2021). 25. A. Feist et al., Science 377, 777–780 (2022). 26. A. Konečná, V. Di Giulio, V. Mkhitaryan, C. Ropers, F. J. García de Abajo, ACS Photonics 7, 1290–1296 (2020). 27. F. J. García de Abajo, E. J. C. Dias, V. Di Giulio, Phys. Rev. Lett. 129, 093401 (2022). 28. S.-W. Huang et al., Phys. Rev. X 7, 041002 (2017). 29. T. Hansson, D. Modotto, S. Wabnitz, Phys. Rev. A 88, 023819 (2013). 30. M. Yu et al., Nat. Commun. 8, 14569 (2017). 31. D. C. Cole, E. S. Lamb, P. Del’Haye, S. A. Diddams, S. B. Papp, Nat. Photonics 11, 671–676 (2017). 32. T. J. Kippenberg, S. M. Spillane, K. J. Vahala, Phys. Rev. Lett. 93, 083904 (2004). 33. S. T. Park, M. Lin, A. H. Zewail, New J. Phys. 12, 123028 (2010). 34. Materials and methods are available as supplementary materials. 35. K. E. Echternkamp, A. Feist, S. Schäfer, C. Ropers, Nat. Phys. 12, 1000–1004 (2016). 36. H. Guo et al., Nat. Phys. 13, 94–102 (2017). 37. M. T. Hassan, H. Liu, J. S. Baskin, A. H. Zewail, Proc. Natl. Acad. Sci. U.S.A. 112, 12944–12949 (2015). 38. X. Fu et al., Nat. Commun. 11, 5770 (2020). 39. Y. Yang et al., Figure data for: Free-electron interaction with nonlinear optical states in microresonators. Zenodo (2023); https://doi.org/10.5281/zenodo.10104801.

y g

RE FERENCES AND NOTES

Discussion and outlook

Yang et al., Science 383, 168–173 (2024)

by use of DKSs instead of mode-locked lasers. Proper spectral filtering will lead to a train of electron pulses with 12.4, the MnV/IV potential was less well defined because of the redox instability of the electrolytes; however, the data fit reasonably well with a pKa-independent redox couple, corresponding to [MnV=O]+/[MnIV=O]. These data, together Hoque et al., Science 383, 173–178 (2024)

monooxygenase reactions. The intermediate MnV=O species can undergo oxygen atom transfer or proton-coupled reduction to H2O. (E) Influence of acid source and NMI on Ph2S oxidation during constant current electrolysis (5 mA) in an H-type divided cell. A 38.6 C charge was passed to attain 20% theoretical conversion. Faradaic yield is the charge used to form product (2 e−/mol Ph2SO, 4 e−/mol Ph2SO2)/total charge passed; TON is the turnover number (mmol product/mmol 1-Cl). FEc, FE in the cathode (sum of the Faradaic yields); RVC, reticulated vitreous carbon. *0.1 mM 1-Cl (0.1 mol%).

with the observed onset potentials for catalytic substrate oxidation, provide the basis for the catalytic mechanism shown in Fig. 3E. The substrate-dependent variations in catalytic onset potentials suggest that OAT can be initiated by either MnIV-(hydr)oxo or MnV-oxo species. Both pathways enable OAT to the or-

12 January 2024

ganic substrate with no competition from H2O oxidation to O2. Net dioxygenase activity

The results presented in Figs. 2 and 3 show that the same Mn(TPP) catalyst supports both monooxygenase reactivity through reductive 3 of 6

,

Fig. 2. Analysis of electrochemical monooxygenase reactivity. (A) Electrochemical reductive activation of O2 for monooxygenase-type oxygenation of Ph2S. (B) Cyclic voltammograms of 1 mM 1-Cl in acetonitrile under N2 (black), 1 atm O2/200 mM AcOH (blue), and 1 atm O2/200 mM AcOH/100 mM Ph2S (red), all with 0.1 M n-Bu4NPF6. Scan rate = 50 mV/s. (C) Constant potential electrolysis of 1 mM 1-Cl (1 mol%) in a divided cell, 1 atm O2/200 mM AcOH/100 mM Ph2S in MeCN (red), and 1 atm O2/200 mM AcOH in MeCN (blue). (D) Proposed mechanism for electrochemical

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

p g y y g y ,

Fig. 3. Analysis of electrochemical dehydrogenase reactivity. (A) Electrochemical oxidative activation of H2O for dehydrogenase-type oxygenation of Ph2S. (B) Cyclic voltammograms of 1 mM 1-Cl in acetonitrile with 1 M H2O in the absence of 100 mM Ph2S (black) and in the presence of 100 mM Ph2S (red), all with 0.1 M n-Bu4NPF6. Scan rate = 50 mV/s. (C) Influence of H2O and NMI on Ph2S oxidation during constant

activation of O2 at the cathode and dehydrogenase reactivity through H2O oxidation at the anode. Because H2O is a stoichiometric byproduct in the cathodic reaction and a substrate in the anodic reaction, pairing of these two reactions within a single electrochemical cell would support net dioxygenase reactivity using one molecule of O2 to support two OAT reactions (Fig. 4A). Paired electrolysis perforHoque et al., Science 383, 173–178 (2024)

current electrolysis (5 mA) in an H-type divided cell. A 38.6 C charge was passed to attain 20% theoretical conversion. FEa, FE in the anode (sum of the Faradaic yields). *0.1 mM 1-Cl (0.1 mol%). (D) Nonaqueous Pourbaix diagram with various organic buffer solutions (see fitting details in section 8.3 of the supplementary materials). (E) Proposed mechanism for electrochemical dehydrogenase reactions.

mance was evaluated by using a “combined FE” metric that accounts for the charge associated with the formation of sulfoxide (2 e–) and sulfone (4 e–) at either electrode, divided by the total charge passed in the electrochemical circuit. FEs (i.e., Faradaic yields) of 200% are theoretically possible according to this metric because passing two electrons through the circuit can generate up to two equivalents of

12 January 2024

sulfoxide. Although this metric is used in the literature (16, 18) and is helpful for “bookkeeping,” there is no consumption or generation of electrons in the net dioxygenase reaction (compare Fig. 1B and see further discussion below). To test this possibility, the Mn(TPP) catalyst 1-Cl and NMI were combined with the Ph2S substrate, H2O (1 M), and AcOH (200 mM) in an undivided cell and placed under 1 atm of O2. 4 of 6

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

p g y y g y ,

Fig. 4. Analysis of electrochemical dioxygenase reactivity. (A) Electrochemical reductive activation of O2 and oxidative activation of H2O for oxygenation of Ph2S. (B) Linear paired electrolysis test for Ph2S oxidation during constant current electrolysis (5 mA) in an H-type divided cell. A 38.6 C charge was passed to attain 20% theoretical conversion in each compartment. The cathode compartment

The AcOH was included to support O2 reduction at the cathode (compare Fig. 2); control experiments showed that AcOH does not interfere with the anodic OAT reaction (see section 7.2 of the supplementary materials). Conducting the reaction with a constant current of 5 mA led to the formation of Ph2SO as the major product with small amounts of Hoque et al., Science 383, 173–178 (2024)

was charged with 1 atm O2. FEtotal, FE of the cell (sum of the Faradaic yields in both the cathode and anode). *Reaction performed in an undivided cell under identical reaction conditions to entry 2. (C) Preparative scale demonstration of linear paired electrolysis. (D) Energetic analysis of dioxygenase reactivity.

Ph2SO2; 140% FE was observed (138:2% Faradaic yield of Ph2SO:Ph2SO2) (Fig. 4B, entry 1). The lack of side product formation suggests that the FE is lowered by redox cycling of the Mn catalyst at the cathode and anode, passing charge without oxidizing the substrate. This problem was avoided by using a divided cell, because 195% FE was observed

12 January 2024

when conducting the same experiment in an H cell equipped with a glass frit divider (Fig. 4B, entry 2, and see photograph of the reaction cell). Further improvement was achieved when NMI was not added to the anode compartment (compare Fig. 3): 200% FE was observed (197:3% for Ph2SO:Ph2SO2), reflecting perfect utilization of electrons in the double-OAT 5 of 6

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

23. G. Passard, D. K. Dogutan, M. Qiu, C. Costentin, D. G. Nocera, ACS Catal. 8, 8671–8679 (2018). 24. N. Kobayashi, H. Saiki, T. Osa, Chem. Lett. 14, 1917–1920 (1985). 25. R. L. Shook et al., J. Am. Chem. Soc. 133, 5810–5817 (2011). 26. E. N. Cook, C. W. Machan, Chem. Commun. 58, 11746–11761 (2022). 27. K.-C. Cheung, W.-L. Wong, D.-L. Ma, T.-S. Lai, K.-Y. Wong, Coord. Chem. Rev. 251, 2367–2385 (2007). 28. A. K. Vannucci, Z. Chen, J. J. Concepcion, T. J. Meyer, ACS Catal. 2, 716–719 (2012). 29. A. Das, J. E. Nutting, S. S. Stahl, Chem. Sci. 10, 7542–7548 (2019). 30. K. Jin et al., J. Am. Chem. Soc. 141, 6413–6418 (2019). 31. B. Chandra, K. M. Hellan, S. Pattanayak, S. S. Gupta, Chem. Sci. 11, 11877–11885 (2020). 32. S. G. Robinson, J. B. C. Mack, S. N. Alektiar, J. Du Bois, M. S. Sigman, Org. Lett. 22, 7060–7063 (2020). 33. C. J. Gagliardi, A. K. Vannucci, J. J. Concepcion, Z. Chen, T. J. Meyer, Energy Environ. Sci. 5, 7704–7717 (2012). 34. Y. Naruta, M. Sasayama, T. Sasaki, Angew. Chem. Int. Ed. 33, 1839–1841 (1994). 35. X. Li et al., J. Am. Chem. Soc. 143, 14613–14621 (2021). 36. D. Wang, J. T. Groves, Inorg. Chem. 61, 13667–13672 (2022). 37. B. D. McCarthy, J. L. Dempsey, Inorg. Chem. 56, 1225–1231 (2017). 38. J. Rosenthal, B. J. Pistorio, L. L. Chng, D. G. Nocera, J. Org. Chem. 70, 1885–1888 (2005). 39. H. Tanaka, H. Nishikawa, T. Uchida, T. Katsuki, J. Am. Chem. Soc. 132, 12034–12041 (2010). AC KNOWLED GME NTS

RE FERENCES AND NOTES

y g y ,

12 January 2024

We thank I. A. Guzei and K. M. Sanders for assistance with x-ray crystallographic characterization and D. T. Hofsommer for helpful discussions. Funding: This work was initially supported by the Center for Molecular Electrocatalysis, an Energy Frontier Research Center funded by the US Department of Energy, Office of Science, Office of Basic Energy Sciences, and continued with support from the National Institutes of Health (NIH grant R35 GM134929). The spectrometers were supported by the National Science Foundation (NSF grant CHE-1048642) and by a generous gift from Paul J. and Margaret M. Bender. The mass spectrometer was supported by the NIH (grant S10 OD020022). The x-ray diffractometer was partially supported by the NSF (grant CHE1919350 to the Univeristy of Wisconsin−Madison Department of Chemistry). Author contributions: This project was conceived by S.S.S. in collaboration with Md.A.H. and J.B.G. All the work was done by Md.A.H. in consultation with J.B.G. and S.S.S. All authors contributed to the preparation of the manuscript. Competing interests: The authors declare no competing interests. Data and materials availability: X-ray structural data with access code CCDC 2289621 for 1-Cl and 2289622 for 1-OH2 are available free of charge from the Cambridge Crystallographic Data Centre (www. ccdc.cam.ac.uk/data_request/cif; Email: [email protected]. ac.uk). All other data are available in the main text or the supplementary materials. License information: Copyright © 2024 the authors, some rights reserved; exclusive licensee American Association for the Advancement of Science. No claim to original US government works. https://www.science.org/about/sciencelicenses-journal-article-reuse

y

1. A. G. Stamoulis, D. L. Bruns, S. S. Stahl, J. Am. Chem. Soc. 145, 17515–17526 (2023). 2. T. Pu, H. Tian, M. E. Ford, S. Rangarajan, I. E. Wachs, ACS Catal. 9, 10727–10750 (2019). 3. J. T. Groves, R. Quinn, J. Am. Chem. Soc. 107, 5790–5792 (1985). 4. B. J. Pistorio, C. J. Chang, D. G. Nocera, J. Am. Chem. Soc. 124, 7884–7885 (2002). 5. S. Koya, Y. Nishioka, H. Mizoguchi, T. Uchida, T. Katsuki, Angew. Chem. Int. Ed. 51, 8243–8246 (2012). 6. T. D. H. Bugg, Tetrahedron 59, 7075–7101 (2003). 7. E. G. Kovaleva, J. D. Lipscomb, Nat. Chem. Biol. 4, 186–193 (2008). 8. F. P. Guengerich, F. K. Yoshimoto, Chem. Rev. 118, 6573–6655 (2018). 9. A. L. Feig, S. J. Lippard, Chem. Rev. 94, 759–805 (1994). 10. I. G. Denisov, T. M. Makris, S. G. Sligar, I. Schlichting, Chem. Rev. 105, 2253–2278 (2005). 11. D. E. Torres Pazmiño, M. Winkler, A. Glieder, M. W. Fraaije, J. Biotechnol. 146, 9–24 (2010). 12. S.-S. Gao, N. Naowarojna, R. Cheng, X. Liu, P. Liu, Nat. Prod. Rep. 35, 792–837 (2018). 13. S. M. Barry, G. L. Challis, ACS Catal. 3, 2362–2370 (2013). 14. M. L. Pegis, C. F. Wise, D. J. Martin, J. M. Mayer, Chem. Rev. 118, 2340–2391 (2018). 15. R. Matheu et al., Nat. Rev. Chem. 3, 331–341 (2019). 16. W. Li, T. Nonaka, J. Electrochem. Soc. 146, 592–599 (1999). 17. J. G. Ibanez, B. A. Frontana-Uribe, R. Vasquez-Medrano, J. Mex. Chem. Soc. 60, 247–260 (2016). 18. J. Strehl, M. L. Abraham, G. Hilt, Angew. Chem. Int. Ed. 60, 9996–10000 (2021). 19. H. Sheng et al., Nat. Catal. 5, 716–725 (2022). 20. S. E. Creager, S. A. Raybuck, R. W. Murray, J. Am. Chem. Soc. 108, 4225–4227 (1986). 21. P. Leduc, P. Battioni, J. F. Bartoli, D. Mansuy, Tetrahedron Lett. 29, 205–208 (1988). 22. M. Mukherjee, A. Dey, ACS Cent. Sci. 5, 671–682 (2019).

g

Hoque et al., Science 383, 173–178 (2024)

The analysis outlined above provides a crucial bridge between the reactivity of molecular oxygen in synthetic and energy conversion applications. Electrochemical metrics such as overpotential and 2- versus 4-e– reaction selectivity are widely used in the development of ORR and OER electrocatalysts for fuel cells and electrolyzers. The present study shows how similar metrics may be used to guide the development of (electro)catalysts for chemical synthesis, enabling a rare demonstration of dioxygenase reactivity. Sulfide oxidation was chosen as a prototypical example of a “highpotential” substrate, but many other oxidation reactions, including C–H hydroxylation and alkene epoxidation, feature similar challenges (compare Fig. 1A). Dioxygenase reactivity with each of these substrate classes is thermodynamically possible but faces substantial kinetic barriers associated with O2 activation and subsequent OAT. The analysis introduced here provides a foundation for the systematic design and optimization of new catalysts that achieve better efficiency and access effective dioxygenase reactivity with diverse substrates.

p

process. Excellent FE was also achieved on a preparative (1 mmol) scale, evident from the 189 to 196% FE observed for oxidation of a diaryl, a dialkyl, and an aryl alkyl sulfide (Fig. 4C). The net stoichiometry of this electrochemical process corresponds to a dioxygenase reaction in which O2 adds an oxygen atom to two equivalents of organic substrate. Electrons are neither generated nor consumed in this net reaction, but the electrochemical potential provides the energy input needed to overcome the kinetic barriers that limit thermal reactivity. This energy input may be quantified by analyzing the cathodic and anodic redox potentials (Fig. 4D). CV analysis reveals an onset potential at –0.59 V versus Fc+/Fc for O2 reductive activation and at +0.65 V versus Fc+/Fc for H2O oxidation (the onset potential was defined as Icat = 20 mA). The difference of 1.24 V corresponds to a free energy of +28.6 kcal/mol used to promote the dioxygenase reaction. The Pourbaix diagram in Fig. 3D suggests that it might be possible to reduce the energy gap by using an electrolyte with a higher pKa. With an acetic acid/acetate (2:1) electrolyte, the cathodic and anodic onset potentials shifted to –0.69 and +0.29 V versus Fc+/Fc, respectively, while retaining quantitative current efficiency (Fig. 4D). This 0.98 V gap reflects a reduced energy input of 22.6 kcal/mol (for similar analysis with buffering electrolytes with intermediate pKa values, see figs. S17 to S20). The thermodynamic potential for Me2SO/Me2S under these buffered conditions is E°′ = –0.15 V. Thus, the overpotentials for thioether oxidation are approximately balanced between the two half reactions, with a slightly higher value for the cathodic monooxygenase step (DE ~ 0.54 V) relative to the anodic dehydrogenase step (DE ~ 0.44 V) (Fig. 4D). This analysis permits comparison to photochemical methods that have been used to promote dioxygenase-like reactivity. An oxo-bridged cofacial bis-FeIII-porphyrin has been shown to support oxidation of dimethyl sulfide by O2 in a 2:1 ratio under irradiation with 365 to 425 nm light, corresponding to an energy input of 66.9 to 78.4 kcal/mol (4, 38). In addition to the three- to fourfold higher energy input, this process features a quantum yield of only 10−4 to 10−8, which may be compared to the quantitative FE of the electrochemical process. Another method for photochemically promoted aerobic oxidation of thioethers features a Ru-salen catalyst under 467 nm irradiation (61.2 kcal/mol), although the lack of a reported quantum efficiency prevents quantitative comparison (39).

SUPPLEMENTARY MATERIALS

science.org/doi/10.1126/science.adk5097 Materials and Methods Figs. S1 to S20 Tables S1 to S15 NMR Spectra References (40–59) Submitted 2 September 2023; accepted 4 December 2023 10.1126/science.adk5097

6 of 6

RES EARCH

MEMBRANES

A molecular sieve with ultrafast adsorption kinetics for propylene separation Jiyu Cui1, Zhaoqiang Zhang3, Lifeng Yang1*, Jianbo Hu2, Anye Jin1, Zhenglu Yang1, Yue Zhao4, Biao Meng5, Yu Zhou5, Jun Wang5, Yun Su6, Jun Wang6, Xili Cui1,2, Huabin Xing1,2* The design of molecular sieves is vital for gas separation, but it suffers from a long-standing issue of slow adsorption kinetics due to the intrinsic contradiction between molecular sieving and diffusion within restricted nanopores. We report a molecular sieve ZU-609 with local sieving channels that feature molecular sieving gates and rapid diffusion channels. The precise cross-sectional cutoff of molecular sieving gates enables the exclusion of propane from propylene. The coexisting large channels constituted by sulfonic anions and helically arranged metal-organic architectures allow the fast adsorption kinetics of propylene, and the measured propylene diffusion coefficient in ZU-609 is one to two orders of magnitude higher than previous molecular sieves. Propylene with 99.9% purity is obtained through breakthrough experiments with a productivity of 32.2 L kg−1.

Equilibrium propylene/propane separation performance

,

12 January 2024

y

Cui et al., Science 383, 179–183 (2024)

y g

*Corresponding author. Email: [email protected] (H.X.); [email protected] (L.Y.)

y

Key Laboratory of Biomass Chemical Engineering of Ministry of Education, College of Chemical and Biological Engineering, Zhejiang University, Hangzhou 310012, P.R. China. 2Engineering Research Center of Functional Materials Intelligent Manufacturing of Zhejiang Province, ZJU-Hangzhou Global Scientific and Technological Innovation Center, Hangzhou 311215, P.R. China. 3 Department of Chemical and Biomolecular Engineering, National University of Singapore, Singapore 117585, Singapore. 4 Coordination Chemistry Institute, State Key Laboratory of Coordination Chemistry, School of Chemistry and Chemical Engineering, Nanjing University, Nanjing 210023, P.R. China. 5 State Key Laboratory of Materials-Oriented Chemical Engineering, College of Chemical Engineering, Nanjing Tech University, Nanjing 211816, P.R. China. 6School of Chemistry and Chemical Engineering, Nanchang University, Nanchang 330031, P.R. China.

