Vol 587, 26 November 2020 Nature

Citation preview

The international journal of science / 26 November 2020

UNESCO must reform to stay relevant At 75, the UN agency with a focus on science cooperation is fighting for its future role.

U

NESCO was born on 16 November 1945, just a few weeks after the end of the Second World War. Its founders had been persuaded that science — along with culture and education — could help to cement peace between countries, protect human rights and improve living standards. Now, as the United Nations and UNESCO turn 75, the Paris-based agency is struggling to determine its future. There’s a lot to show for those 75 years. Today, UNESCO operates the system that has awarded World Heritage status to more than 1,100 important historical sites; the agency has also established a global network of more than 700 biosphere reserves. It holds nations to account on their commitments to get every child into school, and monitors threats to journalists around the world. But among the UN’s family of specialized agencies, UNESCO has never been properly funded — and it has been trying to recover from a funding crisis for the past decade. It spent US$1.1 billion in the 2-year period from 2010 to 2011, but in 2012–13, spending was down by 16% after the Palestinian Authority was granted full membership and the United States and Israel stopped their financial contributions in protest. Although its spending was back to $1.1 billion by 2018–19, inflation has greatly reduced its spending power. UNESCO is now in the middle of a transformation designed, in part, to enable it to live within its means. When Nature spoke to UNESCO’s current and former staff, as well as to researchers who study and collaborate with it, we found immense affection for the organization and respect for its past achievements. However, there was also a sense of frustration over its future. UNESCO needs to put these concerns to rest once and for all.

Pulling together UNESCO’s history is a stellar example of science’s power to advance both knowledge and diplomacy. In the wake of two world wars, and especially during the cold war, the agency helped to unlock the doors to international scientific cooperation, particularly in the physical sciences. In 1951, it hosted the meeting that led to the creation of CERN, Europe’s particle-physics laboratory. Since then, CERN has mushroomed from a project intended to reunite and stimulate Europe’s physicists to a place where scientists from all over the world can collaborate. It has spawned a number of technological spin-offs and has maintained its commitment to global knowledge-sharing.

If UNESCO ceased to exist, the world would need to recreate it.”

When nations were reluctant to share their oceans data, UNESCO hosted the first meeting of the International Oceanographic Commission in 1961. The commission still has a role in international efforts to sustainably manage ocean resources. And UNESCO’s efforts to connect scientists from countries with difficult relationships continued with SESAME, the Middle East’s first synchrotron light source. That project was launched in 1999 and provides an essential tool to researchers in a range of fields, from medicine to materials science. Getting scientists from Iran and Israel, or Cyprus and Turkey — all SESAME member countries — to work together is no small achievement. That same year, UNESCO co-organized the World Conference on Science in Budapest. One of the outcomes was the creation of SciDev.Net, one of the first open-access platforms for sharing the results of scientific research, on which Science and Nature worked together to share some of their content with low- and middle-income countries. And all of this happened in an organization that might never have had an ‘S’ in its title. UNESCO was originally conceived to protect and promote education and culture. It made room for science after leading scientists and science media (including Nature) helped to persuade the UN’s founding nations that their vision of a world at peace could not be a world without science. And yet, for all its external successes, UNESCO has faced difficulties in how it is treated by some of its larger member states. That, in turn, has affected the ability of its staff to get things done. It hasn’t helped that some countries have treated their membership of UNESCO like a revolving door, joining and leaving as they wish, with little regard for the consequences for the agency’s work when their funding stops. The United States has left twice, and the United Kingdom and Singapore have also withdrawn in the past, then returned some years later. When richer countries stop paying, projects on the ground suffer, but so does trust in those nations’ commitment to UNESCO’s goals. It means officials at UNESCO’s headquarters are forced to spend time and energy raising funds from other sources, and reorganizing staff and management structures to fit changing priorities — and end up spreading themselves too thinly. Time spent fire-fighting is time taken away from other priorities. In 2013, UNESCO’s leadership responded to its loss of income with a proposal that would probably have led to most of its work in its communication and information sector being abolished. But this was seen as a step too far and rejected by member states. Now, the director-general Audrey Azoulay is trying a different approach — intended, in part, to take some of the political heat out of UNESCO’s work by focusing on things more countries can agree on, and playing to the agency’s strengths as cultural guardian, ethical compass and laboratory of ideas. Azoulay and her team have initiated a “strategic transformation” to spearhead internal reform and programme change — the latter requiring approval by member states late next year. Meanwhile, she is prioritizing five areas: rebuilding and reviving the devastated Iraqi city of Mosul; promoting open science; working on much-needed

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

Nature | Vol 587 | 26 November 2020 | 521

Editorials

common standards on the ethics of artificial intelligence; a long-term vision for education; and biodiversity. The last of these is a belated, but much-needed recognition of UNESCO’s long-standing experience in the study of Indigenous and local knowledge across research fields. Its importance is bolstered by the results of a UNESCO survey that asked 15,000 people what they saw as the biggest threats to peace — two-thirds of respondents said biodiversity and climate change were their greatest concern. There’s also a strong argument for reviving UNESCO’s earlier science mission. In today’s fractured world, fundamental and applied science could once again be used to help bring people and societies together. In the Middle East, for example, UNESCO could help to reconnect scientists in Qatar with those in neighbouring countries. At present, researchers are unable to collaborate because of a regional dispute. The agency could have a greater role in South Asia’s science, which is affected by the strained relations between India and Pakistan. And UNESCO could do more for researchers in Europe, where fractures are developing between members of the European Union. UNESCO should seek to reconnect people through science, as it has done before. But there can be no illusions about how hard the task will be. After 75 years, UNESCO is facing one of its toughest tests. Member states must make every effort to pull together with the agency’s headquarters and its field staff. UNESCO’s potential in a crisis-ridden world should not be underestimated. If UNESCO ceased to exist, the world would need to recreate it.

The challenges for COVID vaccination efforts As positive results emerge at last, researchers must help the world to address vaccine hesitancy, supply logistics and price.

A

year on from the first known case of COVID-19, the world has been hungry for good news. This month, vaccine makers have provided welcome nourishment. Large clinical trials of four vaccine candidates are showing remarkable promise, with three exceeding 90% efficacy — an unexpectedly high rate — according to results released so far. None reported worrying safety signals and one has shown promise in older adults, a demographic that is particularly vulnerable to SARS-CoV-2 but sometimes responds less well to vaccines. Early studies had shown that these candidate vaccines could stimulate an immune response. The latest trials show that this immune response can protect people against

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

522 | Nature | Vol 587 | 26 November 2020

An outbreak anywhere is an outbreak everywhere.”

COVID-19 — a major achievement. Vaccine development is fraught with possibilities for failure, and even the most ardent optimist might not have expected to have a highly effective vaccine against a new virus less than a year after its genome was sequenced. But there is still much work for researchers and clinicians to do. First, they need to determine how well the vaccines work in people who are at high risk of COVID-19, including older individuals, people with obesity and those with diabetes. Second, it isn’t clear how well some of the vaccines protect against severe COVID-19. Third, it is also not clear to what extent the vaccines prevent those who have been vaccinated from passing the virus on to others. Some people are understandably concerned that the speed of both scientific review and vaccine regulation could compromise safety — despite assurances to the contrary from vaccine developers and regulators. To build confidence in vaccination, it’s important that regulators, companies and their research partners keep promises they have made to ensure transparency, publish data and engage with open discussion of those data as they arrive. Much of what we know about the latest trials has been communicated through press releases and media interviews, rather than papers that have been subject to independent peer review. Such speed of communication is necessary in an emergency. But more-complete data should not be held back, and the teams involved must be prepared to provide access to all relevant data as soon as this is practically possible, to allow others to scrutinize their findings and test their claims. Vaccine distribution poses another challenge, and is accompanied by questions such as how much it will cost and who will pay for it. One of the vaccines that have shown success in late-stage trials was developed by researchers at the University of Oxford, UK, and the pharmaceutical firm AstraZeneca in Cambridge, UK. This vaccine can be stored in a normal refrigerator, which makes rapid distribution more feasible than it would be for the vaccine developed by Pfizer in New York City and BioNTech in Mainz, Germany — which needs to be stored at temperatures below −70 °C. Importantly, AstraZeneca and Oxford have also pledged to provide their vaccine at cost price to all during the pandemic, and to maintain this price for middle- and low-income countries after the pandemic. But, as Nature went to press, neither Pfizer nor Moderna, a drug company in Cambridge, Massachusetts, with a similarly promising vaccine candidate, had committed to keeping prices down once the current pandemic is over. They need to change this stance. A number of countries — most of them wealthy — have already pre-ordered nearly four billion doses. COVAX, a global alliance seeking to ensure that middle- and low-income countries get adequate vaccine provision, has been able to secure vaccines for only around 250 million people — nowhere near enough. Once prices start to rise, the poorest countries will be even less able to pay than they are now. Not making the vaccine affordable for them would be morally wrong. It would also be short-sighted, because, as infectious-disease researchers often say, an outbreak anywhere is an outbreak everywhere.

A personal take on science and society

World view

By Elizabeth Gadd

University rankings need a rethink World league tables for higher education are flawed, poorly used and entrench inequity.

PAUL GADD

R

esearchers often complain about the indicators that hiring and grant committees use to judge them. In the past ten years, initiatives such as the San Francisco Declaration on Research Assessment and the Leiden Manifesto have pushed universities to rethink how and when to use publications and citations to assess research and researchers. The use of rankings to assess universities also needs a rethink. These league tables, produced by the Academic Ranking of World Universities (ARWU) and the Times Higher Education World University Ranking (THE WUR) and others, determine eligibility for scholarships and other income, and sway where scholars decide to work and study. Governments devise policies and divert funds to help institutions in their countries claw up these rankings. Researchers at many institutions, such as mine, miss out on opportunities owing to their placing. Two years ago, the International Network of Research Management Societies (INORMS), a collective of research-management organizations, invited me to chair a new working group on research evaluation with members from a dozen countries. From our first meeting, we were unanimous about our top concern: the need for fairer and more responsible university rankings. When we drew up criteria on what those would entail and rated the rankers, their shortcomings became clear. This week, the Global Research Council, which includes heads of science- and engineering-funding agencies, is gathering experts online to discuss how assessments can improve research culture. This should include how university rankings are constructed and used. The literature on research management is full of critiques of rankings. Rankings are methodologically challenged — often using inappropriate indicators such as counting Nobel-prizewinning alumni as a proxy for offering a quality education. They favour publications in English, and institutions that did well in past rankings. So, older, wealthier organizations in Europe and North America consistently top the charts. Rankings apply a combination of indicators that might not represent universities’ particular missions, and often overlook societal impact or teaching quality. Nonetheless, they have become entrenched, with new rankers cropping up each year. As with the journal impact factor, students, faculty members and funders turn to rankings as a lazy proxy for quality, no matter the flaws. The consequences are all too real: talent deterred, income affected. And inequities quickly become embedded. Our working group combed the literature to develop our

The rankers with the largest audiences were found most wanting.”

Elizabeth Gadd is a researchpolicy manager at Loughborough University, UK, and chair of the Research Evaluation Working Group for the International Network of Research Management Societies. e-mail: e.a.gadd@ lboro.ac.uk

criteria, and asked for feedback through various community discussion lists open to academics, research-support professionals and related groups. We synthesized feedback into 20 principles involving good governance (such as the declaration of financial conflicts of interest), transparency (of aims, methods and data), measuring what matters (in line with a university’s mission) and rigour (the indicators are a good proxy for what they claim to measure). Then we converted these principles into a tool to assess rankings, qualitatively and quantitatively (see go.nature. com/2ioxhhoq). We recruited international specialists to assess six of the world’s highest-profile rankers, and invited rankers to self-assess. (Only one, CWTS Leiden, did so.) Richard Holmes, editor of the University Ranking Watch blog, calibrated the results, which we presented as profiles, not rankings. The rankings with the largest audiences (ARWU, QS World University Ranking, THE WUR and US News & World Report global ranking) were found most wanting, particularly in terms of ‘measuring what matters’ and ‘rigour’. None of these ‘flagship’ rankings considered open access, equality, diversity, sustainability or other society-focused agendas. None allows users to weigh indicators to reflect a university’s mission. Yet all claim to identify the world’s best universities. Rankers might argue that our principles were unrealistic — that it’s impossible to be completely fair in such evaluations, and that simple, overarching metrics have their place. I counter that we derived the principles from community best-practice expectations, and if rankers cannot meet them, perhaps they should stop ranking, or at least be honest about the inherent uncertainty in their conclusions (in our assessment, only CWTS Leiden attempted this). Ultimately, rankers need to be made more accountable. I take heart from new expectations about how researchers are evaluated. From January 2021, UK research funder Wellcome will fund only organizations that present evidence that they conduct fair output assessments for researchers. Similarly, the European Commission’s ‘Towards 2030’ vision statement calls for higher education to move beyond current ranking systems for assessing university performance because they are limited and “overly simplistic”. We hope that drawing attention to their weaknesses will draw in allies to push for change, such as neutral, independent oversight and standards for ethics and rigour as applied to other aspects of academia. Such pressure could lead to greater alignment between the world rankers’ approaches and the higher-education community’s expectations for fair and responsible rankings. It might also help users to wise up to rankings’ limitations, and to exercise due caution when using them for decision-making. Either would be progress.

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

Nature | Vol 587 | 26 November 2020 | 523

The world this week

News in brief AI SUMS UP TL;DR RESEARCH IN A SENTENCE

L TO R: GOPAL MURTI/SPL; R.P. DOHERTY ET AL./SOFT MATTER

IMMUNE RESPONSES TO CORONAVIRUS LAST SIX MONTHS The immune system’s memory of the new coronavirus lingers for at least six months in most people. Sporadic accounts of coronavirus reinfection and reports of rapidly declining antibody levels have raised concerns that immunity to SARS-CoV-2 could dwindle within weeks of recovery from infection. Shane Crotty at the La Jolla Institute for Immunology in California and his colleagues analysed markers of the immune response in blood samples from 185 people who had a range of COVID-19 symptoms; 41 study participants were followed for at least 6 months ( J. M. Dan et al. Preprint at bioRxiv https://doi. org/ghkc5k; 2020). The team found that participants’ immune responses varied widely. But several components of immune memory for SARS-CoV-2 tended to persist for at least 6 months. Among the persistent immune defenders were memory B cells (pictured), which jump-start antibody production when a pathogen is re-encountered, and two important classes of T cell: memory CD4+ and memory CD8+ T cells. The results have not yet been peer reviewed.

The creators of a scientific search engine have unveiled software that automatically generates one-sentence summaries of research papers, which they say could help scientists to skim-read papers faster. The free tool, called TLDR (the common Internet acronym for ‘Too long; didn’t read’), was activated this week for search results at Semantic Scholar, a search engine created by the non-profit Allen Institute for Artificial Intelligence (AI2) in Seattle, Washington. For the moment, TLDR generates sentences only for the ten million computer-science papers covered by Semantic Scholar, but the researchers say that papers from other disciplines should be getting summaries in the next month or so, once the software has been fine-tuned. A preprint describing the tool was first published on the arXiv preprint server in April (I. Cachola et al. Preprint at https://arxiv.org/ abs/2004.15011; 2020), and was accepted for publication after peer review by a natural-language-processing conference taking place this month. The authors have made their code freely available, along with a working demo website where anyone can try the tool (see go.nature.com/3psfs3t).

3D-printed microboat This boat-shaped particle measures just 30 micrometres in length, but is fully equipped with a cabin, chimney and flag post, and is able to propel itself through a solution of 10% hydrogen peroxide. It was 3D printed using a technique called two-photon polymerization, and was then coated with a mixture of platinum and palladium, which catalyses the breakdown of the hydrogen peroxide. This reaction produces bubbles of gas that propel the particle along. Daniela Kraft’s team at Leiden University in the Netherlands made many swimming shapes using the same method — including spheres, spirals, triangles and even a miniature starship (R. P. Doherty et al. Soft Matter https://doi.org/fjrf; 2020). They hope that this work will help them to study the effect of shape in microorganisms that swim, such as bacteria.

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

Nature | Vol 587 | 26 November 2020 | 527

The world this week

ARECIBO OBSERVATORY/UNIV. CENTRAL FLORIDA

News in focus

A main cable broke in half on 6 November and tore large gashes through a central portion of the telescope’s dish.

LEGENDARY ARECIBO TELESCOPE WILL CLOSE FOREVER — SCIENTISTS ARE REELING Cable breaks caused damage too extensive to repair, ending an era in astronomical observation. By Alexandra Witze

O

ne of astronomy’s most renowned tele­scopes — the 305-metre-wide radio telescope at Arecibo, Puerto Rico — is closing permanently. Engineers cannot find a safe way to repair it, after two cables supporting the structure broke suddenly and catastrophically, one in August and one in early November. It is the end of one of the most iconic and scientifically productive telescopes in the history of astronomy — and scientists are

mourning its loss. “I don’t know what to say,” says Robert Kerr, a former director of the observatory. “It’s just unbelievable.” “I am totally devastated,” says Abel Méndez, an astrobiologist at the University of Puerto Rico at Arecibo who uses the observatory. The Arecibo telescope, which was built in 1963, was the world’s largest radio telescope for decades and has historical and modern importance in astronomy. It was the site from which astronomers sent an inter­stellar radio message in 1974, in the hope that any

extraterrestrials might hear it, and from which the first confirmed extrasolar planet was discovered, in 1992. It has also done pioneering work in exploring many phenomena, including near-Earth asteroids and the puzzling celestial blasts known as fast radio bursts. All those lines of investigation have now been shut down for good, although limited science will continue at some smaller facilities on the Arecibo site. The cables that broke helped to support a 900-tonne platform of scientific instruments, which hangs above the main telescope dish.

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

Nature | Vol 587 | 26 November 2020 | 529

The first cable slipped out of its socket and smashed panels at the edge of the dish, but the second broke in half and tore huge gashes in a central portion of the dish. If any more cables fail — which could happen at any time — the entire platform could crash into the dish below. The US National Science Foundation (NSF), which owns the Arecibo Observatory, is working on plans to lower the platform in a safe, controlled fashion. But those plans will take weeks to develop. “Even attempts at stabilization or at testing the cables could result in accelerating the catastrophic failure,” said Ralph Gaume, director of the NSF’s astronomy division, at a 19 November media briefing. So the NSF decided to close the Arecibo dish permanently. “This decision is not an easy one to make, but safety is the number-one priority,” said Sean Jones, head of the NSF’s mathematical- and physical-sciences directorate. The closure comes as a shock to the wider astronomical community. A social-media campaign with the hashtag #WhatAreciboMeans­ ToMe sprung up almost immediately, with astronomers, engineers and other scientists — many from Puerto Rico — sharing stories of how the observatory had shaped their careers. “Losing the Arecibo Observatory would be a big loss for science, for planetary defence and for Puerto Rico,” said Desireé Cotto-Figueroa, an astronomer at the University of Puerto Rico Humacao, in an e-mail before the closure was announced.

What went wrong NSF officials insist that the cable failures came as a surprise. After the first, engineering teams spotted a handful of broken wires on the second cable, which was more crucial to holding up the platform, but they did not see it as a major problem because the weight it was carrying was well within its design capacity. “It was not seen as an immediate threat,” says Ashley Zauderer, programme director for Arecibo at the NSF. But that main cable, which was installed in the early 1960s, had apparently degraded over time. Over the years, external review committees have highlighted the ongoing need to maintain the ageing cables. Zauderer said that maintenance in recent years had been completed according to schedule. Before this year, the last major cable problems at the observatory were in January 2014, when a magnitude-6.4 earthquake caused damage to another of the main cables, which engineers repaired. The ageing structure has sustained other shocks in recent years, including damage to an antenna and the dish caused by Hurricane Maria in 2017. There is no estimate yet for the cost of decommissioning the telescope. The science that has ground to a halt includes Arecibo’s world-leading asteroid

studies. The telescope pinged radio waves at near-Earth asteroids to reveal the shape and spin of these threatening space rocks. Not having it “will be a big loss”, says Alan Harris, an asteroid scientist in La Canada, California. (China’s Five-hundred-meter Aperture Spherical Tele­scope (FAST), which opened in 2016, does not currently have the ability to do such radar studies.) Some of the observatory’s scientific projects could be transferred to other facilities, Gaume said — and he expects scientists to propose where to move their research. Much of the work conducted at Arecibo, however, could be done only with its unique array of astronomical instrumentation. “The Arecibo Telescope is irreplaceable,” said a statement from two major US radio-astronomy organizations,

CAN DOGS SMELL COVID? HERE’S WHAT THE SCIENCE SAYS Canines seem to detect infections accurately, but researchers say large-scale studies are needed. By Holly Else

A

sher is an eccentric, Storm likes sunbathing and Maple loves to use her brain. All three could play a part in controlling the COVID-19 pandemic, but they are not scientists or politicians. They are dogs. And they are not alone. Around the world, canines are being trained to detect the whiff of COVID-19 infections. Dog trainers are claiming extraordinary results — in some cases,

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

530 | Nature | Vol 587 | 26 November 2020

Broken wires on the second failed cable.

the National Radio Astronomy Observatory in Charlottesville, Virginia, and the Green Bank Observatory in West Virginia. Small amounts of science will continue at other portions of the Arecibo observatory, which encompasses more than the 305-metre dish. For instance, two lidar facilities shoot lasers into the skies to study atmospheric phenomena. The Arecibo telescope had been upgraded regularly, with several new instruments slated for installation in the coming years. “The telescope is in no way obsolete,” says Christopher Salter, an astronomer at the Green Bank Observatory, who worked at Arecibo for years. Planned upgrades are now presumably on hold, including a US$5.8-million antenna that was being developed for the telescope’s platform and would have massively increased its sensitivity. Brian Jeffs, an engineer at Brigham Young University in Provo, Utah, who heads the project, says his team expects to discuss options for its future with the NSF eventually. “Our greatest concerns are for the wonderful scientific, technical, management and support staff” of the observatory, he says. The observatory is a major centre for science education in Puerto Rico, where it has fostered the careers of many astronomers and engineers. And it has become a part of the pop-culture lexicon, featuring in major movies such as Contact (1997), which was based on a novel by astronomer Carl Sagan, and the 1995 James Bond film GoldenEye. The most recent major radio-telescope disaster happened in 1988, when a 300-foot-wide antenna at the Green Bank Observatory collapsed one night, owing to structural failure.

they say that dogs can detect the virus with almost perfect accuracy. Scientists involved suggest that canines could help to control the pandemic because they can screen hundreds of people an hour in busy places such as airports or sports stadiums, and are cheaper to use than conventional testing methods such as the RNA-amplification technique RT-PCR. But most of these findings have not yet been peer reviewed or published, making it hard for the wider scientific community to evaluate the claims. Researchers working on more

ARECIBO OBSERVATORY/UNIV. CENTRAL FLORIDA

News in focus

FATEMEH BAHRAMI/ANADOLU AGENCY/GETTY

conventional viral tests say that initial results from dog groups are intriguing and show promise. But some question whether the process can be scaled up to a level that would allow the animals to make a meaningful impact. On 3 November, groups working with the animals took part in an online workshop called International K9 Team to share preliminary results from experiments and to improve how their research is coordinated. “No one is saying they can replace a PCR machine, but they could be very promising,” says veterinary neurologist Holger Volk at the University of Veterinary Medicine Hanover in Germany, who is leading an effort to train and study COVID-sniffing dogs and did not speak at the event.

Sense of wonder Humans have taken advantage of canines’ superior sense of smell for decades. Dogs’ noses bear 300 million scent receptors, compared with humans’ 5 million or 6 million. That enables them to detect tiny concentrations of odour that people can’t. Sniffer dogs are already a familiar sight in airports, where they detect firearms, explosives and drugs. Scientists have also trained dogs to detect some cancers and malaria, but the animals are not routinely used for this purpose. Researchers don’t know for sure what the dogs are smelling, but many suspect that these illnesses cause the human body to let off a distinct pattern of volatile organic compounds (VOCs). These molecules readily evaporate to create scent that dogs can pick up. Previous work with non-COVID viruses has suggested that viral infections might also cause the body to do this. Many sniffer-dog scientists turned their attention to COVID-19 early in the pandemic. They have trained their canines to smell samples, most often of sweat, in sterile containers, and to sit or paw the floor when they detect signs of infection. Trials at airports in the United Arab Emirates, Finland and Lebanon are using dogs to detect COVID-19 in sweat samples from passengers; these are then checked against conventional tests. According to data presented at the K9 meeting, dogs in Finland and Lebanon have identified cases days before conventional tests, suggesting that they can spot infection before symptoms start. Riad Sarkis, a surgeon and researcher at Saint Joseph University in Beirut, is part of a French–Lebanese project that has trained 18 dogs. Sarkis used the best two performers for the airport trial in Lebanon. The dogs screened 1,680 passengers and found 158 COVID-19 cases that were confirmed by PCR tests. The animals correctly identified negative results with 100% accuracy, and correctly detected 92% of positive cases, according to unpublished results. “This is very accurate, feasible, cheap and reproducible,” says Sarkis, who has been approached about

Research groups around the world are testing whether dogs can detect COVID-19 by smell.

using the dogs in schools, banks and prisons. Low-income countries with limited lab space could particularly benefit from the approach, says Isabella Eckerle, a virologist at the University Hospitals of Geneva in Switzerland.

Sample sizes But there is just one published journal article on dogs’ efficacy at sniffing out COVID-19, by Volk’s group1. The researchers trained eight dogs on samples taken from the mouths and windpipes of seven people hospitalized with COVID-19 and seven uninfected people. The dogs identified 83% of positive cases and 96% of negative ones.

“It’s important not to go out too early with grand claims and small data sets.” The false positive and negative rates of the standard PCR lab test vary depending on the brand of test used and the timing of the test. A systematic review published as a preprint2 on medRxiv found the false-negative rate of RT-PCR tests to be 2–33% if the same sample is tested repeatedly. Up to 4% of UK PCR test results could be false positives, according to government documents. Critics say the German dog study used samples from too few patients. The dogs could be learning to identify the specific scent of the samples rather than of COVID-19, says Cynthia Otto, who leads the Penn Vet Working Dog Centre at the University of Pennsylvania in Philadelphia and is also working with COVID‑19 sniffer dogs. In her work, which is unpublished, she has found that the dogs can tell the difference

between samples of either urine or sweat from people with COVID-19 and those from people without the disease. She is working with chemists to understand which VOCs the dogs are picking up; a paper describing this is under review. “The dogs can do it. The challenge is the ignorance that we have as humans as to what can confuse the dogs,” she says. A group led by veterinary scientist Dominique Grandjean at the National Veterinary School of Alfort near Paris, posted its work3 on the preprint server bioRxiv in June. The researchers, who included Sarkis, trained 8 dogs to detect COVID-19 in 198 sweat samples, around half of which were from people with the disease. When these were hidden in a row of negative samples, the dogs identified the positive samples 83–100% of the time. The paper does not say how well the dogs identified negative test results. Those data look promising, says Fyodor Urnov, a gene-editing scientist who is working on COVID testing at the University of California, Berkeley. But he would like to see larger data sets on how well dogs identify positive and negative samples. He also notes that there is variation in how well individual dogs perform. Groups need to boost their sample size, agrees James Logan, an infectious-disease researcher at the London School of Hygiene & Tropical Medicine who is training and studying COVID-19 dogs, including Storm, Maple and Asher. “It’s important not to go out too early with grand claims and small data sets,” he says. 1. Jendrny, P. et al. BMC Infect. Dis. 20, 536 (2020). 2. Arevalo-Rodriguez, I. et al. Preprint at medRxiv https://doi. org/10.1101/2020.04.16.20066787 (2020). 3. Grandjean, D. et al. Preprint at bioRxiv https://doi. org/10.1101/2020.06.03.132134 (2020).

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

Nature | Vol 587 | 26 November 2020 | 531

News in focus

Reanalysis suggests a fainter signal of phosphine gas in the planet’s atmosphere than originally reported. By Alexandra Witze

S

igns of a potential marker of life in Venus’s atmosphere have faded — but they’re still there, according to a new data analysis. In September, an international team of astronomers made headlines when it reported finding the gas phosphine — a potential marker of life — in the planet’s atmosphere1. Several studies questioning the observations and conclusions quickly followed. Now, the same team has reanalysed part of its data, citing a processing error in the original data set. The researchers confirmed the phosphine signal, but say that it’s fainter than before. The work is an important step forward in resolving the most exciting Venus debate in decades. “I’ve waited all my life for this,” says Sanjay Limaye, a planetary scientist at the University of Wisconsin–Madison, who says the debate has reinvigorated the field. The reanalysis 2, based on radio-telescope observations at the Atacama Large Milli­meter/submillimeter Array (ALMA) in Chile, concludes that average phosphine levels across Venus are about one part per

billion — approximately one-seventh of the earlier estimate. Unlike in their original report, the scientists now describe their discovery of phosphine on Venus as tentative. In its September report, the team used data from ALMA and the James Clerk Maxwell Telescope ( JCMT) in Hawaii to make its discovery. Team leader Jane Greaves, an astronomer at Cardiff University, UK, says she and her colleagues redid the work because they had learnt that the original ALMA data contained a spurious signal that could have affected the results. ALMA posted the corrected data on 16 November, and Greaves and her team ran a fresh analysis that night and posted it ahead of peer review on the preprint server arxiv.org. “We’ve been working like crazy,” she told a meeting of the Venus Exploration Analysis Group, a NASA community forum, on 17 November. According to Greaves and her colleagues, the ALMA data show the spectral signature of phosphine, a molecule made of one phosphorus and three hydrogen atoms. They say no other compound can explain the data. Finding phosphine on Venus would be tantalizing because microbes produce the gas on Earth. If

Venus has a thick, acidic atmosphere, once thought unsuitable for life.

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

532 | Nature | Vol 587 | 26 November 2020

Solving the mystery One other new strand of evidence supports phosphine on Venus. Inspired by Greaves’s original report, a team led by Rakesh Mogul, a biochemist at California State Polytechnic University in Pomona, dug through decades-old data from NASA’s 1978 Pioneer Venus mission. This spacecraft dropped a probe that measured the chemistry of clouds in the planet’s atmosphere as it fell. It detected a phosphorus signal that could be attributed to phosphine or another phosphorus compound6. But “we believe the simplest gas that fits the data is phosphine”, Mogul said at the meeting on 17 November. Where the phosphine comes from remains a mystery. The only spacecraft currently orbiting Venus, Japan’s Akatsuki, does not carry instruments that could help settle the debate. The Indian Space Research Organisation is planning a Venus mission that would launch in 2025 and could potentially carry instruments capable of looking for phosphine. In the meantime, Greaves and other researchers are applying for more time on Earth-based telescopes, including ALMA. Researchers are investigating many other aspects of Venus, says David Grinspoon, an astrobiologist at the Planetary Science Institute who is based in Washington DC. “There are 1,001 reasons to go back to Venus, and if the phosphine ‘goes away’ through further observations and analysis, there will still be 1,000 reasons to go.” 1. Greaves, J. S. et al. Nature Astron. https://doi.org/10.1038/ s41550-020-1174-4 (2020). 2. Greaves, J. S. et al. Preprint at https://arxiv.org/ abs/2011.08176 (2020). 3. Bains, W. et al. Preprint at https://arxiv.org/ abs/2009.06499 (2020). 4. Seager, S. et al. Astrobiology https://doi.org/10.1089/ ast.2020.2244 (2020). 5. Villanueva, G. et al. Preprint at https://arxiv.org/ abs/2010.14305 (2020). 6. Mogul, R., Limaye, S. S., Way, M. J. & Cordova, J. A. Jr Preprint at https://arxiv.org/abs/2009.12758 (2020).