ZU-609 is self-assembled with inorganic metal node and organic linkers (EDS2−, 1,2ethanedisulfonate; dps, 4,4’-dipyridyl sulfide) (Fig. 1B). It consists of a 2D network that is coordinated by copper metal nodes and dps ligands, and it extends in a helical-like conformation along the direction that is vertical to the network (Fig. 1B and fig. S1). The EDS2− sulfonic anions coordinate to the bare Cu2+ on the adjacent 2D networks to form 3D coordination networks (Fig. 1C, figs. S1 and S2, and table S1) (46, 47). It possesses large onedimensional channels with a size of 7.5 × 8.1 × 11.1 Å3, connected by the local contraction resulting from the tilted pyridine rings (Fig. 1D). The cross-section size of the contraction is 4.2 Å × 5.1 Å, which falls just between the size of C3H6 (4.1 Å × 5.1 Å) and C3H8 (5.1 Å × 5.3 Å) molecules, suggesting the size/shape sieving potential of ZU-609 (Fig. 1E). The surface area of ZU-609 is explored using both N2 and CO2 as probes. Because of the obvious diffusion barrier of N2 in ZU-609 at 77 K, the CO2 probe is selected and the measured apparent surface area and pore volume of ZU-609 are 380 m2 g−1 and 0.15 cm3 g−1, respectively (fig. S3). Additionally, the thermal stability of ZU-609 is around 200°C and the activated structure remains stable after different treatments, such as air, water, and solutions of different pH (pH 3 to pH 11) (figs. S4 and S5).

g

1

in molecular diffusion rates (18–21). However, as a result of low to moderate selectivity, these methods inevitably suffer from co-adsorption of similar alkanes, and consecutive adsorptiondesorption cycles are needed to further improve the purity of the produced olefins (22–24). Molecular sieving avoids co-adsorption as it only adsorbs molecules within the molecular size or shape cutoff (25–27) and is advantageous in the production of high-purity olefins (28–30). However, the design of ideal molecular sieves with optimal adsorption thermodynamics and kinetics for gas separation remains a daunting task because of two important challenges: (i) Pore size control for molecular size sieving (31) and (ii) pore shape control for fast adsorption kinetics (32). Considering the difficulties in fine-tuning the pore size in 0.2 to 0.4 Å increments within the 3 to 5 Å range, the full exclusion of propylene from propane remains a challenge as a result of their small size difference (< 0.4 Å) (33–35). Adsorption kinetics is scarified in molecular sieves as ultrahigh selectivity is achieved based on restricted channels (2, 36). Thus, a high adsorption temperature is required to reach the diffusion rate threshold; however, this approach sacrifices capacity and energy efficiency. The dilemma is ascribed to the continuous diffusion limitation path along the whole narrow sieving channel (Fig. 1A), a typical pore structure in molecular sieves, which allows only “single-file diffusion” of adsorbates. The development of common cage-type structures with narrow windows and large voids breaks the tradeoff between separation selectivity and capacity, but still suffers from diffusion limitations (37, 38). To rationally utilize the limited pore space within molecular sieves, a sieving channel that features a local limited diffusion path is proposed to solve the challenge. The local contraction along the diffusion path functions as a sieve to exclude large molecules and the coexisting large channel allows the rapid dif-

Propylene molecular sieve design

p

A

s a key feedstock, the production of propylene (C3H6) exceeded 100 metric tonnes (Mt) per year in 2020 and its demand is expected to exceed 150 Mt per year in 2050 (1). Propane (C3H8) is a common byproduct of propylene manufacturing and separation of propylene and propane is essential for production of polymer-grade propylene (2–4). Conventional cryogenic distillation is energy intensive as a result of the close relative volatility and the production of olefin and paraffin accounts for nearly 1% of global carbon emissions (5–7). It is estimated that the development of nonthermally driven alternatives could make the separation ten times more energy efficient (8, 9). Physisorption based on porous materials is recognized as an alternative to cryogenic distillation owing to its moderate energy consumption (10–12). The development of advanced porous materials such as metal-organic frameworks, zeolites, and others has attracted considerable interest (13–15). Adsorbents with open metal sites and cations can selectively interact with propylene through p complexation but show low selectivity (16, 17). Additionally, kineticdriven adsorbents with restricted pore structures realize separation based on the difference

fusion of adsorbed molecules (Fig. 1A). Relative to the whole narrow sieving channel, the local sieving channel shortens the restricted diffusion path of adsorbed molecules. Porous crystalline materials that are self-assembled by secondary building blocks (SBUs) offer the possibility of tailor-made porous materials with designed pore size (39, 40), shape (41, 42), and pore chemistry (43–45). Rational modification using anions based on the original two-dimensional (2D) structure provides the possibility of pore shape and pore size control, from which molecular sieve ZU-609 was designed.

Pure-component equilibrium adsorption isotherms for C3H6 and C3H8 were measured, and ZU-609 showed sieving performance for C3H6 and C3H8 with an uptake ratio of 22.3 at 298 K, 1 bar (Fig. 2A, fig. S6, and table S2). The time-dependent adsorption profile of C3H8 was measured and its uptake was almost invariable in 24 hours, indicating that the low C3H8 uptake of ZU-609 is attributed to the size-sieving effect (fig. S7). The C3H6 adsorption isotherm approaches linear type with a moderate isosteric heat of adsorption (Qst) of 43 kJ mol−1 (figs. S8 and S9), providing a considerable C3H6 equilibrium working capacity of 2.0 mmol g−1 between 0.1 and 1 bar (Fig. 2A). The value is 1 of 5

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

p g y y g

(2.35 × 10−10 cm2 s−1) (fig. S14 and table S4). Such a high diffusion coefficient of ZU-609 enables the possible implementation of the adsorption process under near ambient conditions (fig. S12) without the necessity of increasing temperature to enhance diffusion, as in the case of Zeolite-4A (49).

y

higher than reported C3H6/C3H8 sieving materials such as Co-gallate (1.3 mmol g−1) (36), KAUST-7 (1.2 mmol g−1) (2), Y-abtc (0.58 mmol g−1) (48) and common zeolites 5A (0.41 mmol g−1), 4A (0.15 mmol g−1), ZSM-5 (0.25 mmol g−1) (figs. S10 and S11). The adsorption behavior is attributed to the embedded large pore space within the sieving channel and is ideal for the pressure swing adsorption process (fig. S12). Propylene diffusion performance

In addition to the thermodynamic equilibrium capacity, we also evaluated the C3H6 adsorption kinetics of ZU-609 and the other reported materials. Time-dependent C3H6 uptake profiles were assessed (figs. S13 to S15). Considering the influence of particle size on gas diffusion behavior, the particle size distriCui et al., Science 383, 179–183 (2024)

12 January 2024

bution of the tested materials was measured by laser diffraction (fig. S16). The calculated C3H6 diffusivity coefficients of ZU-609 with different particle size are close, around 1.00 × 10−9 cm2 s−1 (4.8 mm), 1.36 × 10−9 cm2 s−1 (26.8 mm), based on time-dependent adsorption profiles measured volumetrically (Fig. 2B, fig. S13, and table S3). This value is nearly an order of magnitude larger than the C3H6/C3H8 sieving materials Co-gallate (6.47 × 10−11 cm2 s−1), KAUST-7 (9.49 × 10 −11 cm2 s −1 ) and is several orders of magnitude higher than robust Zeolite-4A (1.01 × 10−12 cm2 s−1) (Fig. 2B). The consistent trend is observed based on the concentrationswing frequency response method for micropore diffusion measurement. The value of ZU-609 (2.38 × 10−9 cm2 s−1) is around 10 times that of Co-gallate (2.54 × 10−10 cm2 s−1), KAUST-7

Molecular-level understanding about adsorption and diffusion behavior

An in-situ powder x-ray diffraction experiment was conducted on ZU-609·C3H6 to study its C3H6 adsorption behavior (figs. S17 and S18). Each unit cell could accommodate 5.6 C3H6 molecules, which is close to its adsorption capacity at 298 K (table S5). The adsorbed C3H6 molecules are mainly distributed around the anions and bounded by EDS2− anions 2 of 5

,

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

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

p

Fig. 2. Gas sorption properties of ZU-609 and molecular-level understanding of C3H6 adsorption and diffusion behavior. (A) Gas sorption isotherms of propylene (red) and propane (blue) at 298 K for ZU-609. (B) Comparison of propylene diffusivity coefficient of ZU-609 with reported porous materials using time-dependent adsorption profiles measured volumetrically. Crystal sizes of Zeolite 4A (1.5 mm), Cogallate (1.7 mm), KAUST-7 (1.9 mm), ZU-609 (4.8 mm), and ZU-609 crystal (26.8 mm). (C) The adsorption configurations of C3H6 in ZU-609 obtained by in situ PXRD experiments. (D) The representative minimum energy path (MEP) profile and snapshots for C3H6 in ZU-609. The energy relative to the lowest energy configuration is shown as a function of the position of the central carbon on the C3H6 guest molecule. (Color code: C, gray; H, white; N, blue; Cu, pink; O, red; S, yellow; the carbon of propylene is highlighted in black.)

g y y g

The C3H6/C3H8 (50/50) mixture separation performance of ZU-609 was evaluated by breakthrough experiments. C3H8 breaks out immediately along with the inert gas He, indicating the C3H8 sieving effect exhibited by ZU-609 (Fig. 3A and fig. S21). C3H6 is continuously Cui et al., Science 383, 179–183 (2024)

12 January 2024

alkane impurities (figs. S26 and S27). The rapid C3H6 diffusion behavior of ZU-609 enables the adsorption process to be implemented under high gas velocity without obvious dynamic capacity loss (figs. S28 to S30). Even under a C3H6/C3H8 gas flow rate of 9 NmL min−1, the C3H6 dynamic capacity of ZU-609 preserves 94.1% of the equilibrium capacity; by contrast, the corresponding value on KAUST-7 is only 57.4% (Fig. 3D and fig. S31). In addition, the breakthrough performance of ZU-609 is invariable over the consecutive adsorption-desorption cycles (Fig. 3E and fig. S32). Even when ZU-609 is presaturated with high water content (500 to 3000 ppm) the C3H6 purity is not influenced (>99.9%) (figs. S27 and S33). The pressure swing adsorption (PSA) process is used to evaluate the actual capability of ZU-609 toward C3H6/C3H8 separation, as well as Co-gallate and KAUST-7 (Fig. 3F and figs. S34 to S36). ZU-609 is able to recover 89.7% 3 of 5

,

Propylene mixture separation performance and pressure swing adsorption simulation

trapped for 18.5 min (24.6 min g−1), affording a dynamic C3H6 capacity of approximately 1.64 mmol g−1 (Fig. 3, A and B). The corresponding C3H6 productivity of ZU-609 is about 32.2 L kg−1 (99.9%), higher than that of 21.9 L kg−1 (97.7%) for Co-gallate (36) and 16.6 L kg−1 (93.2%) for KAUST-7 (table S6). The purity of the eluted C3H6 is calculated to be 99.9% with a yield of 87.9% on average (Fig. 3C and figs. S22 and S23), which is higher than those of other materials (for JNU-3, purity was 99.5% and recovery was 51%; for KAUST-7, purity was 93.2% and recovery was 83%) (figs. S24 and S25) (50). Even for complex mixtures mimicking the composition of propane dehydrogenation CH4(2%)/C2H6(5%)/C2H4(5%)/C3H6(44%)/ C3H8(44%)/H2O (2000 ppm), ZU-609 also shows good separation performance with high olefin (C3H6+C2H4) purity (99.72%) and recovery (91.90%) at 283 K, and its separation ability is least influenced by the coexistence of C1-C3

y

through multiple hydrogen-bond interactions C-H···O (2.37 to 2.70 Å) (Fig. 2C). The C3H6 binding energy (DE) revealed by the dispersioncorrected density functional theory (DFT-D) calculation is around 47 kJ mol−1 (fig. S19). The minimum energy path profile using DFT-D calculations indicates the local diffusion energy barriers as the propylene molecules diffuse along the channel (Fig. 2D and fig. S20). These studies provide a molecular-level understanding about confined C3H6 adsorption and diffusion behavior.

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

p g

Cui et al., Science 383, 179–183 (2024)

12 January 2024

1. IEA, The future of petrochemicals: towards more sustainable plastics and fertilisers. IEA Publications; http://www.iea.org (IEA, 2020). 2. A. Cadiau, K. Adil, P. M. Bhatt, Y. Belmabkhout, M. Eddaoudi, Science 353, 137–140 (2016). 3. J. J. H. B. Sattler, J. Ruiz-Martinez, E. Santillan-Jimenez, B. M. Weckhuysen, Chem. Rev. 114, 10613–10653 (2014). 4. W. Wang et al., Science 381, 886–890 (2023). 5. L. Li et al., Science 362, 443–446 (2018). 6. K. J. Chen et al., Science 366, 241–246 (2019). 7. A. Tullo, C&EN 99, i9 (2021). 8. D. S. Sholl, R. P. Lively, Nature 532, 435–437 (2016). 9. G. Liu et al., Nat. Mater. 17, 283–289 (2018). 10. X. Han, S. Yang, M. Schröder, Nat. Rev. Chem. 3, 108–118 (2019). 11. C. Gu et al., Science 363, 387–391 (2019). 12. P. G. Boyd et al., Nature 576, 253–256 (2019). 13. P. Q. Liao, N. Y. Huang, W. X. Zhang, J. P. Zhang, X. M. Chen, Science 356, 1193–1196 (2017). 14. X. Cui et al., Science 353, 141–144 (2016). 15. S. J. Datta et al., Science 350, 302–306 (2015). 16. E. D. Bloch et al., Science 335, 1606–1610 (2012). 17. Y. Chai et al., Science 368, 1002–1006 (2020). 18. P. J. Bereciartua et al., Science 358, 1068–1071 (2017). 19. Q. Liu et al., Angew. Chem. Int. Ed. 62, e202218854 (2023). 20. M. Yang et al., Nat. Commun. 13, 4792 (2022). 21. Y. Su et al., Nature 611, 289–294 (2022). 22. L. Yang et al., Chem. Soc. Rev. 49, 5359–5406 (2020).

4 of 5

,

Progress has been made in the separation performance of molecular sieves in challenging gas separations, as demonstrated by our designed sulfonate microporous material ZU-609 (fig. S37). The local-limited diffusion path ensures the sieving effect and the coexisting relatively unrestricted diffusion path enables fast adsorption kinetics. Such a strategy offers solutions to the contradictory qualities of diffusion, capacity, and selectivity of common molecular sieves and exhibits attractive olefin/ paraffin separation performance in industrial operation processes, which highlights the importance of requisite dual control of pore size and pore shape in molecular sieve designs (figs. S38 to S40). Our work offers proof of the high efficiency of molecular sieves with fast

RE FERENCES AND NOTES

23. Y. Yang et al., Nat. Chem. 13, 933–939 (2021). 24. M. Chang et al., AIChE J. 68, e17794 (2022). 25. Z. Liu, W. Qiu, W. Quan, W. J. Koros, Nat. Mater. 22, 109–116 (2023). 26. D. D. Zhou et al., Nat. Mater. 18, 994–998 (2019). 27. L. Li et al., Science 377, 335–339 (2022). 28. R. B. Lin et al., Nat. Mater. 17, 1128–1133 (2018). 29. L. Wang et al., Angew. Chem. Int. Ed. 62, e202218596 (2023). 30. Q. Wang et al., Nat. Commun. 13, 2955 (2022). 31. X. Zhao, Y. Wang, D. S. Li, X. Bu, P. Feng, Adv. Mater. 30, e1705189 (2018). 32. C. Zhang et al., J. Phys. Chem. Lett. 3, 2130–2134 (2012). 33. Q. Ding, S. Zhang, Energy Fuels 36, 7337–7361 (2022). 34. Y. Xiao et al., J. Am. Chem. Soc. 145, 10980–10986 (2023). 35. D. Liu et al., Angew. Chem. Int. Ed. 62, e202218590 (2023). 36. B. Liang et al., J. Am. Chem. Soc. 142, 17795–17801 (2020). 37. P. Nugent et al., Nature 495, 80–84 (2013). 38. B. Li et al., Adv. Mater. 29, 1704210 (2017). 39. Z. Chen et al., Science 368, 297–303 (2020). 40. T. He et al., Nat. Mater. 21, 689–695 (2022). 41. X. Han et al., Nat. Mater. 17, 691–696 (2018). 42. S. H. Lo et al., Nat. Chem. 12, 90–97 (2020). 43. W. Xu et al., Nat. Rev. Mater. 5, 764–779 (2020). 44. K. Jayaramulu et al., Chem. Rev. 122, 17241–17338 (2022). 45. R. Freund et al., Angew. Chem. Int. Ed. 60, 23946–23974 (2021). 46. G. Y. Zhang, G. F. Wei, Z. P. Liu, R. J. Oliver, H. H. Fei, Chem. Mater. 28, 6276–6281 (2016). 47. J. Wang et al., J. Am. Chem. Soc. 142, 9744–9751 (2020). 48. H. Wang et al., Adv. Mater. 30, e1805088 (2018).

y

Conclusions

adsorption kinetics and provides opportunities to address other challenging gas separation processes.

y g

C3H6 with purity 99.7% (Fig. 3G). The C3H6 production efficiency of ZU-609 is up to 2.28 mol kg−1 hour−1, 345 and 316% more than that of Co-gallate and KAUST-7, respectively (Fig. 3H). Meanwhile, the corresponding consumed energy of ZU-609 is the least, at ~2.35 MJ kg (C3H6)−1 (Fig. 3I).

process with 3 NmL min−1 of N2 at 298 K. (D) Ratios of C3H6 dynamic uptake (Udynamic)/equilibrium uptake (Uequilibrium) of ZU-609 (n = 3 per group and are presented as means ± SD) and KAUST-7 under different superficial gas velocity for C3H6/C3H8 (50/50) mixture. (E) Dynamic uptake of C3H6 in cycling tests on ZU-609 for C3H6/C3H8 (50/50) mixture. (F) Schematic model for 2-bed PSA process. (G) Simulation results of C3H6 recovery for 2-bed PSA process. (H) Simulation results of C3H6 production efficiency for 2-bed PSA process. (I) Simulation results of energy consumption for 2-bed PSA process.

y

Fig. 3. Experimental breakthrough results of ZU-609 and simulated pressure swing adsorption process results. (A) Breakthrough curves of ZU-609 for an equimolar C3H6/C3H8 mixture at 298 K and 1 bar. The breakthrough experiments were carried out in a packed column with 0.75 g sample at a flow rate of 3 NmL min−1 [N represents the standard temperature (273 K) and pressure (1 bar)]. (B) Comparison of C3H6 dynamic uptake of ZU-609 with other reported C3H6/C3H8 sieving materials. (C) Flow rate curve of the desorbed C3H6 and C3H8 from ZU-609 and the purity of eluted C3H6 during the regeneration

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

49. C. A. Grande, F. Poplow, A. E. Rodrigues, Sep. Sci. Technol. 45, 1252–1259 (2010). 50. H. Zeng et al., Nature 595, 542–548 (2021). ACKN OW LEDG MEN TS Funding: This work was supported by the National Natural Science Foundation of China (22122811, 22227812, and 22108240) and Zhejiang Provincial Natural Science Foundation of China (LR20B060001). Author contributions: H.X. and L.Y. initiated and supervised the research. J.C., Z.Z., and A.J. performed the adsorbents preparation. J.C., Z.Z., Y.S., J.W., Y.Z., B.M., Y.Z., and J.W. performed the characterization. Z.Y., J.C., L.Y., and J.H. performed the simulated calculations. J.C., L.Y., H.X., and X.C.

analyzed the data and wrote the paper. Competing interests: The authors declare no competing financial interests. The authors and their affiliated institutions have a patent pending related to the results presented here. Data and materials availability: All data are available in the manuscript or the supplementary materials. The related CIF of ZU-609, ZU-609 after refinement and ZU-609 loaded with C3H6 have been uploaded to the Cambridge crystallographic data center (CCDC 2266913, 2277294, and 2277295). License information: Copyright © 2024 the authors, some rights reserved; exclusive licensee American Association for the Advancement of Science. No claim to original US government works. https://www.science.org/about/sciencelicenses-journal-article-reuse

SUPPLEMENTARY MATERIALS

science.org/doi/10.1126/science.abn8418 Materials and Methods Supplementary Text Figs. S1 to S41 Tables S1 to S14 References (51–55) Data S1 and S2 Submitted 23 December 2021; resubmitted 5 July 2023 Accepted 1 December 2023 Published online 14 December 2023 10.1126/science.abn8418

p g y y g y ,

Cui et al., Science 383, 179–183 (2024)

12 January 2024

5 of 5

RES EARCH

ANTHROPOLOGY

Two thousand years of garden urbanism in the Upper Amazon Stéphen Rostain1*, Antoine Dorison2, Geoffroy de Saulieu3, Heiko Prümers4, Jean-Luc Le Pennec5, Fernando Mejía Mejía6, Ana Maritza Freire7, Jaime R. Pagán-Jiménez8, Philippe Descola9 A dense system of pre-Hispanic urban centers has been found in the Upano Valley of Amazonian Ecuador, in the eastern foothills of the Andes. Fieldwork and light detection and ranging (LIDAR) analysis have revealed an anthropized landscape with clusters of monumental platforms, plazas, and streets following a specific pattern intertwined with extensive agricultural drainages and terraces as well as wide straight roads running over great distances. Archaeological excavations date the occupation from around 500 BCE to between 300 and 600 CE. The most notable landscape feature is the complex road system extending over tens of kilometers, connecting the different urban centers, thus creating a regional-scale network. Such extensive early development in the Upper Amazon is comparable to similar Maya urban systems recently highlighted in Mexico and Guatemala.

1 of 7

,

12 January 2024

y

Rostain et al., Science 383, 183–189 (2024)

y g

*Corresponding author. Email: [email protected]

Stretching along the foothills of the Andes in southern Ecuador, enclosed between the Andes to the west and the Cutucú range to the east, the Upano Valley is a region where Amazonian and Andean ecosystems meet, with an important seismic risk—the earthquake of 1995 reached a moment magnitude of 7.0 (19). Towering above it, the active Sangay stratovolcano, a steep-sided, cone-shaped, snow-capped mountain, peaks at 5230 m above sea level. After descending the Andean slopes, the Upano River goes straight south along the sub-Andean Range (20). Groups of earthen platforms have been archaeologically explored: They form large settlements extending over the 70- to 100-m-high alluvial terraces along the river (Fig. 1). Few sites have been excavated (21–24). One of the largest settlements, Sangay, was discovered in the late 1970s and excavated by three successive teams (17, 25, 26). Started in the mid1990s, the Upano interdisciplinary project brought together archaeologists, geoscientists,

y

Archaeology of the Americas (UMR8096), French National Center for Scientific Research (CNRS), Paris, France. Archaeology of the Americas (UMR8096), Paris-1 University, Paris, France. 3Local Heritage, Environment and Globalization (UMR 208), French National Research Institute for Sustainable Development, Paris, France. 4Kommission für Archäologie Außereuropäischer Kulturen, Bonn, Germany. 5 Geo-Ocean, Brest University, CNRS, Ifremer, Institut Universitaire Européen de la Mer (UMR6538), French National Research Institute for Sustainable Development, Plouzané, France. 6Anthropology and Archaeology Department, Pontifical Catholic University of Ecuador, Quito, Ecuador. 7Museum of the Banco Central of Ecuador, Guayaquil, Ecuador. 8Center for Social Research, University of Puerto Rico, San Juan, Puerto Rico. 9Social Anthropology Laboratory, Collège de France, Paris Sciences et Lettres, Paris, France. 2

Context

g

1

Located in the Upper Amazon on the eastern slopes of the Ecuadorian Andes, its dense population extended across the high alluvial terraces that border the Upano River (17). The data of this study stem from more than two decades of interdisciplinary investigations in the region, the scope of which was recently broadened by light detection and ranging (LIDAR) mapping in a 300-km2 area. This research revealed the largest urban network of erected and excavated features known in Amazonia, whose beginnings date back to 2500 years ago (18). This discovery raises many questions, among them the following: What types of features were built by the pre-Hispanic inhabitants? Were the settlements contemporary and connected to each other? Where did the inhabitants cultivate the large quantity of plants needed for their subsistence?

p

U

rbanism and Amazonia are rarely associated when we evoke pre-Hispanic times. However, Francisco de Orellana, leading an expedition down the great river in 1541–42, witnessed large cities along its banks (1) yet was called a fabulist upon his return. More than four centuries later, in the 1980s, great archaeological sites with thick layers of Amazonian dark earth were discovered in the middle Amazon, confirming the existence of extensive pre-Hispanic settlements along the river (2). Orellana did not lie. Recently, various monumental archaeological sites were brought to light in Amazonia (3, 4). Some of them display earthen platforms of various sizes with a variety of features, including causeways, mounds, canals, and/or fortifications [as in the Barinas llanos in Venezuela (5), the Llanos de Mojos in Bolivia (6), and the Upper Xingu and central Amazon in Brazil (7–9)] or roads [as in southwestern Amazonia (10, 11)]. Pre-Hispanic inhabitants of the Amazon were indeed remarkable land builders who intensively reworked their environment, thus changing the morphology of their territories (12–14) and its vegetation cover (15, 16). Here we present an extensive case of a preHispanic urban system in the Amazon region.

and archaeobotanists to investigate the entire valley. The scope of the fieldwork results, which have led to a better understanding of the valley’s earthworks, has recently been broadened by LIDAR survey. Our fieldwork led to results concerning the pre-Hispanic cultural sequence, habitat and diet, and the ancient volcanic activity and morphogenesis of the valley. Archaeological work focused on mounds, central plazas, and roads, as well as sites without built features (supplementary text S1, figs. S1 to S6, and tables S1 and S2). Large-scale excavations in platforms and plazas at two major settlements (Sangay and Kilamope) revealed domestic floors, with postholes, caches, pits, hearths, large jars, grinding stones, and burnt seeds. The construction methods consisted of cutting the natural slope to form a base on which the mound was built (27). Intentional artifact deposits suggest that the building process was accompanied by ritual activities. The newly established dating sequence indicates the succession of at least five cultural ensembles. The original building of earthen platforms and roads took place between approximately 500 BCE and 300 to 600 CE and was carried out by groups from the Kilamope and later Upano cultures. Some mounds were then reoccupied, after a hiatus, by groups of the Huapula culture between 800 and 1200 CE (supplementary text S2 and fig. S7). All the evidence indicates a local cultural evolution process. There is no a priori reason to think that the pre-Hispanic inhabitants originated from the Andes and not from the Amazon. The Kilamope and Upano people were sedentary agrarian societies that densely occupied the valley, where even today, according to local farmers, fertile volcanic soils still allow up to three harvests per year. Analyses of starch grains from potteries revealed the consumption of maize (Zea mays), beans (Phaseolus sp.), manioc (Manihot esculenta), and sweet potato (Ipomoea batatas). The microtraces on a maize starch grain are identical to those left today by the manufacture of the traditional “chewed” chicha (sweet beer), suggesting that this beverage may have been served in the Upano vessel (28). Upano pottery is well made and comes in many decorated types, the most common being the red-banded incised type with straight or curved painted and incised lines (29). Upano bowls and jars were exported up to the Andes, near the modern town of Cuenca (30). Sangay volcano’s intense activity had a definite impact on pre-Hispanic communities (31). Previous studies demonstrated continuous cone growth and two major flank failures during the Pleistocene (32). The resulting debris avalanches (33) spread as far as 60 km from the volcano (34) and left thick hummocky deposits, on which human settlements developed in late Holocene times. Analyses also indicate that large explosive eruptions occurred

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

p g y Fig. 1. Map of archaeological Upano Valley. (Left) Map of the Middle Upano with five major settlements and 10 secondary sites. m a.s.l., meters above sea level. (Right) Map of pre-Hispanic dug roads.

Rostain et al., Science 383, 183–189 (2024)

Platforms and plazas

The most common features are earthen platforms. Going far beyond the field identification of just some individual platforms, the LIDAR data enabled the detection of >6000 platforms within the 300-km2 area. The standard shape is rectangular (although a few were circular) and about 20 m by 10 m with a preserved elevation of 2 to 3 m (Fig. 3).

12 January 2024

Settlements

The distribution of elements in the study area reflects a settlement pattern in which the buffer zones between residential architectures 2 of 7

,

In 2015, the Ecuadorian National Institute for Cultural Heritage commissioned the LIDAR survey of a 600-km2 area stretching from the Upper Upano to the Pastaza River to highlight anthropogenic features hidden under the canopy (35, 36) (Fig. 1). After a 1-m-resolution digital elevation model had been derived from the point cloud, the Upano project team studied the 300-km2 area constituting the southern part of this landscape (supplementary text S3, figs. S8 to S11, and table S3). The number and size of anthropogenic features led us to target only specific scientific goals. Given that previous fieldwork had already provided information on the internal organization of residential sites (37), we focused on the “gaps” between settlements and the road network rather than on

Rarely isolated, the platforms typically occur in groups—or complexes—of three to six units around a plaza, often with a central platform. The most common complexes measure 40 m by 40 m (~1600 m2) and are interpreted as 1residential (37). However, LIDAR imagery also highlighted monumental complexes probably bearing a civic-ceremonial function, with much larger platforms, but of similar height to the smaller ones. The largest complex, at Kilamope, covers 10 ha and includes a 140 m by 40 m monumental platform (4.5 m high). This pattern is repeated almost evenly throughout the area. Average density is 16.6 platforms/km2, but some agglomerated areas have densities of >100 features/km2. Because of their ubiquity and close relationship with the road network, we consider the complexes to be the elementary built features of the pre-Hispanic landscape.

y

LIDAR results

intrasite features. This objective was also motivated by the identification in the field of various road starts (38). We set up a remote sensing methodology based on multiple LIDAR visualizations (39, 40) (Fig. 2). Pre-Hispanic features were distinguished from modern elements according to their orientation and spatial coherence. The analysis revealed an elaborate anthropogenic landscape. Settlements with standardized archaeological features evidencing a shared cultural background are interconnected by short-distance (intrasite) and longdistance (intersite) roads, which strongly suggests contemporaneity.

y g

in the nearer past (32). The hypothesis that the Upano culture came to an abrupt end after a massive eruption ~400 to 600 CE has been raised (18) but has been called into question by the disparity of radiocarbon dates obtained recently for these levels.

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

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

played an important role. Moreover, clustering trends occur. Clusters of complexes have been identified as settlements on the basis of three criteria: feature density, size of the civicRostain et al., Science 383, 183–189 (2024)

ceremonial platforms, and connections with other complexes. Fifteen settlements were classified into two categories: five major sites (large and/or dense

12 January 2024

centers) and 10 secondary sites (Fig. 1). Among the first group, Sangay stands out for its higher density (>125 platforms/km2) and its ostentatious core. Accessed by a 2.5-km-long straight 3 of 7

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

12 January 2024

4 of 7

,

Perhaps the most notable elements of the landscape are the intrasite pathways and a regional-scale intersite road network (Fig. 4). Roads are considered as such because they systematically link complexes (thus distinguishing them from other features, such as canals,

The LIDAR survey highlighted numerous agrarian features of two main types: drained fields and terraces. Their strong spatial coherence with the rest of the detected remains and, conversely, their location under forested areas support the inference that they were an integral part of the pre-Hispanic anthropogenic landscape. Drained fields extend over hundreds of hectares into orthogonal and continuous plot systems

y

Roads and pathways

Drained fields and terraces

y g

Numerous “truncated” hills—natural reliefs with flat summits—were detected. While their general morphology is probably of volcanic origin (hummocks), a plausible hypothesis, supported by the accesses built on their slopes, is that their flattening is artificial, so that they constitute an integral component of the preHispanic landscape (41).

y

Rostain et al., Science 383, 183–189 (2024)

Modified hills

g

road, it lies at the highest point of the cliff that borders the Upano’s northern bank, thus dominating the valley. Junguna and Kunguints seem to be chiefly residential sites with many small complexes adjoining intrasite circulation axes, whereas Kilamope and Copueno are characterized by large civic-ceremonial complexes. Despite the apparent architectural and spatial homogeneity among these sites, several elements suggest that the settlements were exposed to threats. These include peripheral ditches blocking access to some settlements (e.g., east of Sangay) and obstructed roads near some large complexes (e.g., Copueno). We interpreted these elements as the result of tensions between groups or reinforcement of the sites against external threats.

p

Fig. 3. Earth platforms complex. Earth platforms complex of Nijiamanch on the right bank of the Middle Upano. Drained fields are visible around the artificial mounds on the LIDAR image (bottom).

described below). Some of these complexes, unusually isolated in the landscape, adjoin major roads as if they were layover points along the way. Four types of roads or pathways have been identified. Although we cannot totally discount the possibility that a few paths could be the result of usage (sunken lanes) (42, 43), the morphology of the features detected (e.g., rectilinearity, depth, curbs) strongly suggests that these are excavated roads. Therefore, the generic term “dug roads” is used here. The most widespread are the straight-dug footpaths and roads (thus classified on the basis of their size and length) whose morphology is documented by fieldwork (27). Essentially straight and about 2 to 3 meters deep on average, they probably result from digging and accumulating earth on either side of the path, creating a U-shaped profile bordered by curbs. The width is 4 to 6 m for the smallest paths and up to >15 m for the largest ones, creating a walkable surface that is 2 to 5 m wide in the middle. Despite discontinuities in their layout owing to the heterogeneity of the LIDAR coverage, we can reasonably estimate that the longest roads—Uyunts-Jurumbuno and Kilamope-Kunguints—run for more than 14 and 25 km, respectively. Both are likely to continue beyond the boundaries of the study area. We believe that dug roads were designed to be as straight as possible despite the natural irregularities of the terrain. The second type includes roads running along the interfluve in hilly terrain. Occasionally also dug, they follow topographic relief and connect open spaces (e.g., plateaus, riverbeds) through steep areas, such as the Andean footslopes or the valley cliffs. The third category includes possible elevated causeways with parallel ditches on either side of the road reminiscent of “causeways-canals” in the Llanos de Mojos of Bolivia (44). Finally, dug footpaths regularly end with a descent into one of the gullies (or quebradas) that spread on the Upano alluvial terrace and come out of the same gully a few hundred meters further on. This seems to constitute a fourth type of path, which takes advantage of the natural layout of the quebradas. In line with this, we argue that, beyond connecting spaces, most pathways were intimately related to surface water management and agricultural practices.

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

5 of 7

,

12 January 2024

y

The results of fieldwork and LIDAR analysis demonstrate that the Upano Valley was dense-

y g

Discussion and conclusions

y

Rostain et al., Science 383, 183–189 (2024)

Llanos de Mojos in Bolivia and is still used today, for example, by the Karinya of the Orinoco llanos (45). Less widespread, terraces are found occasionally along the edges of quebradas, along the alluvial terrace of the Upano, perpendicular to concave hillslopes or on the lower Andean slopes, where they are associated with drains parallel to the slope. Agrarian features fill the “gaps” between complexes and settlements noted during the first reading of the LIDAR data. Their ubiquity, their close connection with residential and ceremonial areas, and the variety of geomorphological contexts exploited demonstrate the importance of agricultural activity in the settlement pattern.

g

(Fig. 2). The elementary unit is a rectangular field 10 to 40 m wide and several dozens of meters long. The field limits are ditches of 4-m width and 40-cm depth on average. They are connected to drainage canals, slightly wider and deeper, and mitigate waterlogging in this climate with daily rainfall. These canals then flow into the hydrographic network of the quebradas, whose course has sometimes been modified. Drained fields are linked to the network of dug footpaths often surrounding them, making it sometimes difficult to distinguish between a road and a canal. In such cases, a clear connection to a complex was the discriminating criterion. However, it is very likely that pathways had a dual function of circulation and water management. This agricultural technique resembles those documented for the

p

Fig. 4. Dug roads. Wide and deep dug roads connecting Upano monumental settlements.

ly populated around the beginning of the common era. The intimate link between residential and agricultural areas brings to mind the “garden cities” and “green urbanism” described or theorized by other researchers (13, 46–52). Far from the utopia that these terms imply, however, the garden urbanism of the Upano valley constitutes a concrete, dynamic, and probably contested landscape and provides further proof that Amazonia is not the pristine forest once depicted. The settlement pattern is composed of dense sites with standardized domestic groups of platforms around plazas and monumental civic architecture connected by streets. Establishments are linked to each other over great distances by a vast network of roads intertwined with intensive agricultural layouts. The organizational and architectural homogeneity, as well as the consistent interweaving of monumental-ceremonial features, domestic spaces, and economic areas, strongly suggests that the whole network was at least partly contemporary. Despite variations in the settlement density, it must be emphasized that few areas of the valley are devoid of remains. Apparent empty buffer zones between complexes of platforms were in fact dedicated to agriculture. Two main patterning strategies have been recognized and were seemingly induced by the valley’s geomorphology. In the southern part of the study area, extensive parcels of drained fields are spread within a low-density distribution of complexes in the wetlands of the Upano alluvial terraces. Meanwhile, the hilly terrain of the northern zone is more prone to the clustering of complexes, with agricultural terraces intertwined with this denser grid in a more opportunistic manner. However, these two contrasting “environment-forced” sectors are all connected by the road network and its orthogonal layout, independent of all geomorphic constraints and constituting perhaps the most distinctive characteristic of this built landscape. Straight roads cross at right or nearly right angles without deviating in front of hills or ravines. If these roads facilitated exchanges, it is very likely that they also had a marked symbolic and powerful ritual function and participated in the construction of a cultural landscape. The complexity and dynamic nature of the latter are further demonstrated by the presence of defensive features that raise the question of alliances and tensions or possibly even episodic wars. However, it would be imprudent to infer that these cities were organized around a centralized authority capable of mobilizing labor in a more or less coercive manner. The interlocking of groups of filiation and segmental solidarities, regularly reinforced by ceremonial exchanges, are sufficient to ensure the cohesion and coordination necessary for the

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

p 1. G. de Carvajal, Descubrimiento del río de la Amazonas (Estudios ediciones y medios, Sevilla, [1542] 1992). 2. U. Lombardo et al., Nat. Commun. 13, 3444 (2022). 3. S. Rostain, Islands in the Rainforest: Landscape Management in Pre-Columbian Amazonia (Left Coast Press, 2012). 4. S. Rostain, C. Jaimes Betancourt, Eds., Las Siete Maravillas de la Amazonia precolombina (IV Encuentro Internacional de

6 of 7

,

REFERENCES AND NOTES

y

12 January 2024

and human-built hydraulic features. Indeed, it is tempting to see the Upano valley’s geometrical layout, which cuts across the topography, as a cosmological design rather than a common and practical system of communication. This strong emphasis on abstract lineal connections between sites is more reminiscent of the sacred topography of Andean polities as “analogist” systems than of contemporary “animist” cultures of the Upper Amazon [in the sense of Descola (67)]. For lack of ethnographic analogies in settlement patterns in the contemporary lowlands, one should probably look for current highlands autochthonous polities whose sociocosmic structure is embedded in a topographical grid [for instance, the Chipaya of Bolivia (68)]. Such a discovery is another vivid example of the underestimation of Amazonia’s twofold heritage: environmental but also cultural, and therefore Indigenous. Like many others (69–71), we believe that it is crucial to thoroughly revise our preconceptions of the Amazonian world and, in doing so, to reinterpret contexts and concepts in the necessary light of an inclusive and participatory science.

y g

Rostain et al., Science 383, 183–189 (2024)

scale of landscape anthropization, where urbanism covers hundreds of square kilometers. It echoes more clearly the Upper Xingu, where road networks suggest a lowdensity pre-Hispanic urbanism (52). This suggests that, being denser and less accessible, northern Amazonia may have been underestimated in terms of archaeological potential and promises numerous discoveries with future exploratory work and LIDAR coverage. Broadening perspectives, considering the regionscale density of features, the population density they must have supported, or the highly anthropized landscape—and although the size of the valley does not rival that of the Yucatán biosphere reserves—the human investment in the Upano Valley is comparable to that of contemporary Central Maya Lowlands (61–65). Going further, the major ceremonial cores, with monumental platforms, plazas, and causeways, are comparable in size to those of other great cultures of the past, such as Mexican Teotihuacan or the Egyptian Giza Plateau (Fig. 5). What is very noteworthy is that this example of garden urbanism appears to be associated with a mound-building tradition. One wonders whether such architecture might echo other regions of the Americas and a pre-Hispanic ideology wherein fertility, origin of humanity, and ancestors are key aspects. Although the design of roads and canals is not radial in the Upano Valley, the grid it establishes is reminiscent of the ceque system of Cuzco (66) and its connections between kin-based and segmented territories, ritual places of worship,

y

organization of a highly structured settlement. Contemporary ethnology shows that exchanges are not so much based on economic logic as on the circulation of specialized production within ethnic confederations, each defining a local identity (53–55). Indeed, the overall organization and standardization of architectural features, the road patterns, or the defense systems, suggests, for instance, the existence of advanced engineering (mature architectural tradition relying on sighting and surveying instruments and/or expertise). The Upano sites are quite different from other monumental sites of Amazonia, which are all more recent, considerably less dense in terms of features, and, until proven otherwise, not embedded in such a vast and dense communication network (fig. S12) (13, 14). This original 2000-year-old society of the Upano valley constitutes the earliest and largest lowdensity agrarian urbanism ever documented in the Amazon so far. At a supraregional level, beyond exchanges with the Cuenca area, the relations and potential bilateral influence with the contemporary Andean world, such as Chavín de Huántar’s sphere of influence (56, 57), remain difficult to grasp (58, 59). Yet, at present, there is no reason to believe that this is not an endogenous development. The Upano’s LIDAR coverage confirms the exceptional archaeological potential of Amazonia, recently recalled by extrapolative modeling in Brazil (60). However, whereas the latter article focuses on isolated earthworks (i.e., monumental ditches), the Upano materialized another

g

Fig. 5. Comparison of site sizes. Comparison at the same scale of the core areas of major sites of the Upano, ancient Egypt, and ancient Mesoamerica (for comparison with low-density urbanism sites in Amazonia, see supplementary text S4 and fig. S12).

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

5. 6. 7. 8.

9. 10. 11. 12. 13. 14. 15. 16. 17. 18. 19.

21.

26.

29.

31.

12 January 2024

science.org/doi/10.1126/science.adi6317 Materials and Methods Figs. S1 to S13 Tables S1 to S3 References

,

Rostain et al., Science 383, 183–189 (2024)

SUPPLEMENTARY MATERIALS

y

32. 33. 34.

LIDAR images were provided thanks to an agreement elaborated by J. F. Moscoso Novillo (director of the Ecuadorian National Institute for Cultural Heritage). The following persons helped during the fieldwork conducted by S.R.: I. Cangas, P. Baby, R. Bartone, S. Bes de Berc, J. Burgos, T. Bray, J. Burton, A.M.F., F. Fuentes, Y. Graber, J.-L.L.P., K. Olsen Bruhns, L. Pintér, P. Podwojewski, H.P., A. Rostoker, G.d.S., L. Zlonay, and several Ecuadorian and French students. Local people also contributed, especially J. Vega, Juan Vele, Julio Vele, and several workers. A. Gómez de la Peña, K. Leonard, and J.R.P.-J. did the archaeobotanical analyses. Thanks to D. McKey and F. Savatier for their reading and corrections and to the reviewer for their suggestions. Funding: This work was supported by the French National Center for Scientific Research (CNRS), Laboratory “Archaeology of the Americas” (UMR8096), French Ministry for Europe and Foreign Affairs (MAEE), French National Research Institute for Sustainable Development (IRD), German Kommission für Archäologie Außereuropäischer Kulturen (KAAK), Ecuadorian National Institute for Cultural Heritage (INPC), Banco Central del Ecuador, and National Secretary of Higher Education, Science, Technology and Innovation (SENESCyT). Author contributions: Conceptualization: S.R. and A.D. Methodology: S.R., A.D., H.P., and J.-L.L.P. Investigation: S.R., A.D., G.d.S., H.P., J.-L.L.P., F.M.M., A.M.F., J.R.P.-J., and P.D. Visualization: S.R., A.D., and H.P. Funding acquisition: S.R., G.d.S., H.P., and J.-L.L.P. Project administration: S.R. Supervision: S.R. Writing – original draft: S.R., A.D., G.d.S., H.P., J.-L.L.P., and P.D. Writing – review & editing: S.R., A.D., G.d.S., H.P., J.-L.L.P., and P.D. Competing interests: The authors declare that they have no competing interests. Data and materials availability: All data are available in the main text or the supplementary materials. License information: Copyright © 2024 the authors, some rights reserved; exclusive licensee American Association for the Advancement of Science. No claim to original US government works. https://www.science.org/about/sciencelicenses-journal-article-reuse

y g

30.