DETLEV VAN RAVENSWAAY/SPL

PROSPECTS FOR LIFE ON VENUS FADE — BUT AREN’T DEAD YET

the signal is real and indeed due to phosphine, it’s possible that microbes living in and drifting among the planet’s clouds could be producing the gas3,4 — but it’s also possible there might be a non-living source that scientists have yet to identify. Before they can determine whether either of these scenarios is true, researchers first need to confirm phosphine’s presence. In one critique of the original study, researchers suggested 5 that the signal reported as phosphine might really be coming from sulfur dioxide, a gas that is common in Venus’s clouds but is not produced by life there. Greaves and her team fired back in their latest report that that can’t be the case, because of how the phosphine fingerprint appears in data collected by the second tele­ scope they used, the JCMT. Other critiques have focused on the difficulty of extracting a phosphine signal out of complicated data.

under lockdown, with an average twice as high as the original prediction, and closer to the actual figures. “CovidSim may be vaunted as the most complicated epidemiological model, but it’s almost like a toy compared with the really high-end supercomputing applications,” says Coveney, who was asked to check the model’s performance as part of the Royal Society’s Rapid Assistance in Modelling the Pandemic (RAMP) initiative.

HOLLIE ADAMS/AFP/GETTY

Ensembles of calculations

A second national lockdown began in England on 5 November.

WHAT COVID PANDEMIC FORECASTERS CAN LEARN FROM CLIMATE MODELS Analysis of influential pandemic model suggests how to have better predicted lockdown deaths. By David Adam

E

pidemiologists predicting the spread of COVID-19 should adopt climate-modelling methods to make forecasts more reliable, say computer scientists who have spent months auditing one of the most influential models of the pandemic. In a study that was uploaded to the preprint platform Research Square on 6 November, researchers commissioned by London’s Royal Society used a powerful supercomputer to re-examine CovidSim, a model developed by a group at Imperial College London (W. Edeling et al. Preprint at Research Square https://doi. org/fjq5; 2020). In March, that simulation helped convince British and US politicians to introduce lockdowns, but it has since been scrutinized by researchers who doubt the reliability of its results. The analysis, which has not yet been peer-reviewed, shows that because researchers didn’t appreciate how sensitive CovidSim was to small changes in its inputs, their results overestimated the extent to which a lockdown was likely to reduce deaths, says Peter Coveney, a chemist and computer scientist at

University College London, who led the study. Coveney is reluctant to criticize the Imperial group, led by epidemiologist Neil Ferguson, which he says did the best job possible under the circumstances. And the model correctly showed that “doing nothing at all would have disastrous consequences”, he says. But he argues that epidemiologists should stress‑test

“CovidSim is almost like a toy compared with the really high-end supercomputing applications.” their simulations by running ‘ensemble’ models, in which thousands of versions of the model are run with a range of assumptions and inputs, to provide a spread of scenarios with different probabilities. These ‘probabilistic’ methods are routine in computation-heavy fields, from weather forecasting to molecular dynamics. Coveney’s team has now done this for CovidSim: the findings suggest that if the model had been run as an ensemble, it would have forecast a range of probable death tolls

Coveney’s team used the Eagle supercomputer at the Poznan Supercomputing and Networking Center in Poland to perform 6,000 separate runs of CovidSim, each with a unique set of input parameters. These represent features of the pandemic including the infectiousness and lethality of the virus, the probable number of contacts people make in various settings and the estimated success of measures such as telling people to work from home. Back in March, inputs for many of these parameters were educated guesses, with some drawn from preliminary data on the virus, and others based on experience with diseases such as influenza. Models that predict the spread of disease often rely on hundreds of parameters — but this can introduce uncertainty. “There was a concern among the circles who set up the RAMP initiative that these models the epidemiologists work with have an absurd number of parameters in them,” Coveney says. His team found 940 parameters in the CovidSim code, but whittled these down to the 19 that most affected the output. And up to twothirds of the differences in the model’s results could be put down to changes in just three key variables: the length of the latent period during which an infected person has no symptoms and can’t pass the virus on; the effectiveness of social distancing; and how long after getting infected a person goes into isolation. The study suggests that small variations in these parameters could have an outsize, non-linear impact on the model’s output. For example, the majority of the team’s thousands of runs suggested that the UK death toll under lockdown would be much higher than the Imperial team’s initial projections — 5–6 times higher, in some cases. Averaging the figures still suggested twice as many deaths as the Imperial group had forecast. In one modelled scenario, which assumed that the United Kingdom would lock down when 60 people per week needed to be admitted to hospital, the March report forecast a total of 8,700 deaths in the country. The probabilistic results produced by Coveney’s group put this figure at around 15,000 on average, but said that death tolls of more than 40,000 were possible, depending on what parameters were used. It is hard to compare these projections with the actual figures

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

Nature | Vol 587 | 26 November 2020 | 533

News in focus for COVID-19 deaths in the United Kingdom, because the lockdown started a week later than the results of any of the models assume, by which time higher amounts of the disease were already circulating. “They didn’t get it right,” says Coveney. “They ran the simulation correctly: it’s just that they didn’t know how to extract the correct probabilistic description from it.” Coveney said he couldn’t comment on whether running an ensemble model would have altered policy, but Rowland Kao, an epidemiologist and data scientist at the University of Edinburgh, UK, points out that the government compares and synthesizes the results of several different COVID-19 models. “It would be overly simplified to consider that decision-making is based on a single model,” he says.

Improved models Ferguson accepts most of Coveney’s points about the benefits of performing probabilistic forecasts, but says that “we just weren’t in a position to do that in March”. The Imperial group has significantly improved its models since then, he adds. For example, it now presents the uncertainty in CovidSim inputs using Bayesian statistical tools — already common in some models of illnesses such as the livestock disease foot-and-mouth. And a simpler model, he adds, was used to inform the UK government’s decision to reintroduce lockdown measures in England this month. This model is more agile than CovidSim: “Because we can run it several times a week, it’s much easier to fit the data in real time, allowing for uncertainty,” Ferguson says. “This sounds like a step in the right direction, and is aligned with the conclusions of our paper,” says Coveney. The choice of technique often comes down to a computational trade-off, Ferguson says. “If you want to routinely properly characterize all the uncertainty, then that is much easier with a less computationally intensive model.” Bayesian tools are an improvement, says Tim Palmer, a climate physicist at the University of Oxford, UK, who pioneered the use of ensemble modelling in weather forecasting. But only ensemble modelling techniques that are run on the most powerful computers will deliver the most reliable pandemic projections, he says. Such techniques transformed the reliability of climate models, he adds, helped by the coordination of the Intergovernmental Panel on Climate Change (IPCC). “We need something like the IPCC for these pandemic models. We need some kind of international facilities where these models can be developed properly,” Palmer says. “It has been rushed because of the urgency of the situation. But to take this forward, we need some kind of international organization that can work on synthesizing epidemiological models from around the world.”

WHAT THE DATA SAY ABOUT ASYMPTOMATIC COVID INFECTIONS People without symptoms can transmit the virus, but estimating their contribution to outbreaks is tricky. By Bianca Nogrady

M

any people don’t experience any symptoms after becoming infected with SARS-CoV-2. But how many, and what is their role in spreading COVID-19? These have been key questions since the beginning of the pandemic. Now, evidence suggests that about one in five infected people will experience no symptoms, and they will transmit the virus to significantly fewer people than someone with symptoms. But researchers are divided about whether asymptomatic infections are acting as a silent driver of the pandemic. Although there is a growing understanding of asymptomatic infections, researchers say that people should continue to use measures to reduce viral spread, including social distancing and wearing masks, regardless of whether they have symptoms. The issue with estimating the rate of asymptomatic COVID-19 is distinguishing between people who are asymptomatic and pre-symptomatic, says Krutika Kuppalli, an infectious-disease researcher at the Medical University of South Carolina in Charleston.

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

534 | Nature | Vol 587 | 26 November 2020

Roughly one in five people infected with SARS-CoV-2 don’t experience symptoms.

“Asymptomatic is someone who never developed symptoms ever throughout the course of their disease, and pre-symptomatic is somebody who has mild symptoms before they do go on to develop symptoms,” Kuppalli says. Research early in the pandemic suggested that the rate of asymptomatic infections could be as high as 81%. But a meta-analysis

“These people are not the secret drivers of this pandemic.” published last month1, which included 13 studies involving 21,708 people, calculated the rate of asymptomatic presentation to be 17%. The analysis defined asymptomatic people as those who showed none of the key COVID-19 symptoms during the entire follow-up period, and the authors included only studies that followed participants for at least seven days. Evidence suggests that most people develop symptoms in 7–13 days, says lead author Oyungerel Byambasuren, a biomedical researcher at the Institute for Evidence-Based

EZRA ACAYAN/GETTY

Healthcare at Bond University in Gold Coast, Australia. The review also found that asymptomatic individuals were 42% less likely to transmit the virus than symptomatic people. One reason that scientists want to know how frequently people without symptoms transmit the virus is because these infections largely go undetected. Testing in most countries is targeted at those with symptoms. As part of a large population study in Geneva, Switzerland, researchers modelled viral spread among people living together. In a manuscript posted on medRxiv this month2, they report that the risk of an asymptomatic person passing the virus to others in their home is about one-quarter of the risk of transmission from a symptomatic person. Although transmission risk from asymptomatic people is lower, they might still present a public-health risk because they are more likely to be out in the community than isolated at home, says Andrew Azman, an infectious-disease epidemiologist at the Johns Hopkins Bloomberg School of Public Health in Baltimore, Maryland, who is based in Switzerland and was a co-author on the study. “This massive pool of interacting ‘asymptomatics’ in the community probably suggests that a sizeable portion of transmission events are from asymptomatic transmissions,” he says. But other researchers disagree about the extent to which asymptomatic infections are contributing to community transmission. If the studies are correct in finding that asymptomatic people are a low transmission risk, “these people are not the secret drivers of this pandemic”, says Byambasuren. They “are not coughing or sneezing as much”. Muge Cevik, an infectious-disease researcher at the University of St Andrews, UK, points out that because most people are symptomatic, concentrating on identifying them will probably eliminate most transmission events. To understand what is happening in people with no symptoms, Cevik and colleagues conducted a systematic review and meta-analysis3 of 79 studies on the viral dynamics and transmissibility of SARS-CoV-2. Some studies showed that those without symptoms had similar initial levels of viral particles in a throat swab when compared with people with symptoms. But asymptomatic people seem to clear the virus faster and are infectious for a shorter period. The immune systems of asymptomatic individuals might be able to neutralize the virus more rapidly, says Cevik. 1. Byambasuren, O. et al. J. Assoc. Med. Microbiol. Infect. Dis. Can. https://doi.org/10.3138/jammi-2020-0030 (2020). 2. Bi, Q. et al. Preprint at medRxiv https://doi. org/10.1101/2020.11.04.20225573 (2020). 3. Cevik, M. et al. Lancet Microbe https://doi.org/10.1016/ S2666-5247(20)30172-5 (2020).

FUNDING FOR DISPUTED STEM-CELL INSTITUTE SPARKS DEBATE California agency will receive billions from the state — but some scientists oppose the plan. By Nidhi Subbaraman

V

oters in California have approved US$5.5 billion in funding for stem-cell and other medical research, granting a lifeline to a controversial state agency. But scientists are split over whether the California Institute for Regenerative Medicine (CIRM) in Oakland is a worthwhile investment for the US state — or for the field of stem-cell research. A measure to authorize new funds for CIRM, called Proposition 14, appeared on California ballots in the recent US election. After more than a week of vote counting, on 12 November the Associated Press announced that California had passed the proposal. Critics of CIRM are concerned about oversight at the state agency, which has faced complaints about potential conflicts of interest among its board members for years. They also point out that the field has grown and now receives federal support, making state funding hard to justify — especially amid a pandemic that has imperilled California’s economy. “Unfortunately, Proposition 14 sets a bad example for the use of public money for the advancement of science,” says Zach Hall, a ­neurobiologist who led CIRM as its first president between 2005 and 2007. Launched 16 years ago, CIRM drew top researchers to the state, and put California on the map as a hub for regenerative-medicine research. With CIRM’s original $3 billion in state money running out last year, California property developer Robert Klein — an advocate of stem-cell research after his son was diagnosed with type 1 diabetes, and the agency’s original backer — began canvassing support for new funding. His efforts landed Proposition 14 on this year’s ballot. “It is extraordinary that the patient-advocacy groups and the medical societies and the scientific societies have been able to act as a single coalition to reach millions of California voters,” says Klein, who co-wrote the 2004 ballot measure creating the agency. Some scientists are proponents of the agency. “It is very exciting that Prop. 14 passed and that CIRM will continue its funding,” says Cato Laurencin, a biomedical engineer at the University of Connecticut in Farmington, who

is not funded by the institute. “This field is at a bit of an inflection point in terms of our understanding of stem-cell science.” CIRM emerged in 2004, when stem-cell research was a nascent field. Stem cells’ ability to renew themselves offered the promise of treatments for challenging conditions such as heart disease and stroke, in which cells are irreversibly damaged. Much work at the time relied on stem cells obtained from human embryos donated by fertility clinics. Citing ethical concerns about the destruction of fertilized embryos, in 2001, US president George W. Bush severely restricted research in this area, and the science hit a wall. Three years later, CIRM’s launch was a boon. “It gave a tremendous boost to the field at a time when things looked very bad,” says Hall. CIRM has since handed out (as of June 2020) $2.7 billion in grants to California scientists studying a variety of diseases, including diabetes, AIDS and leukaemia. It has built a dozen research facilities, funded more than 60 clinical trials and, according to an independent, agency-funded report, helped create more than 56,500 jobs in the state.

A worthwhile investment? But the agency has also drawn criticism for poor management of its public funds. A 2012 Institute of Medicine report pointed out that CIRM’s policy of allowing board members to vote on grants or issues benefiting their institutions posed a potential conflict of interest. CIRM attempted to address some of the criticism in 2013, when it asked board members from agency-funded universities to abstain from voting on grants, among other changes. Hall says that Proposition 14 doesn’t describe a clear scientific vision. “You could argue that California would do better, economically and scientifically, to have a CRISPR institute,” he says, arguing that the revolutionary precision gene-editing tool is better placed to benefit from such a huge infusion of cash. Responding to the criticisms, Klein says he crafted the proposal with the guidance of multiple groups of experts, and kept the mandate deliberately broad to allow for flexibility as the field grows. “There’s an intent here,” he says, “to have the agency be responsive to the development of science.”

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

Nature | Vol 587 | 26 November 2020 | 535

HOW ICELAND HAMMERED COVID WITH SCIENCE

Feature

The tiny island nation brought huge scientific heft to its attempts to contain and study the coronavirus. Here’s what it learnt. By Megan Scudellari

D

riving along Reykjavik’s windswept roads on a cold March morning, Kári Stefánsson turned up the radio. The World Health Organization had just announced that an estimated 3.4% of people infected with SARSCoV-2 would die — a shockingly high fatality rate, some 30 times larger than that for seasonal influenza. There was a problem with that estimate, however: it was based on reported cases of COVID-19, rather than all cases, including mild and asymptomatic infections. “I couldn’t figure out how they could calculate it out without knowing the spread of the virus,” recalls Stefánsson, who is the founder and chief executive of deCODE genetics, a human-genomics company in Reykjavik. He became convinced that making sense of the epidemic, and protecting the people of Iceland from it, would require a sweeping scientific response. When Stefánsson arrived at work, he phoned the leadership of Amgen, the US pharmaceutical company that owns deCODE, and asked whether he could offer deCODE’s resources to track the spread of the virus, which had landed

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

536 | Nature | Vol 587 | 26 November 2020

on Icelandic shores only six days earlier. “The response I got from them was, ‘For heaven’s sake, do that,’” says Stefánsson. Over the ensuing nine months, deCODE and Iceland’s Directorate of Health, the government agency that oversees health-care services, worked hand-in-hand, sharing ideas, data, laboratory space and staff. The high-powered partnership, coupled with Iceland’s diminutive size, has put the country in the enviable position of knowing practically every move the virus has made. The teams have tracked the health of every person who has tested positive for SARS-CoV-2, sequenced the genetic material of each viral isolate and screened more than half of the island’s 368,000 residents for infection. Late nights analysing the resulting data trove led to some of the earliest insights about how the coronavirus spreads through a population. The data showed, for example, that almost half of infected people are asymptomatic, that children are much less likely to become sick than adults and that the most common symptoms of mild COVID-19 are muscle aches, headaches and a cough — not

fever. “Scientific activities have been a huge part of the entire process,” says Runolfur Palsson, director of internal-medicine services at Landspitali — The National University Hospital of Iceland. Researchers at deCODE and the hospital worked day in and day out to gather and interpret the data. Their achievements aren’t merely academic. Iceland’s science has been credited with preventing deaths — the country reports fewer than 7 per 100,000 people, compared with around 80 per 100,000 in the United States and the United Kingdom. It has also managed to prevent outbreaks while keeping its borders open, welcoming tourists from 45 countries since mid-June. The partnership again kicked into high gear in September, when a second large wave of infections threatened the nation.

Careful steps COVID-19 is not the first pandemic to reach Iceland’s shores: in October 1918, two ships carrying pandemic influenza docked in Reykjavik’s downtown harbour. Within six weeks, two-thirds of the capital city’s inhabitants were infected1.

JON E. GUSTAFSSON

Iceland’s science has been pivotal in understanding the COVID-19 pandemic.

A century later, the Icelandic government was better prepared, enacting a national pandemic preparedness plan at the beginning of January, two months before COVID-19 arrived. “We decided from the beginning we would use isolation, quarantine and contact tracing,” says Þórólfur Guðnason, chief epidemiologist at the Directorate of Health. As part of that plan, the microbiology laboratory at the university hospital began testing citizens in early February. On 28 February, a man returning from a skiing holiday in northeastern Italy tested positive for the virus. Within a week, the number of cases had climbed from 1 to 47, the opening notes of a coming crescendo. As health-care workers began ordering hundreds of tests per day, one of the hospital’s machines for isolating and purifying RNA broke from overuse. “We were not able to cope with all the specimens coming in,” recalls Karl Kristinsson, the university hospital’s chief of microbiology. By 13 March, deCODE had begun screening the general public and was able to quickly take over much of the hospital’s testing. The company repurposed a large phenotyping centre that it had been using to study the genetics

of Icelanders for more than two decades into a COVID-19 testing centre. “It almost looked like these 24 years preceding COVID-19 had just been a training session,” says Stefánsson. “We dove into this full force.” The company has the staff and machinery to sequence 4,000 whole human genomes per week as part of its regular research activities, says Stefánsson. Throughout the spring, it would set that aside to devote its analytical and sequencing heft to the pandemic response. deCODE’s main activity has been COVID-19

IT ALMOST LOOKED LIKE THESE 24 YEARS PRECEDING COVID-19 HAD JUST BEEN A TRAINING SESSION.”

screening, including open invitations to the general population. Today, any resident with even the mildest symptom can sign up to be tested. Residents sign up online using dedicated COVID software built by deCODE programmers. At a testing centre, they show a barcode from their phone to automatically print a label for a swab sample. Once taken, the sample is sent to a laboratory at deCODE’s headquarters that is run jointly by the university hospital and deCODE and operates from 6 a.m. to 10 p.m. Results are always available within 24 hours, but are often ready in just 4 to 6. “We now have the capacity for about 5,000 samples per day,” says Kristinsson. As a whole, the collaborators have so far screened 55% of the country’s population. If the test is negative, the person receives an all-clear text. If the test is positive, it triggers two chains of action: one at the hospital and one at the lab. At the hospital, the individual is registered in a centralized database and enrolled in a telehealth monitoring service at a COVID outpatient clinic for a 14-day isolation period. They will receive frequent phone calls from a nurse or physician who documents their medical and social history, and runs through a standardized checklist of 19 symptoms. All the data are logged in a national electronic medical record system. A team of clinician-scientists at the hospital created the collection system in midMarch with science in mind. “We decided to document clinical findings in a structured way that would be useful for research purposes,” says Palsson. In the lab, each sample is tested for the amount of virus it contains, which has been used as an indicator for contagiousness and severity of illness. And the full RNA genome of the virus is sequenced to determine the strain of the virus and track its origin. The same approach could work in other countries that have suitable resources, such as the United States, where all the methods deCODE is currently using were developed, says Stefánsson. In fact, early in the pandemic, many US labs pivoted to offer coronavirus testing, but were stymied by regulatory and administrative obstacles, which critics attribute to a lack of federal leadership. “This was a wonderful opportunity for academia in the United States to show its worth, and it didn’t,” Stefánsson says. “I was surprised.”

Viral fingerprints Researchers at deCODE, the university hospital and the Directorate of Health began analysing the wealth of data in early March, and quickly published several early results. “Once we started to generate data, we couldn’t resist the temptation to begin to try to pull something cohesive out of it,” says Stefánsson. Iceland’s COVID-19 results are limited by the fact that cases are occurring in a small and

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

Nature | Vol 587 | 26 November 2020 | 537

Zealand population more or less stayed at home for 7 weeks. After that, we emerged into a virus-free country,” says Michael Baker, a public-health researcher at the University of Otago in Wellington. That’s a feat for a country of 5 million people, more than 13 times larger than Iceland. Genetic analysis of the first New Zealand wave, from March to May, showed that the strict lockdown began working right away. The rate of transmission — the number of people infected by each person with the virus — dropped from 7 to 0.2 in the first week in the largest cluster7. Sequencing data also showed that an August outbreak in Auckland, the source of which remains unknown, was from a single lineage, reassuring public-health authorities that there had only been one breach. “Genomics has played a vital role in tracking the re-emergence of COVID-19 in New Zealand,” says Jemma Geoghegan, a microbiologist at Otago who co-led the project with Joep de Ligt at the Institute of Environmental Science and Research in Porirua. One of the first families in Iceland to be screened with deCODE’s COVID-19 test.

Getting the full picture

genetically homogeneous population compared with other countries, notes Palsson. But in some cases, that small sample size is also a strength, because it has led to detailed, population-wide data. In early spring, most of the world’s COVID-19 studies focused on individuals with moderate or severe disease. By testing the general population in Iceland, deCODE was able to track the virus in people with mild or no symptoms. Of 9,199 people recruited for population screening between 13 March and 4 April, 13.3% tested positive for coronavirus. Of that infected group, 43% reported no symptoms at the time of testing2. “This study was the first to provide high-quality evidence that COVID-19 infections are frequently asymptomatic,” says Jade Benjamin-Chung, an epidemiologist at the University of California, Berkeley, who used the Iceland data to estimate rates of SARS-CoV-2 infection in the United States3. “It was the only study we were aware of at the time that conducted population-based testing in a large sample.” A smaller population study, carried out in an Italian town, came to similar results on asymptomatic infection months later. When a 78-year-old man died in the northern Italian town of Vo’, Italy’s first COVID-19 death, the region’s governor locked the town down and ordered that its 3,300 citizens be tested. After the initial round of government testing, Andrea Crisanti, head of microbiology at the University of Padua in Italy, asked the local government whether his team could run a second round of testing. “Then we could measure the effect of the lockdown and the efficiency of contact tracing,” says Crisanti, who is currently

This summer at the university hospital, Palsson’s team used the clinical data to investigate8 the full spectrum of disease caused by SARS-CoV-2. The most common symptoms among the 1,797 people who tested positive between 31 January and 30 April were muscle aches, headache and a non-productive cough — not fever, a symptom listed in both the US Centers for Disease Control and the World Health Organization case definitions for COVID-19. When used to guide testing, those definitions are likely to miss some symptomatic people, says Palsson. “Hopefully others will come to a similar conclusion and that will result in changes in the criteria,” he says. The results from Palsson’s team led to direct medical intervention in Iceland: individuals showing any sign of a common cold or aches are now encouraged to get tested, and the hospital categorizes new patients into one of three stages according to their symptoms, which dictates their level of care. The most recent study from Iceland focused on a major COVID-19 question: how long does immunity to SARS-CoV-2 last? deCODE’s team found that anti-SARS-CoV-2 antibodies remained high in the blood of 91% of infected people for 4 months after diagnosis9, running counter to earlier results suggesting that antibodies decline quickly after infection10,11. It is possible that the conflicting results represent two waves of antibodies. In an editorial accompanying the paper12, Galit Alter at Harvard Medical School in Boston, Massachusetts, and Robert Seder at the US National Institutes of Health’s Vaccine Research Center in Bethesda, Maryland, suggest that a first wave is generated by short-lived plasma cells in response to acute infection, then a second wave, produced

HOPEFULLY WE CAN START RELAXING OUR RESTRICTIONS SOON.” . d e v r e s e r s t h g i r l l A . d e t i m i L e r u t a N r e g n i r p S 0 2 0 2 ©

538 | Nature | Vol 587 | 26 November 2020

on leave from Imperial College London. The local government agreed. On the basis of the results of the two rounds of testing, the researchers found that lockdown and isolation reduced transmission by 98%, and — in line with Iceland’s results — that 43% of the infections across the two tests were asymptomatic4. In addition to tracking asymptomatic infections, the researchers in Iceland concluded that children under 10 were about half as likely to test positive as people aged 10 and older —­a finding confirmed in Crisanti’s study of Vo’, as well as studies in the United Kingdom5 and United States6. Additionally, the deCODE team analysed the viral genetic material of every positive case, and used that fingerprint to track where each strain of the virus came from and how it spread. Most of the initial cases, the researchers found, were imported from popular skiing destinations, whereas later transmission occurred mainly locally, within families (see ‘Iceland’s three COVID waves’). That genetic-tracing approach, called molecular epidemiology, was similarly applied in New Zealand to good effect. In March, New Zealand’s government implemented a stringent countrywide lockdown aimed at eliminating the virus. “Essentially, the New

JON E. GUSTAFSSON

Feature

SOURCE: HTTP://WWW.COVID.IS/DATA; DECODE GENETICS

New wave On 15 June, Iceland opened its borders to non-essential visitors from 31 European nations. A month later, on 16 July, the country also lifted restrictions on visitors from 12 more countries, including Canada, New Zealand and South Korea. The opening gave a boost to the struggling tourism industry, although numbers of visitors remained low, with about 75–80% fewer summer tourists than in 2019, according to the Icelandic Tourist Board. Then, on 10 August, a pair of tourists at Reykjavik airport tested positive for SARSCoV-2, ignored regulations and went into town. That incursion led to a small bump of cases in August centred on two pubs and a fitness centre visited by the tourists. Then, in mid-September, the number of infections increased abruptly, from 1 to 55 in a week. “This one clone of virus was able to spread around and cause lurking infections all over, especially in Reykjavik, and all of a sudden, we saw this increase,” says Guðnason. “It’s evidence of how difficult the virus is to contain.” By October, coronavirus was more widespread in the community than it had been in the first wave, peaking at 291 infections per day. On 17 October, the number of active infections finally began to decline, which researchers attribute to widespread testing, tracing and quarantine procedures, as well as fresh government restrictions and emphasis on mask wearing. “Hopefully we can start relaxing our restrictions soon,” says Guðnason. Despite the outbreak, the country continues to keep its borders open to tourists from some countries, although entry requirements are now stricter. Travellers must either self-quarantine for 14 days after arrival or participate in two screening tests: one on arrival, followed by five days of quarantine, then a second test. This method has led to the discovery that 20% of people who test negative in the first round will test positive in the second, notes Guðnason. That is a high number, but seems consistent with other analyses13. The new requirement is likely to have caught many strains of virus that would have otherwise entered the country.

ICELAND’S THREE COVID WAVES

The island nation has identified about 5,250 positive cases of COVID-19 through testing, including random screening and double-testing of individuals who come to the country from abroad. 1,200

Positive tests

1,000 After a small surge in infections in August, Iceland has had to contend with what it calls its third wave of infections.

800 600 400 200 0 Feb

Mar

Apr

May

Jun

Jul

Aug

Sep

Oct

Nov

Dec

Contact-tracing data from the current COVID-19 surge reveals where domestic infections are coming from. Transmission within households is a key driver. Household transmission

No source identified

School

Social outings

Work

Health-care settings

1.0

0.8

0.6 Ratio

by longer-lived cells, bestows lasting immunity. And finally, Stefánsson was able to pin down the elusive statistic that first intrigued him — the infection fatality ratio (IFR), or the proportion of infected people who die from the disease. Since the beginning of the pandemic, IFR estimates have ranged from less than 0.1% to a whopping 25%, depending on the size of the study and the age of the population. A growing number of studies are converging at about 0.5 to 1%. In Iceland, where the median age is 37 — relatively young compared with other wealthy nations — and patients have access to good health care, Stefánsson’s team found it to be 0.3%.

0.4

0.2

0

Oct

Unlike New Zealand, which closed its borders, elimination was never supported in Iceland for fears that the country would go bankrupt without tourism. So it is possible that new cases will continue to arise, says Guðnason. Furthermore, he and others think the current outbreak might be in large part due to pandemic fatigue, as people disregard health precautions after months of being careful. “I think we’re going to be dealing with the virus, trying to suppress it as much as possible, until we get the vaccine,” he says. And research continues in any and every spare hour. Palsson’s team is planning to analyse the effect of viral loads on patient outcomes and viral transmission, and to use contact-tracing data to tease out the risk factors for a super-spreading event. “We’ve had households where almost everybody gets infected, then other places where people carry the infection and stay in the workplace and nobody gets infected,” says Palsson. “It’s very difficult to understand.” At deCODE, Stefánsson and his colleagues are investigating cellular immune responses and whether people with COVID-19 who are very sick produce antibodies directed against their own tissues. And together, the deCODE and university-hospital teams are

Nov

collaborating on the long-term effects of COVID and how genetics affects susceptibility and responses to the disease. “We’ve been committed for a long time to take everything we learn about human disease and publish it,” says Stefánsson. “There is no way in which we would have not utilized the opportunity.” Megan Scudellari is a science journalist in Boston, Massachusetts. 1. Gottfredsson, M. et al. Proc. Natl Acad. Sci. USA 105, 1303–1308 (2008). 2. Gudbjartsson, D. F. et al. N. Engl. J. Med. 382, 2302–2315 (2020). 3. Wu, S. L. et al. Nature Commun. 11, 4507 (2020). 4. Lavezzo, E. et al. Nature 584, 425–429 (2020). 5. Docherty, A. B. et al. Br. Med. J. 369, m1985 (2020). 6. Rosenberg, E. S. et al. Clin. Infect. Dis. 71, 1953–1959 (2020). 7. Geoghegan, J. L. et al. Preprint at medRxiv https://doi. org/10.1101/2020.08.05.20168930 (2020). 8. Eythorsson, E. et al. Preprint at medRxiv https://doi. org/10.1101/2020.08.09.20171249 (2020). 9. Gudbjartsson, D. F. et al. N. Engl. J. Med. 383, 1724–1734 (2020). 10. Long, Q.-X. et al. Nature Med. 26, 1200–1204 (2020). 11. Ibarrondo, F. J. et al. N. Engl. J. Med. 383, 1085–1087 (2020). 12. Alter, G. & Seder, R. N. Engl. J. Med. 383, 1782–1784 (2020). 13. Kucirka, L. M., Lauer, S. A., Laeyendecker, O., Boon, D. & Lessler, J. Ann. Intern. Med. 173, 262–267 (2020).