AC KNOWLED GME NTS

y

27. 28.

64. M. A. Canuto et al., Science 361, eaau0137 (2018). 65. T. Beach et al., Proc. Natl. Acad. Sci. U.S.A. 116, 21469–21477 (2019). 66. T. Zuidema, The Ceque System of Cuzco: The Social Organization of the Capital of the Inca (Brill, 1964). 67. P. Descola, Beyond Nature and Culture (Univ. of Chicago Press, 2013). 68. N. Wachtel, Le retour des ancêtres. Les Indiens Urus de Bolivie, XX°-XVI° siècle. Essai d’histoire régressive, Bibliothèque des sciences humaines (Éditions Gallimard, 1990). 69. C. Gnecco, P. Ayala, Eds., Indigenous Peoples and Archaeology in Latin America (Left Coast Press, 2011). 70. B. C. da Rocha, thesis, Institute of Archaeology, University College London (2017). 71. P. Erikson, V. Vapnarsky, Eds., Living Ruins: Native Engagements with Past Materialities in Contemporary Mesoamerica, Amazonia, and the Andes (Univ. Press of Colorado, 2022).

g

22. 23. 24. 25.

35. Technoproject, “Informe del levantamiento aéreo y del análisis preliminar e identificación de las anomalías del relieve relacionadas con las estructuras arqueológicas. Superficie Upano–Mera,” field report, INPC, Quito (2015). 36. A. Sánchez-Polo, R. Á. Litben, STRATA 1, e3 (2023). 37. S. Rostain, in Communities in Contact. Essays in Archaeology, Ethnohistory and Ethnography of the Amerindian CircumCaribbean, C. L. Hofman, A. van Duijvenbode, Eds. (Sidestone Press, 2011), pp. 455–475. 38. S. Rostain, Quat. Int. 249, 31–42 (2012b). 39. Z. Kokalj, R. Hesse, Airborne Laser Scanning Raster Data Visualization: A Guide To Good Practice, vol. 14 of Prostor, kraj, čas (Založba ZRC, 2017). 40. A. Dorison, C. Siebe Grabach, M. Elliott, G. Pereira, Boletín de la Sociedad de Geología Mexicana 74, A180622 (2022). 41. G. de Saulieu, S. Rostain, J.-L. Le Pennec, “El Formativo del Alto Pastaza (Ecuador), entre arqueología y vulcanología” in Antes de Orellana. Actas del 3er Encuentro Internacional de Arqueología Amazónica, S. Rostain, Ed., IFEA, FLACSO, Embajada de los EEUU, Quito (2014), pp. 199–205. 42. M. Cardale de Schrimpff, Caminos prehispánicos en Calima (Fundación de Investigaciones Arqueológicas Nacionales, Santafé de Bogotá, 1996). 43. N. Nir et al., Catena 215, 106312 (2022). 44. C. L. Erickson, in Landscapes of Movement: Trails, Paths, and Roads in Anthropological Perspective, J. E. Snead, C. L. Erickson, J. A. Darling, Eds. (Univ. of Pennsylvania Press, 2011), pp. 204–231. 45. W. M. Denevan, K. H. Schwerin, Antropológica 50, 3–91 (1978). 46. E. Howard, Garden Cities of Tomorrow (2007). 47. C. M. Arensberg, in Urban Life: Readings in Urban Anthropology, G. Gmelch, W. P. Zenner, Eds. (St. Martin’s Press, 1980), pp. 37–47. 48. W. J. Eggeling, J. Ecol. 34, 20–87 (1947). 49. E. Graham, Archeol. Pap. Am. Anthropol. Assoc. 9, 185–194 (1999). 50. C. Isendahl, M. E. Smith, Cities 31, 132–143 (2013). 51. D. Marken, M. C. Arnauld, Eds., Building an Archaeology of Maya Urbanism: Planning and Flexibility in the American Tropics (Univ. Press of Colorado, 2023). 52. M. Heckenberger, in Landscapes of Preindustrial Urbanism, G. Farhat, Ed. (Dumbarton Oaks Research Library and Collection, 2020), pp. 225–262. 53. J. Jackson, The Fish People: Linguistic Exogamy and Tukanoan Identity in the Northwest Amazon (Cambridge Univ. Press, 1983). 54. P. Descola, In the Society of Nature (Cambridge University Press, 1994). 55. B. Franchetto, M. Heckenberger, Eds., Os povos do Alto Xingu - história e cultura (UFRJ, 2001). 56. R. L. Burger, Chavin and the Origins of Andean Civilization (Thames & Hudson, 1992). 57. R. L. Burger, in Handbook of South American Archaeology, H. Silverman, W. H. Isbell, Eds. (Springer, 2008), pp. 681–703. 58. A. J. Pearce, D. G. Beresford-Jones, P. Heggarty, Eds., Rethinking the Andes-Amazonia Divide: A Cross-Disciplinary Exploration (UCL Press, 2020). 59. R. Clasby, J. Nesbitt, Eds., The Archaeology of the Upper Amazon: Complexity and Interaction in the Andean Tropical Forest (Univ. Press of Florida, 2021). 60. V. Peripato et al., Science 382, 103–109 (2023). 61. R. Fletcher, Insights 2, 1–19 (2009). 62. A. F. Chase, D. Z. Chase, Annu. Rev. Anthropol. 45, 361–376 (2016). 63. E. Lemonnier, in Settlement Ecology of the Ancient Americas, L. C. Kellett, E. E. Jones, Eds. (Routledge, 2017), pp. 167–194.

p

20.

Arqueología Amazónica, Bonner Altamerika-Sammlung und Studien, Plural Editores, 2017). E. M. Redmond, C. S. Spencer, Anthropol. Pap. Am. Mus. Nat. Hist. 2007, 86 (2007). U. Lombardo, H. Prümers, J. Archaeol. Sci. 37, 1875–1885 (2010). M. J. Heckenberger et al., Science 301, 1710–1714 (2003). M. Heckenberger, The Ecology of Power: Culture, Place, and Personhood in the Southern Amazon, AD 1000-2000 (Routledge, 2005). E. Neves, Sob os tempos do equinócio: Oito mil anos de história na Amazônia central (Ubu Editora, 2022). J. Iriarte et al., J. Comput. Appl. Archaeol. 3, 151–169 (2020). S. Saunaluoma, J. Moat, F. Pugliese, E. G. Neves, Lat. Am. Antiq. 32, 1–15 (2020). W. Denevan, Cultivated Landscapes of Native Amazonia and the Andes (Oxford Univ. Press, 2001). M. J. Heckenberger et al., Science 321, 1214–1217 (2008). H. Prümers, C. J. Betancourt, J. Iriarte, M. Robinson, M. Schaich, Nature 606, 325–328 (2022). C. R. Clement et al., Proc. R. Soc. London Ser. B 282, 20150813 (2015). W. Balée, T. Swanson, M. G. Zurita-Benavides, J. C. Ruiz Macedo, Lat. Am. Antiq. 34, 1–15 (2023). P. Porras, Natl. Geogr. Res. 5, 374–381 (1989). S. Rostain, Bull. Inst. Fr. Études Andines 28, 53–89 (1999). D. Legrand et al., in La cuenca oriente: Geología y petróleo, P. Baby, M. Rivadeneira, R. Barragán, Eds. (IFEA, IRD, Petroecuador, 2004), pp. 168–182. S. Bès de Berc et al., in La cuenca oriente: Geología y petróleo, P. Baby, M. Rivadeneira, R. Barragán, Eds. (IFEA, IRD, Petroecuador, 2004), pp. 153–167. L. Rampón, in Cuadernos de Investigaciones Científicas No. 1 (Arqueología, Misiones Católicas de la Amazonía, 1959). P. Porras, Natl. Geogr. Res. 5, 374–381 (1989). P. Moncayo Echeverría, Revista Sarance 20, 147–154 (1994). A. Rostoker, thesis, City University of New York (2005). P. Porras, Investigaciones arqueológicas a las faldas del Sangay, Tradición Upano (Centro de Investigaciones Arqueológicas, Universidad Católica del Ecuador, 1987). E. Salazar, in Handbook of South American Archaeology, H. Silverman, W. Isbell, Eds. (Springer, Kluwer, Plenum, 2008), pp. 263–278. S. Rostain, STRATA 1, e2 (2023). J. R. Pagán Jiménez, S. Rostain, “Uso de plantas económicas y rituales (médicinales o energizantes) en dos comunidades precolombinas de la Alta Amazonía ecuatoriana: Sangay (Huapula) y Colina Moravia (400 a.C.1200 d.C.)” in Antes de Orellana. Actas del 3er Encuentro Internacional de Arqueología Amazónica, S. Rostain, Ed., IFEA, FLACSO, Embajada de los EEUU, Quito (2014), pp. 313–322. S. Rostain, in Cerâmicas Arqueológicas da Amazônia: Rumo a uma nova síntese, C. Barreto, H. Lima, C. Jaimes Betancourt, Eds. (MPEG, Belém-do-Pará, 2016), pp. 573–585. K. Bruhns, J. H. Burton, A. Rostoker, in En Tecnología y organización de la producción de cerámica prehispánica en los Andes, I. Shimada, Ed. (PUCP, Fondo Editorial, 1994), pp. 53–66. C. Von Hillebrandt, in El paisaje volcánico de la sierra ecuatoriana, P. Mothes, Ed., vol. 4 of Estudios de Geografía (Corporación Editora Nacional, Colegio de Geógrafos del Ecuador, 1991), pp. 39–53. M. Monzier et al., J. Volcanol. Geotherm. Res. 90, 49–79 (1999). L. Siebert, J. Volcanol. Geotherm. Res. 22, 163–197 (1984). V. Valverde, P. A. Mothes, B. Beate, J. Bernard, J. Volcanol. Geotherm. Res. 411, 107172 (2021).

Submitted 10 May 2023; accepted 15 November 2023 10.1126/science.adi6317

7 of 7

RES EARCH

CANCER IMMUNOLOGY

Hyperglycosylation of prosaposin in tumor dendritic cells drives immune escape Pankaj Sharma1†, Xiaolong Zhang1†, Kevin Ly1†, Ji Hyung Kim1, Qi Wan1, Jessica Kim1, Mumeng Lou1, Lisa Kain2, Luc Teyton2, Florian Winau1* Tumors develop strategies to evade immunity by suppressing antigen presentation. In this work, we show that prosaposin (pSAP) drives CD8 T cell–mediated tumor immunity and that its hyperglycosylation in tumor dendritic cells (DCs) leads to cancer immune escape. We found that lysosomal pSAP and its single-saposin cognates mediated disintegration of tumor cell–derived apoptotic bodies to facilitate presentation of membrane-associated antigen and T cell activation. In the tumor microenvironment, transforming growth factor–b (TGF-b) induced hyperglycosylation of pSAP and its subsequent secretion, which ultimately caused depletion of lysosomal saposins. pSAP hyperglycosylation was also observed in tumor-associated DCs from melanoma patients, and reconstitution with pSAP rescued activation of tumor-infiltrating T cells. Targeting DCs with recombinant pSAP triggered tumor protection and enhanced immune checkpoint therapy. Our studies demonstrate a critical function of pSAP in tumor immunity and may support its role in immunotherapy.

1 of 11

,

12 January 2024

y

Sharma et al., Science 383, 190–200 (2024)

y g

*Corresponding author. Email: [email protected] †These authors contributed equally to this work.

y

Program in Cellular and Molecular Medicine, Boston Children's Hospital, Department of Pediatrics, Harvard Medical School, Boston, MA 02115, USA. 2Department of Immunology and Microbiology, Scripps Research Institute, La Jolla, CA 92037, USA.

First, we investigated the effect of saposins on the integrity of apoptotic bodies derived from tumor cells. For this purpose, we exposed murine MCA fibrosarcoma cells to g irradiation [100 Gy (unit of absorbed dose of ionizing radiation); 1 Gy = 100 rads] to trigger apoptotic cell death (Fig. 1A). Successful induction of apoptosis was controlled by measuring PS exposure with AnnexinV staining. Subsequently, we purified apoptotic vesicles from cell culture supernatants using differential ultracentrifugation (100,000g pellets) and visualized them by transmission electron microscopy (Fig. 1B). We then loaded the fluorescent dye calcein into those apoptotic vesicles using a liposome extruder with a 100-nm pore size. After incubation with different recombinant saposins, we measured calcein release and found that saposins disintegrate tumor cell– derived apoptotic vesicles when compared with control bovine serum albumin (BSA) (Fig. 1B). In addition to leakage of the small molecule calcein, we also tested whether saposins facilitate vesicular release of larger proteins and incubated liposomes enclosing fluorescein isothiocyanate (FITC)–labeled ovalbumin (OVAFITC) with single saposins. In contrast to incubation with BSA, saposins triggered the release of soluble OVA protein, with various saposins exhibiting different membrane-perturbing potency (fig. S1A). We next explored the impact of saposins on the processing of apoptotic bodies in DCs. To this end, we first assessed the subcellular localization of pSAP in the endosomal or lysosomal compartments using DC2.4 cells (murine DC line) transfected with a pSAP expression vector containing a C-terminal FLAG tag. We found maximum colocalization of pSAP with LAMP-1 protein, indicating its predominant distribution in lysosomes (fig. S1B). Next, we pulsed bone marrow–derived DCs with carboxyfluorescein diacetate succinimidyl ester (CFSE)–labeled apoptotic cells derived from g-irradiated MCA101 tumor cells and followed their fate along the endolysosomal compartment. Confocal microscopy revealed colocalization of apoptotic bodies with LAMP-1, which indicated trafficking to saposin-containing lysosomes (Fig. 1C). When comparing the kinetic digestion of apoptotic cells in pSAP-deficient or wild-type (WT) DCs, we found that early uptake of apoptotic material was similar, suggesting that phagocytosis is not affected by saposin deficiency. However, at later time points after endocytosis, pSAP-deficient DCs accumulated CFSE-labeled cells, demonstrating

g

1

tumors often develop mechanisms to evade immune responses; for example, by producing immunosuppressive cytokines such as transforming growth factor–b (TGF-b) (14, 15). Thus, the goal of cancer therapy is to target and overcome immune evasion to restore protective immunity. Prosaposin (pSAP) is a precursor protein that is transported from the Golgi apparatus to the lysosome, assisted by its chaperone, sortilin (16, 17). In the lysosome, cathepsins cleave pSAP into the single saposins A to D. Saposins are also called sphingolipid activator proteins because they function as small, nonenzymatic cofactors for lysosomal hydrolases that are required for sphingolipid degradation (18). Moreover, at the low acidic pH of the endolysosomal compartment, saposins are able to interact with anionic phospholipids, such as phosphatidylserine (PS), exposed on intralysosomal vesicles (19). These membraneperturbing properties facilitate vesicle disintegration and pertain to apoptotic vesicles that characteristically contain PS in their lipid bilayers (20, 21). In this context, tumors, owing to their uncontrolled growth kinetics, produce a substantial amount of dying cells and apoptotic bodies that contain tumor antigens to potentially trigger the immune system. Notably, membrane-associated particulate antigen is more immunogenic than soluble protein, and thus, antigen presentation pathways based on vesicular processing might be central to the induction of protective T cell immunity. In this study, we explored the impact of saposins on presentation of membrane-associated tumor antigen and activation of CD8 T cell responses that protect against cancer growth. This work also describes a mechanism by which the tumor counteracts saposin-mediated processing by triggering pSAP hyperglycosylation and secretion from tumor-associated DCs. Ulti-

Results Saposins promote cross-presentation of membrane-associated tumor antigen

p

A

ntigen-specific T cell responses are central to protection against cancer. Cytotoxic CD8 T lymphocytes recognize tumor antigens presented by major histocompatibility complex I (MHC-I) molecules and subsequently deploy their effector functions, such as target-cell killing and production of inflammatory cytokines (1). However, tumor cells often fail to directly activate T cells because of down-regulation of their MHC-I pathway (2, 3). Therefore, other antigenpresenting cells, such as dendritic cells (DCs), are critical to engulf tumor antigens for subsequent processing and display to MHC-I– restricted CD8 T cells in a process called cross-presentation (4). On a cellular level, this mechanism can be broadly divided into a cytosolic and vacuolar pathway (5, 6). According to cytosolic processing, endosomal antigens are retrotranslocated into the cytosol for degradation by proteasomes and subsequent reimport into the endosome for MHC-I loading. By contrast, in the vacuolar pathway, the endosome is more autonomous and relies on its proteases for antigen processing (7–9). Among the different DC subsets, classical DC1s are especially efficient in cross-presentation and fulfill this function in tumor-draining lymph nodes for T cell priming as well as in the tumor microenvironment to activate tumor-infiltrating T lymphocytes (10, 11). Abundance and proper function of immune cells, including antigenpresenting cells and lymphocytes, at the tumor site are vital for effective immunity and control of cancer growth (12, 13). Unfortunately,

mately, we tested the proof-of-principle of pSAP targeting to DCs as a possible mode of cancer immunotherapy.

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

60 40

BSA

0

SAP-D

20

E

ns

12 10 8 6 4 2 0

MFI (x1000)

10 5

V m A O VA

0

sO

2h 4h 8h

G % CFSElow cells

% CFSElow cells

80 60 40 20 0

60 40 20 0

which allows for targeting of OVA to PSexpressing vesicles (22). We analyzed productive antigen processing by staining for the processed OVA epitope in complex with MHC-I (H-2Kb-SIINFEKL) on the DC surface. Flow 2 of 11

,

microscopy images showing the kinetics of apoptotic cell disintegration in WT and pSAP-KO DCs. DCs were pulsed with CFSE-labeled, g-irradiated apoptotic MCA101 tumor cells for 2 hours, and the numbers of apoptotic bodies (ApoBD) were quantified at the indicated time points with ImageJ software (DAPI, blue; LAMP-1, red; ApoBD, green). (E) Representative histogram overlays and bar graph showing flow cytometry staining and mean fluorescence intensity (MFI) of MHC-I-SIINFEKL peptide on the surface of WT or pSAP-KO DCs after incubation with either soluble OVA (sOVA) or irradiated MCA101-OVA tumor cells (mOVA, membrane-associated OVA) for 4 hours. (F) Histograms and bar graph showing frequencies of proliferating CFSElow CD8 T cells after a three-day coculture with WT or pSAP-KO DCs pulsed with soluble OVA. (G) Histograms and bar graph depicting the frequencies of CFSElow CD8 T cells after a three-day coculture with WT or pSAP-KO DCs pulsed with irradiated MCA101-OVA cells. Additionally, pSAP-KO DCs were reconstituted with 10 mg/ml of recombinant pSAP prior to the T cell assay. Data shown in all graphs represent mean ± SD from three to five independent replicates. P values were determined with one-way analysis of variance (ANOVA) [(B) and (G)] or unpaired Student’s t test [(D), (E) and (F)]. *P < 0.05; **P < 0.01; ***P < 0.001; ns, not significant.

ptotic MCA101 cells expressing a membraneassociated form of the antigen OVA prior to coculture with OVA-specific CD8 T cells. MCA101-OVA cells express the model antigen coupled to the C1C2 domain of lactadherin,

12 January 2024

*** ***

y

the importance of saposins for processing of apoptotic bodies in DCs (Fig. 1D). To test for saposin-dependent cross-presentation and CD8 T cell activation, we pulsed DCs from pSAP-knockout (KO) or WT mice with apo-

80

y g

Fig. 1. Saposins promote cross-presentation of membrane-associated tumor antigens. (A) Diagram depicting the experimental read-outs used in Fig. 1. MCA101 fibrosarcoma cells were g-irradiated (100 Gy) prior to the collection of apoptotic vesicles from supernatant and analysis with electron microscopy (EM) and calcein leakage assay. Apoptotic MCA101 or MCA101-OVA cells were used to pulse bone marrow–derived DCs from WT or pSAP-KO mice prior to the analysis of digestion of apoptotic cells with confocal microscopy and antigen processing and T cell activation with FACS. (B) Calcein leakage assay to quantify the effect of saposins on disintegration of apoptotic bodies. Apoptotic vesicles were prepared with differential ultracentrifugation (100,000g) from the supernatant of irradiated MCA101 cells and visualized with transmission electron microscopy (left). Scale bar, 200 nm. Apoptotic bodies were further loaded with calcein dye prior to incubation with indicated saposins or BSA (negative control), and calcein release was quantified in the supernatant with fluorimetry. (Right) Depicted is percent leakage compared with 100% lysis induced by Triton X-100. SAP, saposin. (C) Representative confocal microscopy image showing colocalization of apoptotic bodies (green) with LAMP1 (red). WT DCs were pulsed with CFSE-labeled, g-irradiated apoptotic MCA101 tumor cells for 2 hours. ApoBD, apoptotic body. (D) Representative confocal

y

ns

Sharma et al., Science 383, 190–200 (2024)

ns

g

F

15

p

Number of ApoBDs/cell

D

80

SAP-C

Confocal microscopy Antigen presentation assay

DC:CD8 T cell coculture 72 hours FACS

SAP-B

EM and calcein leakage assay on vesicles

DCs

100

SAP-A

Irradiate tumor cells

C

B

Apoptotic tumor cells

% Calcein Leakage

A

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

p g y y g y ,

Fig. 2. pSAP is required for tumor immunity and boosts T cells derived from melanoma patient samples. (A) Experimental scheme of tumor challenge studies. WT and pSAP-KO BM chimeric mice were primed with 4 × 105 g-irradiated MCA101-OVA cells subcutaneously (s.c.) and subsequently inoculated with 1 × 106 live MCA101-OVA cells (s.c.) 7 days post priming (D0). (B) Comparison of tumor sizes between WT and pSAP-KO mice on day 17 (left) and the kinetics of the tumors’ growth (right). (C) Representative histogram overlay and bar graph depicting the staining and MFI of MHC-I-SIINFEKL peptide on the surface of tumor DCs from pSAP-KO or WT animals. (D) FACS plots and bar graphs showing frequencies of MHC-I (Kb-SIINFEKL) tetramer- and IFN-g–positive tumor-infiltrating CD8 T cells in pSAP-KO or WT mice. MHC-I tetramer specifically detects CD8 T cells that are reactive with SIINFEKL peptide. (E) Experimental setup for the coculture of myeloid and CD8 T cells isolated from human melanoma. Single-cell suspensions from human melanoma samples were FACS-sorted for CD146+ melanoma cells, CD8+ T cells, Sharma et al., Science 383, 190–200 (2024)