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

Nature | Vol 587 | 26 November 2020 | 539

Science in culture

LISA SALTZMAN

Books & arts

Thoughts, a photograph portrait by Lisa Saltzman.

When is sorrow sickness? A history of depression A book traces the shifting lines between sadness and illness, but not who gets to draw them. By China Mills

W

hen is sorrow sickness? So begins Jonathan Sadowsky’s The Empire of Depression, a history riven with professional turf wars around where to draw the line between normal sadness and something more serious — now, across much of the world, called depression. He argues against reductionism and dogma. Instead of getting stuck in old disagreements about whether depression is caused by a chemical imbalance or by social inequality, Sadowsky urges that depression can be psychological, biological and social, just as it can be a real illness even if it is cultural.

Given that the World Health Organization names depression as a major contributor to the global burden of disease, tracing its history is a significant task. And it is an important The Empire of Depression: A New History Jonathan Sadowsky Polity (2020)

one, given the mental-health crisis attending the COVID-19 pandemic. It is no mean feat to characterize something that has ever-shifting and contested boundaries dependent on time and place. Sadowsky, a historian of medicine, offers three possible reasons for the current boom in diagnoses: that there really is more depression; that the amount is the same but we’re better at detecting it; or that emotional states not previously seen as illness are now being labelled as such. This is no lament on over-diagnosis. Rather, Sadowsky offers a review of how psychiatry has helped people. His tale spans the ‘melancholia’ of Renaissance Europe (said to be caused by too much black bile, and treated by purging) and today’s research on imbalances in neurotransmitters, treated by drugs. It takes us through the Christian Middle Ages and the emergence of questions about whether ill people were to be blamed for their suffering; Sigmund Freud’s psychoanalytic ideas about anger turned inwards; and the 1980s cultural sensation of Prozac (fluoxetine), quantification and globalization. In this positive take, the vast majority of people with depression are being treated voluntarily, and treatment helps them feel better. He makes no claims that people are being duped en masse into chemical cures. But this breezy approach doesn’t reckon with power: a lot goes on between the lines of being forced to take medication and choosing to.

Cultural condition Depression has a cultural life, as, others have argued, do panic disorder, bipolar disorder and suicide (see Jackie Orr’s book Panic Diaries (2006); Emily Martin’s Bipolar Expeditions (2007); and Ian Marsh’s 2010 Suicide). A diagnosis can give validity to feelings, help people find others who share similar experiences and provide hope. But it can also stigmatize, embroil people in coercive treatment regimes and overlook context. That is why it is important to ask: what does depression ‘do’, personally and politically? Individual explanations can divert attention from wider societal factors and how they make some lives unliveable. In other words, the history of depression is also about who decides what is normal and what is not. If life presents many reasons to be depressed — poverty, discrimination, precarious living situations — then should all depression be seen as an illness? This is more than theoretical, as increasing prescriptions of antidepressants in austerity Britain and opioids in rural North America testify. To be fair, this idea is key to Sadowsky’s history. He

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

Nature | Vol 587 | 26 November 2020 | 541

Books & arts

Books in brief Burning the Books Richard Ovenden John Murray (2020) In 1660, when the British monarchy was restored, the University of Oxford ordered the burning of books by anti-royalist, pro-free-speech poet John Milton. This is one of many graphic examples inspired by the collections at the institution’s Bodleian Library, cited in Bodleian librarian Richard Ovenden’s powerful history, which ranges from ancient Mesopotamia and Egypt’s Library of Alexandria to Nazi Germany and the Iraq wars. Attacks on books and archives are, he concludes, a “signal that attacks on humans cannot be far behind”.

Numbers Don’t Lie Vaclav Smil Viking (2020) Vaclav Smil’s publisher claims: “No other living scientist has had more books (on a wide variety of topics) reviewed in Nature.” Smil might regard this as almost unprovable, but certainly Bill Gates calls him a impressive polymath. His latest book, an appraisal of statistics, offers 71 short, thoughtful essays on psychology, globalization, inventions, fuels and electricity, transport, diets and the environment. Discussing the annual World Happiness Report, he notes a “remarkable lack of correlation” between national happiness rankings and suicide rates.

Arctic Eds Amber Lincoln et al. Thames & Hudson/British Museum (2020) Humans have “the right to be cold”, says an Inuit political activist in this accompaniment to a British Museum exhibition. Arctic cultures encompass 4 million people across 8 nations, whose way of life is threatened: 75% of Arctic sea ice has melted over 40 years. Superb illustrations and many essays reveal fascinating accoutrements, such as Russian snow spectacles from 1850–80, crafted from reindeer skin, multicoloured glass beads and uranium beads, with minute metal slits protecting against blindness induced by Arctic spring sunlight.

Is Capitalism Broken? Yanis Varoufakis et al. Oneworld (2020) Two models of capitalism — US democratic and Chinese authoritarian — underlie this record of a 2019 debate. Former Greek finance minister Yanis Varoufakis and US publisher Katrina vanden Heuvel proposed that capitalism is broken. They were opposed by social scientist Arthur Brooks and journalist David Brooks, the narrow winners. But Varoufakis wins on eloquence: he says capitalism liberated us from prejudice and feudalism, but entangled us in “unbearable inequality, unsustainable debt, brazen authoritarianism, and, yes, catastrophic climate change”.

Inscriptions of Nature Pratik Chakrabarti Johns Hopkins Univ. Press (2020) Science historian Pratik Chakrabarti’s idiosyncratic book ponders an 1820s excavation of the Yamuna Canal in India’s Indo-Gangetic Plain, which revealed ancient canal networks, vanished river beds, traces of mythological rivers and prehistoric fossils. Simultaneously, geologists, ethnologists, archaeologists and missionaries dug into ancient texts and legends. From both perspectives, the plain seemed to be the bedrock of Indian civilization — a view complicated by the 1920s discovery of the Indus civilization. Andrew Robinson

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

542 | Nature | Vol 587 | 26 November 2020

explains that choices about where to draw the line risk both medicalizing everyday suffering and disqualifying many people’s suffering from being seen as clinical depression. Depression, then, can’t be separated from unequal power relations — between doctor and patient, and between countries and corporations with unequal power to globalize their ways of viewing distress and its treatments. The power to say who’s rational and who isn’t, and to detain people or treat them without consent, is perhaps the starkest reminder of why treating depression is not just like administering insulin for diabetes, and of why stigma looms large despite (or because of) the understandable appeals to biochemistry. This concept is behind anti­psychiatry, the movement that protests against harmful practices, especially those founded on power differentials. Yet antipsychiatry makes only brief appearances in the book, like a pantomime villain.

False dichotomies Sadowsky points out that era after era grapples with false choices — between a political understanding of depression and a medical one, or between physical and psychological understandings. He is right to call for a move beyond these crude binaries. In my view, to do so, we must face the central roles of racism, sexism and ableism in delineating diagnostic boundaries over the years — not dismiss them as unfortunate. For example, psychiatry has a history of labelling some people as too uncivilized to be mentally ill, yet also diagnosing anticolonialism as mental illness. In fact, many forms of resistance have been deemed symptoms of mental illness, from enslaved Africans fleeing brutality in the nineteenth-century United States to the Black Power movement of the 1960s. It is alarming, then, that apart from using “empire” as an analogy for the global dominance of Western psychiatry in interpreting distress, Sadowsky devotes little attention here to power and politics — especially given his previous work on colonialism (in the 1999 book Imperial Bedlam). The book ends with the wise injunction: “listen to patients”. Yet, apart from illness memoirs, the voices (and research) of people who experience depression, those who become patients, those who refuse to become patients, and service users or psychiatric survivors are almost completely absent from the book. These people (and the collectives they have formed) contribute to production of knowledge about depression by leading research and global movements fighting for their rights. A history of depression without these voices will always be one-sided. China Mills is a senior lecturer in public health at City, University of London, and author of Decolonizing Global Mental Health. e-mail: [email protected]

Setting the agenda in research

KATE​ HOLT/EYEVINE

Comment

A nurse at a health centre in Cape Coast, Ghana, talks to women about family planning.

Reboot contraceptives research — it has been stuck for decades Sarah G. Chamberlain, Kirsten M. Vogelsong, Michelle Weinberger, Emily Serazin, Sarah Cairns-Smith & Stephen E. Gerrard

There is a huge global market, and exciting tools are ready to help develop what women want.

W

orldwide, almost half of women who are of reproductive age use contraception 1. Another 171 million women — around 1 in 11 aged between 15 and 49 — do not use it, yet want to avoid pregnancy1. Several factors contribute to this unmet need. New, effective and more-desirable contraceptive options are urgently needed. Family planning

is a key aspect to meeting United Nations Sustainable Development Goals 3 and 5. Many women and men are highly dissatisfied with the contraceptives available2. Male condoms fail too often — in the first year of condom use, about 13% of women become pregnant3 — and women must rely on men. Implants and intra-uterine devices (IUDs) require medical procedures and can be invasive; pills have to be taken every day. Hormonal methods and non-hormonal IUDs can have side effects, including irregular or unpredictable menstrual bleeding, headaches, acne and weight gain, as well as depression and other mood changes4. For many women worldwide, contraception has been difficult to obtain or afford, even before the COVID-19 pandemic.

All of this has serious consequences. Around 40% of pregnancies globally are unintended, and about half of those end in induced abortion5. A high proportion of unintended pregnancies occur even where contraception is relatively accessible and cultural stigma against it is generally low, for example in North America (48%) and Europe (43%)5. Those pregnancies can happen because women aren’t using contraception, because their method failed or because it was used incorrectly. Nearly 25% of unintended pregnancies in low- and middle-income countries (LMICs) occur in women who were using modern forms of contraception6, and globally it’s an even higher share (see ‘Needs gap’). Unintended pregnancies can have

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

Nature | Vol 587 | 19 November 2020 | 543

Lack of satisfaction The contraceptives available do not fully meet many women’s changing needs and preferences through their reproductive lives. Among women in LMICs who do not want to get pregnant but are not using modern contraceptives, more than one-quarter cite side effects as the main reason11. And the same types of side effect accompany many different products. Globally, about one-third of women discontinue their hormonal method of contraception in the first year of use, many citing side effects or health concerns as the main reason4. Large-scale, detailed data are extremely sparse, especially from women who continue to use a method of contraception despite being dissatisfied with it. More than 100,000 women from nearly 200 countries completed a survey on contraceptive preferences within 1  month of its opening. (The survey was released through the reproductive-health app

Current methods of contraception do not fulfil many women's requirements.

Unintended pregnancies 40% Improper use or method failure 16%

for International Development. Over the past few years, other organizations have funded or invested in specific non-hormonal technologies, including the BioInnovation Institute in Copenhagen and US-based RHIA Ventures and Adjuvant Capital. There’s room for many more.

Cycle of neglect Intended pregnancies 60%

Not using contraception 24%

Clue and online at http://shapefuturect.org; it was developed by Avenir Health, where M.W. works, and funded by the Bill & Melinda Gates Foundation in Seattle, Washington, with S.E.G. as program officer). Early analysis suggests that GAP a range of side effects would lead FUNDING Contraceptives needs more respondentsresearch to stop using a investment method, espeglobally. Private firms and public sourceschanges such as the US cially changes to mood, physical such National Institutes of Health are key to future funding. as acne, weight gains of 2–4.5 kilograms, loss Women’s health of hair and lowered sex drive. These findings 2019 US NIH spend and STIs* 1.1% echo others (go.nature.com/35hgqsm). A 2018 review showed that changes to heaviness or frequency of menstrual bleeding have all been associated with reported dissatisfaction with contraceptives12. All of this probably helps to explain the enthusiastic responses to product launches Contraceptive over the past decade. For example, the Mirena and reproduction family of products, a hormone-releasing 21.3% intra-uterine system made by Bayer in LeverWomen’s kusen, Germany, has maintained blockbuster 2019 biopharma R&D† spend health 7.2% sales of more than $1 billion for each of the past 5 years. An oral contraceptive introduced in 2011, Lo Loestrin, which offers the lowest amount of daily oestrogen available (with the potential for fewer side effects than for those of related products), captured a significant share of the market9 and net revenues have Contraception seen double-digit growth over time. In 2018, 0.7% an app called Natural Cycles was approved as a *STIs, sexually transmitted infections; †R&D, research and development. contraceptive, and relies on body temperature to inform users when they are fertile. Earlier this year, Evofem launched Phexxi — a firstin-class vaginal pH modulator that works as a non-hormonal contraceptive. The impact of these two latest products will become clear over the next few years. Yet there are few truly innovative and highly effective contraceptive products in development. According to ClinicalTrials. gov, there have been 20–25 industry-funded clinical trials between 2017 and 2020. The majority focus on incremental revisions to existing hormonal products. By comparison, in 2019 there were about 3,100 trials for cancer drugs, 600 for cardiovascular drugs and 140 for treatments for eye disorders13. Funding of R&D for female contraception comes from just a handful of players. These include the Eunice Kennedy Shriver National Institute of Child Health and Human Development in Rockville, Maryland, the Bill & Melinda Gates Foundation (where K.M.V. works and S.E.G. recently worked) and the US Agency

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

544 | Nature | Vol 587 | 19 November 2020

NEEDS GAP

Why are funding and R&D so limited for female contraceptives? One reason is that they are given to healthy women of reproductive age, so the safety requirements for regulatory approval are (appropriately) very high: serious or severe adverse effects are not acceptable. Efficacy requirements are also extremely stringent. These regulations act as commercial disincentives for trying something new. In addition, there are unique liability concerns for new products in reproductive health — especially in the United States, which is a litigious market. There have been a number of high-profile cases against leading contraceptive manufacturers, resulting in multimillion-dollar settlements NEEDS GAP(see go.nature.com/3ncb7jv). Vaccines one ofofthe only other product classes Currentare methods Unintended contraception do not pregnancies 40% that are administered to a healthy population. fulfil many women's However, in the United States, vaccine manrequirements. Improper use or ufacturers are protected from liability under method failure 16% the National Childhood Vaccine Injury Act, established by an Act of Congress in 1986. This is unlikely to happen for contraception. Another a business perIntendedproblem is that,from Not using spective, the contraceptive market seems24% to be pregnancies 60% contraception healthy and growing. It was valued at $24 billion in 2018 (ref. 14). Yet the demand from women for transformational change is not reflected as a reduction in sales. Furthermore, women’s health issues, and their preferences, are simply

FUNDING GAP

Contraceptives research needs more investment globally. Private firms and public sources such as the US National Institutes of Health are key to future funding.

2019 US NIH spend

Women’s health and STIs* 1.1%

Contraceptive and reproduction 21.3%

2019 biopharma R&D† spend

Women’s health 7.2%

Contraception 0.7% *STIs, sexually transmitted infections; †R&D, research and development.

SOURCES: NIH/EVALUATEPHARMA/BCG ANALYSIS

lasting effects on women, children, families and society. The direct health-care costs in LMICs amounted to US$10 billion6 in 2019 alone; the indirect, longer-term economic costs could be 40 times that7. Estimates suggest that US unplanned births resulted in $21 billion in publicly funded medical costs in 2010 (ref. 8). Progress in contraceptives research and development (R&D) has been slow in recent decades. Pharmaceutical companies typically spend around 20% of their sales revenue on R&D for new products9. For contraception, that figure is just 2%. We estimate that most of this spending has focused on incremental improvements to classes of hormonal contraceptive compound that have been in use for decades. Encouragingly, there are now more opportunities than ever for innovation. Many scientific advances of the past decade can now be applied to developing non-hormonal drugs that target the egg, sperm or processes along the journey to conception. Such products could have fewer and less-severe side effects than current ones. And alongside daily oral pills, to meet the needs of women who want contraception for different lengths of time, various non-hormonal products with months or years of action could be delivered — through injections, implants, IUDs and other user-responsive systems that are currently in development. At the same time, the COVID-19 pandemic is changing health-care services, possibly forever. Many of today’s contraceptives, such as implants and IUDs, require an in-person appointment10. New methods could be delivered remotely, directly to users. A coalition of innovators, researchers, biopharmaceutical firms, donors and investors needs to come together now to produce better contraception for women. Many of the steps needed might catalyse innovation in male contraception, too.

SOURCE: INTERNAL ANALYSIS BY AVENIR HEALTH USING MULTIPLE SOURCES OF GLOBAL PREGNANCY DATA; FUNDED BY THE BILL & MELINDA GATES FOUNDATION.

Comment

under-studied and under-funded, and unmet needs are ignored and misunderstood by those who could work to address these issues. These barriers — tight regulations, high liability risk and the lack of a strong market signal — fuel a false perception of low return on investment in contraceptive R&D. So, over the past 20 years, many global biopharma companies have sold off, reduced or closed divisions that were developing non-hormonal contraceptives and other women’s health products, such as those to support menopause. Companies have instead focused on therapeutic areas that are evidently fast-growing, such as oncology. When drug firms step away from a field, it can start a cycle of neglect. Venture capitalists become wary of supporting technologies with unclear opportunities for exit strategies. Academics become cautious about pursuing an area with reduced commercial outlets and financial support. Private companies have few promising avenues to explore, and potential funders cannot easily identify where to invest. For contraception, this has led to missed opportunities, because the scientific tools for R&D have mushroomed. Public-sector funding has been one of the key reasons the field has dodged dormancy (see ‘Funding gap’).

and applied genomics, for example, engage industry in early stages. Such collaboration focuses research on saleable products targeted at consumer need. It also increases reproducibility of results, breaks down silos and brings in diverse perspectives to improve robustness. Without early buy-in, innovation efforts typically fall outside biopharma’s tolerance for risk. Fresh thinking will also be needed to ensure that the latest contraceptive products get to those who need them most. For example, new non-hormonal contraceptives might require new manufacturing processes and are therefore likely to be priced higher than existing products. Revenue from high-income markets could subsidize affordable prices in LMICs15. Vaccine development is a good example of collaboration on product development that enables access by LMICs. Research communities such as the Collaboration for AIDS Vaccine Discovery focus on discovery. These bring the best scientists together, prioritize high-impact

“Women’s health issues are simply under-studied and under-funded, and unmet needs are ignored.”

Prime time For the first time in a generation, a coalition of stakeholders could revolutionize the sector. For example, it is now possible to use genomics tools in a way that was not available 20 years ago. Operating costs have plummeted, and analytical methods and data sets have rapidly expanded in sophistication and size. Biostatisticians can comb for genes or proteins key to egg or follicle maturation, fertilization or gamete function. This can isolate targets for non-hormonal pharmaceutical interventions in a way that is much more efficient than previous, failed approaches. The neighbouring fields of gynaecological oncology and infertility have seen industry funding increase over the past decade. Advances in those fields could help contraceptive R&D. For example, progress in understanding the mechanisms underpinning ovulation could help to identify potential drugs that could temporarily affect the same biological pathways. Online tools also offer opportunities. Between 2015 and 2018, investors ploughed more than $1 billion into digital and diagnostic products and services that aid family planning, including menstruation and fertility-tracking apps (go.nature.com/2toj2vp). The sector is expected to be worth $50 billion by 2025 (go. nature.com/3pcswpt). Other apps and social media could help to create large-scale data sets articulating women’s needs, as long as privacy can be protected.

Key collaborations Public–private partnerships will be key. The best innovation models in oncology, immunology

research and support the development of assays and model systems. To aggregate public, private and philanthropic funding, product-development partnerships have come together, such as the International AIDS Vaccine Initiative, the International Vaccine Institute and PATH Center for Vaccine Innovation and Access. These mechanisms drive innovation and significantly reduce the financial risk of early-stage investment. This type of infrastructure and collaboration has been crucial to the fast pace of innovation and development for COVID-19 vaccines and therapeutics. Regulators, too, can help to lower the barriers to innovation. For example, once contraceptive drugs are well studied for safety and efficacy, developers might be allowed to use modelling, alongside clinical-trial data, in support of future products using the same drug16. Regulators might also need to consider how they assess effectiveness. A new non-hormonal product might have lower efficacy in clinical settings but have fewer side effects than some existing products, leading to higher acceptance and use, for example. New funding is crucial to catalyse innovation in any sector. Venture funds, biopharma, biotechnology firms and universities should assess opportunities to apply their technologies and expertise to contraception, which could accelerate and increase innovations in R&D across multiple fields. The public needs to speak up about its desires and demands, so that we can move from methods that women tolerate to those

that actually satisfy their needs. Success stories of new, reliable contraceptives with fewer side effects will create a virtuous cycle, spurring more funding and research and better options for consumers. A thriving contraceptive R&D ecosystem might also catalyse innovation in other sectors of women’s health: infertility, endometriosis and sexually transmitted infections, to name a few. What are we waiting for?

The authors Sarah G. Chamberlain is a partner at Boston Consulting Group, Seattle, Washington, USA. Kirsten M. Vogelsong is a senior program officer at the Bill & Melinda Gates Foundation, Seattle, Washington, USA. Michelle Weinberger is a senior associate at Avenir Health, Washington DC, USA. Emily Serazin is a partner and managing director at Boston Consulting Group, Washington DC, USA. Sarah Cairns-Smith is a senior adviser at Boston Consulting Group, Boston, Massachusetts, USA. Stephen E. Gerrard co-authored this article while a program officer at the Bill & Melinda Gates Foundation, Seattle, Washington, USA. e-mail: [email protected] S.G.C., E.S., S.C-S. & S.E.G. declare competing and competing financial interests. For details and further reading, see go.nature.com/3najukk. 1. United Nations Population Division. Estimates and Projections of Family Planning Indicators 2020: Regions (UN, 2020). 2. Moreau, C., Cleland, K. & Trussell, J. Contraception 76, 267–272 (2007). 3. Trussell, J., Aiken, A. R. A., Micks, E. & Guthrie, K. A. in Contraceptive Technology 21st edn (eds Hatcher, R. A. et al.) (Ayer, 2018). 4. Ali, M. M., Cleland, J. G., Shah, I. H. & World Health Organization. Causes and Consequences of Contraceptive Discontinuation (WHO, 2012). 5. Bearak, J., Popinchalk, A., Alkema, L. & Sedgh, G. Lancet Glob. Health 6, e380–e389 (2018). 6. Sully, E. et al. Adding It Up: Investing in Sexual and Reproductive Health (Guttmacher Inst., 2019). 7. Kohler, H. & Behrman, J. Benefits and Costs of the Population and Demography Targets for the Post-2015 Development Agenda (Copenhagen Consensus Center/ Univ. Pennsylvania, 2014). 8. Sonfield, A. & Kost, K. Public Costs from Unintended Pregnancies and the Role of Public Insurance Programs in Paying for Pregnancy-Related Care: National and State Estimates for 2010 (Guttmacher Inst., 2015). 9. EvaluatePharma Database; available at https://go.nature. com/2khvqot (Evaluate, 2019). 10. Weinberger, M., Hayes, B., White, J. & Skibiak, J. Glob. Health Sci. Practice 8, 169–175 (2020). 11. Guttmacher Institute. Reasons for Unmet Need For Contraception in Developing Countries: Fact Sheet (Guttmacher Inst., 2016). 12. Polis, C. B., Hussain, R. & Berry, A. Reprod. Health 15, 114 (2018). 13. Shin, D. Clinical Trials 2019 Round-up (PharmaIntelligence Informa, 2019). 14. Ugalmugle, S. & Swain, R. Contraceptives Market Size By Product (Global Market Insights, 2019). 15. Kaslow, D. C. in Global Innovation Index 2019. Creating Healthy Lives — the Future of Medical Innovation Ch. 11 (WIPO, 2019). 16. Li, L. et al. Annu. Rev. Pharmacol. Toxicol. https://doi. org/10.1146/annurev-pharmtox-031120-015212 (2020).

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

Nature | Vol 587 | 19 November 2020 | 545

Comment fundamental, sometimes counter-intuitive changes that will strengthen the use of science in US policy and by the research community more broadly.

President-elect Joe Biden has a very different approach from that of his electoral opponent.

Memo for President Biden: Five steps to getting more from science Roger Pielke Jr & Neal Lane

Going back to normal is not enough. A revamp is required.

A

s things look now, the US presidency of Donald Trump will soon be in the rear-view mirror, but the damage his administration leaves behind will require a sustained effort to repair. That’s especially true when it comes to restoring competency and trust in federal

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

546 | Nature | Vol 587 | 19 November 2020

research agencies. President-elect Joe Biden needs to do this as soon as possible, not least to quell a pandemic that is setting records for the numbers of new cases and is on track to kill more Americans than died in the Second World War. The country cannot continue to bear the ad hoc, ineffective and incoherent pandemic response it has endured under Trump. The list of needed actions is long, but here we highlight five that the Biden administration should take swiftly. We call not for a return to business as usual but for

Trump’s coronavirus task force, which ostensibly guided the administration’s response to the pandemic, had little authority and no accountability, had to fight for attention against other priorities, and was deliberately politicized. The task force usurped the leading role of the Department of Health and Human Services, and sidelined its Centers for Disease Control and Prevention, damaging public trust in both. A better, albeit less-obvious, option to lead the pandemic response under Biden is the White House Office of Science and Technology Policy (OSTP, which one of us, N.L., led from 1998 to 2001). It was established in 1976 to advise the president and coordinate federal science agencies. Although the OSTP has predominantly focused on deciding priorities for research funding, its history and mandate make it ideally poised to coordinate a national effort for responding to COVID-19. In February, as the pandemic was just beginning to spread in the United States, the Government Accountability Office warned that the nation’s biodefence strategy needed “to move away from traditional mission stovepipes toward a strategic enterprise-wide approach”. The OSTP has the perspective needed to work across agencies, and it has coordinated policy before. Former president Ronald Reagan relied on it to advance his ‘Star Wars’ ballistic-missile defence programme. What’s more, the OSTP would offer a fresh start to the pandemic response. Under Trump, it had little visible role and so, unlike the federal public-health agencies, has been less politicized. Finally, the OSTP sits in the White House but is also accountable to Congress, with a director confirmed by the Senate. That keeps it both close to the president and subject to congressional oversight, unlike Trump’s coronavirus task force. Leadership will require working across branches of government, and having the OSTP in charge would boost legitimacy, because the Democrat-led office will be working with a Republican-led Senate. At the same time, the head of the OSTP — the White House science adviser — should also be elevated to the president’s cabinet. This guarantees a seat at the table when the most important, consequential decisions are made. It will also signify the importance of the role to federal agencies, to Congress and to the public.

Make advisory processes more independent A tenet of effective advisory bodies is that advisers advise and decision makers decide.

KEVIN LAMARQUE/REUTERS

Let an oft-overlooked White House office lead the pandemic response

Advice might take the form of narrow technical guidance on scientific matters (does a particular drug improve COVID-19 health outcomes?), presentation of policy alternatives (what are the risk-reduction options for reopening schools?), or recommendation of a specific action (should masks be mandatory indoors?). Under Trump, scientific advice was typically ignored or, worse, manipulated for political expediency. That is easier to do when responses are managed by ad hoc groups. For example, radiologist Scott Atlas was selected as Trump’s top pandemic adviser to counter government staff scientists and support the political agenda of the president. The advisory mechanisms available to draw on are broad and deep. The US government lists more than 1,000 bodies currently active under the Federal Advisory Committee Act. Biden and the OSTP must ensure that advisory committees consist of independent experts selected for competency, that their role is clear, and that their advice reaches decision makers in the field — from public health to environmental protection. The White House will also need to reject Trump-era policies that keep the government from drawing on competent expertise. First in line should be reversal of an executive order signed last month that removed civil-service protections from positions usually filled by career employees, making them easier to fire for political reasons. Advisory committees, such as those leading the US National Climate Assessment, should comprise independent experts, selected by bipartisan panels (as is typically done for committees linked to politicized issues), and not political appointees. And political appointees should never alter or edit science advisory-committee reports or recommendations. The main criticism of such reforms might be that they would empower independent experts over administration officials. Indeed — we see that as a feature, not a flaw. Also, having independent advice doesn’t mean decision makers will always heed it; the administration of former president Barack Obama decided, contrary to recommendations of its expert advisers, to limit distribution of the morning- after pill in 2011; it similarly rejected expert advice in 2016 to strengthen ozone regulations. Still, as Biden has said, decision makers have an obligation to “listen to the scientists”.

Expedite scientific-integrity legislation The Obama administration instigated an effort to implement scientific-integrity policies across federal agencies; some 24 agencies developed relevant administrative policies in response. But several subsequent reviews, including one by the Government Accountability Office,

found these scientific-integrity policies to be unevenly interpreted and applied. Some agencies, such as the Department of Defense, were not included under the mandate. Others, including the National Institutes of Health and the Department of Labor, did not develop policies. Agencies that did develop policies defined ‘scientific integrity’ in different ways, and created conflicting guidelines for topics such as media relations and how to handle disparate scientific perspectives. And the Trump administration rode roughshod over these rules anyway, for instance by barring a Department of State analyst from including

“The White House will need to reject Trump-era policies that keep the government from drawing on competent expertise.” information about climate change in written testimony to a congressional committee. Harmonized legislation that allows congressional oversight would be more difficult to ignore or evade. Several proposals exist that would promote scientific integrity, protect agency officials and strengthen the ability of Congress to keep the executive branch in check. Presidents rarely advocate restricting their own power, but Biden should. One relevant bill was introduced in the House of Representatives in 2019 and has more than 200 co-sponsors.

Give public universities tough love and lots of support The US public-university system has suffered deep budget cuts during the pandemic, with no relief in sight. And state governments had been cutting support in the decades before that. On average, according to the American Academy of Arts and Sciences, states cut funding per student by 30% between 2000 and 2014 — leading to tuition and fee hikes, a greater reliance on out-of-state tuition to replace those state funds, and drastically increased student debt. Some students are particularly disadvantaged: a recent report from the Education Trust gave failing grades to more than 75% of the nation’s top 101 universities for their accessibility to Black students, with about 50% receiving failing grades for accessibility to Latino students (see go.nature.com/2i7pidk). The federal government should help public universities with long-term financial sustainability, and perhaps even provide temporary recovery funding. Strings attached should include plans to boost diversity among students, faculty members and researchers. Critics might argue that such issues are not the concern of the federal government.

However, the data indicate that these issues are a systemic, national concern. There is ample precedent for a federal role in higher-education policy, dating back to the 1965 Higher Education Act.