12 January 2024

and CD11c/b+ myeloid cells. CD146+ cells were g-irradiated and incubated with DCs, which were further cocultured with CD8 T cells in the presence or absence of recombinant pSAP. (F) FACS plots and bar graph showing the frequencies of IFN-g– positive CD8 T cells following the indicated culture conditions. (G) Representative histogram overlay and bar graph demonstrating surface staining and MFI of LAMP-1 on CD8 T cells according to the indicated culture conditions. Graph colors represent the same as in (F). (H) Flow cytometry analysis and summarizing bar graph depicting the frequencies of antigen-specific CD8 T cells reactive with HLA-A*0201 tetramers loaded with epitopes from gp100, MART-1, tyrosinase, and NY-ESO-1 following the indicated culture setups. Amino acid residues are depicted on top, and percentages of gated cells are shown as mean ± SD in the dot plots. Data shown in (B) to (D) are representative of three independent experiments, whereas (F) to (H) depict mean ± SD from seven independent subjects. P values were determined by unpaired Student’s t test [(B) to (D)] or one-way ANOVA [(F) to (H)]. **P < 0.01; ***P < 0.001; ****P < 0.0001. 3 of 11

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

4 of 11

,

12 January 2024

y

To study pSAP in the context of human cancer, we used dissociated tumor cell samples from melanoma patients. A detailed description of the patient samples can be found in table S1. Briefly, the majority of specimens were isolated from primary melanoma lesions of the skin, assigned as clinical stage III, including those from white female and male patients older than 50 years before treatment (table S1). To purify antigen-presenting cells, responder T cells, and tumor cells as source of antigen, we used fluorescence-activated cell sorting (FACS) on CD146+ melanoma cells, CD11b/c+ myeloid cells, and CD8+ T cells (fig. S4). We then irradiated CD146+ melanoma cells and pulsed them onto sorted myeloid cells prior to coculture with autologous CD8 T cells (Fig. 2E). In parallel, we also treated DCs and T cells with human recombinant pSAP. Five days after culture, we analyzed effector functions of CD8 T cells and found that recombinant pSAP was able to boost IFN-g production (Fig. 2F) and cytolytic activity, indicated by surface LAMP-1 staining as a sign of cytotoxic degranulation (Fig. 2G). Furthermore, we measured the frequencies of tumor antigen–

Beyond the use of pSAP deficiency in a mouse model, we aimed at understanding the regulation of pSAP in tumor-associated DCs in a pathophysiological context. To this end, we inoculated WT mice with live MCA101-OVA cells subcutaneously, and after tumor outgrowth, we isolated DCs from the TME, tumordraining lymph nodes, and spleen (fig. S5A). We FACS-purified the two main classical DC subsets on the basis of their established markers as cDC1 (XCR1+ or CD103+) and cDC2 (SIRP1a+ or CD11b+) prior to performing an array of antigen processing and presentation assays. Accordingly, pulsing with FITC-dextran showed that the phagocytosis rate of tumor DCs was not altered when compared with lymph node and spleen (fig. S5B). Incubation with a self-quenched antigen conjugate (DQOVA), which exhibits fluorescence upon proteolytic degradation, demonstrated that mainly cDC2s in the tumor are compromised to process soluble antigen (fig. S5C). These findings were in line with the ultimate epitope expression on surface MHC-I following pulsing with soluble OVA, which revealed hampered antigen presentation by tumor cDC2 (fig. S5D). In sharp contrast, the presentation capacity of cDC1 in the TME was only affected after incubation with irradiated MCA101-OVA cells, which contain antigen in membrane-associated form (fig. S5D). This phenomenon was reflected in functional T cell experiments because DCs isolated from tumors were severely perturbed to induce T cell responses reactive to membraneassociated antigen (fig. S5E). Because we found that saposins are critical for presentation of particulate antigen, we then hypothesized that pSAP function might be modulated in tumor DCs as a basis for poor T cell induction in the TME. Indeed, analysis by immunoblot revealed the expression of a 75-kDa high molecular weight form of pSAP in tumor DCs when compared with pSAP-65, which was predominant in DCs from spleen (Fig. 3A). Moreover, the small, single saposins were severely depleted in DCs from the TME (Fig. 3A). When we cultured the respective DC subsets ex vivo, we observed secretion of pSAP

y g

Sharma et al., Science 383, 190–200 (2024)

T lymphocytes from melanoma patients are boosted by pSAP

Hyperglycosylation of pSAP in tumor DCs leads to its secretion

y

Next, we investigated how pSAP function affects T cell activation in vivo and protection against cancer. To assess T cell priming, we transferred naïve CFSE-labeled, OVA-specific CD8 T cells into pSAP-deficient or WT recipients prior to subcutaneous administration of apoptotic MCA101-OVA cells (fig. S2A). Four days after tumor cell injection, we isolated DCs from skin-draining lymph nodes and analyzed antigen processing. Expression of H-2Kb-SIINFEKL proved to be reduced in DCs from pSAP-KO when compared with that of WT mice (fig. S2B). Moreover, antigenspecific proliferation and interferon (IFN)–g production by CD8 T cells from lymph nodes were severely hampered in the absence of pSAP (fig. S2C). Thus, pSAP facilitates CD8 T cell priming in response to particulate antigen. Because straight pSAP-KO mice have a reduced life span, we generated chimeric mice by transferring pSAP-KO or WT bone marrow to WT recipients to allow for tumor challenge experiments. In this context, we immunized mice with irradiated tumor cells and challenged them with a higher number of live MCA101-OVA cells one week later (Fig. 2A). Subsequently, we monitored tumor growth in the skin and found a drastic expansion of

specific CD8 T cells by staining with MHC-I tetramers loaded with dominant melanoma antigens, including MART, gp100, tyrosinase, and NY-ESO-1. The abundance of melanomaspecific CD8 T cells was greatly increased when tumor DCs were treated with pSAP (Fig. 2H). Notably, the patient samples were human leukocyte antigen (HLA)–typed by flow cytometry beforehand to select the proper haplotype for MHC-I tetramer analysis (HLA-A02). Overall, the impact of pSAP on tumor DCs is able to rescue T cell activation from melanoma patients.

g

pSAP is required for tumor immunity

cancer in pSAP deficiency (Fig. 2B). Flow cytometry analysis of isolated DCs from the tumor site showed a pSAP-dependent decrease in antigen processing and presentation (Fig. 2C). Furthermore, MHC-I tetramer–mediated detection of antigen-specific CD8 T cells showed reduced frequency of tumor-infiltrating T cells as well as cytokine production in pSAP-deficient mice (Fig. 2D). Additionally, we also challenged bone marrow–chimeric mice with live tumor cells without prior vaccination (fig. S3A). As a result, cancer protection, antigen processing in tumor DCs, frequency of tumor-infiltrating, antigenspecific T cells, and cytokine production and cytotoxicity were all greatly reduced when pSAP was lacking (fig. S3, B to E). To exclude that pSAP-dependent tumor immunity was influenced by differential DC migration, we examined migratory DCs in tumor-draining lymph nodes and found no alteration in DC frequencies and expression of the homing marker CCR7 in pSAP deficiency (fig. S3, F and G). Moreover, we analyzed DC-attracting chemokines at the tumor site and observed comparable expression of CCL19 and CCL21 in pSAP-KO and WT mice (fig. S3H). Because pSAP has been associated with macrophage function and inflammation (23), we controlled our mouse model for a possible impact on tumor-associated macrophages. However, pSAP deficiency neither affected the number of macrophages nor their expression of inflammatory genes in the tumor microenvironment (TME) (fig. S3, I and J). Altogether, these findings demonstrate that tumor immunity critically depends on pSAP function.

p

cytometry demonstrated that processing of soluble OVA was equally efficient in pSAP-KO and WT DCs (Fig. 1E). However, processing and MHC-I loading of membrane-associated antigen derived from tumor cells was significantly hampered in the absence of pSAP (Fig. 1E). These findings were in accordance with our T cell data, as pSAP-deficient DCs efficiently activated CD8 T cells in response to soluble antigen as well as OVA-coated latex beads (Fig. 1F and fig. S1C). In sharp contrast, when DCs were pulsed with tumor cells containing membraneassociated antigen, CD8 T cell activation by pSAP-KO DCs was greatly reduced, as highlighted by their truncated CFSE dilution profile in flow cytometry (Fig. 1G). Notably, incubation of pSAP-deficient DCs with recombinant pSAP fully reconstituted CD8 T cell activation (Fig. 1G). Similarly, we pulsed pSAP-KO or WT DCs with irradiated B16F10 melanoma cells prior to coculture with antigen-specific CD8 T cells that recognize the integral membrane protein Pmel-1 (gp100). Impaired T cell activation by pSAP-deficient DCs revealed that saposins are equally important for processing of membraneassociated melanoma antigens (fig. S1D). In addition, pSAP deficiency also compromised MHC-II–restricted presentation of tumor antigens derived from apoptotic bodies, as indicated by diminished stimulation of antigen-specific CD4 T cells (fig. S1E). Taken together, saposins disintegrate apoptotic vesicles and process membrane-associated antigens for presentation to CD4 and CD8 T cells.

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

p g y y g y ,

Fig. 3. Hyperglycosylation of pSAP in tumor DCs leads to its secretion. WT mice were inoculated with 1 × 106 live MCA101 cells, and 18 days post tumor Sharma et al., Science 383, 190–200 (2024)

12 January 2024

inoculation, cDC1 and cDC2 populations were FACS-sorted from tumor and spleen. (A) Immunoblot showing the abundance of pSAP and saposins in 5 of 11

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

TGF-b induces pSAP hyperglycosylation in tumor DCs and drives immune evasion

,

6 of 11

y

We mechanistically addressed the question how hyperglycosylation of pSAP is regulated. To this end, we screened a panel of recombinant cytokines upon incubation of a murine DC line (DC2.4) and found that TGF-b can induce pSAP secretion (fig. S6A). Treatment of DCs with TGF-b triggered dose-dependent induction of pSAP-75 and subsequent secretion into the culture supernatant, as detected by immunoblot and enzyme-linked immunosorbent assay (ELISA), without affecting the expression of the pSAP chaperone, sortilin (Fig. 4, A to D). Moreover, we measured the gene expression of a panel of enzymes involved in the glycosylation pathway in DC2.4 cells treated with TGF-b (fig. S6B). We observed that the up-regulated genes in TGF-b–treated DCs correlated well with the enzyme signature detected in tumor DCs (Fig. 4E), suggesting that TGF-b is responsible for triggering the respective glycosylation program. To test whether TGF-b signaling is indeed required for pSAP hyperglycosylation in vivo, we used mice that lacked TGF-b receptor II specifically in DCs (CD11c-Cre x Tgfbr2flox/flox) for challenge with live MCA101-OVA tumor cells (Fig. 4F). As anticipated, the lack of TGF-b downstream signaling in DCs caused better tumor protection, increased antigen presentation in tumor DCs, and stronger IFN-g production by tumorinfiltrating CD8 T cells (Fig. 4, G to I). More notably, when we isolated DCs, macrophages, and other CD45+ leukocytes from the tumor site for analysis by immunoblot, we found that

y g

12 January 2024

the compromised antigen presentation capacity in the tumor microenvironment.

y

Sharma et al., Science 383, 190–200 (2024)

cient lysosomal delivery of pSAP. For this purpose, we subjected tumor and splenic DCs to a proximity ligation assay (PLA), in which antibodies against pSAP and sortilin are coupled to oligonucleotide probes that allow for subsequent ligation and amplification in case the two target proteins are in close vicinity (10 to 80 nm). Confocal microscopy revealed these PLA events as discrete spots and showed that the spatial relationship between pSAP and sortilin was perturbed in DCs from the TME (Fig. 3G). To assess the direct physical interaction between pSAP and sortilin through biochemistry, we performed immunoprecipitation of pSAP and subsequently resolved sortilin using immunoblot. As a result, the abundance of sortilin recovered from the pSAP precipitate was clearly reduced in tumor DCs, indicating that the interaction of pSAP with its chaperone sortilin is hampered in the TME (Fig. 3H). To translate these findings to the human system, we explored pSAP-sortilin interaction in DCs isolated from the tumor site of melanoma patients compared with that of monocytederived DCs. The resulting PLA signals demonstrated that the spatial interaction of pSAP with sortilin was reduced in melanoma DCs (Fig. 3I). Furthermore, we examined the abundance of the different molecular weight forms of pSAP and found that tumor DCs from melanoma patients exclusively expressed hyperglycosylated pSAP-75 (Fig. 3J). Taken together, our results demonstrate that pSAP is hyperglycosylated in tumor DCs, fails to interact with its chaperone sortilin, and follows instead a secretory route (Fig. 3K). This mechanism of pSAP hyperglycosylation leads to a depletion of intracellular saposins available for antigen processing, which might explain

g

into the cell culture supernatant mainly by tumor DCs (Fig. 3, A and B). To demonstrate that the occurrence of pSAP-75 was due to glycosylation, we tested endoglycosidase H (Endo H) sensitivity of pSAP. Endo H cleaves N-linked glycans between the two proximal N-acetylglucosamine residues only in highmannose carbohydrate chains, but not in complex glycans. After treatment of protein lysates with Endo H, pSAP-65 from splenic DCs was cleaved to lower molecular weight forms, whereas pSAP-75 from tumor DCs proved to be Endo H–resistant, suggesting that it contains complex glycans (Fig. 3C). To corroborate this finding, we performed deeper molecular analysis using mass spectrometry of sugar structures based on purified pSAP bands derived from splenic (pSAP-65) or tumor DCs (pSAP-75). Tandem mass spectrometry showed that the glycan of pSAP-65 mainly consisted of mannose residues, whereas pSAP-75 exhibited complex glycans involving additions of N-acetylglucosamine, galactose, and sialic acid (Fig. 3D). Because glycan structures are synthesized by a diverse set of glycosyltransferases, we compared the expression of glycosyltransferases between tumor and splenic DCs using a quantitative reverse transcription polymerase chain reaction (qRT-PCR) array (Fig. 3E). In tumor DCs, we found up-regulation of several enzymes that facilitate the attachment of complex glycan residues, such as N-acetylglucosaminyltransferases, galactosyltransferases, and sialyltransferases (Fig. 3F). Because hyperglycosylation of pSAP leads to its secretion and therefore, the reduced generation of single saposins, we next investigated the interaction of pSAP with its chaperone sortilin, which is normally required for effi-

splenic DC1. (G) PLA of pSAP and sortilin. Confocal microscopy images of tumor and splenic DC subsets reveal PLA signal between pSAP and sortilin. Blue indicates cell nucleus (DAPI) and magenta represents ligation signal. The violin plot shows quantification of PLA signal, where 200 cells from each sample were analyzed for statistics. (H) Coimmunoprecipitation of sortilin and pSAP in tumor and splenic DCs. (Top) Blot of sortilin pulled down by anti-pSAP antibody. (Bottom) Immunoblot of total sortilin in corresponding DC populations. Bar graphs depict the densitometric and statistical analysis of sortilin abundance in the immunoblots. Red, tumor; blue, spleen. IP-pSAP, immunoprecipitation of pSAP. (I) PLA of pSAP and sortilin in human melanoma and monocyte-derived DCs (MoDCs). Melanoma DCs were sorted as CD11c+ cells from viable CD45+ cells isolated from human melanoma samples, whereas MoDCs were generated by culturing monocytes with interleukin 4 (IL-4) and granulocyte-macrophage colony-stimulating factor (GM-CSF) for 4 days. Blue indicates cell nucleus (DAPI), and magenta represents ligation signal. The violin plot shows quantification of PLA signal, where 200 cells from each sample were analyzed for statistics. (J) Immunoblot of pSAP in human melanoma DCs and MoDCs. (K) Illustration visualizing glycosylation mechanisms that control pSAP trafficking in tumor DCs. Hyperglycosylation of pSAP compromises its interaction with sortilin and reroutes it to the secretory pathway. Data shown in all graphs are representative of three independent experiments, and P values were determined with unpaired Student’s t test. *P < 0.05; **P < 0.01; ***P < 0.001; ****P < 0.0001; ns, not significant.

p

DCs as well as pSAP secretion. The left blot shows the expression of pSAP and saposins in FACS-sorted DC subsets from tumor and spleen, whereas the right blot shows secreted pSAP in the culture supernatant. pSAP-75, hyperglycosylated pSAP; pSAP-65, glycosylated pSAP; SAPs, saposins. (B) Quantification of pSAP secreted by DCs. FACS-sorted splenic and tumor DC subsets were cultured in cRPMI for 48 hours, and pSAP in culture supernatant was quantified with ELISA. (C) Immunoblot of Endo H–treated pSAP. (Left) Mechanism of Endo H that leads to cleavage of high-mannose, but not complex, glycans. (Right) FACS-sorted DCs from tumor and spleen were lysed in radioimmunoprecipitation assay (RIPA) buffer and cell lysates were treated with Endo H for 12 hours at 37°C prior to analysis with immunoblot. (D) Matrix-assisted laser desorption/ ionization–time-of-flight (MALDI-TOF) mass spectrometry analysis of permethylated N-linked glycans of pSAP immunoprecipitated from FACS-sorted CD11c+ DCs. Enzymatically released N-glycans from pSAP of splenic (top) and tumor (bottom) DCs were analyzed. Glycan compositions were assigned based on m/z values. x axis, mass to charge ratio (m/z). y axis, signal intensity of the ions. green circle, mannose; yellow circle, galactose; red triangle, fucose; blue square, N-acetylglucosamine; magenta diamond, sialic acid. (E) Heat map of differentially expressed genes in tumor DCs involved in glycosylation, as analyzed by real-time RT2 profiler PCR array. Splenic DCs were used as the control to calculate fold change in gene expression. (F) Bar graph depicting glycosyltransferase and glycosidase gene expression in tumor compared with that of

C

E β

2.0 1.5 1.0

0.0

10 ng/ml

0.5

Unst.

β

D

2.5 ng/ml

B

pSAP (ng/ml)

A

mRNA fold change in TGF- β stimulated DC2.4 cells

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

B3gnt4 St8sia6 St8sia2 Mgat4a&5 Galnt3&9 B3gnt3 Neu1

101

Galnt6 B4galt1 B4galt5

100 10-1 0.25

1

4

16

H 2000

MFI (x1000)

3 ***

1500 ***

1000 ***

500

**

2 1 p

0

0 0 6 8 10 12 14 16 18 20

Tumor volume (mm 3)

G

102

B3gnt8

R2=0.56

mRNA fold change in tumor DCs compared to splenic DCs

TGF- β

F

103

Days

J

I

K

10 5 0

0 -5 -10 -15

,

12 January 2024

5

y

Sharma et al., Science 383, 190–200 (2024)

y g

Fig. 4. TGF-b induces pSAP hyperglycosylation in DCs and compromises tumor immunity. (A) Experimental setup of murine DC (DC2.4) cells treated with TGF-b and read-outs shown in (B) to (E). (B) Immunoblot showing dose-dependent induction of pSAP hyperglycosylation and secretion by DC2.4 cell line incubated with recombinant TGF-b for 48 hours. (C) Quantification of pSAP in the culture supernatant of DC2.4 cells after incubation with recombinant TGF-b for 2 days as measured with ELISA. Unst., xxx. (D) Immunoblot of sortilin in DC2.4 cells after incubation with recombinant TGF-b for 2 days. (E) Scatter plot showing the correlation of gene expression of glycosyltransferases and glycosidases between tumor DCs and TGF-b–stimulated DC2.4 cells. mRNA fold changes were quantified by real-time RT2 profiler PCR array. CD11c+ tumor DCs were analyzed with splenic DCs as the control, whereas TGF-bstimulated DC2.4 cells were compared with sham-treated DC2.4 cells. (F) Experimental scheme of tumor cell challenge. Tgfbr2f/f and CD11c–Cre x Tgfbr2f/f (Tgfbr2DDC) BM chimeric mice were inoculated with 1 × 106 live MCA101-OVA cells (s.c.). (G) Kinetics of tumor growth in Tgfbr2f/f and Tgfbr2DDC mice. (H) Histogram overlay and bar graph depicting H-2Kb-SIINFEKL staining and MFI on tumor DCs from Tgfbr2DDC mice or Tgfbr2f/f controls on day 20 after tumor cell injection. (I) FACS plots and bar graph showing frequencies of IFN-g+ tumor-infiltrating CD8 T cells in Tgfbr2DDC or Tgfbr2f/f animals on day 20 post tumor challenge.