Refocus science funding In spite of the Trump administration’s efforts to slash investment, Congress ensured that federal funding of research and development increased by more than 20% between 2017 and 2020. Still, the United States ranks tenth among member nations of the Organisation for Economic Co-operation and Development in national investment (public and private, as a percentage of gross domestic product) in research and development, and the federal government’s share of that has fallen steadily over recent decades. Policy proposals from Biden’s team, and several bipartisan bills in Congress, suggest that federal research and development funding will grow substantially. That growth must come with shifts in priorities. It should no longer be based on incremental changes to legacy budgets, as presidents often put forward. Instead, it must give higher priority to achieving national policy goals, beyond fundamental scientific knowledge. For instance, achieving net-zero carbon dioxide emissions from electricity generation will require a new era of federal–industry partnerships supporting sustained energy-technology innovation. Other priorities should include research and development to help Americans recover from the pandemic, the economic catastrophe, the ‘infodemic’ and the ravages of systemic inequality. The academic research community conventionally emphasizes basic research over science directed at solving societal challenges, because the former occurs mainly in academia and the latter in federal laboratories. To gain researchers’ support for ‘mission science’, the Biden administration will need to assure them of its continuing support for basic research. The challenges the Biden administration faces are daunting. Yet they create opportunities to make ‘build back better’ a reality, not just a bumper sticker.

The authors Roger Pielke Jr is a professor of environmental studies at the University of Colorado Boulder, USA. His books include The Honest Broker: Making Sense of Science in Policy and Politics. Neal Lane is a senior fellow in science and technology policy at Rice University’s Baker Institute in Houston, Texas, and former director of the White House Office of Science and Technology Policy and the National Science Foundation. e-mails: [email protected]; [email protected]

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

Nature | Vol 587 | 19 November 2020 | 547

Readers respond

Correspondence COVID-19: Panama stockpiles unproven drugs

Explain ESA’s late ditching of new space telescope

COVID-19: students caught in Pakistan’s digital divide

Giant tortoises make a comeback in Madagascar

Panama has gone one step further than other Latin American countries that use the unproven drug ivermectin against COVID-19 (see Nature 586, 481–482; 2020). The government has approved this drug and hydroxychloroquine — despite the lack of efficacy data for either — and is stockpiling both. We find this alarming because the drugs’ side effects could be toxic in a significant proportion of the population. In March 2020, hydroxychloroquine was prescribed only to people with COVID-19 who were taking part in a clinical-trial protocol. Panama’s ministry of health followed the advice of its Scientific Advisory Committee and two months later suspended the trial on the basis of growing evidence of the drug’s ineffectiveness (see M. R. Mehra et al. Lancet 395, 1820; 2020). The World Health Organization subsequently declared that, according to the available data, hydroxychloroquine does not reduce fatality in people hospitalized with COVID-19, nor does it ameliorate symptoms of mild or moderate illness. Despite these developments, Panama has since purchased 2,900,000 doses of hydroxychloroquine and 450,000 doses of ivermectin, to be distributed as part of a self-treatment kit to people selfisolating with the virus.

The European Space Agency (ESA) has cancelled its proposed Space Infrared Telescope for Cosmology and Astrophysics (SPICA; go.nature.com/2jfp8fw) just months before the final mission-selection review. The decision — made by the executive on the grounds of undisclosed costs, not by the Science Programme Committee on the basis of peer review — has left many in the astronomy community with no confidence in the decision-making processes at the agency’s highest levels (see public letter to ESA’s director of science signed by almost 300 scientists, at https://spicarebelalliance.com). ESA member states and collaborating countries have invested heavily in developing SPICA and expected a fair, transparent process for all competing projects. This cancellation was imposed without negotiation or communication with the SPICA team, and no details were given about the underlying costings. Why these suddenly became a problem after the completion of several comprehensive reviews is a mystery. The project team was given no opportunity to find a solution. Without transparency and accountability in the making of such decisions, nothing will prevent other ESA projects from experiencing a similar fate.

Just two years after their reintroduction as part of a bold conservation strategy, giant tortoises (Aldabrachelys gigantea) have hatched in the wild in Madagascar for the first time in around 600 years. This milestone in rewilding could provide insight into the structure and dynamics of Madagascar’s unique ecosystems, which were shaped by megafauna extirpated centuries ago. For us, some of the conservation biologists involved, it feels like a timetravel bonanza. Overexploited and driven to extinction in Madagascar after humans arrived on the island 1,500 years or so ago (see B. E. Crowley Quat. Sci. Rev. 29, 2591–2603; 2010), giant tortoises survived because they colonized remote islands in the Seychelles. With the support of the Madagascan government, 12 were released in 2018 into a secure nature reserve. Two hatchlings appeared there in 2019, followed by another 25 in October this year. Now in a nursery, these juveniles will be returned to the wild once their carapaces can protect them from predators. This type of innovative approach could help stop the catastrophic decline in the island’s biodiversity, particularly in a changing climate (see also B. B. N. Strassburg et al. Nature 586, 724–729; 2020).

Stephen Serjeant The Open University, UK.

Education in developing countries with patchy Internet coverage is particularly hardhit by the COVID-19 pandemic (see Nature 585, 482; 2020), compromising the future of students unable to access online teaching. A United Nations’ resolution emphasizes access to the Internet as a means of bridging the digital divide and facilitating the fundamental human right to education (see go.nature.com/2kcjp1p). In Pakistan, for example, nationwide Internet availability must be accelerated if the country’s potential is not to be irreversibly compromised. Take the remote mountainous Gilgit-Baltistan region, which has a record of high literacy. This will plummet without proper Internet connectivity because local schools can no longer teach. University students returning home in lockdown cannot access their institutional online classes. In December 2019, the government started the Digital Pakistan initiative (DPI) to prioritize “access and connectivity”. After a year, an unacceptable proportion of the population is still without electricity, let alone broadband. A further blow to Pakistan’s education and research has been dealt by the government’s continual axing of the country’s Higher Education Commission (HEC) budget. As a researcher from Pakistan, I urge the government and the leadership of the DPI and the HEC to speed up plans to rectify this digital divide so that education can flourish again (see also Nature 582, 162–164; 2020).

Shoko Jin SRON Netherlands Institute for Space Research, the Netherlands.

Aziz Khan Stanford University, California, USA. [email protected]

Ivonne Torres-Atencio University of Panama, Panama. Amador Goodridge INDICASATAIP, City of Knowledge, Panama. [email protected] Silvio Vega Hospital Doctor Arnulfo Arias Madrid, Panama.

David L. Clements Imperial College London, UK. [email protected]

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

548 | Nature | Vol 587 | 26 November 2020

Miguel Pedrono CIRAD, University of Montpellier, France. [email protected] Elodi Rambeloson Anjajavy Nature Reserve, Madagascar. Alison Clausen UNESCO, Paris, France.

Expert insight into current research

News & views

Perhaps one of the biggest scientific challenges of our time is explaining how intelligent behaviour arises in both natural and artificial systems. Resolving this question will have practical applications. For natural systems, it could allow us to describe and correct behavioural disorders with unprecedented precision. For artificial systems, it would enable safe distribution of agents that will enhance many aspects of our lives, from controlling self-driving cars to fighting misinformation. Many parallels can be drawn between the two system types, but there are also many differences. For instance, unlike a typical artificial system, the mammalian brain contains organized networks of tight reciprocal connections between two distinct components — the

thalamus and the cortex. These two components have different internal structures: neurons in the cortex are highly interconnected, whereas thalamic neurons are not. In artificial systems, recurrent neural networks can produce short-term memory patterns5. The cortex, at some level of abstraction, can be considered as a collection of recurrent networks that handles different types of short-term memory. So the question arises: why is a thalamus needed in the midst of all of this? An understanding of the molecular mechanisms that regulate thalamocortical circuits might help us to tackle this question. But identifying genes associated with cognitive processes is hard, because genetic mapping requires many repeated measurements, which can be difficult to obtain from behavioural studies. Hsiao et al. used an innovative approach to overcome this obstacle, making use of a method called quantitative trait locus (QTL) analysis that can link traits (such as eye colour, height or propensity to develop a given disease) to specific locations in the genome, or even to specific genes6. The team tested the working memory of mice using a simple behavioural task — a maze test, in which the animals explored arms of a T-shaped maze at will. If they chose to explore arms they had not previously visited, they passed the test, whereas if they returned to familiar arms, they failed. The authors found that performance varied between mouse strains, which they reasoned might be partly

b Delay

c Choice

Cognition

Deep brain control of memory Michael M. Halassa

An innovative approach has been used to link genetics to behaviour in mice. The analysis reveals that the gene Gpr12 underpins the role of the brain’s thalamus region in maintaining short-term memory. The brain’s thalamus has historically been thought of as a relay centre that transmits sensory and motor inputs to the cortex for processing, or that transmits information from one part of the cortex to another. In 2017, three groups made the unexpected discovery that the thalamus also has a key role in short-term memory — specifically, in maintaining the recurrent patterns of cortical activity that underlie memory1–3. However, the genetic basis of this role for the thalamus remained unexplored. Writing in Cell, Hsiao et al.4 reveal that the gene Gpr12 is key to thalamic maintenance of short-term memory. Their findings will have relevance for many fields, from cognitive therapeutics to artificial intelligence. a Sample

MDT

Prefrontal cortex

Signalling

High Gpr12 Food

Gate

Low Gpr12

Figure 1 | A gene involved in short-term memory. Hsiao et al.4 report that variability in expression of the gene Gpr12 in the thalamus of the mouse brain leads to variability in how well animals can keep short-term memory patterns in mind. The authors verified this finding using a working-memory task. a, Animals were placed in a T-shaped maze. In the initial sample phase of the test, one arm was gated off at random, allowing the animals to enter the other arm. b, During the delay between the first and second parts of the task, reciprocal

signalling between the brain’s prefrontal cortex and mediolateral thalamus (MDT) becomes synchronized. c, The animals had been trained to know that, in the second part of the task, they could retrieve a food reward by visiting the previously unexplored arm. Those that expressed high levels of the gene Gpr12 in the MDT were good at remembering which arm they had visited, and so choosing the arm that contained the reward. By contrast, those with low Gpr12 expression performed poorly.

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

Nature | Vol 587 | 26 November 2020 | 549

News & views explained by the ability of individual animals to keep previous actions in mind as short-term memory patterns. The researchers performed QTL analysis, and found one genetic region that stood out as different between the various strains of mice; they named this region Smart1 (short for spontaneous T-maze alternation QTL 1). In particular, animals that had one particular DNA sequence at Smart1 (dubbed Smart1CAST) were especially good at the exploratory task, and those with another (Smart1B6) were especially poor. Having identified this region, Hsiao and colleagues confirmed their findings from the high-throughput behavioural test using a similar but more-complex maze assay designed to test spatial working memory. In this assay, which used fewer animals, mice had to remember which arm of a maze they had visited on a first visit, and choose to visit the other arm to get a reward on a second visit (Fig. 1). Again, Smart1CAST and Smart1B6 animals performed better or worse, respectively, than the group as a whole. Next, Hsiao et al. examined gene-expression patterns across several brain regions in these two mouse strains. The most significant differences between the two were in the mediodorsal thalamus, in expression of a gene called Gpr12 that is located in Smart1. This brain region is strongly connected to the pre­ frontal cortex, which is involved in higher-level cognitive functions such as working memory. The authors found that reducing expression of Gpr12 led to poorer task performance in Smart1CAST mice, whereas overexpressing the gene improved the performance of Smart1B6 animals. Gpr12 encodes a protein belonging to a family known as orphan receptors, in which no ligand molecule that binds to and activates each receptor has been identified. Gpr12 probably enhances the activity of mediodorsal thalamus neurons once they are engaged by external inputs (such as those from the prefrontal cortex). Indeed, Hsiao et al. found that patterns of neuronal activity in the medio­ dorsal thalamus became much more in-sync with those in the prefrontal cortex during those parts of the maze test when animals were presumably remembering where they had been on the previous maze run. Hsiao and colleagues’ work provides key evidence to reinforce the conclusions of the 2017 papers1–3. Their findings also indicate that coordinated thalamocortical activity patterns depend on the version of Smart1 present: the more Gpr12 is expressed from this region, the more thalamocortical coordination occurs and the better the performance of spatial working memory. The discovery of this role for Gpr12 could lead to the development of pharmacological agents that boost working-memory

both. It is exciting to think about the many possibilities ahead as we continue to draw biological inspiration from innovative work such as that of Hsiao and colleagues. Michael M. Halassa is at the McGovern Institute and in the Department of Brain and Cognitive Sciences, Massachusetts Institute of Technology, Cambridge, Massachusetts 02139, USA. e-mail: [email protected]

Bolkan, S. S. et al. Nature Neurosci. 20, 987–996 (2017). Guo, Z. V. et al. Nature 545, 181–186 (2017). Schmitt, L. I. et al. Nature 545, 219–223 (2017). Hsiao, K. et al. Cell 183, 522–536 (2020). Barak, O. Curr. Opin. Neurobiol. 46, 1–6 (2017). Miles, C. M. & Wayne, M. Nature Educ. 1, 208 (2008). Akrami, A., Kopec, C. D., Diamond, M. E. & Brody, C. D. Nature 554, 368–372 (2018). 8. Wimmer, R. et al. Nature 526, 705–709 (2015). 9. Chakraborty, S., Kolling, N., Walton, M. E. & Mitchell, A. S. eLife 5, e13588 (2016). 10. Rikhye, R. V., Gilra, A. & Halassa, M. M. Nature Neurosci. 21, 1753–1763 (2018).

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

This article was published online on 16 November 2020.

Solar physics

Neutrino detection gets to the core of the Sun Gabriel D. Orebi Gann

The first detection of neutrinos produced by the Sun’s secondary solar-fusion cycle paves the way for a detailed understanding of the structure of the Sun and of the formation of massive stars. See p.577 On page 577, the Borexino Collaboration1 reports results that blast past a milestone in neutrino physics. They have detected solar neutrinos produced by a cycle of nuclear-­ fusion reactions known as the carbon–nitrogen–oxygen (CNO) cycle. Measurements of these neutrinos have the potential to resolve uncertainties about the composition of the solar core, and offer crucial insights into the formation of heavy stars. Neutrinos are tiny, subatomic particles. They were first postulated to exist by Wolfgang Pauli in 1930, to account for the energy that was apparently missing during β-decay, a process in which energetic electrons are emitted from an atomic nucleus. The presence of a massless particle that could carry any fraction of the energy from the decay would explain why the spectrum of emitted electron energies is continuous. Pauli’s explanation for why neutrinos had never been observed was that

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

550 | Nature | Vol 587 | 26 November 2020

performance. However, it would be important to first determine the types of cortical activity pattern that are enhanced by thalamic Gpr12. For example, in tasks in which animals have to withhold actions while remembering a task-relevant piece of information2,7,8, would we see the same type of effect? It is also intriguing to speculate on what other types of cognitive function could be linked to genetic underpinnings using a QTL approach. The mediodorsal thalamus is known to be involved in switching between tasks9,10; could one find a simple and scalable behavioural test that could be used to assess this process and probe its genetic underpinnings? Finally, to return to the comparison between natural and artificial systems, is the lack of a thalamus-like architecture in most artificial models of intelligence a missed opportunity? On the one hand, artificial recurrent neural networks require no such structure to maintain memory patterns or switch them across tasks. On the other, perhaps incorporating this biological inspiration into artificial-­intelligence systems would enable us to expand their computational capabilities, power efficiency or

they interact incredibly weakly with matter. Subsequent decades of research have yielded a wealth of information about Pauli’s ‘ghost particle’, including the Nobel-prizewinning discovery that neutrinos do, in fact, have a mass2–4, albeit one so small as to be beyond the reach of current measurements. Fusion reactions in the Sun produce an astonishing number of neutrinos: roughly 100 billion solar neutrinos pass through each of your thumbnails every second. Because of the weakness of their interactions, they are barely deterred from their path even when they have to pass through the entire body of the Earth: cutting-edge experiments5 ­(see also go.nature.com/​36sktyj) have struggled to observe a difference in the measured neutrino flux between daytime and night-time, owing to the vanishingly small scale of this effect. Neutrinos are therefore both challenging to observe and yet able to offer insights into

REF. 1

Figure 1 | The Borexino neutrino detector. The Borexino experiment detects light produced when solar neutrinos scatter off electrons in a large vat of liquid scintillator — a medium that produces light in response to the passage of charged particles. The Borexino Collaboration wrapped the detector in thermal insulation to control temperature variations in the detector. This helped the team to take the highly precise measurements needed to detect solar neutrinos produced by the Sun’s secondary solar-fusion cycle1.

other­wise unreachable regions of the Universe, such as distant supernovae or the interiors of stars. Energy produced in the centre of the Sun in the form of photons takes tens of thousands of years to escape, but a solar neutrino can escape the Sun and reach Earth in just eight minutes. This gives us a unique window into the core of this blazing star. The Sun is powered by fusion reactions that occur in its core: in the intense heat of this highly pressurized environment, protons fuse together to form helium. This occurs in two distinct cycles of nuclear reactions. The first is called the proton–proton chain (or pp chain), and dominates energy production in stars the size of our Sun. The second is the CNO cycle, which accounts for roughly 1% of solar power, but dominates energy production in heavier stars6. The first experiment to detect solar neutrinos was carried out using a detector in Homestake Mine, South Dakota. This used measurements of pp-chain solar neutrinos to probe the Standard Solar Model (SSM), which describes nuclear fusion in the Sun. The surprising result from this experiment was that only approximately one-third as many neutrinos of the expected type (flavour) were detected7. A decades-long campaign of experiments followed, seeking to resolve this ‘solar neutrino problem’. Nobel-prizewinning results from the Sudbury Neutrino Observatory in

Ontario, Canada, eventually explained the deficit: the neutrinos were changing flavour between their production and detection3. The Borexino experiment at the Gran Sasso National Laboratory in Italy followed up this result with a full spectral analysis of neutrinos from many stages of the pp chain8. This analysis finally allowed the field to come full circle, re­opening the possibility of using solar neutrinos as a means of probing the Sun’s interior. The Borexino Collaboration now reports another groundbreaking achievement from its experiment: the first detection of neutrinos from the CNO cycle. This result is a huge leap forward, offering the chance to resolve the mystery of the elemental composition of the Sun’s core. In astrophysics, any element heavier than helium is termed a metal. The exact metal content (the metallicity) of a star’s core affects the rate of the CNO cycle. This, in turn, influences the temperature and density profile — and thus the evolution — of the star, as well as the opacity of its outer layers. The metallicity and opacity of the Sun affect the speed of sound waves propagating through its volume. For decades, helio­seismological measurements were in agreement with SSM predictions for the speed of sound in the Sun, giving confidence in that model. However, more-recent spectroscopic measurements of solar opacity produced results that were significantly lower than previously thought,

leading to discrepancies with the helio­ seismological data9. Precise measurements of CNO-cycle neutrinos offer the only independent handle by which to investigate this difference. Such measurements would also shed further light on stellar evolution. The chief obstacles to making these measurements are the low energy and flux of CNO neutrinos, and the difficulty of separating the neutrino signal from sources of background signals, such as radioactive-decay processes. The Borexino experiment detects light produced when solar neutrinos scatter off electrons in a large vat of liquid scintillator — a medium that produces light in response to the passage of charged particles. A precise measurement of the energy and time profile of the detected light allows the scintillation caused by solar neutrinos to be differentiated from light resulting from other sources, such as radio­ active contamination in the scintillator itself and in surrounding detector components. The Borexino Collaboration carried out a multi-year purification campaign to ensure unprecedentedly low levels of radioactive contaminants in the scintillator. Even so, minor convection currents caused by temperature variations allowed radioactive contaminants to diffuse from the outer edges of the detector. The researchers mitigated this effect by establishing exquisitely fine control of thermal variations in the detector (Fig. 1), thus allowing them to achieve the extremely challenging

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

Nature | Vol 587 | 26 November 2020 | 551

News & views feat of detecting CNO neutrinos. The resulting measurements are not yet precise enough to resolve the question of solar metallicity, but they offer a path towards this objective. Future experiments will seek to improve on the precision achieved by Borexino, by developing innovative methods to identify and reject background noise caused by radioactive contamination. In the meantime, the Borexino Collaboration’s tremendous accomplishment moves us closer to a complete understanding of our Sun, and of the formation of massive stars, and is likely to define the goal in this field for years to come.

Gabriel D. Orebi Gann is in the Department of Physics, University of California, Berkeley, Berkeley, California 94720, USA, and at the Lawrence Berkeley National Laboratory. e-mail: [email protected] The Borexino Collaboration. Nature 587, 577–582 (2020). Ahmad, Q. R. Phys. Rev. Lett. 87, 071301 (2001). Ahmad, Q. R. Phys. Rev. Lett. 89, 011301 (2002). Ashie, Y. et al. Phys. Rev. D 71, 112005 (2005). Aharmim, B. et al. Phys. Rev. C 72, 055502 (2005). Adelberger, E. G. et al. Rev. Mod. Phys. 83, 195 (2011). Davis, R. Jr, Harmer, D. S. & Hoffman, K. C. Phys. Rev. Lett. 20, 1205 (1968). 8. Agostini, M. et al. Nature 562, 505–510 (2018). 9. Asplund, M., Grevesse, N., Sauval, A. J. & Scott, P. Annu. Rev. Astron. Astrophys. 47, 481–522 (2009). 1. 2. 3. 4. 5. 6. 7.

Genetics

Neanderthal DNA raises risk of severe COVID Yang Luo

A genetic analysis reveals that some people who have severe reactions to the SARS-CoV-2 virus inherited certain sections of their DNA from Neanderthals. However, our ancestors can’t take all the blame for how someone responds to the virus. See p.610 A key part of tackling COVID-19 is understanding why some people experience more-severe symptoms than do others. Earlier this year, a segment of DNA 50,000 nucleotides long (corresponding to 0.002% of the human genome) was found to have a strong association with severe COVID-19 infection and hospitalization1. Zeberg and Pääbo2 report on page 610 that this region is inherited from Neanderthals. Their results not only shed light on one reason that some people are more susceptible to severe disease, but also provide insights into human evolutionary biology. DNA sequences that are physically close to one other in the genome are often inherited (linked) together. These blocks of DNA, known as haplotypes, therefore contain tightly linked variants — DNA sequences or nucleotides that vary between individuals in a population. For example, the COVID-19 risk haplotype described earlier this year1 harbours variants across its entire 50,000-nucleotide span that are inherited together more than 98% of the time. Long haplotypes such as this could be a result of positive selection, maintained in our genomes because they contributed to our species’ chances of survival and reproductive success. They could also be introduced as a result of interbreeding with archaic hominin species such as the Denisovans and Neanderthals. Some 1–4% of the modern human genome comes from these ancient relatives3. Many

35% 10 1 0

Figure 1 | Uneven global spread of a genetic risk factor for COVID-19. Zeberg and Pääbo2 report that a long sequence of DNA that is associated with severe COVID-19 infection and hospitalization is derived from Neanderthals. The sequence is unevenly distributed across modern human populations. This map shows the frequency at which the risk factor is found in various populations from around the world. The sequencing data for these populations were gathered by the 1000 Genomes Project10. (Adapted from Fig. 3 of ref. 2.)

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

552 | Nature | Vol 587 | 26 November 2020

of the surviving archaic genes are harmful to modern humans, and are associated with infertility and an increased risk of disease4. But a few are beneficial. Examples include the Denisovan-like version of a gene called EPAS1 that helps modern Tibetans to cope with life at extremely high altitudes5, a Neanderthal gene that increases our sensitivity to pain6 and others that help us fend off viruses7.

To investigate whether the COVID-19 risk haplotype might have been introduced from our ancient relatives, Zeberg and Pääbo compared the region with an online database of archaic genomes from around the world. They found the region to be closely related to that in the genome of a Neanderthal individual that lived in modern-day Croatia around 50,000 years ago, but it was not related to any known Denisovan genomes. The authors next checked the prevalence of the Neanderthal-derived haplotype in the modern human population. They report that it is rare or completely absent in east Asians and Africans. Among Latin Americans and Europeans, the risk haplotype is maintained at a modest frequency (4% and 8%, respectively). By contrast, the haplotype occurs at a frequency of 30% in individuals who have south Asian ancestry, reaching as high as 37% in those with Bangladeshi heritage (Fig. 1). The researchers therefore speculate that the Neanderthal-derived haplotype is a substantial contributor to COVID-19 risk in specific groups. Their hypothesis is supported by hospital data8 from the Office for National Statistics in the United Kingdom, which indicates that individuals of Bangladeshi origin in the country are twice as likely to die from COVID-19 as are members of the general population (although other risk factors will, of course, contribute to these statistics). Why has this haplotype been retained in some populations? The authors posit that it might be protective against other ancient pathogens, and therefore positively selected for in certain populations around the world9. But when individuals are infected with the SARS-CoV-2 coronavirus, the protective immune response mediated by these ancient genes might be overly aggressive, leading

to the potentially fatal immune response observed in people who develop severe COVID-19 symptoms. As a result, a haplotype that at times in our past might have been beneficial for survival could now be having an adverse effect. Despite the correlation between this risk haplotype and clinical outcomes, genetics alone do not determine a person’s risk of developing severe COVID-19. Our genes and their origins clearly influence the development and progression of COVID-19 (and other infectious diseases), but environmental factors also have key roles in disease outcomes. For example, although the Neanderthalderived risk haplotype is almost completely absent in people with African ancestry, this population has a higher COVID-19 mortality rate than do people of other ethnic backgrounds, even after adjusting for geography and socio-economic factors (see go.nature. com/3jcxezx (‘Demographics’ tab) and go. nature.com/2h4qfqu, for example). Social inequality and its repercussions seem likely to account for a larger proportion of the risk of COVID-19 death than does Neanderthal-derived DNA. It is fascinating to think that our ancestor’s genetic legacy might be playing a part in the current pandemic. However, the underlying impact of the inherited DNA on the body’s response to the virus is unclear. Ongoing global efforts to study associations between our genetics and COVID-19 by analysing more individuals from diverse populations, such as that being undertaken by the COVID-19 Host Genetics Initiative (www.covid19hg.org), will help us to develop a better understanding of the disease’s aetiology. It is important to acknowledge that, although genes involved in the COVID-19 response might be inherited, social factors and behaviours (such as social distancing and mask wearing) are in our control, and can effectively reduce the risk of infection. Yang Luo is in the Department of Genetics, Harvard Medical School, Boston, Massachusetts 02115, USA. e-mail: [email protected]

1. The Severe Covid-19 GWAS Group. N. Engl. J. Med. 383, 1522–1534 (2020). 2. Zeberg, H. & Pääbo, S. Nature 587, 610–612 (2020). 3. Green, R. E. et al. Science 328, 710–722 (2010). 4. Sankararaman, S. et al. Nature 507, 354–357 (2014). 5. Huerta-Sánchez, E. et al. Nature 512, 194–197 (2014). 6. Zeberg, H. et al. Curr. Biol. 30, 3465–3469 (2020). 7. Enard, D. & Petrov, D. A. Cell 175, 360–371 (2018). 8. Public Health England. Disparities in the Risk and Outcomes of COVID-19 (PHE Publs, 2020). 9. Browning, S. R., Browning, B. L., Zhou, Y., Tucci, S. & Akey, J. M. Cell 173, 53–61 (2018). 10. The 1000 Genomes Project Consortium. Nature 526, 68–74 (2015). This article was published online on 26 October 2020.