10

St3Gal1 St8sia2 St8sia6 St8sia4 Galnt6 B3gnt 8 B3gnt 3 B3gnt 4 B3gnt 2 B4galt 5 B4galt 1 Mgat4a Mgat5 Mgat5b Man1b1 Man1a2 Man2b1 Man1c1 Galnt13 Galnt14 Galnt7 Uggt1 Fuca 2 Fuca 1

L

mRNA fold change in tumor DC1 compared to splenic DC1

% IFN-γ+ CD8+ T cells

15

y

γ

g

20

(J) Immunoblot of pSAP and saposins (SAPs) in tumor DCs, macrophages, and other CD45+ leukocytes from Tgfbr2DDC or Tgfbr2f/f mice on day 20 after tumor cell inoculation. (K) Immunoblot of sortilin in tumor DCs from Tgfbr2DDC or Tgfbr2f/f mice on day 20 after tumor cell inoculation. (L) Differentially expressed glycosyltransferase and glycosidase genes in tumor DCs isolated from Tgfbr2DDC or Tgfbr2f/f mice on day 20 after tumor challenge. Splenic DCs from Tgfbr2f/f animals were used as control to calculate mRNA fold change. Data shown in all graphs are representative of three independent experiments, and P values were determined with one-way ANOVA (C) or unpaired Student’s t test [(G) to (I)]. *P < 0.05; **P < 0.01; ***P < 0.001; ****P < 0.0001. 7 of 11

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

,

8 of 11

y

12 January 2024

y g

Based on the importance of pSAP for antigen presentation in the tumor microenvironment, we then aimed at targeting recombinant pSAP to tumor DCs. For this purpose, we coupled pSAP to anti-DEC205, an antibody well established to target the endocytic receptor DEC205 on DCs, using chemical conjugation as previously described (24, 25). Briefly, we incubated recombinant pSAP with tris (2-carboxyethyl) phosphine hydrochloride to expose its sulfhydryl groups, and in parallel, the anti-DEC205 antibody or isotype control IgG were activated for chemical conjugation by sulfosuccinimidyl 4-[N-maleimidomethyl] cyclohexane-1-

y

Immunotherapeutic targeting of tumor DCs with recombinant pSAP

carboxylate. Following overnight incubation, the pSAP-antibody conjugates were concentrated, and successful coupling was analyzed by immunoblot (fig. S8A). In addition, we controlled that the amount of pSAP coupled to anti-DEC205 or isotype control was comparable (fig. S8B). Moreover, we verified that pSAP conjugation still preserved the fine specificity of the DEC205 antibody by showing that the pSAP–anti-DEC205 conjugate stained a similar percentage of DCs when compared to separate anti-DEC205 detection by flow cytometry (fig. S8C). Furthermore, we incubated pSAPKO DCs with the pSAP-antibody conjugates and tested OVA epitope expression and CD8 T cell activation after the pulsing of DCs with apoptotic MCA101-OVA cells (fig. S8D). Accordingly, we observed reconstitution of antigen presentation (H-2Kb-SIINFEKL) and CD8 T cell priming (CD69) specifically by pSAP coupled to anti-DEC205 in a dose-dependent manner (fig. S8, E and F). To assess the impact of glycosylation on pSAP’s biological activity, we purified pSAP-65 and pSAP-75 from DC2.4 cells. Treatment of pSAP-65 with peptide N-glycosidase (PNGase) F generated nonglycosylated pSAP-55 (fig. S8G). Subsequently, we coupled the differentially glycosylated pSAP forms with anti-DEC205 and incubated pSAP-KO DCs with the conjugates prior to performing functional T cell assays. In contrast to pSAP65, hyperglycosylated and nonglycosylated pSAP mediated significantly lower CD8 T cell activation (fig. S8H), which highlights that optimal glycosylation of pSAP is required for deploying its functions in the context of endosomal delivery by pSAP-based therapeutics. After validation of the targeting tool in vitro and ex vivo, we inoculated pSAP-KO bone marrow–chimeric mice with MCA101-OVA tumor cells and injected 100 mg of pSAP coupled with anti-DEC205 or isotype IgG on days 9 and 13 after tumor challenge (Fig. 5A). On day 14, we isolated and sorted intratumoral DCs, focusing on the two major classical DC subsets, cDC1 and cDC2. Staining for intracellular pSAP in the respective DC subsets revealed effective delivery of pSAP targeted through DEC205 when compared with the isotype control (Fig. 5B). Additionally, immunoblot analysis revealed that pSAPdeficient tumor DCs took up recombinant pSAP-65 and generated massive amounts of single saposins upon lysosomal cleavage (Fig. 5C). This demonstrated successful reconstitution of pSAP-deficient DCs in the TME, especially in the cDC1 subtype that is central to crosspriming of CD8 T cells. We then followed a similar experimental schedule using pSAP targeting of tumor-inoculated WT animals (Fig. 5D). Treatment with pSAP coupled to antiDEC205 greatly reduced tumor burden in WT mice (Fig. 5E). Correspondingly, antigen presentation of OVA peptide by tumor DCs as

g

Sharma et al., Science 383, 190–200 (2024)

lation failed to affect pSAP secretion and T cell induction (fig. S7, E and F). Notably, gene silencing of St6Gal enzymes was sufficient to reconstitute the pSAP-sortilin PLA signal, implicating terminal sialic acid residues as critical blockers of this interaction. To confirm these findings, we used an established inhibitor of sialyltransferase (3F-NeuAc) upon incubation of TGF-b–treated DC2.4 cells (fig. S7G), which completely abrogated the addition of sialic acid to the pSAP glycan, as indicated by ELISA-based binding of a lectin specific for sialic acid (fig. S7H). Consistent with the siRNA results, inhibition of pSAP sialylation reinstated its interaction with sortilin and prevented eventual secretion, thereby promoting activation of CD8 T cells (fig. S7, I to K). These data indicate that TGF-b–regulated Mgat and B4galt1 enzymes are crucial for the sequential synthesis of the complex pSAP glycan and that terminal sialylation by St6Gal impedes the pSAP and sortilin interaction. Furthermore, we examined the molecular mechanism of TGF-b downstream signaling and its regulation of signature enzymes involved in pSAP hyperglycosylation. For this purpose, we performed chromatin immunoprecipitation of the signal transducers Smad2/3 using genomic DNA derived from TGF-b– treated DC2.4 cells (fig. S7L). When compared with isotype control immunoglobulin G (IgG) and the housekeeping gene b-actin, subsequent amplification by PCR revealed significant enrichment of promoter regions containing Smadresponsive elements in the genes coding for Mgat4a, Mgat5, and B4galt1 (fig. S7L). Taken together, these findings demonstrate (i) the constitutive and TGF-b–regulated enzyme pathway required for the generation of hyperglycosylated pSAP, (ii) the critical function of terminal sialic acid to hamper interaction with sortilin, and (iii) the presence of TGF-b–responsive gene promoter sequences in signature enzymes involved in pSAP glycan synthesis. Overall, TGF-b is essential for hyperglycosylation of pSAP in tumor DCs, a mechanism associated with immune escape.

p

pSAP-75 was specifically absent in DCs that lacked TGF-b signaling, whereas the expression of sortilin in tumor DCs was not dependent on TGF-b (Fig. 4, J and K). Additionally, tumor DCs lacking TGF-b receptor exhibited an abundance of single saposins, which was in great contrast to the saposin depletion and exclusive expression of hyperglycosylated pSAP (pSAP-75) in WT DCs as well as tumor-associated macrophages and other CD45+ leukocytes (Fig. 4J). Furthermore, the enzyme signature involved in glycosylation proved to be reduced when TGF-b signaling was deficient in tumor DCs (Fig. 4L). Next, we investigated how TGF-b mechanistically mediates the glycan complexity of pSAP and its effects on sortilin interaction and pSAP secretion. Our mass spectrometry data revealed that the complex glycan of pSAP-75 is generated through the addition of N-acetylglucosamine, galactose, and sialic acid to the ternary mannose core (Fig. 3D and fig. S7A). On the basis of this observation, we hypothesized that TGF-b controls pSAP hyperglycosylation by modulating the expression of specific enzymes involved in glycan synthesis. Analyzing our gene expression array, we selected TGF-b–induced N-acetylglucosaminyltransferases (Mgat4a and Mgat5) and galactosyltransferase (B4galt1) based on their relevance in the corresponding pSAP glycan pathway identified by mass spectrometry (fig. S7A). In addition to the TGF-b– regulated enzymes, the pSAP glycan pathway is complemented by constitutive glycosyltransferases, such as Mgat1/2, which synthesize precursor molecules for Mgat4a and Mgat5, as well as the sialyltransferases St6Gal1 and St6Gal2 that add terminal sialic acid to the carbohydrate chain generated by B4galt1 (fig. S7A). Subsequently, we performed small interfering RNA (siRNA)–mediated knockdown of those enzyme candidates in TGF-b–treated DC2.4 cells (fig. S7, B and C) prior to testing the impact on pSAP interaction with sortilin using PLA, pSAP secretion based on ELISA, and T cell activation using antigen presentation assays. As a negative control, we chose the TGFb–regulated enzyme Mgat3 that produces a bisecting N-acetylglucosamine on the trimannosyl core, which represents a carbohydrate structure absent from our mass spectrometry data on pSAP-75 and therefore is not expected to play a role in the pSAP glycan pathway (fig. S7B). Notably, the silencing of all pSAP-75– associated glycosyltransferases restored the interaction of pSAP with sortilin, as indicated by drastically amplified PLA signal (fig. S7D). By contrast, knockdown of negative control Mgat3 had no effect on pSAP-sortilin engagement (fig. S7D). In a complementary manner, downregulation of the respective enzyme candidates prevented the release of pSAP into the supernatant of DCs and facilitated proper CD8 T cell activation, whereas Mgat3 modu-

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

B

C

% pSAPpos DCs

A

50 40 30 20 10

F *** ***

1500 1000 ***

500 0

15 10 5

p

MFI (x1000)

2000

0 0 6 8 10 12 14 16 18 20

E

Tumor volume (mm3)

D

D C 2

D C 1

0

Days

G

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

40 20 0

Anti-DEC205/pSAP + anti-PD-L1 Anti-DEC205/pSAP Anti-PD-L1 IgG/pSAP

2000 1500 1000

2500 2000 1500 1000 500 0

32

28

24

20

16

8

12

4

0 0

0

,

500

5

y

Tumor volume (mm 3)

2500

10

y g

I

Tumor volume (mm 3)

% IFN-γ+ T cells

γ

60

10 8 6 4 2 0

y

80

g

% IFN-γ+ T cells

H 10 8 6 4 2 0

Day 32

Days Fig. 5. Reconstitution of tumor DCs with recombinant pSAP drives protection from cancer. (A) Regime of pSAP targeting to tumor DCs. pSAPKO BM-chimeric mice were inoculated with 1 × 106 live MCA101-OVA cells. On days 9 and 13 after tumor cell injection, mice were intravenously treated with pSAP coupled with either anti-DEC205 or isotype control antibodies. (B) FACS plots and bar graph showing the amount of pSAP uptake by DCs at the tumor site as analyzed on day 14 after tumor challenge. (C) Immunoblot of delivered pSAP and single saposins in pSAP-deficient tumor DCs on day 20 after tumor cell inoculation. (D) Experimental setup depicting tumor cell inoculation and pSAP targeting through DEC205 in WT Sharma et al., Science 383, 190–200 (2024)

12 January 2024

mice. (E) Comparison of tumor sizes on day 20 (left) and kinetics of tumor growth (right). (F) Histogram overlay and bar graph showing H-2Kb-SIINFEKL peptide surface staining and MFI on tumor DCs on day 20 post tumor challenge. (G) FACS plots and bar graphs showing percentages of IFN-g– positive CD8 T cells in tumors and dLNs in mice treated with pSAP coupled to anti-DEC205 or isotype control. (H) Flow cytometry and bar graphs showing MHC-I (Kb-SIINFEKL) tetramer (MHC-I-Tet)–positive CD8 T cells in tumors and dLNs. (I) Experimental setup depicting B16F10 melanoma cell inoculation, pSAP targeting through DEC205, and tumor growth kinetics. WT mice were inoculated with 3 × 106 live melanoma cells and were treated 9 of 11

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

with pSAP coupled with anti-DEC205 or isotype control antibodies, either alone or in combination with anti–PD-L1 antibodies. Statistical analysis of tumor volume across all treatment groups is shown on day 32 after tumor challenge. Percentages of gated cells are shown as mean ± SD in the dot plots in (B), (G),

Discussion

,

10 of 11

y

1. M. D. Vesely, M. H. Kershaw, R. D. Schreiber, M. J. Smyth, Annu. Rev. Immunol. 29, 235–271 (2011). 2. M. S. Rooney, S. A. Shukla, C. J. Wu, G. Getz, N. Hacohen, Cell 160, 48–61 (2015). 3. M. Sade-Feldman et al., Nat. Commun. 8, 1136 (2017). 4. A. Alloatti et al., J. Exp. Med. 214, 2231–2241 (2017). 5. J. S. Blum, P. A. Wearsch, P. Cresswell, Annu. Rev. Immunol. 31, 443–473 (2013). 6. J. M. Blander, Annu. Rev. Immunol. 36, 717–753 (2018). 7. L. Shen, L. J. Sigal, M. Boes, K. L. Rock, Immunity 21, 155–165 (2004). 8. L. Saveanu et al., Science 325, 213–217 (2009). 9. T. L. Tang-Huau et al., Nat. Commun. 9, 2570 (2018). 10. K. Hildner et al., Science 322, 1097–1100 (2008). 11. S. T. Ferris et al., Nature 584, 624–629 (2020). 12. T. F. Gajewski, H. Schreiber, Y. X. Fu, Nat. Immunol. 14, 1014–1022 (2013).

y g

12 January 2024

REFERENCES AND NOTES

y

In this work, we demonstrated the critical function of saposins in the processing of corpuscular, membrane-associated antigens required for efficient CD8 T cell activation. Notably, saposins were not essential for cross-presentation of soluble antigen, which mechanistically involves a rather early endosomal or phagosomal compartment (6). In cancer immunity, the provision of particulate antigen derived from dying tumor cells represents a physiologic and potent route of antigen delivery. In this context, it has been shown that only DCs that contain tumor-derived vesicles are able to induce T cell responses (28). Thus, our findings highlight the importance of the lysosome in the cross-priming of T cells and underscore the impact of saposins on the immunogenic pathway of vesicular processing. In tumor biology, previous reports suggested a trophic function of pSAP stimulating the proliferation of cancer cells (29, 30). However, pSAP and a pSAP-derived synthetic cyclic peptide were able to prevent tumor metastasis in a mouse model (31, 32). Our work established a genuine immunological function of pSAP in tumor DCs. Thus, pSAP-driven antigen processing and presentation at the tumor site amplified abundance and functionality of tumor-infiltrating T lymphocytes, ultimately leading to cancer control. TGF-b is a pleiotropic cytokine with a diverse set of immunosuppressive functions and is also produced by most cell types. Therefore, tumors themselves, as well as the infiltrating cells of the TME, can serve as the source and target of TGF-b. For example, tumor-derived TGF-b is capable of restricting T cell infiltration or functionally blocking differentiation of protective T lymphocyte populations (15, 33). Our results revealed that TGF-b acts on tumor DCs to trigger hyperglycosylation of pSAP and its subsequent secretion, depleting the lysosomal pool of saposins required for proper antigen processing and presentation (fig. S10). The secretion of pSAP has been shown to be caused by pSAP oligomerization in a cell line (34). However, in primary DCs in the tumor microenvironment, we did not observe large oligomers and instead found perturbed inter-

action between sortilin and hyperglycosylated pSAP. An additional lysosomal player called progranulin might be involved in this mechanism, as it has been shown to bridge the interaction between pSAP and sortilin (35). Overall, tumors are prone to immune escape, and our work describes a further cancerous strategy to manipulate an antigen processing molecule through glycosylation. Because hyperglycosylation of pSAP occurs along the secretory pathway, an approach to overcome tumor-induced saposin deficiency entails the feeding of recombinant pSAP into the endocytic route of DCs (fig. S10). In this context, our targeting experiments using pSAP coupled to anti-DEC205 demonstrated that reconstitution with fully functional pSAP can restore antigen presentation in tumor DCs, leading to amplified T cell responses and eventual tumor protection. DCs are not only crucial for T cell priming in draining lymph nodes; a growing body of evidence indicates that tumor-associated DCs have vital functions regarding recruitment of effector T cells and stimulation of tumor-infiltrating T lymphocytes, which are overall required for effective immunity to cancer (36–39). Current therapeutic options, such as immune checkpoint blockade, aim at reinvigorating exhausted T cells to augment antitumor responses (40). However, these boosted T lymphocytes need to encounter their respective antigens to deploy their functions in a tumor-specific manner. Consistent with that notion, we showed that pSAP combination treatment can override the resistance of immunologically cold tumors to immune checkpoint inhibitors. Taken together, a pSAP-based therapy might help to restore powerful antigen presentation by tumor DCs with the goal of driving protective immune responses at the site of the cancer.

g

Sharma et al., Science 383, 190–200 (2024)

responsiveness to immune checkpoint blockade (fig. S9E). Taken together, these results highlight the crucial impact of pSAP on antigen presentation by tumor DCs to trigger powerful intratumoral T cell responses and point to a viable future strategy for immunotherapy of cancer.

p

well as frequency of IFN-g-producing, antigenspecific T cells at the tumor site and in draining lymph nodes (dLNs) were significantly amplified upon pSAP treatment (Fig. 5, F to H). Moreover, we investigated whether pSAP therapy is able to amplify antitumor memory. To this end, we first subcutaneously administered 1 × 106 MCA101-OVA tumor cells into the left flank of WT mice prior to treatment with 100 mg of pSAP coupled to anti-DEC205 or isotype IgG (fig. S8I). When tumors reached a volume of ~600 mm3, mice were anesthetized and tumors were surgically removed, followed by tumor cell rechallenge into the contralateral side (fig. S8I). Monitoring tumor volume after rechallenge revealed that DC-targeted pSAP treatment significantly reduced tumor growth (fig. S8J). This increased protection was reflected in a drastic expansion of antigen-specific CD8 T cells at the tumor site (fig. S8K). These data suggest that pSAP-based immunotherapy drives immunological memory toward tumor protection. Because pSAP delivery to tumor DCs promoted an immunologically active TME, we next asked whether pSAP could also rescue immune suppression in immunologically cold tumors. For this purpose, we used the B16F10 melanoma model that displays limited T cell infiltration and low susceptibility to treatment with immune checkpoint inhibitors such as anti–PD-1 despite strong expression of PD-L1 (26, 27). When compared with mice treated with anti–PD-L1 alone, the tumor growth kinetics revealed that pSAP combination therapy was able to overcome the resistance of B16F10 melanoma to immune checkpoint blockade in order to enable protection (Fig. 5I). Furthermore, pSAP combination therapy led to significantly increased frequencies of total CD4 and CD8 T cells among various subsets of tumor-infiltrating immune cells (fig. S9, A and B). Notably, treatment with recombinant pSAP and immune checkpoint blockade activated CD8 T cells, amplified the frequencies of T lymphocytes specific for the melanoma antigen Pmel-1 (gp100), and facilitated robust cytokine production (fig. S9, C and D). On the basis of our results that genetic ablation of TGF-b signaling in DCs fully abrogates pSAP hyperglycosylation and secretion, we investigated whether restored antigen presentation could increase the susceptibility of Tgfbr2DDC mice to anti–PD-L1 treatment in the B16F10 melanoma model. Our tumor challenge data show that the intrinsic loss of TGF-b signaling in DCs indeed enhances the

and (H). Data shown in all graphs are representative of three independent experiments, and P values were determined with unpaired Student’s t test [(B) and (E) to (H)] or one-way ANOVA (I). **P < 0.01; ***P < 0.001; ****P < 0.0001.

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

13. S. K. Wculek et al., Nat. Rev. Immunol. 20, 7–24 (2020). 14. R. A. Flavell, S. Sanjabi, S. H. Wrzesinski, P. Licona-Limón, Nat. Rev. Immunol. 10, 554–567 (2010). 15. S. Mariathasan et al., Nature 554, 544–548 (2018). 16. S. Lefrancois, J. Zeng, A. J. Hassan, M. Canuel, C. R. Morales, EMBO J. 22, 6430–6437 (2003). 17. L. Yuan, C. R. Morales, J. Histochem. Cytochem. 58, 287–300 (2010). 18. A. Darmoise, P. Maschmeyer, F. Winau, Adv. Immunol. 105, 25–62 (2010). 19. T. Kolter, K. Sandhoff, Annu. Rev. Cell Dev. Biol. 21, 81–103 (2005). 20. F. Ciaffoni et al., J. Biol. Chem. 276, 31583–31589 (2001). 21. F. Winau et al., Immunity 24, 105–117 (2006). 22. I. S. Zeelenberg et al., Cancer Res. 68, 1228–1235 (2008). 23. M. M. T. van Leent et al., Sci. Transl. Med. 13, eabe1433 (2021). 24. W. Jiang et al., Nature 375, 151–155 (1995). 25. J. Volckmar et al., J. Vis. Exp. (168): (2021). 26. I. J. Fidler, Cancer Res. 35, 218–224 (1975). 27. S. A. Quezada et al., J. Exp. Med. 205, 2125–2138 (2008). 28. M. K. Ruhland et al., Cancer Cell 37, 786–799.e5 (2020). 29. Y. Jiang et al., J. Pathol. 249, 26–38 (2019). 30. S. Ishihara, D. R. Inman, W. J. Li, S. M. Ponik, P. J. Keely, Cancer Res. 77, 6179–6189 (2017). 31. S. Y. Kang et al., Proc. Natl. Acad. Sci. U.S.A. 106, 12115–12120 (2009).

32. 33. 34. 35. 36. 37.