Biochemistry

Isoforms combine for diverse signalling Joshua C. Snyder & Sudarshan Rajagopal

Many receptor proteins of the GPCR family exist in multiple isoforms. A comprehensive analysis of different combinations of GPCR isoforms that produce diverse signalling patterns in cells has implications for drug development. See p.650 With more than 800 members1, the G-proteincoupled receptor (GPCR) superfamily is the largest family of cell-surface receptor proteins in humans. GPCRs trigger intra­cellular signalling pathways in response to activation by extracellular factors. In doing so, they determine how a cell responds to and interacts with its environment, thereby influencing nearly every aspect of physiology. As such, they are excellent drug targets — at least 475 drugs approved by the US Food and Drug Administration (FDA) are aimed at GPCRs2. But many GPCRs exist in multiple isoforms, or variants, complicating attempts to find drugs that can bind to them. On page 650, Marti-Solano et al.3 describe a catalogue of the structure and expression of GPCR isoforms in humans. This resource has been added to a GPCR database, called GPCRdb, and is already openly available to the scientific community4 (https://gpcrdb.org/protein/isoforms). One common hurdle when attempting to design drugs that control GPCR signalling is that the same GPCR can activate multiple intracellular signalling pathways5. Pharmacologically altering the receptor’s activity can therefore lead to unforeseen side effects. Drugs called biased agonists that target just one pathway downstream of GPCRs have shown great promise6,7. However, they are effective in only some cases — perhaps because the genes that encode GPCRs can be processed in different ways during transcription, producing multiple versions of the final messenger RNA, called splice variants. Through this splicing mechanism, specific domains can be excluded from a GPCR or atypical ones added, producing a range of isoforms. Each one might preferentially activate alternative downstream signalling pathways. So far, our understanding of this key aspect of GPCR biology has been limited to studies of a few isoforms in unnatural settings8,9. Marti-Solano and colleagues set out to determine how the presence of various isoforms affects the signalling of around 350  GPCRs across tissues of the human

body. First, they made use of information about GPCR structures and DNA sequences from GPCRdb to help them identify candidate GPCRs in a database called GTex — a catalogue of gene expression in human tissues. This produced a list of 625 GPCR isoforms, with 38% of GPCRs having more than one. The group then systematically organized these GPCR isoforms according to their topology. They developed a set of ‘structural finger­prints’ for GPCR isoforms, based on the specific extracellular, intracellular and transmembrane domains present in each one (Fig. 1a). The most common structural fingerprints preserved GPCR topology, and the most frequent changes were seen only in the protein’s extracellular amino terminus or intracellular carboxy terminus. The N-terminal alterations typically caused changes in the binding of ligand molecules or efficacy. By contrast, C-terminal alterations led to changes in the ability of the receptor to couple with other receptor monomers, or in alterations in the internalization or transport of receptors through the cell inside vesicles — all of which are key to downstream signalling. The authors also found a few truncated isoforms, in which transmembrane domains were eliminated. They propose that these decrease receptor signalling. The truncated isoforms might be expressed only inside the cell, where they bind to more-complete versions — isoforms internalized in this way are unable to signal. Next, to model the potential tissue-specific effects of different isoforms, Marti-Solano et al. generated tissue-expression signatures — maps of the expression of each isoform for each receptor across 30 tissues. This revealed different combinations across tissues. The authors confirmed that co-expressing various combinations of isoforms of a given receptor in cells in culture resulted in different patterns of downstream signalling (Fig. 1b). It is not surprising that isoforms have different signalling properties. Nonetheless, the demonstration that co-expression of different isoforms alters

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

Nature | Vol 587 | 26 November 2020 | 553

News & views a Structural fingerprints

b Differential expression

Structural segment

Isoform

1

1 234

Kidney

Lung

16 A

B

B

C

2 3 Signalling pathways A+B

Signalling pathways B+C

Differential drug responses

Figure 1 | Cataloguing G-protein coupled receptor (GPCR) isoforms. Marti-Solano et al.3 analysed 625 GPCR isoforms expressed across 30 human tissues, and show that there is often more than one isoform of the same receptor in a tissue. a, The authors catalogued structural variation between GPCR isoforms by generating structural fingerprints — descriptions of the structural segments included in each isoform (dark circles indicate segments that are part of a given isoform; empty circles indicate segments that are missing). This simplified schematic shows three isoforms for one imagined GPCR; many more are possible (dashed arrow). b, The group shows that these isoforms are expressed in different combinations across tissues. Each combination might activate different downstream signalling pathways, and so respond differently to drugs.

signalling suggests a broad mechanism for generating ‘systems bias’10, in which various tissue-expression signatures promote differential activation of intracellular signalling pathways. The authors next checked that the tissue-expression signatures they observed truly reflected co-expression of multiple isoforms of a receptor, rather than expression of different isoforms in different cell types within a tissue. They analysed isoform-level expression in various cell lines, as well as data from single-cell RNA sequencing. These assays confirmed that single cells expressed multiple receptor isoforms. Lastly, Marti-Solano et al. showed that 42% of the 111 GPCRs that are targets of FDA-approved drugs had more than one isoform — and that in many cases each of the isoforms for a given receptor has a different tissue distribution. The authors also found that specific single-nucleotide DNA mutations in some were associated with disease. This finding suggests that isoform-selective drugs might be useful for treating human diseases. The search for these drugs recalls the long and continuing process of developing subtype-selective drugs for many GPCRs11 (drugs that modulate just one of the 13 GPCRs activated by the chemical serotonin, for instance). The current finding indicates that drugs might need to be both subtype- and isoform-selective. Time will tell whether thoughtful design of isoform-selective drugs will lead to increased specificity and fewer off-target effects than occur with current drugs. It is important to note the limitations inherent in this study. Isoform expression could be assessed only at the level of gene expression, whereas the gold standard in receptor biology is to measure protein expression using

“Time will tell whether thoughtful design of isoform-selective drugs will lead to increased specificity and fewer off-target effects.” cell-specific. Their behaviour might depend on spatial localization of receptors, on cell-specific cofactors, or on the isoforms’ ability to control cell-intrinsic responses to an external microenvironment. As such, an in vitro setting might not fully reveal how each isoform would act in vivo. It will also be of interest to determine whether plasticity in isoform expression serves as a mechanism by which to dynamically regulate system-level responses. Could tissue-expression signatures change over time or in response to signals from other regions of the body, enabling a tissue to respond to the same signal in different ways under different conditions? A related avenue for future research will be to systematically determine how splicing affects the expression and activity of the protein ligands that bind to GPCRs. New ligand

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

554 | Nature | Vol 587 | 26 November 2020

an approach called radioligand binding. However, radioligands might not differentiate between receptor isoforms. Furthermore, Marti-Solano et al. did analyse mass spectro­ metry data to confirm protein expression for some isoforms. Another caveat is that the authors’ experi­ ments on how isoforms act in combination involved expressing the proteins in an atypical setting, in cells in culture. However, the biology of these isoforms is complex and

isoforms can arise in cancer, as a result of gene fusions12. For example, the expression of these ‘oncogenic fusion ligands’ leads to changes in the stem-cell microenvironment of the colon that enable the spread of precancerous stem cells13. It seems likely that splice isoforms of GPCRs could alter system-level responses in disease, separate from or together with ligand fusions. Going forward, the same possibility could be investigated for all transmembrane receptors, because biased signalling has also been described for receptor tyrosine kinases14 — these receptors also have a range of isoforms that have roles in health and disease. For example, splice isoforms of the tyrosine kinase ERBB2 are drivers of breast and lung cancers15,16. Finally, we need to consider whether we should update our approach to GPCR drug development. The current study clearly indicates that tissue-specific expression of receptor isoforms could complicate attempts to target a given GPCR. This is especially true when trying to design a biased agonist that blocks only one downstream signalling pathway — such drugs might have different effects in different tissues. A better understanding of the system bias induced by different isoform combinations will be needed to overcome this obstacle. Experiments should move beyond cell lines to analyse systems bias in cells taken directly from tissues, or in vivo. Marti-Solano and colleagues’ valuable resource will be essential for informing such studies. Joshua C. Snyder is in the Departments of Surgery and Cell Biology, Duke University School of Medicine, Durham, North Carolina 27710, USA. Sudarshan Rajagopal is in the Departments of Medicine and Biochemistry, Duke University School of Medicine. e-mail: [email protected]

1. Fredriksson, R., Lagerström, M. C., Lundin, L.-G. & Schiöth, H. B. Mol. Pharmacol. 63, 1256–1272 (2003). 2. Hauser, A. S., Attwood, M. M., Rask-Andersen, M., Schiöth, H. B. & Gloriam, D. E. Nature Rev. Drug Discov. 16, 829–842 (2017). 3. Marti-Solano, M. et al. Nature 587, 650–656 (2020). 4. Munk, C. et al. Br. J. Pharmacol. 173, 2195–2207 (2016). 5. Galandrin, S. & Bouvier, M. Mol. Pharmacol. 70, 1575–1584 (2006). 6. Slosky, L. M. et al. Cell 181, 1364–1379 (2020). 7. Smith, J. S., Lefkowitz, R. J. & Rajagopal, S. Nature Rev. Drug Discov. 17, 243–260 (2018). 8. Kendall, R. T. & Senogles, S. E. Neuropharmacology 60, 336–342 (2011). 9. Smith, J. S. et al. Mol. Pharmacol. 92, 136–150 (2017). 10. Kenakin, T. & Christopoulos, A. Nature Rev. Drug Discov. 12, 205–216 (2013). 11. Lindsley, C. W. et al. Chem. Rev. 116, 6707–6741 (2016). 12. Seshagiri, S. et al. Nature 488, 660–664 (2012). 13. Boone, P. G. et al. Nature Commun. 10, 5490 (2019). 14. Ho, C. C. M. et al. Cell 168, 1041–1052 (2017). 15. Turpin, J. et al. Oncogene 35, 6053–6064 (2016). 16. Smith, H. W. et al. Proc. Natl Acad. Sci. USA 117, 20139–20148 (2020). This article was published online on 4 November 2020.

Review

Fibrosis: from mechanisms to medicines https://doi.org/10.1038/s41586-020-2938-9

Neil C. Henderson1,2, Florian Rieder3,4 & Thomas A. Wynn5 ✉

Received: 4 May 2020 Accepted: 14 September 2020 Published online: 25 November 2020 Check for updates

Fibrosis can affect any organ and is responsible for up to 45% of all deaths in the industrialized world. It has long been thought to be relentlessly progressive and irreversible, but both preclinical models and clinical trials in various organ systems have shown that fibrosis is a highly dynamic process. This has clear implications for therapeutic interventions that are designed to capitalize on this inherent plasticity. However, despite substantial progress in our understanding of the pathobiology of fibrosis, a translational gap remains between the identification of putative antifibrotic targets and conversion of this knowledge into effective treatments in humans. Here we discuss the transformative experimental strategies that are being leveraged to dissect the key cellular and molecular mechanisms that regulate fibrosis, and the translational approaches that are enabling the emergence of precision medicine-based therapies for patients with fibrosis.

Fibrosis is not a disease but rather an outcome of the tissue repair response that becomes dysregulated following many types of tissue injury, most notably during chronic inflammatory disorders. The formation of fibrotic tissue, which is defined by the excessive accumulation of extracellular matrix (ECM) components such as collagen and fibronectin, is in fact a normal and important phase of tissue repair in all organs. When tissues are injured, local tissue fibroblasts become activated and increase their contractility, secretion of inflammatory mediators, and synthesis of ECM components; together, these changes initiate the wound healing response. When damage is minor or non-repetitive, wound healing is efficient, resulting in only a transient increase in the deposition of ECM components and facilitating the restoration of functional tissue architecture. However, when the injury is repetitive or severe, ECM components continue to accumulate, which can lead to disruption of tissue architecture, organ dysfunction and ultimately organ failure. Notably, studies of tissue repair in embryonic and fetal mice and human fetal surgery have shown that before the onset of the wound inflammatory response, immature tissues are capable of scarless healing, suggesting that inflammation might be a cause of fibrosis1. However, in adult mammalian tissue, ageing, the response to invading microorganisms, and the changing character of the inflammatory response over time influence whether wound healing responses lead to progressive fibrosis or end in efficient repair. Genetics is also important; specific mutations and rare variants that are associated with fibrosis have revealed antifibrotic targets and core pathways that might be druggable. The genes involved include MUC5B in pulmonary fibrosis2, MYH7 in cardiac fibrosis3, and DMD in Duchenne muscular dystrophy-associated skeletal muscle fibrosis4. Such genetic alterations suggest the involvement of non-fibroblast cell types that act upstream of mesenchymal cell activation. These findings emphasize the importance of multicellular interactions in the pathogenesis of fibrosis. In this review, we provide an update on recent research into the mechanisms of fibrosis and discuss how this information is enabling the development of antifibrotic treatments.

Single-cell genomics of fibrosis Single-cell multi-omics approaches are transforming our understanding of disease pathogenesis across medicine, making it possible to study cell populations in health and disease at unprecedented resolution. This ‘resolution revolution’ allows the powerful unbiased exploration of cell states and types at single-cell level, resulting in unexpected insights into tissue biology and disease mechanisms (Fig. 1). These cutting-edge single-cell approaches have already been avidly adopted by the fibrosis research community to deepen our understanding of the complex, multicellular interplay that drives lung fibrosis5. Mesenchymal cells are the key source of pathological ECM deposition during lung fibrosis, which ultimately leads to architectural disruption and reduced lung function. Spatial transcriptional maps of the mouse lung mesenchyme have been generated by combining single-cell RNA sequencing (scRNA-seq) and signalling lineage reporters6. Each mesenchymal lineage demonstrated a distinct spatial address and transcriptome, in turn conferring distinct fibrotic niche regulatory functions across these mesenchymal subpopulations. Examples include mesenchymal cells in the alveolar niche that express Pdgfra and respond to Wnt signalling, and are critical for the growth and self-renewal of alveolar epithelial cells. By contrast, Axin2+ myofibrogenic progenitor cells preferentially generated pathologically deleterious myofibroblasts after lung injury6. Further studies using the mouse model of bleomycin-induced lung injury have also identified lung mesenchymal cell heterogeneity in both healthy and fibrotic mouse lungs7–9. The analysis of more than 70,000 cells of multiple lineages from eight lung explants from patients with pulmonary fibrosis (of varying aetiologies) and eight lung samples from healthy donors10 identified a distinct population of pro-fibrotic alveolar macrophages, which had previously been characterized in mice10,11, in the samples from patients with fibrosis. This study and others10,12 have suggested that alveolar type 2 (AT2) cells, which secrete pulmonary surfactant and serve as alveolar stem cells, have a pathological role. These results identified

University of Edinburgh Centre for Inflammation Research, The Queen’s Medical Research Institute, Edinburgh BioQuarter, Edinburgh, UK. 2MRC Human Genetics Unit, Institute of Genetics and Molecular Medicine, University of Edinburgh, Edinburgh, UK. 3Department of Inflammation and Immunity, Lerner Research Institute, Cleveland Clinic Foundation, Cleveland, USA. 4Department of Gastroenterology, Hepatology and Nutrition, Digestive Diseases and Surgery Institute, Cleveland Clinic Foundation, Cleveland, Ohio, USA. 5Inflammation & Immunology Research Unit, Pfizer Worldwide Research, Development & Medical, Cambridge, MA, USA. ✉e-mail: [email protected] 1

Nature | Vol 587 | 26 November 2020 | 555

Review Integrated single-cell maps of organ fibrosis

Fibrotic organ system Multi-modal single-cell approaches • Transcriptome • Whole genome

Integration

• Epigenome • Cell ontogeny • Proteomics (cells and ECM) • Spatial transcriptomics

,GHQWLÀFDWLRQDQGLQWHUURJDWLRQ RIFHOOVXESRSXODWLRQVDQG DQWLÀEURWLFWDUJHWV

Fig. 1 | Deconvolving fibrosis using multi-modal single-cell approaches. Cutting-edge single-cell approaches are transforming our understanding of the complex cellular and molecular mechanisms that regulate fibrosis and making it possible to assess the transcriptome, genome, epigenome and proteome at a single-cell level, in addition to spatial profiling. Furthermore, combined readouts from the same single cell are now possible (for example, the simultaneous profiling of transcriptome and chromatin accessibilty), and

integration of these multi-modal single-cell omics readouts has allowed ever more powerful, comprehensive assessments of cell state, ontogeny, phenotype and function during human fibrotic disease. The new biological insights gained from these integrated approaches should enable the identification of novel and tractable therapeutic targets to treat patients with a broad range of fibrotic diseases.

a distinct population of AT2 cells in fibrotic lungs13 and established a direct mechanistic link between elevated TGFβ signalling induced by mechanical tension caused by impaired alveolar regeneration, and progressive lung fibrosis. Decreasing the mechanical tension on alveoli could be a therapeutic approach for treating progressive lung fibrosis13. A study that profiled 312,928 lung cells from 32 patients with idiopathic pulmonary fibrosis (IPF), 18 patients with chronic obstructive pulmonary disease (COPD) and 29 healthy control individuals identified a population of aberrant basaloid epithelial cells located at the edge of myofibroblast foci that were enriched in the lungs of patients with IPF14. Within the vascular endothelial cell compartment, samples from patients with IPF contained an expanded cell population that was transcriptomically identical to vascular endothelial cells that are normally restricted to the bronchial circulation. Furthermore, diffusion map and pseudotemporal trajectory analyses (computational techniques used in single-cell transcriptomics to determine the pattern of a dynamic process experienced by cells, and then to arrange cells according to their progression through the process) made it possible to infer the origins of activated myofibroblasts in IPF14. scRNA-seq has also been used to comprehensively profile the cellular and molecular landscape in liver homeostasis and regeneration15–18. Since their discovery as major collagen-producing cells in the liver19, hepatic stellate cells (HSCs) have been considered a homogenous population, with equal potential to transition to the activated, myofibroblast phenotype. However, scRNA-seq has shown that mouse HSCs can be divided into functional zones, allowing high-resolution identification of the critical pathogenic collagen-producing cells in livers with centrilobular injury20. Pseudotemporal trajectory and RNA velocity, another computational approach that predicts the future state of individual cells on a timescale, demonstrated that central vein-associated HSCs are the dominant source of pathogenic collagen-producing cells following centrilobular liver injury20. Furthermore, the use of scRNA-seq to interrogate retinol-positive myofibroblasts isolated from fibrotic mouse livers has also shown that liver myofibroblasts are heterogeneous and functionally diverse21.

The profiling of more than 100,000 human liver cells yielded molecular definitions for non-parenchymal cell types that are found in the healthy and cirrhotic human liver and identified a scar-associated subpopulation of macrophages that express triggering receptor expressed on a myeloid cell-2 (TREM2) and CD9, which expands in liver fibrosis, differentiates from circulating monocytes and is pro-fibrogenic22. Endothelial cell subpopulations that express disease-associated atypical chemokine receptor-1 (ACKR1) and plasmalemma vesicle-associated protein (PLVAP), which are topographically restricted to the fibrotic niche and enhance the transmigration of leucocytes, were also defined. Multi-lineage modelling23,24 of ligand–receptor interactions among the scar-associated macrophages, endothelial cells and PDGFRα+ collagen-producing mesenchymal cells revealed that several pro-fibrogenic pathways were active in scars, including TNF receptor superfamily (TNFRSF) 12A, PDGFR and NOTCH signalling, which provides a conceptual framework for the discovery of rational therapeutic targets in cirrhotic livers22. Macrophages associated with liver injury in mice show substantial overlap of marker genes with human scar-associated macrophages, including the expression of TREM2 and CD9 in both species22. Unbiased cross-species mapping of scRNA-seq data using canonical correlation analysis (CCA)25 confirmed that mouse and human scar-associated macrophages represent corollary populations. This shows that scRNA-seq approaches can be useful for defining ‘core’ fibrotic injury-induced populations and therapeutic targets across species, thereby increasing precision in the interrogation of putative targets across the translational pipeline, from preclinical rodent models to human liver primary cell or organoid-based systems. Although the treatment of liver fibrosis has met with disappointing failures over the past few years26, including late-stage readouts for Elafibranor, a dual PPARα/δ agonist (NCT02704403), and Selonsertib, an ASK1 inhibitor (NCT03053050), clinical studies of Ocaliva (obetacholic acid; NCT03836937), Cenicriviroc (CCR2/5 dual antagonist; NCT03028740), Aramchol (a fatty acid bile acid conjugate; NCT04104321), MGL-3196 (liver-directed, thyroid hormone receptor (THR) β-selective agonist; NCT03900429), granulocyte

556 | Nature | Vol 587 | 26 November 2020

colony-stimulating factor (GCSF; NCT03911037), and a combination study of diacylglycerol O-acyltransferase-2 (DGAT2) and acetyl-CoA carboxylase (ACC) inhibitors (NCT04321031) have shown promise and are undergoing further testing. In addition, omics and genetic analyses are beginning to better inform patient selection and stratification, and it is hoped that this will facilitate improved outcomes. In the gastrointestinal tract27, a single-cell census of the human colonic mesenchyme revealed four subsets of fibroblasts in addition to pericytes and myofibroblasts, and identified a fibroblast subpopulation proximal to the colonic crypt niche that expressed SOX6, F3 (also known as CD142), and WNT genes, which are essential for colonic epithelial stem cell function. In colitis, this niche became dysregulated; an activated mesenchymal population emerged that expressed TNFSF14, genes associated with fibroblastic reticular cells, IL33, and lysyl oxidase (LOX). These cells led to impaired epithelial proliferation and maturation, thus illustrating how the colonic mesenchyme remodels to drive inflammation and barrier dysfunction in inflammatory bowel disease (IBD)27. In the context of arthritis, deletion of fibroblasts expressing fibroblast activation protein-α (FAPα) suppressed inflammation and bone erosion in mouse models of resolving and persistent arthritis. Single-cell RNA-seq identified two anatomically distinct fibroblast subsets within the FAPα+ population: FAPα+ thymus cell antigen (THY1)+ immune ‘effector’ fibroblasts in the synovial sub-lining, and FAPα+THY1− ‘destructive’ fibroblasts restricted to the synovial lining layer. Adoptive transfer of FAPα+THY1− fibroblasts into the joint selectively mediated bone and cartilage damage with little effect on inflammation, whereas transfer of FAPα+ THY1+ fibroblasts resulted in a more severe and persistent inflammatory arthritis, with minimal effect on bone and cartilage. The discovery of these anatomically discrete, functionally distinct subsets of fibroblasts with non-overlapping functions has important implications for the rational design of therapies aimed at precisely modulating inflammation, fibrosis and tissue repair28. Single-cell RNA-seq studies are also beginning to shed new light on the mechanisms that regulate kidney injury and fibrosis29–31. For example, recent work using single-nucleus RNA-seq in a mouse model of acute kidney injury identified a distinct pro-inflammatory and pro-fibrotic proximal tubule cell state that fails to repair. Deconvolution of bulk RNA-seq data sets showed that this failed-repair proximal tubule cell (FR-PTC) state can be detected in other models of kidney injury, and that it increases during ageing in rat kidney and over time in human kidney allografts29. Furthermore, a recent study has used scRNA-seq to profile the transcriptomes of proximal and non-proximal tubule cells in healthy and fibrotic human kidneys, enabling mapping of all matrix-producing cells at high resolution. This revealed distinct subpopulations of pericytes and fibroblasts as the major cellular sources of scar-forming myofibroblasts during human kidney fibrosis. Genetic fate-tracing, time-course scRNA-seq and assay for transposase-accessible chromatin (ATAC)–seq experiments in mice, and spatial transcriptomics in human kidney fibrosis, were then used to functionally interrogate these findings, identifying Nkd2 as a myofibroblast-specific target in human kidney fibrosis30. Recent studies in which scRNA-seq has been used to investigate mechanisms of fibrosis in various organ systems are summarized in Supplementary Table 1.

Fibroblast heterogeneity and plasticity Functional fibroblast heterogeneity Increasingly sophisticated experimental approaches have revealed substantial diversity and functional heterogeneity within the fibroblast population during organ fibrosis27,28,32–35. A combination of fate mapping and live imaging showed that a specialized subset of fibroblasts, fascia fibroblasts, rise to the surface of the skin after wounding36. These fascia fibroblasts gather their surrounding ECM (including blood vessels, macrophages and peripheral nerves) to form the provisional matrix, and

ablation of these fibroblasts inhibits matrix homing into wounds and leads to defective scars. Notably, the placement of an impermeable film beneath the skin (preventing upward migration of fascia fibroblasts) led to chronic open wounds. Thus, the fascia contains a specialized prefabricated kit of sentry fibroblasts, which are embedded within a movable sealant. Whether similar fibroblast subpopulations exist in other organs and use analogous mechanisms to promote wound healing remains to be determined. There is also substantial functional diversity among myofibroblasts during skin injury and ageing37. Lineage tracing and flow cytometry identified distinct subsets of wound bed myofibroblasts, including CD26-expressing adipocyte precursors and a CD29high subpopulation. Wound beds in aged mice or in bleomycin-induced fibrotic mouse skin showed a decrease in adipocyte precursors and an increase in CD29high cells compared to young healthy mice, suggesting that the fibrotic microenvironment alters the composition and function of myofibroblasts. Senesecence, a state in which cells cease to divide, also influences the fate and function of fibroblasts. However, whether fibroblast senescence plays a benefical or detrimental role in inflammation, tissue repair or fibrosis remains unclear, and its effects may vary in different tissues and types of disease38–40. Some studies have suggested that senescent fibroblasts become resistant to apoptosis, thereby sustaining inflammation and fibrosis through their production of inflammatory cytokines, immune modulators, growth factors and proteases. Consequently, senotherapeutic and senolytic drugs have emerged as potential new treatments for fibrosis and related ageing-associated diseases39,41. Recent studies have implicated a range of mesenchymal progenitor cells (MPCs) in the initiation and propagation of fibrosis42,43. In particular, two studies, focusing on populations of MPCs expressing HIC1, PDGFRα and LY6A in the heart and skeletal muscle, have demonstrated a hierarchy of MPCs, diversity in the pathophysiological roles of their progeny, and how fate determination of MPCs is context dependent44,45. Conditional genetic inactivation of Hic1 in mice led to activation and expansion of MPCs in both heart and skeletal muscle, demonstrating that HIC1 is required for maintaining MPC quiescence. In the heart, HIC1 deficiency (in PDGFRα-expressing cells) led to activation of MPCs and accumulation of cardiac fibroadipogenic progenitor cells, with epicardial thickening, interstitial fibrosis and fibrofatty depositions resulting in pathological features that are pathognomonic of arrhythmogenic cardiomyopathy45. However, although inactivation of Hic1 in skeletal muscle also resulted in a marked expansion of PDGFRα+LY6A+ cells during homeostasis, this did not increase skeletal muscle fibrosis and had no substantial effect on skeletal muscle regeneration44, highlighting the diverse pathophysiological roles of these cells in different organs. Fibrosis, secondary to age-associated chronic low-grade inflammation, is increasingly recognized as an important cause of morbidity and mortality. Increasing age is a driver of functional heterogeneity in fibroblasts, with ‘old’ fibroblasts demonstrating variability in their ability to reprogram and heal wounds46. Old mice showed varying wound healing rates in vivo, and scRNA-seq identified distinct subpopulations of fibroblasts with differing cytokine expression profiles in the wounds of old mice with slow versus fast healing rates. This increased variability in wound healing with increasing age may reflect distinct stochastic ageing trajectories between individuals, which will need to be considered when designing personalized antifibrotic therapies for the elderly population46.

Fibroblast targeting and reprogramming Fibroblasts, as well as displaying substantial functional heterogeneity, are capable of remarkable plasticity and phenotype switching during the progression and regression of fibrosis43,47,48. For example, the transcription factor PU.1 (also known as SPI1) has a major role in fibroblast polarization and fibrogenesis49. PU.1 both polarizes resting fibroblasts and repolarizes ECM-degrading inflammatory fibroblasts to an ECM-producing fibrotic phenotype. Furthermore, inactivation Nature | Vol 587 | 26 November 2020 | 557

Review Fibroblast functional heterogeneity

Lineage switching Dermal adipocyte

FAPα+THY1+ immune ‘effector’ fibroblasts

Rheumatoid arthritis

FAPα+THY1– ‘destructive’ fibroblasts

Repolarization to resting fibroblasts

5HVWLQJÀEUREODVW

Matrix-secreting myofibroblast

Tissue-dependent functional diversity of HIC1+PDGFRα+LY6A+ MP progeny Fibrogenic

Cardiac muscle

5HSURJUDPPLQJWR KHSDWRF\WHV

Non-fibrogenic

Skeletal muscle

Reversion to quiescence in absence of ongoing injury

Fig. 2 | Functional heterogeneity and plasticity of fibroblasts. Fibroblast populations show substantial functional heterogeneity and plasticity during fibrosis. In the context of arthritis, scRNA-seq combined with adoptive transfer experiments has been used to identify two anatomically distinct fibroblast subsets within the FAPα+ population: FAPα+ thymus cell antigen (THY1)+ immune ‘effector’ fibroblasts located in the synovial sub-lining, and FAPα+THY1− ‘destructive’ fibroblasts that are restricted to the synovial lining layer. Studies of MPC populations (HIC1+PDGFRα+LY6A+) in heart and skeletal muscle have demonstrated a hierarchy of MPCs, diversity in the pathophysiological roles of their progeny, and how fate determination of MPCs is tissue-dependent.

Fibroblasts are also capable of remarkable plasticity and phenotype switching during the progression and regression of fibrosis. Myofibroblasts can revert to a quiescent state in the absence of ongoing injury, or undergo full lineage switching with adipocytes (as observed during cutaneous wound healing in mice). Furthermore, genetic and pharmacological inactivation of the transcription factor PU.1 can reprogram fibrotic fibroblasts into resting fibroblasts, resulting in the regression of fibrosis in several organs. Finally, viral vector-mediated expression of specific transcription factors in myofibroblasts in the liver has been used to reprogram myofibroblasts into hepatocyte-like cells in fibrotic mouse livers, thereby reducing liver fibrosis and increasing liver function.

of PU.1 enabled fibrotic fibroblasts to be reprogrammed into resting fibroblasts, leading to the regression of fibrosis49. There is also remarkable inter-lineage plasticity between myofibroblasts and other cell types during fibrosis. During cutaneous wound healing in mice, adipocytes were regenerated from myofibroblasts. This reprogramming required neogenic hair follicles, which triggered bone morphogenetic protein (BMP) signalling and the activation of adipocyte transcription factors that are expressed during development50. Furthermore, adipocytes were generated from human keloid fibroblasts when treated with BMP in vitro, or when placed with human hair follicles50. Viral vector-mediated expression of specific transcription factors in liver myofibroblasts has been used to reprogram myofibroblasts into hepatocyte-like cells in fibrotic mouse livers, thereby reducing liver fibrosis and increasing liver function51,52. The ability to selectively target scar-producing myofibroblasts during fibrosis and to reprogram these cells into other lineages that support organ function opens up exciting new avenues for antifibrotic and pro-regenerative therapies. Together, these studies highlight the profound diversity, functional heterogeneity and plasticity of fibroblasts during fibrosis, both within and between organs (Fig. 2). More precise delineation of fibroblast heterogeneity and phenotype in different disease settings should facilitate the design of rational, highly targeted antifibrotic therapies, ultimately allowing the specific inhibition, ablation, or reprogramming of pathological fibroblast subpopulations while preserving essential, homeostatic fibroblast function. For example, adoptive transfer into mice of CD8+ T cells expressing a chimaeric antigen receptor against FAP led to the selective ablation of pathogenic fibroblasts and a substantial decrease in cardiac fibrosis following injury53. Boosting natural killer (NK) cell responses by blocking the NK cell receptor NKG2A has also been proposed as a mechanism to eliminate senescent fibroblasts in

skin54. Finally, one of the first trials to explore fibroblast cellular therapy was a phase 2 study that used allogeneic human dermal fibroblasts to remodel contracted scars (NCT01564407; Supplementary Table 2).

558 | Nature | Vol 587 | 26 November 2020

The dynamic matrisome During wound healing, the ECM is critical for mechanically stabilizing injured tissue, immobilizing growth factors and acting as a scaffold for the migration of fibroblasts, immune cells and endothelial cells into areas of tissue injury and repair55. As such, the ECM is increasingly appreciated as a highly dynamic entity that can influence the progression and resolution of fibrosis via a range of mechanisms. The fibrotic matrix directly promotes myofibroblast activation through mechanotransduction pathways, which convert mechanical signals (changes in stiffness) into biochemical responses. For example, in a pig model of incisional skin wounding, mechanical loading of wounds upregulated the expression of genes associated with fibrosis, whereas mechanical offloading of these wounds reversed this effect56. The increased mechanical strain within the stiffened matrix also provides a direct mechanism for the conversion of latent TGFβ1 into its active form57. Finally, through the actions of matrix metalloproteinases (MMPs) and tissue inhibitors of metalloproteinases (TIMPs), myofibroblasts continuously regulate matrix deposition and turnover58,59. Advances in mass spectrometry, proteomics, and spatial proteomics60 should greatly accelerate our understanding of how changes within the matrisome itself maintain tissue fibrogenesis independently of inflammatory signals.