S. Wang et al., Sci. Transl. Med. 8, 329ra34 (2016). M. Liu et al., Nature 587, 115–120 (2020). L. Yuan, C. R. Morales, Exp. Cell Res. 317, 2456–2467 (2011). X. Zhou et al., Nat. Commun. 8, 15277 (2017). H. Salmon et al., Immunity 44, 924–938 (2016). S. Spranger, D. Dai, B. Horton, T. F. Gajewski, Cancer Cell 31, 711–723.e4 (2017). 38. J. E. Jang et al., Cell Rep. 20, 558–571 (2017). 39. K. C. Barry et al., Nat. Med. 24, 1178–1191 (2018). 40. P. Sharma et al., Cancer Discov. 11, 838–857 (2021). ACKN OWLED GMEN TS

We thank C. Théry for providing us with OVA-expressing MCA101 fibrosarcoma cells. We are grateful to R. Cummings, S. Lehoux, and all members of the National Center for Functional Glycomics (NCFG) at Harvard Medical School for performing glycan mass spectrometry. Illustrations used in the manuscript were created using BioRender.com. Funding: This work was supported by National Institutes of Health grant R01 AI136939 to F.W. Author contributions: P.S. and F.W. designed the study and performed data analysis. P.S., X.Z., K.L., J.H.K., Q.W., J.K., M.L., and L.K. performed experiments. L.T. produced recombinant pSAP and saposins. P.S. and F.W. wrote the manuscript. Competing interests: F.W. and P.S. are listed as inventors on a patent application filed by Boston Children's Hospital (provisional application no. 63/350,734),

which covers the use of pSAP in cancer therapy. The authors declare no conflict of financial interests. Data and material availability: Gene expression data have been deposited to Gene Expression Omnibus and are available under accession number GSE248278. Mass spectrometry data are available from GlycoPOST (ID: GPST000381). All data are available in the manuscript or the supplementary materials. License information: Copyright © 2024 the authors, some rights reserved; exclusive licensee American Association for the Advancement of Science. No claim to original US government works. https://www.science.org/about/science-licensesjournal-article-reuse

SUPPLEMENTARY MATERIALS

science.org/doi/10.1126/science.adg1955 Materials and Methods Figs. S1 to S10 Tables S1 and S2 References (41–46) MDAR Reproducibility Checklist Submitted 7 December 2022; resubmitted 29 August 2023 Accepted 27 November 2023 10.1126/science.adg1955

p g y y g y ,

Sharma et al., Science 383, 190–200 (2024)

12 January 2024

11 of 11

RES EARCH

TRIBOLOGY

Designing metainterfaces with specified friction laws Antoine Aymard, Emilie Delplanque, Davy Dalmas, Julien Scheibert* Many devices, including touchscreens and robotic hands, involve frictional contacts. Optimizing these devices requires fine control of the interface’s friction law. We lack systematic methods to create dry contact interfaces whose frictional behavior satisfies preset specifications. We propose a generic surface design strategy to prepare dry rough interfaces that have predefined relationships between normal and friction forces. Such metainterfaces circumvent the usual multiscale challenge of tribology by considering simplified surface topographies as assemblies of spherical asperities. Optimizing the individual asperities’ heights enables specific friction laws to be targeted. Through various centimeterscaled elastomer-glass metainterfaces, we illustrate three types of achievable friction laws, including linear laws with a specified friction coefficient and unusual nonlinear laws. This design strategy represents a scale- and material-independent, chemical-free pathway toward energy-saving and adaptable smart interfaces.

,

1 of 5

y

12 January 2024

Although the generic flowchart shown in Fig. 1 is expected to be applicable to a large variety of frictional interfaces, the following scientific and technological choices represent only one example of how our design strategy can be implemented. First, we chose to work with contacts between polydimethylsiloxane (PDMS) and glass (26), a widely used pair of materials in contact mechanics and friction studies (1, 11). The behavior of sheared PDMS-glass, single sphereplane contacts has been extensively characterized recently (27–30), which provides useful insights into the single-asperity laws that should be used in the friction model. In particular, the friction force, F, is found to be proportional to the contact area, A, through F = sA, where s is the frictional strength of the PDMS-glass interface. Predicting the macroscale friction force F is thus substantially simplified because, for a given s, predicting F reduces to predicting the total contact area A of all microcontacts. Second, as illustrated in Fig. 1, we chose to use a square lattice of 64 asperities as spherical caps, with a common radius of curvature R = 526 ± 5 mm and distributed summit heights hi. We obtained them by molding a PDMS slab onto an aluminum mold prepared with deterministic spherical holes using a sphereended cutting tool in a micromilling machine (26) (for a typical image of a PDMS textured sample, see Fig. 1). This method of preparing populations of spherical PDMS asperities has been successfully applied in the literature (31–33) but not used to design interfaces with predefined friction laws, as is done in this work. We have calibrated the indentation and shear behavior of single such microcontacts (26). Their normal

y g

Aymard et al., Science 383, 200–204 (2024)

Applying the strategy to centimetric elastomer-glass interfaces

y

*Corresponding author. Email: [email protected]

To bypass the multiscale and multiphysics challenges of friction along rough interfaces, we propose a compromise between simplicity and richness in the surface description, by considering flat-flat contact interfaces between a smooth and a rough surface. Once the material pair is fixed, the designable feature of the interface is the topography of the rough surface, which is built as an ensemble of individual microasperities, with both well-controlled geometrical properties and calibrated contact and friction behaviors against the smooth counter surface. The richness of the emerging macroscale behavior stems from the countless possible combinations of geometrical properties of all individual asperities. Just as the microstructures of the materials can be engineered to provide metamaterials with macroscale properties that are rarely found in nature [see (16–19) for mechanical metamaterials], we propose a design strategy (Fig. 1) to prepare contact interfaces with complex predefined frictional behavior. We denote these as metainterfaces. The strategy starts with a target friction law expressed as the macroscopic relationship Ftarget(P) between the normal load P applied to the interface and the target macroscopic friction force Ftarget. This law is the input of an inversion step, the output of which is a geometrical description of the surface’s topography, including the number of necessary asperities and the list of their individual properties (shape, size, height, position). The inversion is based on two main ingredients. First, the indentation and frictional behaviors of a single microcontact are obtained through a preliminary calibration (topright illustration in Fig. 1), which may, but need not, be captured by an existing tribological model. Crucially, those calibrated behaviors contain any specific effect due to the manufacturing process, interface physicochemistry,

g

Université de Lyon, École Centrale de Lyon, CNRS, ENTPE, LTDS, UMR5513, 69130 Ecully, France.

Results General interface design strategy

p

D

espite centuries of investigation, a comprehensive understanding of friction is still lacking. For instance, predicting from first principles the value of the friction force, F, of a given dry contact interface remains out of reach, mainly because of the multiscale character of surfaces and the multiphysics nature of contact interactions (1). Thus, time- and resource-consuming experimental tests remain necessary to calibrate the frictional behavior of a contact interface as soon as any change is brought to the material pair, shape of the solids, loading conditions, surface finish, or environmental conditions. This inability to master friction is a major obstacle to the optimization of devices whose function relies on dry contact interfaces. For soft interfaces, these devices include sport accessories [e.g., racket coatings (2), shoe soles (3, 4)], robotic grasping devices (5–7), haptic feedback tools for virtual reality (8), and conveyor belts (9). At the present time, surface functionalization is the main approach that is followed worldwide to provide contact interfaces with improved frictional capabilities (1, 10). It often consists of creating a certain surface topography at various length scales or adding a homogeneous or heterogeneous thin coating at the solid surface. Unfortunately, despite many successes in a variety of specific cases (11–15), this approach is still based on trial and error. It does not offer a general, systematic design strategy to convert a set of frictional specifications into an actual interface that offers the expected behavior. By circumventing the main pitfalls that make it difficult to understand friction in natural interfaces, we propose and validate a general design strategy to prepare multicontact interfaces with on-demand frictional features.

or surface contamination. Second, a suitable asperity-based friction model, able to predict the global frictional behavior as the collective response of the population of asperities, is identified. Depending on the expected relevant physics, the model can range from analytical (20–22) to numerical (23) through artificial intelligence [e.g., by extending the approaches of previous work (24, 25) to asperity-based descriptions of surface topography]. Based on the inverted asperities’ geometries, a corresponding sample can be manufactured. The material and characteristic size of the prescribed asperities contribute to the selection of the relevant manufacturing method. Finally, shearing tests against the smooth counter surface enable determination of the resulting friction law, F(P), and, through direct comparison with Ftarget(P), assessment of the overall reliability of the workflow. Discrepancies with respect to the target friction law may arise from an incomplete calibration of the single asperity behavior, incorrect assumptions in the friction model, or manufacturing imperfections.

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

the fraction of the initial contact area that remains when the contact is sheared and just about to slide. We found B to be independent of a0 (Fig. 2C). We also confirmed the proportionality between the contact area and static friction force, f = saf (Fig. 2D). The third choice we made was the type of friction model to be inverted. For the present proof of concept, we found that a very simple linear-elastic asperity-based model was sufficient. We considered N spherical linear-elastic

indentation is well captured by the classical Hertz model of a linear-elastic sphere–rigid plane contact (34) (Fig. 2B), with a composite elastic E modulus E* (E
¼ 1n 2 , where E and n are the Young’s modulus and Poisson’s ratio of the PDMS, respectively). Consistent with previous work (27–30), the contact region is initially circular with an area a0 but decreases anisotropically under shear, by ~10 to 15%, down to an area af (Fig. 2A) at the onset of sliding (static friction). The ratio B = af /a0 thus represents

Target macroscale friction law: Ftarget(P)

P

Inversion of the friction model

R

hi

Number of asperities + list of individual geometrical properties

1cm

Actual macroscale friction law: F(P)

Building a friction law from operating points

For our first example, we consider specificain tions  terms of a list of operating points   Pi Fi F P E
; sB through which the friction law sB E
must pass. An infinity of suitable lists of asperity heights hi exist as the solution to this

y g

Fig. 1. Flowchart of the design strategy. Shown at the top right is an illustration of a single microcontact submitted to a normal force p and a friction force f. An illustration of a metainterface submitted to normal and friction forces, P and F, is shown at the middle right. Dark gray ellipses represent real contact regions. The topography is made of N asperities, each with specifically designed geometrical properties (here, spherical caps of height hi and curvature radius R). At the bottom right is an image of a typical centimetric elastomer-based realization of such a textured sample. Photographs of metainterfaces can be found in Figs. 3 to 5.

y

Surface manufacturing & friction test

g

BA0. The model parameter values were experimentally calibrated (26). Inverting the above friction model consists in determining a suitable list of heights hi such that the predicted friction law satisfies specifications set a priori. Interestingly, in the model, the friction law F F(P) can be rescaled as sB versus EP
, where the F P function relating sB and E
depends only on the geometrical parameters of the asperities, R and hi, and not on the material parameters, s, B, and E*. Thus, in the following three examples, we illustrate the success of our design strategy using  the material-independent relaF P tionship sB E
. This rescaled friction law characterizes solely the effect of surface topography, which is the actual interfacial quantity on which the design is performed.

F

p

Calibration of indentation & friction of single microcontacts

Definition of relevant asperity-based friction model

p

f

DESIGN STRATEGY

asperities with common curvature radius, R, and composite elastic modulus, E*, but distributed summit heights, hi. This population of asperities is brought into contact with a rigid smooth plane, forming a number of microcontacts, which are each assumed to obey both Hertz’s model (34) under pure compression and the calibrated friction behavior under shear. Microcontacts are assumed to be independent [see (26) for justification] such that their individual in-plane locations become irrelevant. The area of the ith microcontact is expressed as a0,i = pR(hi – d) if hi ≥ d and a0,i = 0 if hi < d, where d is the separation between the rigid indenting plane and the base plane that supports the asperities. The macroscopic XN a , quantities are simple sums: A0 ¼ i¼1 0;i XN XN
3=2 4E Af ¼ a ,P ¼ 3=2 a0;i (Hertz’s i¼1 f;i XN i¼1 3p R model), and F ¼ f ¼ sAf , where Af = i¼1 i

y

B

C

0.2

0.2

0

0 0

0.05

p

2/3

0.1

(N

2/3

0.15

)

Fig. 2. Tribological calibration of single microcontacts. Illustrations for PDMS batch 1. (A) Images of a microcontact under normal load p = 0.061 N with no shear, initial area a0 (dark central region) (left) and at the onset of sliding, area af (tangential force applied upward in the image plane) (right). (B) a0 versus p2/3, for three indentation experiments using three different samples (data points). The  2=3 solid line is Hertz’s prediction a0 ¼ p 3Rp for R = 526 mm and E* = 1.36 MPa 4E
Aymard et al., Science 383, 200–204 (2024)

B

0.5

12 January 2024

0.4

f (N)

af (mm2 )

0.1

0

0.5 mm

D 1

a0 (mm2 )

af

a0

,

A

0

0.5

a0 (mm2)

1

0

0.5

1

af (mm2)

[see (26) for information about how those values were determined]. (C) Typical evolution of af as a function of a0 (data points). The solid line is the linear fit showing the existence of a constant area reduction ratio B = af /a0. (D) Typical friction force f versus af (data points). The solid line is the linear fit showing the existence of a constant friction strength s = f/af. Error bars are ±5 mN for p, pffiffiffiffiffi ±(1 mN + 0.01f) for f, and ±2:8 pa  106 m2 for a0 and af. 2 of 5

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

5

h1 > h 2 > h3

F/ B (mm2)

4

h2 h1 & ly h 1 on

3

2

1

Real contact 0

0. 2

0.4

0. 6

P/E* (mm2 )

0.8

µ~ ~

3

B m = 2.41 E* f

mf

2.5

µ~ ~ 0

0.1

B m = 1.51 E* f

0.2

P/E* (mm2 )

0.3

0.4

F/ B (mm2)

Fig. 4. Specifying the friction coefficient at a constant material pair. Experimental friction laws (data points) of two metainterfaces with exponential-like distributions of asperity heights (table S1). Error bars are as described in Fig. 3. Lines are linear fits of the data [over the last 12 (7) points for the blue (red) data]. The red and blue shaded regions around the lines represent the 68% confidence interval on the linear regression. The fitted slopes mf of the rescaled laws, 6.73 ± 0.22 (red) and 9.53 ± 0.25 (blue) [which correspond to friction coefficients m ≈ sB E
mf ¼ 1:51 and 2.41 that are typical of PDMS-glass rough contacts (27, 29)] match the targeted slopes (dashed lines), mt = 6.87 (red) and 9.43 (blue), to better than 2.1%. The insets show typical photographs of the two metainterfaces under pure normal force P = 0.39 N (blue) or 0.93 N (red). Scale bars are 1.5 mm. Aymard et al., Science 383, 200–204 (2024)

12 January 2024

3 of 5

,

0.5

y

mt

1

For our second example, we acknowledged that the most classical descriptor of the frictional properties of an interface is the value of its friction coefficient, m. Commonly considered to be a characteristic of a pair of materials in contact, the friction coefficient does not relate to a particular operating point but rather to the global shape of the friction law. Indeed, unlike the curve shown in Fig. 3, in most natural or human-made interfaces, the friction law F(P) is found to be linear [AmontonsCoulomb friction (1, 35)], with m being the slope of the law. Hence, we targeted linear friction laws with tunable slopes. For surface topographies made of spherical asperities that obey Hertz’s indentation law, as in our metainterfaces, a statistical framework relating the contact area A0 to the normal load P exists [see Greenwood and Williamson’s model (20)]. This framework predicts a linear A0(P) relationship when the asperities’ heights, hi, follow a probability distribution that is exponential (e−h/l/l, where l is the scale parameter

y g

mf

1.5

0

Specifying the friction coefficient at a constant material pair

mt

2

y

3.5

g

Fig. 3. Designing an interface that reaches three preset frictional operating points. The green, blue, and F versus EP
. Error red disks represent targeted operating points. The black points represent measured friction law sB bars are calculated as described in (26). The solid line is the model prediction. Dashed lines are continuations of the two first branches of the solid line if the subsequent height levels of the asperities had not been populated. The top-left inset shows a full image of the unsheared contact under P = 0.92 N. Asperities at each of the three height levels (h1 > h2 > h3) are circled with the same color as their associated operating point (main panel). Scale bar is 1.5 mm. The bottom-right inset shows a magnified view of the real contact of nine microcontacts (the region outlined by the dashed line in the top-left inset).

p

0

problem, each giving a different shape of the friction law between operating points. For pedagogical purposes, we adopted an inversion strategy (26) in which each operating point is reached using a single level of asperity height (fig. S1). To pass from the first (trivial) operating point (0, 0) to the next, one evaluates the maximum number of asperities with identical heights (first height level) required to approach the operating point from below and adds a single adjustment asperity whose height is selected to reach exactly the desired operating point. The procedure is repeated as many times as the number of additional operating points. As a result, upon normal loading, a prescribed number of new asperities enter the contact as soon as the previous operating point is reached. We applied this procedure to design a metainterface with a predicted three-branched friction behavior passing through three nonaligned operating points (Fig. 3). Our measurement points delineate a friction law that closely approaches the three target friction forces to better than 5% (2.5% for the first two points). This agreement shows that the design strategy can be used successfully to prepare real-life metainterfaces that target a nontrivial list of friction forces. The latter design strategy can be extended to an arbitrary number of operating points (although sufficiently smaller than the number N of available asperities), opening the way to the design of interfaces with friction laws defined by many operating points. We expect that inversion procedures that are more versatile than the pedagogical one we present will be developed, for instance, those that invert not only hi but also other geometrical quantities such as the curvature radius of each asperity.

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

F/ B (mm2)

1

0.5

Pc,f /E*

m 1,t

,

Aymard et al., Science 383, 200–204 (2024)

m 1,f

y

Understanding friction in its generality remains a formidable challenge and has been

m 2,f

m 2,t

1.5

y g

Discussion

y

2

Building friction laws with multiple specified linear branches

For our third example, we show that unnatural, nonlinear-shaped friction laws can also be designed, enabling one to go beyond AmontonsCoulomb–like friction. As an illustration, we targeted a succession of two linear branches with specified slopes, m1 and m2, and a crossover at a specified critical normal load, Pc/E*. We first describe how such a friction law can be obtained [see (26)] using a weighted exponential distribution of heights, which generates two subpopulations of asperities. We then explain how to invert the law and find a suitable list of asperity heights, hi [see (26)]. We demonstrate that such bilinear-shaped friction laws can indeed be achieved (Fig. 5) through two different realizations. As targeted, both realizations share the same two slopes, which are different by a ratio of 1.75, but have two different crossover normal loads, Pc/E* (ratio 1.94). The slopes values are fulfilled by better than 5.2% (10% for the ratio of Pc/E*), further demonstrating experimentally the success of our design strategy in satisfying quantitatively nontrivial friction prescriptions.

approaches proposed here as pedagogical illustrations. Static friction is not the only possible target functionality, because the design strategy can presumably be adapted to control other topography-related quantities, including interfacial stiffnesses and adhesion forces. Such extensions require different types of calibrations of individual microcontacts that are dedicated to those target properties. In this respect, note that in our experiments, the friction design strategy can be applied indifferently to the static friction force (as in Figs. 3 to 5) or sliding friction force [see (26)]. Our design strategy should be applicable to other material pairs than PDMS and glass. For materials whose contact remains elastic [see (26) for a criterion], the approach developed in our examples should hold, with material changes being accounted for by the rescaled, material-independent version of the friction   F P law, sB E
. For nonelastic materials, qualitative changes might need to be brought to the calibrated microcontact behavior laws, friction model, and related inversions. For instance, for elastoplastic contacts, enriched asperity-based friction models are required (40, 41). In addition, the plasticity-induced irreversible modifications of the topography will likely affect all

g

for centuries. Thus, controlling friction based on generic principles is often considered unreachable. Our work demonstrates that the friction of dry elastomer-based metainterfaces can actually be finely tuned when the surface topography is designed according to our proposed strategy (Fig. 1). We emphasize that our three examples far from exhaust the myriad possible variations around our design strategy, which provides a simple framework that enables friction-control opportunities, starting from, but not limited to, the universally used friction coefficient. Such topography-based control confirms, in particular, that the static friction coefficient is not a constant for a given pair of materials, as predicted by models (22, 23, 36) and observed in stick-slip experiments (37). The potentially accessible friction laws F(P) are countless, in terms of both their nonlinear shape and the way specifications are expressed (forces, slopes, etc.). Those alternative, richer prescriptions likely require the inversion of more topographic features than the individual asperities’ heights, including their individual radii of curvature and in-plane locations. Such additional degrees of freedom should prompt the development and use of advanced friction models (22, 38) and inversion tools (25, 39) that are not limited to the analytical

p

of the exponential). We adapted this model to account for the existence of a maximum asperity height hm due to the finite number of asperities and extended it to predict linear-like friction laws [details in (26)]. In particular,   we derived the F P value of the slope of sB E
for large P: m ¼ qffiffiffiffi x pR 1e pffiffi pffiffi l erf ð xÞp2ffip xex , where x = hm/l and erf is the error function. m is a rescaled estimator for E
the friction coefficient of the interface (m ≈ sB m). For illustrative purposes, we fixed R and hm and targeted two different slopes m. Using the above expression of m, we identified the two l that were suitable to generate two metainterfaces, each with one of the target slopes (26). The two friction laws that we obtained are linear, and their slopes match the predictions to better than 2.1% (Fig. 4). This agreement validates that our design strategy quantitatively enables the prescription of the slope of a linear friction law [after a nonlinear part at small P that is due to the finite value of hm/l and is responsible for a nonvanishing intercept; see (26)]. The slopes m that  we  achieved for the rescaled friction law F P differ by a ratio of 1.42 (1.60 for the fric
sB E tion coefficients m≈ sB E
m). For a discussion about the range of achievable friction coefficients, see (26). We emphasize that the results provide a practical way to tune the friction coefficient of a contact interface by manipulating the surface topography only, that is, without bringing any change to the bulk materials or physicochemistry of their surfaces.

Pc,t /E* 0

0

0.05

0.1

0.15

0.2

2

P/E* (mm ) Fig. 5. Friction law made of two specified linear branches. Two experimental realizations of bilinear friction laws (data points) based on height distributions given by eq. S6 (table S2). Error bars are as described in Figs. 3 and 4. Solid lines are linear fits of each branch [six (four) first data points for the first branch of the blue (red) curve; eight (seven) last data points for the second branch of the blue (red) curve]. The red and blue shaded regions around the lines are as in Fig. 4. The fitted slopes mf of both linear-like branches in each curve match the target slopes mt (dashed black lines), 7.15 and 12.52, to better than 5.2%. Vertical lines show target (Pc,t/E*, dashed lines) and observed (Pc,f/E*, solid lines) crossover normal loads (in the same colors as the corresponding curves). Insets show typical photographs of the metainterfaces that underlie both friction laws under pure normal force P = 0.23 N (red) or 0.34 N (blue). The frame color refers to that of the corresponding friction law. Scale bars are 1.5 mm. See movie S1 for a qualitative illustration experiment on analogous metainterfaces.

12 January 2024

4 of 5

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

bring relevant real-time changes to the topography, these metainterfaces could also become an asset in the development of smart systems (46) that incorporate functional solid contacts. RE FERENCES AND NOTES

1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12. 13. 14. 15. 16. 17. 18.

21. 22.

25. 26. 27. 28. 29. 30. 31.

SUPPLEMENTARY MATERIALS

science.org/doi/10.1126/science.adk4234 Materials and Methods Figs. S1 to S3 Tables S1 and S2 References (47–51) Movie S1

y g

32. 33.

We thank A. Malthe-Sørenssen, K. Thøgersen, H. A Sveinsson, C. Oliver, M. Guibert, D. Bonamy, and G. Pallares for discussions. We thank D. Roux and N. Morgado for their help with the experimental device used in movie S1 and N. Morgado for the sketches in Fig. 1. We thank Z. Lin and X. Qi for their help in developing the inversion scheme for friction laws defined by operating points. Funding: This work was funded by LABEX iMUST (grant ANR-10-LABX-0064) of the Université de Lyon within the program “Investissements d’Avenir” (grant ANR-11-IDEX-0007) operated by the French National Research Agency (ANR) (J.S.), the ANR through grant ANR-18-CE08-0011 (PROMETAF project) (J.S.), and the Institut Carnot Ingénierie@Lyon (PREGLISS project) (D.D.). Author contributions: Conceptualization: J.S.; Methodology: A.A., J.S., D.D.; Investigation: A.A., J.S., E.D.; Visualization: A.A., J.S., D.D.; Writing – original draft: J.S.; Writing – review and editing: A.A., D.D., J.S.; Funding acquisition: J.S., D.D. Competing interests: The authors declare that they have no competing interests. Data and materials availability: All data are available in the main text or the supplementary materials. License information: Copyright © 2024 the authors, some rights reserved; exclusive licensee American Association for the Advancement of Science. No claim to original US government works. https://www.science.org/about/science-licensesjournal-article-reuse

y

23. 24.

AC KNOWLED GME NTS

g

19. 20.

A. I. Vakis et al., Tribol. Int. 125, 169–199 (2018). M. Varenberg, A. Varenberg, Tribol. Lett. 47, 51–56 (2012). L. Damm et al., Footwear Sci. 6, 155–164 (2014). J. Worobets, J. W. Wannop, Sports Biomech. 14, 351–360 (2015). M. R. Cutkosky, P. K. Wright, Int. J. Robot. Res. 5, 20–37 (1986). A. J. Spiers, B. Calli, A. M. Dollar, IEEE Robot. Autom. Lett. 3, 4116–4123 (2018). Y. Golan, A. Shapiro, E. Rimon, IEEE Robot. Autom. Lett. 5, 2889–2896 (2020). G. Millet, M. Otis, D. Horodniczy, J. R. Cooperstock, Mechatronics 46, 115–125 (2017). H.-W. Zhu et al., Powder Technol. 367, 421–426 (2020). T. Ibatan, M. S. Uddin, M. A. K. Chowdhury, Surf. Coat. Tech. 272, 102–120 (2015). J. Scheibert, S. Leurent, A. Prevost, G. Debrégeas, Science 323, 1503–1506 (2009). B. Murarash, Y. Itovich, M. Varenberg, Soft Matter 7, 5553–5557 (2011). M. J. Baum, L. Heepe, E. Fadeeva, S. N. Gorb, Beilstein J. Nanotechnol. 5, 1091–1103 (2014). N. Li, E. Xu, Z. Liu, X. Wang, L. Liu, Sci. Rep. 6, 39388 (2016). J. Dufils et al., Surf. Coat. Tech. 329, 29–41 (2017). M. Kadic, T. Bückmann, R. Schittny, M. Wegener, Rep. Prog. Phys. 76, 126501 (2013). B. Florijn, C. Coulais, M. van Hecke, Phys. Rev. Lett. 113, 175503 (2014). K. Bertoldi, V. Vitelli, J. Christensen, M. van Hecke, Nat. Rev. Mater. 2, 17066 (2017). J. U. Surjadi et al., Adv. Eng. Mater. 21, 1800864 (2019). J. A. Greenwood, J. B. P. Williamson, Proc. R. Soc. London Ser. A 295, 300–319 (1966). O. M. Braun, M. Peyrard, Phys. Rev. Lett. 100, 125501 (2008). K. Thøgersen, J. K. Trømborg, H. A. Sveinsson, A. Malthe-Sørenssen, J. Scheibert, Phys. Rev. E Stat. Nonlin. Soft Matter Phys. 89, 052401 (2014). J. K. Trømborg et al., Proc. Natl. Acad. Sci. U.S.A. 111, 8764–8769 (2014). G. Yang, Q. J. Li, Y. Zhan, Y. Fei, A. Zhang, J. Comput. Civ. Eng. 32, 04018052 (2018). P. Cinat, G. Gnecco, M. Paggi, Front. Mech. Eng. 6, 29 (2020). Materials and methods are available as supplementary materials. R. Sahli et al., Proc. Natl. Acad. Sci. U.S.A. 115, 471–476 (2018). J. C. Mergel, R. Sahli, J. Scheibert, R. A. Sauer, J. Adhes. 95, 1101–1133 (2019). R. Sahli et al., Phys. Rev. Lett. 122, 214301 (2019). J. Lengiewicz et al., J. Mech. Phys. Solids 143, 104056 (2020). V. Romero, E. Wandersman, G. Debrégeas, A. Prevost, Phys. Rev. Lett. 112, 094301 (2014). S. Yashima et al., Soft Matter 11, 871–881 (2015). V. Acito, M. Ciavarella, A. M. Prevost, A. Chateauminois, Tribol. Lett. 67, 54 (2019).

34. J. Barber, Contact Mechanics, Solid Mechanics and Its Applications Series, vol. 250 (Springer, 2018). 35. E. Popova, V. L. Popov, Friction 3, 183–190 (2015). 36. J. Scheibert, D. K. Dysthe, Europhys. Lett. 92, 54001 (2010). 37. O. Ben-David, J. Fineberg, Phys. Rev. Lett. 106, 254301 (2011). 38. Y. Xu, J. Scheibert, N. Gadegaard, D. M. Mulvihill, J. Mech. Phys. Solids 164, 104878 (2022). 39. T. Djourachkovitch, N. Blal, N. Hamila, A. Gravouil, Comput. Struc. 255, 106574 (2021). 40. W. R. Chang, I. Etsion, D. B. Bogy, J. Tribol. 109, 257–263 (1987). 41. H. Ghaednia et al., Appl. Mech. Rev. 69, 060804 (2017). 42. R. Aghababaei, D. H. Warner, J.-F. Molinari, Nat. Commun. 7, 11816 (2016). 43. R. Sahli et al., Sci. Rep. 10, 15800 (2020). 44. B. Mao, A. Siddaiah, Y. Liao, P. L. Menezes, J. Manuf. Process. 53, 153–173 (2020). 45. Z. Zhang, C. Geng, Z. Hao, T. Wei, Q. Yan, Adv. Colloid Interface Sci. 228, 105–122 (2016). 46. M. A. McEvoy, N. Correll, Science 347, 1261689 (2015).

p

successive uses of the metainterface such that the friction law is no longer a constant but rather a loading history–dependent function. Irreversible changes of the topography may also arise as a result of wear-induced loss of material. Upon sliding, various wear mechanisms may occur, including progressive material removal that leads to blunted asperities and asperity-sized debris formation due to fracture processes (42). With such alterations of the surface topography, the behavior of metainterfaces will progressively deviate from the target one. In this context, evaluating the lifetime of metainterfaces as a function of the materials, roughness, and loading conditions will be an important step before use in any specific engineering application. We expect our design strategy to be applicable over a large variety of length scales. We used asperities with a submillimetric radius of curvature that was obtained using micromilling. Larger asperities can be obtained in the same way or, for example, through 3D printing (43). For smaller scales, other well-established methods might be suitable, from laser ablation for asperities down to the micrometer scale (44) to micro- or nanolithography techniques at submicrometer scales (45). In all cases, experimental challenges include the reproducibility of the shapes of asperities (for microcontact calibration to be relevant), the level of accuracy of the manufactured asperity heights with respect to their prescribed values, and the potential misalignment between the two surfaces during the friction measurements [see (26) for a discussion of the latter two issues]. Many different types of frictional metainterfaces can be developed by using our design strategy, constituting a useful toolbox for designers of friction-based devices. By operating in optimized conditions, these devices should benefit from increased energy efficiency and lifetime. If further equipped with suitable sensors or actuators that

Submitted 21 August 2023; accepted 22 November 2023 10.1126/science.adk4234

y ,

Aymard et al., Science 383, 200–204 (2024)

12 January 2024

5 of 5

RES EARCH

IMMUNOLOGY

Antibody production relies on the tRNA inosine wobble modification to meet biased codon demand Sophie Giguère1*, Xuesong Wang1, Sabrina Huber2†, Liling Xu1, John Warner1, Stephanie R. Weldon1, Jennifer Hu2, Quynh Anh Phan1, Katie Tumang1, Thavaleak Prum1, Duanduan Ma3, Kathrin H. Kirsch1, Usha Nair1, Peter Dedon2,4, Facundo D. Batista1,5,6,7* Antibodies are produced at high rates to provide immunoprotection, which puts pressure on the B cell translational machinery. Here, we identified a pattern of codon usage conserved across antibody genes. One feature thereof is the hyperutilization of codons that lack genome-encoded Watson-Crick transfer RNAs (tRNAs), instead relying on the posttranscriptional tRNA modification inosine (I34), which expands the decoding capacity of specific tRNAs through wobbling. Antibody-secreting cells had increased I34 levels and were more reliant on I34 for protein production than naïve B cells. Furthermore, antibody I34-dependent codon usage may influence B cell passage through regulatory checkpoints. Our work elucidates the interface between the tRNA pool and protein production in the immune system and has implications for the design and selection of antibodies for vaccines and therapeutics.

1 of 7

,

12 January 2024

tRNA decoding capacity is determined by the genome-encoded primary sequences of the

y

Giguère et al., Science 383, 205–211 (2024)

tRNA modifications in antibody-secreting cell lines expand decoding capacity

y g

*Corresponding author. Email: [email protected] (S.G.); [email protected] (F.D.B.) †Present Address: ETH Zürich, Laboratory of Toxicology, 8092 Zürich, Switzerland.

The human immunoglobulin heavy chain (IgH) constant region comprises the bulk of the antibody protein and, consequently, its production demand. Unlike the antigen-recognizing variable region, IgH has not evolved for hypermutability and remains relatively consistent across cells and individuals. In human IGHM—the first isotype expressed and, at 49 kDa, the largest—the frequencies of the 61 sense codons varied substantially, including within synonymous codon sets (Fig. 1A and fig. S1A). Seven codons were not used, despite the presence of all standard amino acids (Fig. 1A and fig. S1, A and B). Each of the 10 most common codons in IGHM encoded a different amino acid and was far more frequent than its synonyms, suggesting strong bias (Fig. 1A and fig. S1C). IGHM synonymous codon usage order (SCUO), which quantifies gene-wide codon bias, was within the top 10% of human coding sequences (CDS) (Fig. 1B), suggesting potential selective pressure on IGHM codon use.

y

The Ragon Institute of Mass General, Massachusetts Institute of Technology (MIT), and Harvard, Cambridge, MA 02139, USA. 2Department of Biological Engineering, MIT, Cambridge, MA 02139, USA. 3BioMicro Center, MIT, Cambridge, MA 02139, USA. 4Singapore-MIT Alliance for Research and Technology, Singapore 138602. 5Department of Immunology, Harvard Medical School, Boston, MA 02115, USA. 6Department of Microbiology, Harvard Medical School, Boston, MA 02115, USA. 7Department of Biology, MIT, Cambridge, MA 02139, USA.

Antibody constant-region genes demonstrate a biased codon pattern that does not predict the corresponding tRNA

g

1

proteotoxic stress and energetic costs (7–10). Although a correlation between codon usage of abundant proteins and tRNA gene copy numbers—often used as a proxy for abundance but weakly correlated—is well documented in single-celled organisms, the interactions between tRNA supply and codon demand may be less straightforward in multicellular eukaryotes (11–14). The tRNA pool of metazoan cells is dynamic, and changes in tRNA expression and modifications affect cell differentiation versus proliferation, as well as tumor metastasis (12, 15–17). Given the importance of antibody production to immunity, we investigated whether antibody-secreting cells have tRNA pools tailored to immunoglobulin production.

p

P

rotein production is among the most energetically costly and stressful processes that cells undertake (1). Efficient translation is critical in cell fitness and function, particularly for highly expressed proteins (2). Antibodies must be synthesized in substantial quantities to provide systemic protection against infection (3). Initially expressed in the hundreds of thousands per cell as membranebound immunoglobulin B cell receptors (BCRs), antibodies are secreted by plasma cells (PCs) at rates of up to 100 million molecules per hour, per cell, for years (4, 5). Antibody production in PCs is tightly regulated, requiring dramatic alterations in cellular biology. Key processes include the coordinated activity of multiple transcription factors, metabolic changes, and expansion of the endoplasmic reticulum (ER) and Golgi complex because cell survival requires the up-regulation of unfolded protein response factors and ER chaperones to prevent toxic antibody aggregation (6). However, little is known about how the translational machinery adapts to antibody production. tRNAs, the primary interpreters of the genetic code, are central and dynamic components of translation. The adaptation of codon usage to the abundance of corresponding tRNAs can affect translational efficiency and accuracy. Perturbations of this balance induce

We next interrogated the usage of each codon within IGHM relative to the human proteincoding genome by calculating the probability of finding a CDS with a lower usage frequency. Overall, IGHM used few codons at a frequency comparable with that of the average human gene: P(Usage) of 0.5 indicates standard usage; 1.0 indicates a higher usage in IGHM than in all other CDS (Fig. 1C). Thus, the pattern of codon usage in IGHM is distinct from the general biases of human protein-coding genes. All members of the IgH family have high SCUO scores (fig. S1D). We found the pattern identified in immunoglobulin M (IgM) throughout: IgH isotype P(Usage) scores were highly correlated, indicating a conserved pattern of codon bias within the family (Fig. 1, C and D). To interrogate whether this pattern was associated with protein structure, we examined the T cell receptor (TCR) constant-region family. Like IgHs, TCRs are antigen receptors with a cell-specific variable region and universal constant region and structurally resemble the Fab portion of immunoglobulins (18). The codon usage patterns of human TCR constant regions did not correlate with those of IgH isotypes, except for TCRb, which was still less similar to any IgH isotype than the IgH isotypes were to each other (Fig. 1, D and E). To determine whether this specific pattern of codon bias was conserved, we compared human and murine IgH constant regions by using hierarchical clustering analysis of codon P(Usage) for all IgH isotypes (Fig. 1F). Across all isotypes, human and murine IgH sequences shared a set of 12 hyperutilized codons, as well as 11 hypoutilized relative to the corresponding genome. Most notably, ACC (Thr) had high P(Usage) in both mice and humans. ACC (Thr) appeared, on average, more frequently in IgH sequences than in nearly any human [P(usage) = 0.995] or murine (0.96) CDS. Murine and human IgH genes shared a characteristic codon bias pattern. To assess the ability of the tRNA pool to match immunoglobulin codon patterns, we examined gene copy numbers for tRNAs corresponding to codons hyperutilized in IgH genes in both species. Eight had multiple directly corresponding tRNA genes for the anticodon (Fig. 1, G and H, and fig. S1F). However, four hyperutilized codons lacked corresponding tRNA genes according to classic Watson-Crick base pairing rules: GTC (Val), TCC (Ser), CCC (Pro), and ACC (Thr) (Fig. 1, G and H, and fig. S1F). Genomeencoded tRNAs were incongruous with the IgH pattern of codon bias. Thus, to translate antibody genes, cells must reprogram available tRNAs to meet codon demand.

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

A

B

C

0.02

P(Usage) Score

0.04

0.4

0.2

0.00

D 1.00

R2 = 0.82 P < 2e-16

IGHA

0.75

0.50

0.25

M

0.00

IG

H

Codon

G

en

om

e

ATA CGA CGT CTA GGA GTA TTA AGA CAA CAT GTT TTT ACT ATT CCG GCG TAT TCA TGT AGT GGT CTT GCA TCG TTG GCT CCT TCT AAT CGC GAA GAT ACA ATG AGG CTC CGG TGG TAC AAA ACG CAC GGG TGC CCA AAC ATC GGC AAG GTC AGC GAC CCC TTC CAG CTG GAG GCC TCC GTG ACC

0.0

Codon

P(Usage) Score

R2 = 0.90 P < 2e-16

IGHM

0.6

SCUO

Frequency

0.06

1.00

R2 = 0.59 P < 3e-13

IGHD

R2 = 0.71 P < 2e-16

IGHE

R2 = 0.90 P < 2e-16

IGHG

0.75

0.50

0.25

0.00

R2 = 0.0091 P = 0.21

TCR alpha

TCR beta

R2 = 0.020 P = 0.14

TCR delta

R2 = 0.00070 P = 0.31

TCR gamma

0.75

0.50

g

P(Usage) Score

1.00

Codon

Codon R2 = 0.41 P < 7e-9

p

Codon

Codon

E

0.25

0.00

IgG2 IgG4 IgG1 IgG3

10

Average P(Usage) Score

IgG2

Species

IgG3 IgE

Human Mouse

IgG1

0 F S

S

S

T

P

N

S

P V

T

Codon

H

Asn (N)

Val (V)

20

GTC

GTG

GTT

GTA

Codon

Fig. 1. Biased codon usage patterns across IgH constant regions. (A) Frequency of codon usage in the human IGHM gene. Mean ± SD. (B) Quartile distribution of SCUO scores, representing total codon usage bias across the human IGHM gene compared with all genes from the human-genome Consensus Coding Sequence (CCDS) dataset. (C) Mean of codon P(Usage) scores for human IGHM. The codons on the x axis are ranked by the combined mean scores for all five human IgH isotypes, using a linear correlation model. (D) Mean of codon P(Usage) scores for all other human IgH isotypes. The x axis is ordered as in (C), and codons were analyzed with a linear correlation model. (E) Same as (C) and (D) but showing scores for each human TCR constant-region chain with the x axis Giguère et al., Science 383, 205–211 (2024)

AAC

Hypoutilized Codons

0

AAT

ACA AAC TGG TTC GTC ACC AGC CCC TGC CCA TCC GTG AGG TAT CAA ATG ATT GCA TTT CTT GTT CGA TTA CGC CGG GCG CAG CGT GCC GGC CTG TTG GCT GGA GAT CTA AAT GAA CAT ATA ACG GGG GAC CTC AAG CAC TCG TAC GAG CCG CCT TCA AGA ATC ACT TCT TGT AAA GTA AGT GGT Hyperutilized Codons

5

,

1.00 0.75 0.50 0.25 0.00

10

12 January 2024

y

Average P(Usage) Score

15

IgA

1.00 0.75 0.50 0.25 0.00

S

P

y g

1.0 0.8 0.6 0.4 0.2 0.0

W

T

V

5

V

F

VSPT

TTT AGT GTA AAT GTT TCT CCT ACT TCA ACA TCG TGG AAC CCA TTC CCG TCC GTG ACG CCC GTC AGC ACC

P(Usage) Score

tRNA Gene Count

IgA1 IgE IgD

IgH Isotype

N

20

15

IgA2

IgD IgE

G

tRNA Gene Count

TTT ATG ATT GCT TAT CTT GAA TTG GGA GAT CA A GCA AGT GTA AGA CAT AAT AGG CTA GTT TCT ATA CCT GGT A CT CGA TGT TTA ATC AAA CGC TCA CGT CGG GGG A CA GCC GCG CTC CAG AAG GAC GAG CTG GGC CAC TCG TAC TAA TGG AAC TGA CCA TAG TTC CCG TCC TGC GTG A CG CCC GTC AGC ACC

F

Codon

y

Codon

Codon

Codon Codon Order:

ordered as in (C), analyzed with a linear correlation model. (F) Hierarchical clustering analysis of codon P(Usage) scores for human and mouse IgH genes. Each row represents one gene isoform [as delimited by the International Immunogenetics Information System (IMGT)]; each column represents a codon. (G and H) tRNA gene copy number in the human genome for the amino acids corresponding to hyperutilized codons bracketed in (F). Codons and tRNAs are matched by Watson-Crick base pairing. Codons lacking genomeencoded Watson-Crick complementary tRNAs are highlighted in red. Bars are color-coded by the average codon probability score for all IgH isotypes across humans. Source data for this figure can be found in data S1 and S2. 2 of 7

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

3 of 7

,

To assess the functional relevance of I34modified tRNA decoding to protein production, we designed a fluorescent reporter expression assay (fig. S3A). We generated plasmids with enhanced green fluorescent protein (GFPWT)— which hyperutilized I34-dependent codons compared with all murine CDSs—or I34 codonmodified eGFP (GFPI34mod), which dramatically underutilized these codons (fig. S3B), and we used mCerulean as an internal control (fig. S3C). In GFPI34mod, all I34-dependent codons (NNC)

y

12 January 2024

I34-decoding efficiency is critical to effective protein production, particularly in antibodysecreting cells

y g

Changes in tRNA modification abundance detected with MS may have been caused by differential expression of modifiable tRNAs, the extent to which available tRNAs were modified, or both. For a more comprehensive comparison of tRNA expression in B cells, we performed tRNA sequencing (tRNA-seq) (29) on the sRNA extractions (