Metabolic regulation of mesenchymal cells It is now apparent that cells involved in the progression and resolution of fibrosis are metabolically ‘reprogrammed’ to perform distinct

Glucose

Extracellular matrix Myofibroblast

Mitochondria Acetyl-CoA ↓ pH

↑

↑ Glutaminase

Lactate

Latent TGFβ

α-KG

Glucose 6-P

+ Active TGFβ

↑

Glutamine

TCA

Hexokinase

Succinate

LDH

Glutamate

Pyruvate

TGFβR

↑α-SMA ↑ Collagen ↑ Proliferation

Nucleus Collagen ↑ Stabilization ↓ Apoptosis

Fig. 3 | Metabolomic reprogramming of activated fibroblasts. Pro-fibrotic fibroblasts increase glycolysis via hexokinase, which leads to increased pyruvate and lactate. Lactate decreases extracellular pH and activates latent TGFβ1. Pyruvate also feeds the tricarboxylic acid (TCA) cycle after conversion into acetyl-CoA and thereby increases succinate levels. Both mechanisms lead

to an increase in α-SMA, collagen production and cell proliferation. Glutaminase activity is increased, and this converts glutamate to glutamine, which is converted into α-ketoglutarate (α-KG) via the TCA cycle; this decreases apoptosis and enhances collagen stabilization. LDH, lactate hydrogenase; P, phosphate.

functions during tissue repair. The effects of metabolism have been explored extensively in the context of non-alcoholic steatohepatitis (NASH)-driven fibrosis, in which dysregulated hepatic lipid metabolism serves as a key driver of liver injury and cirrhosis61. As discussed earlier, fibroblasts are the key source of ECM deposition during fibrosis. Hence, metabolic alterations of local tissue mesenchymal cells may offer future therapeutic avenues that include the major carbohydrate, amino acid and lipid metabolism pathways (Fig. 3). Following tissue injury, mesenchymal cells undergo profound metabolic changes to facilitate energy-consuming cellular functions such as proliferation and protein synthesis62. In fibroblasts, aerobic glycolysis is increased by upregulating rate-limiting glycolytic enzymes62. As well as providing a rapid energy-generating mechanism compared to oxidative phosphorylation, glycolysis produces by-products such as lactate that regulate fibrosis. A reduction in extracellular pH, combined with an increase in lactic acid, promotes myofibroblast differentiation by activating TGFβ1, and lactate itself may serve as an additional source of energy for mesenchymal cells63,64. Activation of fibroblasts increases key glycolytic enzymes, such as hexokinase 2 and lactate dehydrogenase65, which in turn increase cell proliferation65 and collagen synthesis66. During enhanced glycolysis, increased amounts of pyruvate are converted into acetyl-CoA in the mitochondrial matrix65 before entering the citric acid cycle. This yields intermediate metabolites, such as succinate, which promote fibrosis67. Disregulated glycolysis has been implicated in experimental models of lung, liver and kidney fibrosis, and inhibition of glycolysis reduces ECM accumulation68–70. During fibrogenesis, mesenchymal cells also exploit changes in amino acid metabolism through glutaminolytic reprogramming. Glutaminolysis and levels of the key enzyme glutaminase are increased in

TGFβ1-stimulated fibroblasts71. This leads to enhanced conversion of glutamine to glutamate, which confers resistance to apoptosis72 and promotes the stabilization of collagen71 via mTOR signalling. In vivo, the inhibition of glutaminase 1 ameliorates bleomycin- and TGFβ1-induced pulmonary fibrosis73. Changes in fatty acid oxidation have also been linked to fibrogenesis. Intracellular fatty acid oxidation is downregulated in tubulointerstitial fibrosis in mice and humans, and its restoration protects against fibrosis74. It has been reported that glycolysis is upregulated to compensate for reduced fatty acid oxidation during kidney injury, which could result in enhanced progression to fibrosis70 (Fig. 3). Key metabolic pathways such as increased glycolysis, upregulation of glutaminolysis and enhanced fatty acid oxidation are emerging as important drivers of fibroblast activation. No drugs that target these metabolic pathways have yet reached the clinic as antifibrotic therapies. However, drugs with known safety profiles that target metabolic pathways have been approved or are in clinical trials for the treatment of cancer, fuelling hope that in the future they could also be used to treat fibrotic diseases75.

Macrophage-mediated regulation of fibrosis Inflammatory monocytes and tissue-resident macrophages are key regulators of tissue fibrosis, playing important roles in the initiation, maintenance and resolution of tissue injury76–78. Furthermore, monocytes and macrophages can undergo remarkable functional plasticity, displaying diverse phenotypes during wound healing that depend on multiple cues including the environmental niche79,80 and the temporal stage of tissue injury and repair81,82. Tissue macrophages are also Nature | Vol 587 | 26 November 2020 | 559

Review important producers of T cell- and fibroblast-recruiting chemokines, which orchestrate the development of the fibrotic niche83.

Functional heterogeneity of monocytes and macrophages Several studies have identified subpopulations of monocytes that can regulate fibrosis and tissue remodelling 22,79,84–87. For example, a population of atypical monocytes characterized by carcinoembryonic antigen-related adhesion molecule-1 (CEACAM1+ MSR1+LY6C−F4/80−MAC1+ monocytes)84, which have been termed segregated-nucleus-containing atypical monocytes (SatMs) and share granulocyte characteristics, have a key role in lung fibrogenesis. SatMs are regulated by CCAAT/enhancer binding protein β (C/EBPβ), and Cebpb deficiency leads to a complete lack of SatMs. Bleomycin-induced fibrosis, but not inflammation, was inhibited in chimaeric mice with Cebpb−/− haematopoietic cells, and adoptive transfer of SatMs into Cebpb−/− mice resulted in fibrosis. Notably, SatMs are derived from Ly6C−FcεRI+ granulocyte/macrophage progenitors, but not from macrophage/dendritic cell progenitors. Single-cell RNA-seq approaches were used to investigate macrophage heterogeneity and function in the context of lung fibrosis. Although the main tissue-resident macrophage populations have been intensively studied, much less is known about the role of interstitial macrophages in fibrosis. Two independent subpopulations of interstitial macrophages that are conserved across lung, fat, heart, and dermis have been identified: LYVE1loMHCIIhiCX3CR1hi (LYVE1loMHCIIhi) and LYVE1hiMHCIIloCX3CR1lo (LYVE1hiMHCIIlo) monocyte-derived interstitial macrophages. In a mouse model of inducible macrophage depletion (Slco2b1flox/DTR), the absence of LYVE1hiMHCIIlo interstitial macrophages exacerbates experimental lung fibrosis, thereby showing that two independent populations of interstitial macrophages coexist across tissues with conserved niche-dependent functional programs79. In addition, a pathological subgroup of transitional macrophages is required for the fibrotic response to injury in bleomycin-induced lung fibrosis. A computational approach that allows scRNA-seq data to be annotated by reference to bulk transcriptomes (SingleR) enabled macrophage subclustering and uncovered a disease-associated subpopulation with a transitional gene expression profile that is intermediate between monocyte-derived and alveolar macrophages. These CX3CR1+SIGLECF+ transitional macrophages localized to the fibrotic niche and were pro-fibrotic in vivo. This appears to be relevant to human disease, because human orthologues of genes expressed by these transitional macrophages were upregulated in samples from patients with IPF86. Research into the regulatory roles of monocytes and macrophages during tissue injury and repair has largely focused on blood-derived monocytes and macrophages. However, there is emerging evidence that resident cavity macrophages are also key contributors to fibrosis and tissue remodelling. For example, a reservoir of mature F4/80hiGATA6+ peritoneal cavity macrophages rapidly invades the liver via direct (avascular) recruitment across the mesothelium in response to sterile liver injury88. These recruited macrophages dismantle necrotic cell nuclei, releasing DNA and forming a cover across the site of injury88. Similarly, following myocardial infarction in mice, GATA6+ macrophages in mouse pericardial fluid invade the epicardium and lose GATA6 expression but maintain antifibrotic properties, and loss of this macrophage population enhances interstitial fibrosis after an ischaemic injury. GATA6+ macrophages are also found in human pericardial fluid, suggesting that this immune cardioprotective role for the pericardial tissue compartment may be relevant in human disease89. Macrophage and fibroblast cross-talk Irrespective of how monocytes and macrophages are recruited into areas of tissue injury, pro-fibrotic macrophages commonly coordinate scar formation through a range of interactions with fibroblasts90, which are the main cellular source of pathological ECM deposition during fibrosis22,37,55,91–93. For example, macrophage-derived amphiregulin 560 | Nature | Vol 587 | 26 November 2020

has recently been shown to induce the differentiation of mesenchymal stromal cells into myofibroblasts via integrin-αV-mediated activation of TGFβ94. Previous work has shown that proximity is crucial to allow cross-talk between macrophages and contractile fibroblasts37,92,93; however, until recently it remained unclear how proximity between these two cell types is established. In an elegant study, contracting fibroblasts were shown to generate deformation fields in fibrillar collagen matrix that provided far-reaching physical cues to macrophages95. Within the collagen deformation fields created by fibroblasts or actuated microneedles, macrophages migrated towards the source of the force from distances of several hundred micrometres, and the presence of a dynamic force source within the matrix was required to initiate and direct macrophage migration. Notably, and counter to traditional views on how macrophages migrate within fibrotic tissues, the authors proposed that macrophages mechanosense the velocity of local displacements of their substrate, allowing contractile fibroblasts to attract macrophages over distances that exceed the range of chemotactic gradients95.

Integrin-mediated activation of TGFβ Secreted TGFβ is a major pro-fibrogenic cytokine, and therefore potentially represents an attractive antifibrotic target. Sustained systemic inhibition of TGFβ1, however, has undesired effects including cardiac valve problems, and TGFβ1-knockout mice develop systemic autoimmunity96. This is relevant to all mucosal surfaces, especially the intestine, where TGFβ1 activity is believed to control tissue homeostasis97,98. In addition, pan-TGFβ1 blockade has been found to induce carcinogenesis, perhaps owing to the role of TGFβ1 as an anti-proliferative mediator for most epithelial cell types. Some clinical trials using antibody-based pan-TGFβ blockade (for example, Fresolimumab) or TGFβ1 blockade (for example, Metelimumab) were terminated because of dose-limiting adverse events. Thus, strategies to avoid these deleterious effects could involve choosing the correct magnitude or duration of inhibition, co-administering anti-inflammatory therapies, or inhibiting TGFβ1 at specific sites in the tissue by blocking integrins and other mediators that locally activate latent TGFβ1. The pericellular fibrotic matrix is a remarkably dynamic environment that exerts profound influences on cell behaviour, and many of the key cell–cell and cell–matrix interactions that regulate fibrosis are mediated by members of the integrin family (noncovalent α–β heterodimers with18 different α-subunits and 8 β-subunits, resulting in 24 known members in humans)99. Importantly, integrins can mediate the translation of spatially fixed extracellular signals into a wide variety of changes in cell behaviour, including alterations in cell adhesion, migration, proliferation, differentiation and apoptosis99,100. Of key relevance to fibrosis, integrins can also potentiate signals from soluble pro-fibrogenic growth factors such as TGFβ1. Nearly all TGFβ1 is secreted and bound to the ECM in a latent form, and therefore the majority of the regulation of TGFβ function during fibrosis depends on site-specific regulation of TGFβ activation, rather than its synthesis or secretion101. The most intensively studied mechanism for activation of TGFβ1 is the interaction of the TGFβ1 latent complex with the αv-containing subset of integrins. Specifically, the integrins αvβ1, αvβ3, αvβ5, αvβ6 and αvβ8 have all been shown to bind to an N-terminal fragment of the TGFβ1 gene product called the latency associated peptide (LAP), which forms a noncovalent complex with the active cytokine, preventing latent TGFβ from binding to its cognate receptors and inducing biological effects102–104. When a mechanical force is applied to the latent complex by contraction of αvβ6 integrin-expressing cells, the resultant conformational change leads to the release of active TGFβ1105–107. Notably, a recent study has shown that in contrast to this αvβ6-mediated mechanism of TGFβ activation, αvβ8-dependent activation of TGFβ can occur independently of actin-cytoskeletal force and does not require

the release of mature TGFβ108, further highlighting the complexity of αv integrin-mediated TGFβ activation. There are now abundant preclinical data across a range of fibrotic disease models demonstrating critical regulatory roles for αv-containing integrins expressed on various different cell lineages. Mice lacking the αvβ6 integrin are protected in mouse models of lung, kidney and biliary fibrosis103,109–111. This protection is secondary to local inhibition of TGFβ, and antibody-mediated inhibition of αvβ6-mediated TGFβ1 activation decreased lung fibrosis in preclinical models112,113. TGFβ1 activation by the αvβ8 integrin represents a further potential therapeutic target114,115. Conditional depletion of αvβ8 integrin in lung fibroblasts inhibited experimental airway fibrosis115, and, in mice genetically engineered to replace the mouse β8 subunit with its human orthologue, a blocking antibody against human αvβ8 blocked TGFβ1 activation and protected against allergic airway inflammation and remodelling induced by cigarette smoke116. Furthermore, depletion of the αv integrin subunit on mesenchymal cells also inhibited fibrosis in models of liver, lung and kidney fibrosis117. The depletion of αv integrins on hepatic myofibroblasts in Pdgfrb-Cre mice protected the mice against hepatic fibrosis, whereas global loss of β3, β5 or β6 integrins, or conditional loss of β8 integrins in myofibroblasts, did not; this highlights the context dependency of the regulation of fibrosis in different organs by the various αv-containing integrins. Pharmacological blockade of αv-containing integrins by a small-molecule inhibitor (CWHM 12) attenuated both liver and lung fibrosis, even when fibrosis was already established117. Tissue fibroblasts can express four αv-containing integrins—αvβ1, αvβ3, αvβ5 and αvβ8. Selective small-molecule inhibitors of αvβ1 have been used to investigate the role of this integrin, with studies demonstrating that αvβ1 blockade has an antifibrotic effect in models of lung and liver fibrosis102. Given the abundance of preclinical data, this remains a very active area of research and development in the fibrosis field, with multiple small-molecule and antibody-based approaches undergoing assessment in clinical trials, including inhibitors designed to selectively target multiple αv-containing integrins simultaneously. This includes phase 2 trials of inhibitors of αvβ6 (NCT01371305), αvβ1 and αvβ6 (NCT04072315), and αvβ1, αvβ3 and αvβ6 (NCT03949530), all in pulmonary fibrosis (Supplementary Table 2). Patient safety will be an important consideration in these trials, as Biogen recently terminated their trial of a selective anti-αvβ6 antibody in patients with IPF owing to safety concerns.

Cytokine-mediated regulation of fibrosis Other than TGFβ, several additional cytokines that are secreted from multiple cellular sources have been identified as triggers of fibrosis118. The pro-inflammatory cytokine interleukin 17A (IL-17A) can induce fibrosis in different organ systems, including the lung, liver, kidney, heart and skin119–124. In a study of bleomycin-induced pulmonary fibrosis, IL-17A produced by γδ and CD4+ T cells induced lung inflammation, neutrophil recruitment, and production of TGFβ120. Neutrophils and mast cells are also important sources of IL-17A125. Experiments with mice lacking IL-17A or its receptor IL-17RA, as well as therapeutic studies using IL-17A-neutralizing antibodies, confirmed that IL-17A signalling is involved in fibrosis in multiple tissues120–122,126–128. As well as promoting TGFβ production126, IL-17A increases and stabilizes the expression of TGFβRII on fibroblasts, thereby enhancing their sensitivity to TGFβ129. The TH17-associated cytokine IL-22 similarly enhances TGFβ signalling in fibroblasts125. TGFβ in turn induces the expression of IL-17A when produced concurrently with the pro-inflammatory cytokines IL-1, IL-6, or TNF120,130,131, suggesting that a feed-forward mechanism that involves acute-phase cytokines, IL-17A, and TGFβ is responsible for the development of fibrosis following acute injury119,120,132 (Fig. 4); IL-17A exhibits similar activity in animal studies and human cells133,134.

Tissue injury

Pro-inflammatory stimuli IL-1, IL-6, TNF

Integrin-mediated TGFβ activation

ROS Alarmins IL-25, IL-35, TSLP Neutrophil ILC2 IL-5

Macrophage IL-17A TGFβ

CD4+ TH2

Eosinophil IL-13

G ↑ TGFβR IL-11R IL-11 Fibroblast

Myofibroblast ECM

Fibrosis Fig. 4 | Divergent cytokine pathways drive fibrosis. The innate acute-phase pro-inflammatory cytokines IL-1, IL-6, and TNF, together with TGFβ, which are produced by macrophages, tissue fibroblasts, and other local cell populations, promote the development of IL-17-secreting cells. IL-17A potentiates neutrophil responses that contribute to tissue injury through the production of reactive oxygen species (ROS), while increasing the expression of TGFβ receptors on fibroblasts and thereby facilitating the production of ECM in response to TGFβ. TGFβ is a key driver of fibrosis that is produced and activated locally through integrin-mediated mechanisms. A second and distinct cytokine-mediated pathway that can promote fibrosis independently of TGFβ is the type 2 cytokine axis. Here, the alarmin cytokines IL-25, IL-35, and thymic stromal lymphopoietin (TSLP), secreted by epithelial cells and other damaged tissues, drive the expansion and activation of type 2 innate lymphoid cells (ILC2s) that secrete large amounts of IL-5 and IL-13. IL-5 in turn drives the recruitment and activation of local tissue eosinophils, which provide an additional source of type 2 cytokines and other pro-fibrotic mediators. IL-13, which is derived from eosinophils, CD4+ type 2 T helper (TH2) cells, and ILC2s, exhibits potent pro-fibrotic activity that is independent of TGFβ. Finally, the cytokine IL-11, which is produced by activated myofibroblasts, stimulates ECM production by myofibroblasts in response to multiple pro-fibrotic mediators, including TGFβ and type 2 cytokines.

Caspase 1, the NOD, LRR and pyrin domain-containing (NLRP) 3 inflammasome, and NFκB were identified as important upstream activators of the IL-17A–TGFβ axis131. The mechanisms responsible for the sustained activation of NFκB and NLRP3 inflammasome signalling remain unclear, although commensal microorganism stimulation of Toll-like receptors on myeloid cells and tissue fibroblasts has been hypothesized to be an important activating mechanism, with the resulting pro-inflammatory cytokine and chemokine production exacerbating inflammation and the progression of fibrosis135,136. Of note, stimulation of TLR4 or NFκB in hepatic stellate cells enhances TGFβ signalling by directly downregulating the TGFβ pseudoreceptor BMP and the activin membrane-bound inhibitor Bambi136. A related study showed that sustained activation of the NLRP3 inflammasome is associated with increased chemokine expression, recruitment of neutrophils and macrophages, and persistent production of IL-17A and TNF131. Thus, activation of the pro-fibrotic TGFβ signalling pathway is driven by several collaborating mechanisms, with IL-17A having a prominent role. Whereas the TGFβ superfamily of ligands are well-known drivers of fibrosis137, the type 2-associated cytokines IL-4 and IL-13 have also emerged as distinct but important inducers of fibrosis. Here, the fibrotic response is associated with predominant infiltration of Nature | Vol 587 | 26 November 2020 | 561

Review eosinophils and M2-like macrophages, rather than the neutrophil and M1-like monocyte/macrophage phenotype that characterizes the IL-1– IL-17A–TGFβ axis138. Also, instead of acute-phase cytokines serving as co-inducers, the alarmin cytokines thymic stromal lymphopoietin, IL-25, and IL-33 function as key initiators of type 2-dependent fibrosis by triggering the production of IL-4 and IL-13 in innate lymphoid cells, T cells, eosinophils, and other type 2-associated leukocytes139–142. Although IL-13 can induce and activate TGFβ in macrophages143, it may promote fibrosis independently of TGFβ144 in part by directly targeting stromal and parenchymal cells, including epithelial populations and collagen-producing myofibroblasts145. Mice deficient in IL-13, IL-4R, or IL-13Rβ1, as well as animals treated with neutralizing antibodies to IL-13 or IL-4R, show reduced fibrosis after many types of tissue injury146–150, confirming that type 2 cytokine signalling is critical in the progression of fibrosis (Fig. 4). The mechanisms that dictate whether the IL-1–IL-17A–TGFβ axis or the type 2 cytokine response dominates as the key driver of fibrosis remain unclear, although the type of cellular damage or duration of the injury are likely to be important. For example, studies with the commonly used ‘single hit’ bleomycin model of pulmonary fibrosis revealed a prominent role for the IL-1–IL-17A–TGFβ axis but little to no contribution for type 2 cytokines, despite substantial upregulation of IL-4 and IL-13 in the lungs120. Nevertheless, a modified version of this model in which bleomycin was injected intradermally rather than intratracheally over several weeks uncovered a substantial role for IL-4R signalling in the development of pulmonary fibrosis151. Different stimuli or types of injuries that lead to the preferential production of alarmin cytokines versus the activation of NF-κB and inflammasome signalling are also likely to have key roles. For example, integrin receptors that interact with the ECM preferentially activate TGFβ signalling and the production of IL-17A while antagonizing the production of type 2 cytokines152. Consistent with these observations, several studies have revealed substantial cross-regulation between IL-17A and IL-13120,153, with marked upregulation of the opposing pathway when one mechanism was targeted therapeutically146,152,154,155. Consequently, a successful antifibrotic strategy may need to target the dominant mechanism or reduce both pathways simultaneously. IL-11, a member of the IL-6/gp130 cytokine family, may be a promising target, as it was recently shown to integrate pro-fibrotic signals emanating from both pathways156–158. Not surprisingly, given the robust preclinical data, inhibitors of IL-13 alone (NCT01266135, NCT00987545, NCT00581997, NCT01872689 and NCT01629667) or a combination of IL-4 and IL-13 inhibitors (NCT02921971 and NCT01529853) have been tested in phase 2 trials for pulmonary fibrosis, skin keloids, and systemic sclerosis. Although the results so far have been mostly negative or mixed, Romilkimab (SAR156597), a bi-specific antibody against IL-4 and IL-13, did have a significant effect on modified Rodnan skin score in a 24-week study of diffuse cutaneous systemic sclerosis (NCT02921971). Adalimumab, an antibody against TNF, is also being tested in Dupuytren’s disease, a complex fibroproliferative disease of the hand (NCT03180957). Additional cytokine, chemokine or growth factor inhibitors in development for fibrosis in phase II or III trials are inhibitors of CCR2 and CCR5 (NCT02217475, NCT03028740, NCT03059446 and NCT02330549) for liver fibrosis and NASH, an inhibitor of IL-1 (NCT01538719) for systemic sclerosis, an inhibitor of IL-6 (NCT02453256) for scleroderma, an inhibitor of CCL2 (NCT00786201) for pulmonary fibrosis, and follistatin (an activin antagonist) for Beckers muscular dystrophy (NCT 01519349) (Supplementary Table 2).

Contribution of the microbiome to fibrosis Most human-associated microorganisms are found in the gut, and during homeostasis these microbial populations are essential for maintaining gut health. However, when the balance between healthy and pathogenic microorganisms shifts towards pathogenic subsets, disease 562 | Nature | Vol 587 | 26 November 2020

can ensue. Inflammatory bowel disease (Crohn’s disease or ulcerative colitis) represents a prototypical pathology in which dysbiosis is thought to be a key driver of disease pathogenesis, with evidence that there is a strong link between the microbiota and the development of fibrosis. For example, patients with Crohn’s disease carrying variants of the NOD2 gene, which encodes an intracellular pattern recognition receptor, are at increased risk of stricture formation, which is the major manifestation of intestinal fibrosis159. Furthermore, serologic antimicrobial antibodies are common in patients with Crohn’s disease and are associated with and predictive of intestinal strictures160,161, and almost all mouse models of intestinal fibrosis are influenced by the microbiota162. For instance, global deletion of the bacterial signalling adaptor molecule MyD88163 reduced intestinal fibrosis in a mouse model of Salmonella-induced colitis164. Microbiota-driven intestinal fibrosis may be mediated by induction of the IL-33 receptor ST2 on epithelial cells165 or by the pro-fibrotic action of TL1A166. On a cellular level, TLR2 or TLR4 ligands induce secretion of cytokines and chemokines from cultured intestinal myofibroblasts167. Interestingly, although human intestinal mesenchymal cells express multiple TLRs and NLRs, one study found that a pro-fibrogenic phenotype was triggered exclusively by flagellin, a broad activator of innate and adaptive immunity and a TLR5 ligand. This occurred in a TGFβ1-independent manner and via post-transcriptional regulation168. The role of myofibroblasts in directly sensing pathogen-associated molecular patterns in intestinal fibrosis has been confirmed in vivo, as selective deletion of MyD88 in cells expressing α-smooth muscle actin (α-SMA) ameliorated intestinal fibrosis168. Dysbiosis in the gut also influences liver fibrosis. Translocation of bacteria and their products across the intestinal barrier owing to intestinal barrier disruption is common in patients with chronic liver disease. Increased levels of the microorganism-derived ligand lipopolysaccaride (LPS) in the portal vein or translocation of whole bacteria or their products to the liver activates inflammation that leads to fibrosis169. Blocking TLR4 signalling in mice or reducing hepatic exposure to intestinal microorganisms by reducing microbial load with antibiotics ameliorates experimental liver fibrosis136. HSCs express all known human TLRs and respond to TLR4 ligands136,170, which also downregulate a TGFβ1 decoy receptor and thereby sensitize HSCs to the action of TGFβ1136. A comparable mechanism has been described in pancreatic fibrosis in rats171. TLR4 signalling includes an additional signalling adaptor, called TRIF (also known as TICAM1). Deletion of TRIF in a mouse model of diet-induced NASH reduced hepatic steatosis but increased hepatic fibrosis, and Trif−/− HSCs expressed higher levels of CXCL1 and C-C motif chemokine ligands in response to LPS, highlighting a potential mechanism for this unexpected effect172. Conversely, distinct gut microbiota may be hepatoprotective in liver fibrosis. ECM deposition in the liver was higher in germ-free mice than in conventionally housed mice173, and MyD88- and TRIF-deficient mice showed the same effect. In the kidney, pericytes (a myofibroblast precursor) activate a TLR2–TLR4–MyD88-dependent pro-inflammatory program in response to tissue injury174. The downstream kinase IRAK4 controls the conversion of pericytes into myofibroblasts in vitro, and pharmacological inhibition of MyD88 signalling with an IRAK4 inhibitor reduced fibrosis by attenuating tissue injury in vivo174. Global TLR4 knockout175 and a small-molecule inhibitor of MyD88 ameliorate renal fibrosis in mice176. In human systemic sclerosis, TLR4 and its co-receptors lymphocyte antigen 96 (MD2) and CD14 are overexpressed in lesional skin and chronic dermal LPS exposure leads to overexpression of TGFβ signature genes177. Several TLRs are promiscuous and can also sense damage-associated molecular patterns, lipids or ECM. For example, in lung fibrosis TLR4 and the glycosaminoglycan hyaluronan are important for type 2 alveolar epithelial cell renewal, which limits lung injury and fibrosis178. Notably, TLR4 was protective in the lung, as opposed to its pathogenic effect in gut and liver fibrosis136,167. Nasal polyposis is a disease characterized

by remodelling of the sinonasal mucosa. Short single-stranded DNA molecules (CpG oligonucleotides) can activate fibroblasts derived from patients with nasal polyposis via TLR9 stimulation, providing an additional example whereby multiple pattern recognition receptors, activated by distinct ligands, can contribute to aberrant wound healing and fibrosis. The pathophysiological relevance of TLR4 in inflammation has led to clinical trials of a TLR4 inhibitor for treating rheumatoid arthritis (NCT03241108). Furthermore, pentraxin 2 (PTX2), also known as serum amyloid protein 2, has demonstrated anti-inflammatory and antifibrotic properties in multiple preclinical fibrosis models179,180 and recombinant PTX2 (PRM-151) has entered phase II trials for pulmonary fibrosis and myelofibrosis (NCT02550873; NCT01981850; Supplementary Table 2).

Future directions Fibrosis is a major global healthcare burden. Consequently, the discovery of key therapeutic targets with high relevance to human fibrotic disease and the subsequent development of effective antifibrotic therapies directed against these targets continues to be a research priority. The only two drugs that have been approved in several countries so far for the treatment of a fibrotic disease are nintedanib and pirfenidone, both for patients with IPF181. Nintedanib also recently received approvals for the treatment of systemic sclerosis-associated interstitial lung disease and progressive fibrosing interstitial lung diseases. Thus, drug development in this important field remains limited, has been restricted to only one organ system, and continues to progress slowly. Single-cell genomics methodologies have already yielded new discoveries that would previously have been unattainable. This field continues to evolve rapidly, and emerging technologies are now able to measure multiple omic readouts (genomes, epigenomes, transcriptomes and proteomes) in single cells182–184. Spatially resolved molecular profiling is expanding our understanding of how these populations interact in situ185–187. The convergence and integration of these multi-modal single-cell technologies24,188, alongside global initiatives such as the Human Cell Atlas189, represent an extraordinary opportunity to decode the cellular and molecular mechanisms of fibrosis at unprecedented resolution, which should in turn help to drive a new era of precision medicine in the treatment of fibrotic disease. Novel therapeutics developed for one fibrotic disorder may be applicable to a wide range of fibrotic diseases because of the shared pathways across organs that are uncovered by this work. Drug repositioning efforts may also be assisted by these studies190. Despite impressive progress over the past few years in our understanding of the pathogenesis of fibrosis, multiple challenges need to be overcome to translate this information into effective antifibrotic therapies (Fig. 5). Prognostic animal models and ex vivo primary human tissue culture systems need to be developed that allow better translation of novel mechanisms from the bench to the bedside. Patient heterogeneity, together with the fact that fibrosis progression is typically slow, makes the selection of patients for clinical trials difficult26. Hence, accurate and validated predictors of fibrotic disease progression are needed to stratify patients into high-risk populations before their inclusion in trials. In fact, groups of human fibrotic diseases may be subdivided on a mechanistic basis using analysis of tissue samples. Subsets of patients could then be targeted with personalized antifibrotic therapies. Work in this area should be a research priority. At present, trial end-points are highly variable and often lack the sensitivity needed to predict favourable responses over a short period of time, which necessitates the inclusion of large numbers of patients in clinical trials. Consequently, end-points for fibrosis clinical trials continue to evolve and may require a more global approach involving scientists, industry leaders, patients and regulatory partners, as shown for liver fibrosis and intestinal fibrosis191,192. Ideally, non-invasive end-points

Challenge

Solution

Poor translatability from bench to bedside

Novel animal models and primary human tissue culture systems

Slow progression of disease and patient heterogeneity

Accurate and validated predictors of fibrotic disease courses

Variable and insensitive trial end-points; regulatory uncertainty

Global consensus on end-point development using state-of-the-art index methodology

High number of compounds versus low number of patients to test them in

Adaptive and innovative trial design; acceptance of ‘real-world’ evidence

Fig. 5 | Challenges and solutions in the translation of antifibrotic mechanisms into drugs. The red boxes on the left describe some of the major challenges in the development of antifibrotic drugs that have been described in this Review, with the arrows pointing to potential solutions in the blue boxes.

that better correlate with clinically meaningful outcomes are needed. Recent research in the field has been fueled by the discovery of robust biomarkers and cutting-edge imaging modalities such as PET imaging of collagen and molecular imaging of fibrosis193,194, which allows fast, non-invasive, and whole-organ-quantitative and longitudinal readouts of drug efficacy in antifibrotic clinical trials. Furthermore, combining molecular imaging of fibrosis193 with cutting-edge omic approaches, such as single-cell genomics182–184, could markedly improve patient diagnostics, staging, prognostication, stratification and cohort enrichment, which would in turn optimize clinical trial design and maximize the number of trials that could be run quickly and efficiently185–187. Innovative approaches to trial design are being developed that allow the incorporation of adaptive strategies and the use of ‘bucket’ trials that include patients with different types of fibrosis, as well as the inclusion of ‘real-world’ evidence into the regulatory approval process. Similar to the major advances seen in cancer therapy and the successful treatment of HIV and viral hepatitis, it is likely that we will see increasing numbers of clinical trials testing combinations of drugs to treat fibrosis, as fibrosis is increasingly recognized as a highly complex disorder, with multiple mechanisms collaborating to drive disease progression. These antifibrotic drug cocktails will probably target a variety of orthogonal mechanisms, including a range of receptors, signalling pathways, and cell types that have been shown to function as core drivers of fibrosis in multiple disease states. These multifaceted approaches should pave the way towards the delivery of effective antifibrotic therapies in the future.

Online content Any methods, additional references, Nature Research reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at https://doi.org/10.1038/s41586-020-2938-9. 1. 2.

3.

Eming, S. A., Martin, P. & Tomic-Canic, M. Wound repair and regeneration: mechanisms, signaling, and translation. Sci. Transl. Med. 6, 265sr6 (2014). Allen, R. J. et al. Genetic variants associated with susceptibility to idiopathic pulmonary fibrosis in people of European ancestry: a genome-wide association study. Lancet Respir. Med. 5, 869–880 (2017). Kim, H. Y. et al. Genotype-related clinical characteristics and myocardial fibrosis and their association with prognosis in hypertrophic cardiomyopathy. J. Clin. Med. 9, E1671 (2020).

Nature | Vol 587 | 26 November 2020 | 563

Review 4.

5. 6. 7. 8. 9. 10.

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

20. 21.

22.

23. 24. 25.

26. 27. 28.

29.

30. 31.

32. 33. 34.

35. 36.

37. 38.

Young, C. N. J. et al. Total absence of dystrophin expression exacerbates ectopic myofiber calcification and fibrosis and alters macrophage infiltration patterns. Am. J. Pathol. 190, 190–205 (2020). Schiller, H. B. et al. The Human Lung Cell Atlas: a high-resolution reference map of the human lung in health and disease. Am. J. Respir. Cell Mol. Biol. 61, 31–41 (2019). Zepp, J. A. et al. Distinct mesenchymal lineages and niches promote epithelial self-renewal and myofibrogenesis in the lung. Cell 170, 1134–1148 (2017). Xie, T. et al. Single-cell deconvolution of fibroblast heterogeneity in mouse pulmonary fibrosis. Cell Rep. 22, 3625–3640 (2018). Peyser, R. et al. Defining the activated fibroblast population in lung fibrosis using single-cell sequencing. Am. J. Respir. Cell Mol. Biol. 61, 74–85 (2019). Tsukui, T. et al. Collagen-producing lung cell atlas identifies multiple subsets with distinct localization and relevance to fibrosis. Nat. Commun. 11, 1920 (2020). Reyfman, P. A. et al. Single-cell transcriptomic analysis of human lung provides insights into the pathobiology of pulmonary fibrosis. Am. J. Respir. Crit. Care Med. 199, 1517–1536 (2019). Misharin, A. V. et al. Monocyte-derived alveolar macrophages drive lung fibrosis and persist in the lung over the life span. J. Exp. Med. 214, 2387–2404 (2017). Xu, Y. et al. Single-cell RNA sequencing identifies diverse roles of epithelial cells in idiopathic pulmonary fibrosis. JCI Insight 1, e90558 (2016). Wu, H. et al. Progressive pulmonary fibrosis is caused by elevated mechanical tension on alveolar stem cells. Cell 180, 107–121 (2020). Adams, T. S. et al. Single cell RNA-seq reveals ectopic and aberrant lung resident cell populations in idiopathic pulmonary fibrosis. Sci. Adv. 6, eaba1983 (2019). Halpern, K. B. et al. Single-cell spatial reconstruction reveals global division of labour in the mammalian liver. Nature 542, 352–356 (2017). MacParland, S. A. et al. Single cell RNA sequencing of human liver reveals distinct intrahepatic macrophage populations. Nat. Commun. 9, 4383 (2018). Aizarani, N. et al. A human liver cell atlas reveals heterogeneity and epithelial progenitors. Nature 572, 199–204 (2019). Halpern, K. B. et al. Paired-cell sequencing enables spatial gene expression mapping of liver endothelial cells. Nat. Biotechnol. 36, 962–970 (2018). Friedman, S. L., Roll, F. J., Boyles, J. & Bissell, D. M. Hepatic lipocytes: the principal collagen-producing cells of normal rat liver. Proc. Natl Acad. Sci. USA 82, 8681–8685 (1985). Dobie, R. et al. Single-cell transcriptomics uncovers zonation of function in the mesenchyme during liver fibrosis. Cell Rep. 29, 1832–1847 (2019). Krenkel, O., Hundertmark, J., Ritz, T. P., Weiskirchen, R. & Tacke, F. Single cell RNA sequencing identifies subsets of hepatic stellate cells and myofibroblasts in liver fibrosis. Cells 8, E503 (2019). Ramachandran, P. et al. Resolving the fibrotic niche of human liver cirrhosis at single-cell level. Nature 575, 512–518 (2019). This study dissected unanticipated aspects of the cellular and molecular basis of human liver fibrosis at a single-cell level, providing a framework for the discovery of rational therapeutic targets in liver cirrhosis. Vento-Tormo, R. et al. Single-cell reconstruction of the early maternal–fetal interface in humans. Nature 563, 347–353 (2018). Efremova, M. & Teichmann, S. A. Computational methods for single-cell omics across modalities. Nat. Methods 17, 14–17 (2020). Butler, A., Hoffman, P., Smibert, P., Papalexi, E. & Satija, R. Integrating single-cell transcriptomic data across different conditions, technologies, and species. Nat. Biotechnol. 36, 411–420 (2018). Ratziu, V. & Friedman, S. L. Why do so many NASH trials fail? Gastroenterology https://doi.org/10.1053/j.gastro.2020.05.046 (2020). Kinchen, J. et al. Structural remodeling of the human colonic mesenchyme in inflammatory bowel disease. Cell 175, 372–386 (2018). Croft, A. P. et al. Distinct fibroblast subsets drive inflammation and damage in arthritis. Nature 570, 246–251 (2019). This study uncovered anatomically discrete, functionally distinct subsets of fibroblasts in the context of arthritis. Kirita, Y., Wu, H., Uchimura, K., Wilson, P. C. & Humphreys, B. D. Cell profiling of mouse acute kidney injury reveals conserved cellular responses to injury. Proc. Natl Acad. Sci. USA 117, 15874–15883 (2020). Kuppe, C. et al. Decoding myofibroblast origins in human kidney fibrosis. Nature (in the press, 2020). Der, E. et al. Tubular cell and keratinocyte single-cell transcriptomics applied to lupus nephritis reveal type I IFN and fibrosis relevant pathways. Nat. Immunol. 20, 915–927 (2019). Driskell, R. R. et al. Distinct fibroblast lineages determine dermal architecture in skin development and repair. Nature 504, 277–281 (2013). Rinkevich, Y. et al. Skin fibrosis. Identification and isolation of a dermal lineage with intrinsic fibrogenic potential. Science 348, aaa2151 (2015). Guerrero-Juarez, C. F. et al. Single-cell analysis reveals fibroblast heterogeneity and myeloid-derived adipocyte progenitors in murine skin wounds. Nat. Commun. 10, 650 (2019). Bergmeier, V. et al. Identification of a myofibroblast-specific expression signature in skin wounds. Matrix Biol. 65, 59–74 (2018). Correa-Gallegos, D. et al. Patch repair of deep wounds by mobilized fascia. Nature 576, 287–292 (2019). This work identified a specialized subset of fibroblasts, fascia fibroblasts, which gather the surrounding ECM and then rise to the surface of the skin after wounding. Shook, B. A. et al. Myofibroblast proliferation and heterogeneity are supported by macrophages during skin repair. Science 362, eaar2971 (2018). Montero-Melendez, T. et al. Therapeutic senescence via GPCR activation in synovial fibroblasts facilitates resolution of arthritis. Nat. Commun. 11, 745 (2020).

564 | Nature | Vol 587 | 26 November 2020

39. Schafer, M. J., Haak, A. J., Tschumperlin, D. J. & LeBrasseur, N. K. Targeting senescent cells in fibrosis: pathology, paradox, and practical considerations. Curr. Rheumatol. Rep. 20, 3 (2018). 40. Amor, C. et al. Senolytic CAR T cells reverse senescence-associated pathologies. Nature 583, 127–132 (2020). 41. Hickson, L. J. et al. Senolytics decrease senescent cells in humans: preliminary report from a clinical trial of Dasatinib plus Quercetin in individuals with diabetic kidney disease. EBioMedicine 47, 446–456 (2019). 42. Schneider, R. K. et al. Gli1+ mesenchymal stromal cells are a key driver of bone marrow fibrosis and an important cellular therapeutic target. Cell Stem Cell 23, 308–309 (2018). 43. El Agha, E. et al. Two-way conversion between lipogenic and myogenic fibroblastic phenotypes marks the progression and resolution of lung fibrosis. Cell Stem Cell 20, 261–273 (2017). 44. Scott, R. W., Arostegui, M., Schweitzer, R., Rossi, F. M. V. & Underhill, T. M. Hic1 defines quiescent mesenchymal progenitor subpopulations with distinct functions and fates in skeletal muscle regeneration. Cell Stem Cell 25, 797–813 (2019). 45. Soliman, H. et al. Pathogenic potential of Hic1-expressing cardiac stromal progenitors. Cell Stem Cell 26, 459–461 (2020). 46. Mahmoudi, S. et al. Heterogeneity in old fibroblasts is linked to variability in reprogramming and wound healing. Nature 574, 553–558 (2019). 47. Kisseleva, T. et al. Myofibroblasts revert to an inactive phenotype during regression of liver fibrosis. Proc. Natl Acad. Sci. USA 109, 9448–9453 (2012). 48. Troeger, J. S. et al. Deactivation of hepatic stellate cells during liver fibrosis resolution in mice. Gastroenterology 143, 1073–1083 (2012). 49. Wohlfahrt, T. et al. PU.1 controls fibroblast polarization and tissue fibrosis. Nature 566, 344–349 (2019). 50. Plikus, M. V. et al. Regeneration of fat cells from myofibroblasts during wound healing. Science 355, 748–752 (2017). 51. Song, G. et al. Direct reprogramming of hepatic myofibroblasts into hepatocytes in vivo attenuates liver fibrosis. Cell Stem Cell 18, 797–808 (2016). 52. Rezvani, M. et al. In vivo hepatic reprogramming of myofibroblasts with AAV vectors as a therapeutic strategy for liver fibrosis. Cell Stem Cell 18, 809–816 (2016). 53. Aghajanian, H. et al. Targeting cardiac fibrosis with engineered T cells. Nature 573, 430–433 (2019). 54. Pereira, B. I. et al. Senescent cells evade immune clearance via HLA-E-mediated NK and CD8+ T cell inhibition. Nat. Commun. 10, 2387 (2019). 55. Pakshir, P. & Hinz, B. The big five in fibrosis: macrophages, myofibroblasts, matrix, mechanics, and miscommunication. Matrix Biol. 68-69, 81–93 (2018). 56. Januszyk, M. et al. Mechanical offloading of incisional wounds is associated with transcriptional downregulation of inflammatory pathways in a large animal model. Organogenesis 10, 186–193 (2014). 57. Froese, A. R. et al. Stretch-induced activation of transforming growth factor-β1 in pulmonary fibrosis. Am. J. Respir. Crit. Care Med. 194, 84–96 (2016). 58. Lindsey, M. L., Iyer, R. P., Jung, M., DeLeon-Pennell, K. Y. & Ma, Y. Matrix metalloproteinases as input and output signals for post-myocardial infarction remodeling. J. Mol. Cell. Cardiol. 91, 134–140 (2016). 59. Craig, V. J., Zhang, L., Hagood, J. S. & Owen, C. A. Matrix metalloproteinases as therapeutic targets for idiopathic pulmonary fibrosis. Am. J. Respir. Cell Mol. Biol. 53, 585–600 (2015). 60. Lundberg, E. & Borner, G. H. H. Spatial proteomics: a powerful discovery tool for cell biology. Nat. Rev. Mol. Cell Biol. 20, 285–302 (2019). 61. Schwabe, R. F., Tabas, I. & Pajvani, U. B. Mechanisms of fibrosis development in nonalcoholic steatohepatitis. Gastroenterology 158, 1913–1928 (2020). 62. Xie, N. et al. Glycolytic reprogramming in myofibroblast differentiation and lung fibrosis. Am. J. Respir. Crit. Care Med. 192, 1462–1474 (2015). 63. Faubert, B. et al. Lactate metabolism in human lung tumors. Cell 171, 358–371 (2017). 64. Kottmann, R. M. et al. Lactic acid is elevated in idiopathic pulmonary fibrosis and induces myofibroblast differentiation via pH-dependent activation of transforming growth factor-β. Am. J. Respir. Crit. Care Med. 186, 740–751 (2012). 65. Liu, G. & Summer, R. Cellular metabolism in lung health and disease. Annu. Rev. Physiol. 81, 403–428 (2019). 66. Nigdelioglu, R. et al. Transforming growth factor (TGF)-β promotes de novo serine synthesis for collagen production. J. Biol. Chem. 291, 27239–27251 (2016). 67. Park, S. Y., Le, C. T., Sung, K. Y., Choi, D. H. & Cho, E. H. Succinate induces hepatic fibrogenesis by promoting activation, proliferation, and migration, and inhibiting apoptosis of hepatic stellate cells. Biochem. Biophys. Res. Commun. 496, 673–678 (2018). 68. Lian, N. et al. Curcumin regulates cell fate and metabolism by inhibiting hedgehog signaling in hepatic stellate cells. Lab. Invest. 95, 790–803 (2015). 69. Ding, H. et al. Inhibiting aerobic glycolysis suppresses renal interstitial fibroblast activation and renal fibrosis. Am. J. Physiol. Renal Physiol. 313, F561–F575 (2017). 70. Wei, Q. et al. Glycolysis inhibitors suppress renal interstitial fibrosis via divergent effects on fibroblasts and tubular cells. Am. J. Physiol. Renal Physiol. 316, F1162–F1172 (2019). 71. Ge, J. et al. Glutaminolysis promotes collagen translation and stability via α-ketoglutarate-mediated mTOR activation and proline hydroxylation. Am. J. Respir. Cell Mol. Biol. 58, 378–390 (2018). 72. Bai, L. et al. Glutaminolysis epigenetically regulates antiapoptotic gene expression in idiopathic pulmonary fibrosis fibroblasts. Am. J. Respir. Cell Mol. Biol. 60, 49–57 (2019). 73. Cui, H. et al. Inhibition of glutaminase 1 attenuates experimental pulmonary fibrosis. Am. J. Respir. Cell Mol. Biol. 61, 492–500 (2019). 74. Kang, H. M. et al. Defective fatty acid oxidation in renal tubular epithelial cells has a key role in kidney fibrosis development. Nat. Med. 21, 37–46 (2015). This study elegantly links abnormal fatty acid oxidation to fibrogenesis. 75. Luengo, A., Gui, D. Y. & Vander Heiden, M. G. Targeting metabolism for cancer therapy. Cell Chem. Biol. 24, 1161–1180 (2017). 76. Wynn, T. A. & Vannella, K. M. Macrophages in tissue repair, regeneration, and fibrosis. Immunity 44, 450–462 (2016).

77. Vannella, K. M. & Wynn, T. A. Mechanisms of organ injury and repair by macrophages. Annu. Rev. Physiol. 79, 593–617 (2017). 78. Krenkel, O. & Tacke, F. Liver macrophages in tissue homeostasis and disease. Nat. Rev. Immunol. 17, 306–321 (2017). 79. Chakarov, S. et al. Two distinct interstitial macrophage populations coexist across tissues in specific subtissular niches. Science 363, eaau0964 (2019). 80. Guilliams, M., Thierry, G. R., Bonnardel, J. & Bajenoff, M. Establishment and maintenance of the macrophage niche. Immunity 52, 434–451 (2020). 81. Lavine, K. J. et al. Distinct macrophage lineages contribute to disparate patterns of cardiac recovery and remodeling in the neonatal and adult heart. Proc. Natl Acad. Sci. USA 111, 16029–16034 (2014). 82. Duffield, J. S. et al. Selective depletion of macrophages reveals distinct, opposing roles during liver injury and repair. J. Clin. Invest. 115, 56–65 (2005). 83. Borthwick, L. A. et al. Macrophages are critical to the maintenance of IL-13-dependent lung inflammation and fibrosis. Mucosal Immunol. 9, 38–55 (2016). 84. Satoh, T. et al. Identification of an atypical monocyte and committed progenitor involved in fibrosis. Nature 541, 96–101 (2017). 85. Bajpai, G. et al. The human heart contains distinct macrophage subsets with divergent origins and functions. Nat. Med. 24, 1234–1245 (2018). 86. Aran, D. et al. Reference-based analysis of lung single-cell sequencing reveals a transitional profibrotic macrophage. Nat. Immunol. 20, 163–172 (2019). 87. Dick, S. A. et al. Self-renewing resident cardiac macrophages limit adverse remodeling following myocardial infarction. Nat. Immunol. 20, 29–39 (2019). 88. Wang, J. & Kubes, P. A reservoir of mature cavity macrophages that can rapidly invade visceral organs to affect tissue repair. Cell 165, 668–678 (2016). 89. Deniset, J. F. et al. Gata6+ pericardial cavity macrophages relocate to the injured heart and prevent cardiac fibrosis. Immunity 51, 131–140 (2019). 90. Adler, M. et al. Principles of cell circuits for tissue repair and fibrosis. iScience 23, 100841 (2020). 91. Henderson, N. C. et al. Galectin-3 expression and secretion links macrophages to the promotion of renal fibrosis. Am. J. Pathol. 172, 288–298 (2008). 92. Zhou, X. et al. Circuit design features of a stable two-cell system. Cell 172, 744–757 (2018). 93. Lodyga, M. et al. Cadherin-11-mediated adhesion of macrophages to myofibroblasts establishes a profibrotic niche of active TGFβ. Sci. Signal. 12, eaao3469 (2019). 94. Minutti, C. M. et al. A macrophage-pericyte axis directs tissue restoration via amphiregulin-induced transforming growth factor beta activation. Immunity 50, 645–654 (2019). 95. Pakshir, P. et al. Dynamic fibroblast contractions attract remote macrophages in fibrillar collagen matrix. Nat. Commun. 10, 1850 (2019). 96. Diebold, R. J. et al. Early-onset multifocal inflammation in the transforming growth factor beta 1-null mouse is lymphocyte mediated. Proc. Natl Acad. Sci. USA 92, 12215–12219 (1995). 97. McEntee, C. P., Gunaltay, S. & Travis, M. A. Regulation of barrier immunity and homeostasis by integrin-mediated transforming growth factor β activation. Immunology 160, 139–148 (2020). 98. Kelly, A. et al. Human monocytes and macrophages regulate immune tolerance via integrin αvβ8-mediated TGFβ activation. J. Exp. Med. 215, 2725–2736 (2018). 99. Barczyk, M., Carracedo, S. & Gullberg, D. Integrins. Cell Tissue Res. 339, 269–280 (2010). 100. Hynes, R. O. Integrins: bidirectional, allosteric signaling machines. Cell 110, 673–687 (2002). 101. Robertson, I. B. & Rifkin, D. B. Regulation of the bioavailability of TGFβ and TGFβ-related proteins. Cold Spring Harb. Perspect. Biol. 8, a021907 (2016). 102. Reed, N. I. et al. The αvβ1 integrin plays a critical in vivo role in tissue fibrosis. Sci. Transl. Med. 7, 288ra79 (2015). 103. Munger, J. S. et al. A mechanism for regulating pulmonary inflammation and fibrosis: the integrin αvβ6 binds and activates latent TGFβ1. Cell 96, 319–328 (1999). 104. Wipff, P. J., Rifkin, D. B., Meister, J. J. & Hinz, B. Myofibroblast contraction activates latent TGFβ1 from the extracellular matrix. J. Cell Biol. 179, 1311–1323 (2007). 105. Shi, M. et al. Latent TGFβ structure and activation. Nature 474, 343–349 (2011). 106. Dong, X. et al. Force interacts with macromolecular structure in activation of TGFβ. Nature 542, 55–59 (2017). 107. Dong, X., Hudson, N. E., Lu, C. & Springer, T. A. Structural determinants of integrin β-subunit specificity for latent TGFβ. Nat. Struct. Mol. Biol. 21, 1091–1096 (2014). 108. Campbell, M. G. et al. Cryo-EM reveals integrin-mediated TGFβ activation without release from latent TGFβ. Cell 180, 490–501 (2020). 109. Hahm, K. et al. αvβ6 integrin regulates renal fibrosis and inflammation in Alport mouse. Am. J. Pathol. 170, 110–125 (2007). 110. Wang, B. et al. Role of αvβ6 integrin in acute biliary fibrosis. Hepatology 46, 1404–1412 (2007). 111. Peng, Z. W. et al. Integrin αvβ6 critically regulates hepatic progenitor cell function and promotes ductular reaction, fibrosis, and tumorigenesis. Hepatology 63, 217–232 (2016). 112. Horan, G. S. et al. Partial inhibition of integrin αvβ6 prevents pulmonary fibrosis without exacerbating inflammation. Am. J. Respir. Crit. Care Med. 177, 56–65 (2008). 113. Puthawala, K. et al. Inhibition of integrin αvβ6, an activator of latent transforming growth factor-β, prevents radiation-induced lung fibrosis. Am. J. Respir. Crit. Care Med. 177, 82–90 (2008). 114. Araya, J. et al. Squamous metaplasia amplifies pathologic epithelial-mesenchymal interactions in COPD patients. J. Clin. Invest. 117, 3551–3562 (2007). 115. Kitamura, H. et al. Mouse and human lung fibroblasts regulate dendritic cell trafficking, airway inflammation, and fibrosis through integrin αvβ8-mediated activation of TGFβ. J. Clin. Invest. 121, 2863–2875 (2011). 116. Minagawa, S. et al. Selective targeting of TGFβ activation to treat fibroinflammatory airway disease. Sci. Transl. Med. 6, 241ra79 (2014). 117. Henderson, N. C. et al. Targeting of αv integrin identifies a core molecular pathway that regulates fibrosis in several organs. Nat. Med. 19, 1617–1624 (2013).

118. Barron, L. & Wynn, T. A. Fibrosis is regulated by Th2 and Th17 responses and by dynamic interactions between fibroblasts and macrophages. Am. J. Physiol. Gastrointest. Liver Physiol. 300, G723–G728 (2011). 119. Park, M. J. et al. IL-1–IL-17 signaling axis contributes to fibrosis and inflammation in two different murine models of systemic sclerosis. Front. Immunol. 9, 1611 (2018). 120. Wilson, M. S. et al. Bleomycin and IL-1β-mediated pulmonary fibrosis is IL-17A dependent. J. Exp. Med. 207, 535–552 (2010). 121. Wang, B. Z. et al. Interleukin-17A antagonist attenuates radiation-induced lung injuries in mice. Exp. Lung Res. 40, 77–85 (2014). 122. Meng, F. et al. Interleukin-17 signaling in inflammatory, Kupffer cells, and hepatic stellate cells exacerbates liver fibrosis in mice. Gastroenterology 143, 765–776 (2012). 123. Sun, B. et al. Role of interleukin 17 in TGFβ signaling-mediated renal interstitial fibrosis. Cytokine 106, 80–88 (2018). 124. Feng, W. et al. IL-17 induces myocardial fibrosis and enhances RANKL/OPG and MMP/ TIMP signaling in isoproterenol-induced heart failure. Exp. Mol. Pathol. 87, 212–218 (2009). 125. Fabre, T. et al. Type 3 cytokines IL-17A and IL-22 drive TGFβ-dependent liver fibrosis. Sci. Immunol. 3, eaar7754 (2018). 126. Tan, Z. et al. IL-17A plays a critical role in the pathogenesis of liver fibrosis through hepatic stellate cell activation. J. Immunol. 191, 1835–1844 (2013). 127. Zhang, S. et al. Neutralization of interleukin-17 attenuates cholestatic liver fibrosis in mice. Scand. J. Immunol. 83, 102–108 (2016). 128. Zhang, X. W. et al. Antagonism of interleukin-17A ameliorates experimental hepatic fibrosis by restoring the IL-10/STAT3-suppressed autophagy in hepatocytes. Oncotarget 8, 9922–9934 (2017). 129. Fabre, T., Kared, H., Friedman, S. L. & Shoukry, N. H. IL-17A enhances the expression of profibrotic genes through upregulation of the TGFβ receptor on hepatic stellate cells in a JNK-dependent manner. J. Immunol. 193, 3925–3933 (2014). 130. Oh, K. et al. Epithelial transglutaminase 2 is needed for T cell interleukin-17 production and subsequent pulmonary inflammation and fibrosis in bleomycin-treated mice. J. Exp. Med. 208, 1707–1719 (2011). 131. Wree, A. et al. NLRP3 inflammasome driven liver injury and fibrosis: roles of IL-17 and TNF in mice. Hepatology 67, 736–749 (2018). 132. Gasse, P. et al. IL-1 and IL-23 mediate early IL-17A production in pulmonary inflammation leading to late fibrosis. PLoS One 6, e23185 (2011). 133. Lemmers, A. et al. The interleukin-17 pathway is involved in human alcoholic liver disease. Hepatology 49, 646–657 (2009). 134. Macek Jilkova, Z. et al. Progression of fibrosis in patients with chronic viral hepatitis is associated with IL-17+ neutrophils. Liver Int. 36, 1116–1124 (2016). 135. Yang, D. et al. Dysregulated lung commensal bacteria drive interleukin-17b production to promote pulmonary fibrosis through their outer membrane vesicles. Immunity 50, 692–706 (2019). 136. Seki, E. et al. TLR4 enhances TGFβ signaling and hepatic fibrosis. Nat. Med. 13, 1324–1332 (2007). 137. de Kretser, D. M. et al. Serum activin A and B levels predict outcome in patients with acute respiratory failure: a prospective cohort study. Crit. Care 17, R263 (2013). 138. Gieseck, R. L., III, Wilson, M. S. & Wynn, T. A. Type 2 immunity in tissue repair and fibrosis. Nat. Rev. Immunol. 18, 62–76 (2018). 139. Hams, E. et al. IL-25 and type 2 innate lymphoid cells induce pulmonary fibrosis. Proc. Natl Acad. Sci. USA 111, 367–372 (2014). 140. Jessup, H. K. et al. Intradermal administration of thymic stromal lymphopoietin induces a T cell- and eosinophil-dependent systemic Th2 inflammatory response. J. Immunol. 181, 4311–4319 (2008). 141. McHedlidze, T. et al. Interleukin-33-dependent innate lymphoid cells mediate hepatic fibrosis. Immunity 39, 357–371 (2013). 142. Vannella, K. M. et al. Combinatorial targeting of TSLP, IL-25, and IL-33 in type 2 cytokine-driven inflammation and fibrosis. Sci. Transl. Med. 8, 337ra65 (2016). 143. Lee, C. G. et al. Interleukin-13 induces tissue fibrosis by selectively stimulating and activating transforming growth factor β1. J. Exp. Med. 194, 809–821 (2001). 144. Kaviratne, M. et al. IL-13 activates a mechanism of tissue fibrosis that is completely TGFβ independent. J. Immunol. 173, 4020–4029 (2004). 145. Gieseck, R. L. III et al. Interleukin-13 activates distinct cellular pathways leading to ductular reaction, steatosis, and fibrosis. Immunity 45, 145–158 (2016). 146. Hart, K. M. et al. Type 2 immunity is protective in metabolic disease but exacerbates NAFLD collaboratively with TGFβ. Sci. Transl. Med. 9, eaal3694 (2017). This study identified opposing roles for type 2 immunity in metabolic syndrome and liver fibrosis in an experimental model of NASH. 147. Chiaramonte, M. G., Donaldson, D. D., Cheever, A. W. & Wynn, T. A. An IL-13 inhibitor blocks the development of hepatic fibrosis during a T-helper type 2-dominated inflammatory response. J. Clin. Invest. 104, 777–785 (1999). 148. Xue, J. et al. Alternatively activated macrophages promote pancreatic fibrosis in chronic pancreatitis. Nat. Commun. 6, 7158 (2015). 149. Liu, L. et al. CD4+ T lymphocytes, especially Th2 cells, contribute to the progress of renal fibrosis. Am. J. Nephrol. 36, 386–396 (2012). 150. Chung, S. I. et al. IL-13 is a therapeutic target in radiation lung injury. Sci. Rep. 6, 39714 (2016). 151. Singh, B., Kasam, R. K., Sontake, V., Wynn, T. A. & Madala, S. K. Repetitive intradermal bleomycin injections evoke T-helper cell 2 cytokine-driven pulmonary fibrosis. Am. J. Physiol. Lung Cell. Mol. Physiol. 313, L796–L806 (2017). 152. Sciurba, J. C. et al. Fibroblast-specific integrin-alpha V differentially regulates type 17 and type 2 driven inflammation and fibrosis. J. Pathol. 248, 16–29 (2019). 153. Wang, M. et al. Cross-talk between TH2 and TH17 pathways in patients with chronic rhinosinusitis with nasal polyps. J. Allergy Clin. Immunol. 144, 1254–1264 (2019). 154. Choy, D. F. et al. TH2 and TH17 inflammatory pathways are reciprocally regulated in asthma. Sci. Transl. Med. 7, 301ra129 (2015). 155. Ramalingam, T. R. et al. Enhanced protection from fibrosis and inflammation in the combined absence of IL-13 and IFN-γ. J. Pathol. 239, 344–354 (2016).

Nature | Vol 587 | 26 November 2020 | 565

Review 156. Tang, W. et al. Targeted expression of IL-11 in the murine airway causes lymphocytic inflammation, bronchial remodeling, and airways obstruction. J. Clin. Invest. 98, 2845–2853 (1996). This paper identified autocrine IL-11–IL-11R signaling in fibroblasts as a key mechanism driving cardiovascular fibrosis in response to a variety of stimuli. 157. Schafer, S. et al. IL-11 is a crucial determinant of cardiovascular fibrosis. Nature 552, 110–115 (2017). 158. Ng, B. et al. Interleukin-11 is a therapeutic target in idiopathic pulmonary fibrosis. Sci. Transl. Med. 11, eaaw1237 (2019). 159. Abreu, M. T. et al. Mutations in NOD2 are associated with fibrostenosing disease in patients with Crohn’s disease. Gastroenterology 123, 679–688 (2002). 160. Rieder, F. et al. Association of the novel serologic anti-glycan antibodies anti-laminarin and anti-chitin with complicated Crohn’s disease behavior. Inflamm. Bowel Dis. 16, 263–274 (2010). 161. Rieder, F. et al. Serum anti-glycan antibodies predict complicated Crohn’s disease behavior: a cohort study. Inflamm. Bowel Dis. 16, 1367–1375 (2010). 162. Rieder, F., Kessler, S., Sans, M. & Fiocchi, C. Animal models of intestinal fibrosis: new tools for the understanding of pathogenesis and therapy of human disease. Am. J. Physiol. Gastrointest. Liver Physiol. 303, G786–G801 (2012). 163. Moresco, E. M., LaVine, D. & Beutler, B. Toll-like receptors. Curr. Biol. 21, R488–R493 (2011). 164. Månsson, L. E. et al. MyD88 signaling promotes both mucosal homeostatic and fibrotic responses during Salmonella-induced colitis. Am. J. Physiol. Gastrointest. Liver Physiol. 303, G311–G323 (2012). 165. Imai, J. et al. Flagellin-mediated activation of IL-33-ST2 signaling by a pathobiont promotes intestinal fibrosis. Mucosal Immunol. 12, 632–643 (2019). 166. Jacob, N. et al. Inflammation-independent TL1A-mediated intestinal fibrosis is dependent on the gut microbiome. Mucosal Immunol. 11, 1466–1476 (2018). 167. Otte, J. M., Rosenberg, I. M. & Podolsky, D. K. Intestinal myofibroblasts in innate immune responses of the intestine. Gastroenterology 124, 1866–1878 (2003). 168. Zhao, S. et al. Selective deletion of MyD88 signaling in α-SMA positive cells ameliorates experimental intestinal fibrosis via post-transcriptional regulation. Mucosal Immunol. 13, 665–678 (2020). This study highlights a selective mechanism by which bacteria activate myofibroblasts through flagellin. 169. Chan, C. C. et al. Prognostic value of plasma endotoxin levels in patients with cirrhosis. Scand. J. Gastroenterol. 32, 942–946 (1997). 170. Seki, E. & Brenner, D. A. Toll-like receptors and adaptor molecules in liver disease: update. Hepatology 48, 322–335 (2008). 171. Sun, L. et al. Lipopolysaccharide enhances TGFβ1 signalling pathway and rat pancreatic fibrosis. J. Cell. Mol. Med. 22, 2346–2356 (2018). 172. Yang, L. et al. TRIF differentially regulates hepatic steatosis and inflammation/fibrosis in mice. Cell. Mol. Gastroenterol. Hepatol. 3, 469–483 (2017). 173. Mazagova, M. et al. Commensal microbiota is hepatoprotective and prevents liver fibrosis in mice. FASEB J. 29, 1043–1055 (2015). 174. Leaf, I. A. et al. Pericyte MyD88 and IRAK4 control inflammatory and fibrotic responses to tissue injury. J. Clin. Invest. 127, 321–334 (2017). 175. Jialal, I., Major, A. M. & Devaraj, S. Global Toll-like receptor 4 knockout results in decreased renal inflammation, fibrosis and podocytopathy. J. Diabetes Complications 28, 755–761 (2014). 176. Liu, J. H. et al. A novel inhibitor of homodimerization targeting MyD88 ameliorates renal interstitial fibrosis by counteracting TGFβ1-induced EMT in vivo and in vitro. Kidney Blood Press. Res. 43, 1677–1687 (2018). 177. Stifano, G. et al. Chronic Toll-like receptor 4 stimulation in skin induces inflammation, macrophage activation, transforming growth factor beta signature gene expression, and fibrosis. Arthritis Res. Ther. 16, R136 (2014). 178. Liang, J. et al. Hyaluronan and TLR4 promote surfactant-protein-C-positive alveolar progenitor cell renewal and prevent severe pulmonary fibrosis in mice. Nat. Med. 22, 1285–1293 (2016). This work discovered an anti-fibrotic mechanism for hyaluronan in pulmonary fibrosis, revealing a novel function for TLR4. 179. Pilling, D. et al. Reduction of bleomycin-induced pulmonary fibrosis by serum amyloid P. J. Immunol. 179, 4035–4044 (2007).

566 | Nature | Vol 587 | 26 November 2020

180. Nakagawa, N. et al. Pentraxin-2 suppresses c-Jun/AP-1 signaling to inhibit progressive fibrotic disease. JCI Insight 1, e87446 (2016). 181. Rogliani, P., Calzetta, L., Cavalli, F., Matera, M. G. & Cazzola, M. Pirfenidone, nintedanib and N-acetylcysteine for the treatment of idiopathic pulmonary fibrosis: a systematic review and meta-analysis. Pulm. Pharmacol. Ther. 40, 95–103 (2016). 182. Cao, J. et al. Joint profiling of chromatin accessibility and gene expression in thousands of single cells. Science 361, 1380–1385 (2018). 183. Peterson, V. M. et al. Multiplexed quantification of proteins and transcripts in single cells. Nat. Biotechnol. 35, 936–939 (2017). 184. Stoeckius, M. et al. Simultaneous epitope and transcriptome measurement in single cells. Nat. Methods 14, 865–868 (2017). 185. Rodriques, S. G. et al. Slide-seq: a scalable technology for measuring genome-wide expression at high spatial resolution. Science 363, 1463–1467 (2019). 186. Vickovic, S. et al. High-definition spatial transcriptomics for in situ tissue profiling. Nat. Methods 16, 987–990 (2019). 187. Eng, C. L. et al. Transcriptome-scale super-resolved imaging in tissues by RNA seqFISH+. Nature 568, 235–239 (2019). 188. Stuart, T. et al. Comprehensive integration of single-cell data. Cell 177, 1888–1902 (2019). 189. Regev, A. et al. The Human Cell Atlas. eLife 6, e27041 (2017). 190. Dudley, J. T. et al. Computational repositioning of the anticonvulsant topiramate for inflammatory bowel disease. Sci. Transl. Med. 3, 96ra76 (2011). 191. Torok, N. J., Dranoff, J. A., Schuppan, D. & Friedman, S. L. Strategies and endpoints of antifibrotic drug trials: summary and recommendations from the AASLD Emerging Trends Conference, Chicago, June 2014. Hepatology 62, 627–634 (2015). 192. Rieder, F. et al. An expert consensus to standardise definitions, diagnosis and treatment targets for anti-fibrotic stricture therapies in Crohn’s disease. Aliment. Pharmacol. Ther. 48, 347–357 (2018). 193. Montesi, S. B., Désogère, P., Fuchs, B. C. & Caravan, P. Molecular imaging of fibrosis: recent advances and future directions. J. Clin. Invest. 129, 24–33 (2019). 194. Montesi, S. B. et al. Type I collagen-targeted positron emission tomography imaging in idiopathic pulmonary fibrosis: first-in-human studies. Am. J. Respir. Crit. Care Med. 200, 258–261 (2019). Acknowledgements N.C.H. is supported by a Wellcome Trust Senior Research Fellowship in Clinical Science (ref. 219542/Z/19/Z), the Medical Research Council, a Chan Zuckerberg Initiative Seed Network Grant, the British Heart Foundation and Tenovus Scotland. F.R. is supported by grants from the National Institutes of Health (T32DK083251, P30DK097948 Pilot, K08DK110415 and R01DK123233), the Crohn’s and Colitis Foundation, the Cleveland Clinic, the Rainin Foundation and the Helmsley Charitable Trust through the Stenosis Therapy and Anti-Fibrotic Research (STAR) Consortium. Author contributions N.C.H., F.R. and T.A.W contributed equally to the writing and editing of all aspects of this review. Competing interests N.C.H. has received research funding from AbbVie, Pfizer, Gilead and Galecto, and is an advisor or consultant for Galecto, Indalo Therapeutics, Pliant Therapeutics, GSK and Boehringer-Ingelheim. F.R. is an advisor or consultant for AbbVie, Allergan, BMS, Boehringer-Ingelheim, Celgene, Falk Pharma, Gilead, Genentech, Gossamer, GSK, Receptos, Thetis, UCB, Samsung, Koutif, Pliant Therapeutics, Metacrine, Takeda, Theravance, Pfizer, Agomab, Helmsley, RedX and Roche. T.A.W. is employed by Pfizer. Additional information Supplementary information is available for this paper at https://doi.org/10.1038/s41586-0202938-9. Correspondence and requests for materials should be addressed to T.A.W. Peer review information Nature thanks Christopher Buckley and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Reprints and permissions information is available at http://www.nature.com/reprints. Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. © Springer Nature Limited 2020

Review

Discoveries in structure and physiology of mechanically activated ion channels https://doi.org/10.1038/s41586-020-2933-1

J. M. Kefauver1,2,3, A. B. Ward2 ✉ & A. Patapoutian1 ✉

Received: 1 April 2020 Accepted: 19 August 2020 Published online: 25 November 2020 Check for updates

The ability to sense physical forces is conserved across all organisms. Cells convert mechanical stimuli into electrical or chemical signals via mechanically activated ion channels. In recent years, the identification of new families of mechanosensitive ion channels—such as PIEZO and OSCA/TMEM63 channels—along with surprising insights into well-studied mechanosensitive channels have driven further developments in the mechanotransduction field. Several well-characterized mechanosensory roles such as touch, blood-pressure sensing and hearing are now linked with primary mechanotransducers. Unanticipated roles of mechanical force sensing continue to be uncovered. Furthermore, high-resolution structures representative of nearly every family of mechanically activated channel described so far have underscored their diversity while advancing our understanding of the biophysical mechanisms of pressure sensing. Here we summarize recent discoveries in the physiology and structures of known mechanically activated ion channel families and discuss their implications for understanding the mechanisms of mechanical force sensing.

From the sound of a whisper to the strike of a hammer on a finger, many familiar environmental cues occur as mechanical forces. Mechanotransduction, the conversion of mechanical perturbations into electrochemical signals, is conserved across all domains of life. It is possibly the most ancient sensory process, and may have protected early protocells from osmotic and mechanical forces that threatened to break their membranes1. The primary sensors that mediate many such rapid responses to mechanical signals are ion channels2. Research on these channels has been impeded by their sparse expression, their need (in some cases) for specialized cellular structures and the difficulty in producing physiologically relevant mechanical stimuli that can be used across multiple cellular and experimental contexts2–5. A lack of evolutionary conservation in mechanically activated ion channels has delayed the discovery of the channels responsible for mammalian mechanotransduction in particular. In spite of these difficulties, several families of ion channels have been identified in a variety of organisms, from bacteria and flies to humans2–4. Each channel family is structurally distinct from the others, suggesting that each arose independently2–4. Several recent advancements have prompted considerable excitement about the mechanotransduction field (also reviewed in refs. 6,7). The discovery of new families of mechanically activated ion channels such as PIEZOs, which have important in vivo physiological roles in mammals, has opened new avenues of inquiry into the roles of mechanotransduction in human health and disease8,9. Additionally, mechanosensitive channels from the K2P (two-pore potassium) and OSCA/ TMEM63 (hyperosmolality-gated calcium-permeable) channel families have been validated as bona fide mechanically activated ion channels10,11. Recent technical advances in single-particle cryo-electron microscopy (cryo-EM) have led to published structures for almost every known family of mechanically activated ion channel12–25 (Fig. 1). Finally,

new insights from the best-studied mechanically activated channels have bolstered our mechanistic understanding of mechanotransduction at the molecular level13,16,26–28 (Fig. 2).

Mechanosensitive ion channel families Ion channel families from a variety of organisms have been discovered, each with disparate structures and functions.

MscL, MscS and MscS-like channels The first mechanosensitive ion channels to be discovered were the prokaryotic channels mechanosensitive channel large conductance (MscL) and mechanosensitive channel small conductance (MscS), and their homologues in archaea and plants4,29–31. In bacteria, these channels respond to osmotic shock by permeating ions and osmolytes to prevent cell lysis4. MscL is non-selective and opens its large pore like the iris of a camera in response to membrane tension32,33. Its five-subunit structure is composed of an N-terminal amphipathic helix (S1) anchored to the cytosolic leaflet followed by two transmembrane (TM) domains (TM1 and TM2) and a single cytosolic helix (S3) at the C terminus34 (Fig. 1a). The pore is lined by TM1 from each subunit and the narrowest part of the pore is formed by a junction of TM1 helices near the cytoplasmic side of the membrane34. The cytosolic S3 helices form a bundle below the pore that probably acts as a selectivity filter, preventing leakage of metabolites35. Under membrane tension, the amphipathic S1 helix slides along the membrane at the lipid–solvent interface and drives a tilt in the pore-lining TM1 helix, producing an increase in pore diameter36–39 (Fig. 2a). This tension-induced change in free energy overcomes the stability of the closed state, illustrating a fundamental principle in the biophysics of mechanically gated channels40,41.

Howard Hughes Medical Institute, Department of Neuroscience, The Scripps Research Institute, La Jolla, CA, USA. 2Department of Integrative Structural and Computational Biology, The Scripps Research Institute, La Jolla, CA, USA. 3Present address: Department of Molecular Biology, University of Geneva, Geneva, Switzerland. ✉e-mail: [email protected]; [email protected]

1

Nature | Vol 587 | 26 November 2020 | 567

Review a MscL

b MscS

Out

Out

Previous model of MscS: Previou Out

In

In

In

c TREK-1 f NOMPC

e OSC OSCA1.2 CA1.2 Out

Out

In

In

Out

In

d PIEZO1 In

Out

Fig. 1 | Structures of mechanically activated ion channels. Many mechanically activated channels seem to share a common feature: amphipathic helices (dark red on subunit A, rose on all other subunits) connected directly or indirectly to pore-lining regions (dark blue on subunit A, cornflower on all other subunits). a, Cartoon model of MscL (Protein Data Bank (PDB): 2OAR). Pore-lining TM1 (blue) is connected to the amphipathic S1 helix (red). b, Cartoon models of MscS. Main, a recent structure of MscS in nanodiscs (PDB: 6PWP), with the amphipathic anchor domain (red) sitting on the external membrane leaflet. Inset, the previous model of MscS (PDB: 2OAU), with the pore-lining TM3a helix (blue) completely embedded within the membrane and TM3b (red) predicted to be an amphipathic segment at the cytoplasmic leaflet.

c, Cartoon model of TREK-1 (PDB: 6CQ6). The pore domains (blue) are gated by a C-type mechanism61. The amphipathic C-tail (red) extends below the M4 helix. d, Cartoon model of PIEZO1 (PDB: 5Z10). Beneath the extracellular cap, two TM helices from each subunit line the pore (blue). In the domain-swapped blades, several amphipathic helices (red) line the cytoplasmic leaflet. e, Cartoon model of OSCA1.2 (PDB: 6MGV). Five helices (blue) line each of the two putative pores of OSCA1.2 and an amphipathic helix (red) sits on the opposite face of each subunit. f, Cartoon model of NOMPC (PDB: 5VK4). Each NOMPC subunit has an amphipathic TRP domain (red), a pore helix (blue) and a large spring-like ankyrin repeat domain (green).

MscS is structurally distinct from MscL42. It is a homo-heptamer with each protomer composed of three TM helices, with an extracellular N terminus and a cytoplasmic C terminus42. For many years, it was thought that MscS gating was analogous to that of MscL, whereby the third TM helix (TM3) served as the main pore-facing component with a cytosolic amphipathic helix (termed TM3b) formed by a kink in TM3 (Fig. 1b, insert)42–44. However, recent structures of MscS solved in lipidic nanodiscs show that this region sits below the membrane, within the cytoplasm12,13 (Fig. 1b). The flexible N terminus, now designated the anchor domain, occupies the periplasmic half of the TM region and includes an amphipathic portion that sits on the periplasmic leaflet12,13. This region is important for gating by tension, and spectroscopic data suggest that it moves deeper into the membrane as the channel opens13,45,46. These structures also reveal several bound lipids that may be important for gating by mechanical stimuli12,13 (Fig. 3a) and underscore the importance

of structural determination in lipidic environments. Another distinct feature of the MscS family is the large C-terminal cytoplasmic chamber with eight portals that open to the cytoplasm, which serves as the primary selectivity filter47,48; it may also act as a sensor of cytoplasmic crowding to prevent excessive draining of the cell49 (Fig. 1b). Homologues of MscS channels are found in plants and some fungi and protists, but not in animals50. Land plants encode several MscS-like (MSL) genes that are grouped into three categories on the basis of their subcellular localization50. MSLs in group I and group II are expressed in mitochondria and plastids, where they have an osmoregulatory role30,51. Group III MSLs reside at the plasma membrane, where their roles remain an active area of research. In Arabidopsis thaliana, MSL8 and MSL10 have been shown to be bona fide mechanosensitive ion channels, with roles in pollen survival52, stress-induced cell death53 and cell swelling in seeds54. Recent structures of MSL1 reveal a shared architecture with

568 | Nature | Vol 587 | 26 November 2020

a Dragging model

Membrane lipids

MscL

f

Inactive channels Lateral pressure – +

Resting membrane Bound lipids

b Entropy model

K2P

Attraction

Repulsion

Active channels Lateral pressure – +

Planar membrane expansion

c Membrane dome model

PIEZO

G H F E D

I

Attraction Induced membrane curvature

Repulsion

Lateral pressure – +

C B A

Attraction

Repulsion

MET channel complex

d Emerging models

e Tether model

NOMPC

PCDH15

OSCA

TMHS (LHFPL5)

Fig. 2 | Mechanistic models of mechanically activated ion channel gating. Proposed mechanistic models with channel family examples. Amphipathic helices (violet), TM helices (blue), bound lipids (red), beam-like features (gold), tethers (emerald), ions (orange) and membrane lipids (grey) are indicated. a, Left, the dragging model. Lipids interact with an amphipathic helix and drag it outwards upon membrane expansion4. Right, MscL, for example, has an amphipathic helix on the internal leaflet (helix S1, violet) that drives a tilt to the pore-lining helix (TM1, blue) as it is ‘dragged’ outward under tension38. b, Left, entropy model. Lipids reside in hydrophobic pockets in the closed state and exit these pockets under membrane tension, inducing a conformational change132. Right, K2P, for example, has a fenestration occupied by lipid acyl tails (red) when inactive, whereas in active channels, this fenestration is closed and lipids are absent. c, Left, membrane dome model. Channel curvature within the membrane stores energy18. Right, PIEZOs, for example, expand and flatten134, gating the pore via interactions between the beam domain, the anchor domain and the CTD (gold)90. d, Emerging models. OSCA channels have lipid-occupied pores and an

intersubunit cleft (red), an amphipathic helix (violet) on the inner leaflet, and a beam-like domain (gold) connected to pore-lining helices, which terminates a membrane-entrant hook domain (gold); all of these domains could have a role in gating23. e, Left, the tether model. Force is transmitted to the channel via a tether to the extracellular matrix, the cytoskeleton or both. For example, NOMPC (middle) is tethered to microtubules via its ankyrin repeat domain (emerald)112. Right, the MET channel complex is tethered to the neighbouring stereocilium via the tip link (PCDH15; emerald)16. f, Top, resting membrane tension. The transbilayer pressure profile reflects the lateral pressure experienced through the bilayer as a consequence of repulsion (positive pressure) of the lipid head groups, attraction (negative pressure) due to surface tension at the glycerol backbone, and steric hindrance (positive pressure) between the lipid tails32. Bottom, model membranes under tension. Planar membrane expansion thins the bilayer and increases the area occupied by each lipid. Membrane curvature is induced when suction is applied to the membrane or conical-shaped amphipathic compounds insert into the bilayer130,148.

MscS, although two additional TM domains present in MSL1 sit at an angle within the membrane to create a bowl-shaped TM region that is dilated and flattened in open structures of the channel55,56.

TRAAK and TREK channels can be activated by a variety of mechanical stimuli, including stretching, poking, swelling and fluid jet stimulation, as well as temperature and a diverse group of chemicals, including lysolipids, volatile anaesthetics and antidepressants57. Both TRAAK and TREK channels are sensitive to a wide range of tension, from 0.5 mN m−1 to the membrane lytic point of 12 mN m−1, and their open probabilities are proportional to the applied membrane tension57. Although these

Two-pore potassium channels Three members of the two-pore potassium channel (K2P) family are inherently mechanosensitive ion channels: TREK-1, TREK-2 and TRAAK10.

Nature | Vol 587 | 26 November 2020 | 569

Review a MscS 6PWN 1 6RLD 3 2

b TRAAK

c PIEZO1

2

1

Fig. 3 | Lipids observed in structures of mechanosensitive ion channels. a, Lipids are observed in three locations in MscS structures. (1) One lipid per subunit is ‘hooked’ at the periplasmic leaflet12,13; (2) densities ascribed to lipid acyl chains reside inside the pore12,13 (PDB: 6PWN). (3) Two additional lipids per protomer are observed parallel to TM3b, below the membrane leaflet12 (PDB: 6RLD). b, In inactive structures of TRAAK (PDB: 4WWF), an acyl tail (green) occupies a fenestration below the selectivity filter62. c, Two lipid-like densities are observed in the PIEZO1 structure (PDB: 6BPZ): (1) in the region between the anchor domain and piezo repeat A, and (2) between piezo repeats B and C (second and third from the pore, respectively)19.

channels are expressed in sensory neurons, they are not involved in generating action potentials; instead, they dampen the transduction currents of a non-selective cationic mechanosensor by hyperpolarizing cells10. Knockout mice with deletions of each of these three genes exhibit hypersensitivity to mechanical stimuli57. A recent study using monoclonal antibodies specific to TRAAK showed that it is not localized at nerve terminals (where sensory transduction initiates), but is instead present exclusively in the nodes of Ranvier in myelinated neurons58. This raises the possibility that one of its roles is to prevent misfiring in the event of neuronal stretch, or perhaps compensate the mechanical force induced by an action potential58. However, because these ion channels respond to a variety of activators, a role of mechanical forces in the nodes of Ranvier cannot yet be attributed with certainty. K2Ps have two concatenated pore-facing domains per subunit, which dimerize to form a pseudotetramer with two final amphipathic 570 | Nature | Vol 587 | 26 November 2020

C-terminal tails59 (Fig. 1c). In contrast to many K+ channels, K2Ps rely on a C-type gating mechanism60. In C-type gating, the mobility of residues in the selectivity filter determines whether ions are able to permeate the channel60. In crystal structures of TREKs and TRAAK, two main conformations are observed: a ‘down’ state in which a fenestration below the selectivity filter opens towards the membrane, and an ‘up’ state in which the final TM helix (M4) bends upward to occlude this opening27,61–63 (Fig. 2b). Assigning a functional state to these conformations has been difficult, but evidence from studies with the state-dependent blocker norfluoxetine indicate that, for mechanical activation at least, the up state is most probably the active state of the channel27,64,65. Several mechanisms of mechanosensitivity for K2P channels have been proposed. The first is similar to that of MscL, in which an increased cross-sectional area of the protein is more energetically favourable under membrane tension40,41. Because the selectivity filters of TREK and TRAAK channels must maintain structural integrity to retain selectivity for potassium ions, only the portion of the channel below the selectivity filter (in the cytoplasmic leaflet of the membrane) expands in-plane under tension65,66. An alternative mechanism has been suggested on the basis of the observation that in structures of the presumptive inactive down state, lipids or detergent molecules occupy the fenestration below the selectivity filter62,67, but this lipid binding site is occluded in the presumptive active up state, suggesting that bound lipids have a role in gating62,65 (Figs. 2b, 3b). How these conformational changes affect the C-type gate remains unknown.

PIEZO1 and PIEZO2 The PIEZO family is conserved from protozoa to humans and was the first identified class of non-selective cationic mechanotransducers shown to be physiologically relevant in mammals9. Relative to other known channel families, PIEZOs have roles in a broad and varied set of mechanotransduction processes3,9. PIEZOs are involved in well-characterized mechanosensory roles such as touch9, mechanical allodynia (a clinically relevant form of pain)68,69 and the baroreceptor reflex70, as well as several unexpected functions, including developmental processes (such as lymphatic valve development71,72, heart valve development73,74, angiogenesis75 and stem cell differentiation76) and regulatory processes (such as bone formation77,78, cell migration79, axon regeneration80, the inflammatory response of innate immune cells81 and red blood cell (RBC) volume regulation82). Additionally, one-third of the human population of African descent harbours a relatively mild gain-of-function mutation in PIEZO1 that causes RBC dehydration (consistent with hereditary xerocytosis) and confers resistance to malaria82,83. PIEZOs are large trimeric proteins with a triskelion or three-blade propeller architecture18–21,84. The three blade domains extend outward within the lipid bilayer and an extracellular cap domain resides above the central pore18–21 (Fig. 1d). The cap domain appears to have a major role in channel inactivation85–87. PIEZOs have an unusually large number of TM passes per protomer; 38 TM helices per subunit are resolved in the PIEZO2 structure21, matching previous membrane topology predictions for PIEZO188. The first 36 TM helices that form the flexible blades of PIEZO assemble into nine repeating elements of four-helix bundles preceded by an N-terminal amphipathic helix, termed Piezo repeats18–21 (Fig. 2c). Notably, amphipathic helices are known to sense and/or induce membrane curvature19,89. The Piezo repeats spiral away from the centre of the channel in a helical conformation, resulting in an overall puckered architecture of the channel18–21,84 (Fig. 1d). The curved form of PIEZOs could cause a local distortion to the native cell membrane or result in its preferred localization to a membrane domain of similar curvarture18–21. While it is possible that detergent solubilization of PIEZOs alters its native conformation19,21, when purified PIEZO1 is reconstituted into lipid vesicles, the protein deforms the vesicle, producing a similar curvature to that observed in the detergent-solubilized structures18. Furthermore, interactions that occur between the cap

domain and the blades in the curved conformation are important for PIEZO1 inactivation, suggesting that this architecture represents the closed or inactivated state of PIEZOs18,87. The pore of PIEZO is lined by the final two C-terminal TM helices, termed the inner and outer helices18–21,84. The central pore and the domain-swapped extracellular cap resemble the architecture of P2X channels and acid-sensing ion channels (ASICs), but the C-terminal domain (CTD), the anchor domain and the large blades are unique to PIEZOs18. PIEZO structures provide clues regarding how these three features might contribute to mechanosensitivity. The intracellular CTDs form a vestibule with three portals that are probably a continuation of the ion conduction pathway20,90. Juxtaposed against the CTD, the wedge-shaped anchor domain is inserted into the inner leaflet of the membrane, lodged between the pore-lining helix pair and the nearest piezo repeat (piezo repeat A)18–21,84 (Fig. 2c). Beneath each blade, a long beam-like helix extends under the first three piezo repeats (piezo repeats A–C), then hinges at a conserved motif18–20. The beam terminates near the central pore, just below the CTD18–21,84 (Figs. 1d, 2c). One possibility for mechanical gating is that movements of the beam can be transmitted to the pore by a network of interactions between the CTD, the anchor domain and the pore-lining helices19,20,90–92. Because the blades are also domain-swapped relative to the inner or outer helix pair, a flattening or lever-like motion in the blades could produce a lateral dilation or unraveling of the CTD region via its interactions with the beam20,90–92. Despite the large size of PIEZO proteins, relatively few pharmacological tools have been found to modulate their activity. PIEZO1 and PIEZO2 can both be inhibited by the tarantula toxin GsMTx493,94; however, this amphipathic peptide toxin is suggested to interact with the membrane, relieving the effects of tension in the outer bilayer leaflet rather than binding specifically to PIEZOs95. The small molecule Yoda1 can activate PIEZO1, but not PIEZO296. An additional small molecule, Dooku1, acts as an antagonist of Yoda1 with no agonist capability97. PIEZO1 probably harbours a specific binding site for Yoda1 in the region between the anchor domain and piezo repeat A98. Of note, a lipid density is observed between these two domains in cryo-EM structures, along with an additional lipid density between piezo repeats B and C19 (Fig. 3c).

OSCA/TMEM63 channels OSCA/TMEM63 proteins constitute the largest family of mechanosensitive channels99. Initially described as hyperosmolarity sensors in A. thaliana, OSCA proteins have recently been shown to be pore-forming, inherently mechanosensitive channels that are conserved across plants and animals11,100. OSCAs and their mechanosensitive animal homologues TMEM63A and TMEM63B are stretch-activated at a high threshold relative to PIEZO channels11,22. In plants, mutations in OSCA1.1 impair guard cell responses to stress and inhibit root growth under hyperosmotic conditions100. In humans, mutations in TMEM63A are associated with a myelination defect in infants101, although it is currently unclear whether this is related to a mechanosensory role. Several structures of different OSCA family members in detergent and lipidic nanodiscs have been solved by cryo-EM22–25. These dimeric channels have two pores and 11 TM helices per subunit22–25 (Fig. 1e). They are predicted to share structural homology with the TMEM16 family of ion channels or scramblases, as well as potentially transmembrane channel-like protein 1 (TMC1), which is involved in hereditary deafness (see below), on the basis of modelling and mutagenesis studies23,25,102,103. Multiple structural features have been identified that might confer mechanosensitivity to these channels (Fig. 2d). One is the hydrophobic cleft at the dimeric interface that is occupied by lipid molecules in molecular dynamics simulations23,24 (Fig. 2d). Flexibility between the two protomers may allow an increase in cross-sectional area in response to membrane tension23. A hydrophobic groove in the cytosolic pore vestibule opens towards the membrane22–24, similar to the fenestration observed in mechanosensitive K2Ps. In molecular dynamics

simulations, this region is occupied by lipids23. Upon membrane stretch, lipids residing in the cytosolic pore vestibules or the hydrophobic cleft could unbind or dissociate. One of the pore-lining helices (TM6) also lines this fenestration and is bent at a glycine residue that, in a manner analogous to gating in MscL, might straighten during channel activation22,24,26. At its C terminus, TM6 interacts with a distinct feature of OSCAs—a cytosolic domain composed of two parallel helices connected by a membrane-anchored loop22–25 (Fig. 2d). This beam-like domain is predicted to be one of the most dynamic regions of the channel24,25. It may act as a stretch sensor that coordinates the movements of the TM helices, similar to a model for PIEZO mechanosensing90,91. Experimental data to support these possibilities is lacking at the moment, and an understanding of gating will require structure–function analysis as well as structures of the open state.

Transient receptor potential channels Transient receptor potential (TRP) channels are involved in a variety of sensory processes including chemosensation, thermosensation, mechanosensation and osmosensation, often exhibiting polymodal gating by chemicals, signalling lipids (for example, phosphatidylinositol-4 ,5-bisphosphate (PtdIns(4,5)P2)), and physical stimuli104. They all share a canonical tetrameric architecture, with cytosolic N and C termini and six TM domains104. Several members of the TRP ion channel family have been proposed to be mechanosensitive2. However, with the exception of the Drosophila melanogaster mechanosensor, no mechanoreceptor potential C (NOMPC)105,106, proving that TRP channels are directly activated by mechanical stimuli has been difficult. nompC was first cloned from fruit fly sensory bristles105. It is selectively expressed in the ciliated mechanosensory organs of D. melanogaster, Caenorhabditis elegans105 and zebrafish107 at locations where mechanical forces impinge on the sensory cilia108. NOMPC is a bona fide mechanically activated ion channel in Drosophila, but it has no mammalian homologues109. The N terminus of NOMPC contains a cytosolic 29-ankyrin repeat domain (ARD), which is required for microtubule association, proper localization and the in vivo function of NOMPC106. This ARD can be observed by transmission electron microscopy as a filament that is able to adopt variable lengths (20–80 nm) connecting the membrane to microtubules in Drosophila campaniform mechanoreceptors110. A cryo-EM structure revealed that the ARDs of NOMPC form a flexible helical bundle with several intersubunit interactions14 (Fig. 1f). Atomic force microscopy measurements had previously shown that isolated NOMPC ARD acts analogously to a linear and fully reversible spring111. Molecular dynamics simulations using the NOMPC structure show that compression of the ARDs can produce a clockwise movement of the TRP domain and linker helices, similar to the TRPV1 closed-to-open gating transition28,112 (Fig. 2e). The data are consistent with the tether model of mechanotransduction, in which the ARD of NOMPC acts as a spring that transduces the movement of the microtubules relative to the membrane into channel gating. The most compelling case for direct mechanical activation of a mammalian TRP channel has been made for TRPV4113. TRPV4 has been implicated in several processes that may rely on mechanosensitive ion channels, including osmoregulation, control of vascular tone and nociception15,113. Mechanically activated currents induced by a pilus-deflection stimulus in mouse primary chondrocytes are dependent on TRPV4, and heterologous expression of TRPV4 in hamster embryonic kidney cells also induces these currents113. However, like all other tested mammalian TRP channels, TRPV4 is not activated by stretch, suggesting that mammalian TRPs are not inherently mechanosensitive113,114. Tethering of TRPV4 to the extracellular matrix (ECM), the cytoskeleton or some other localized factor is postulated to enable TRPV4 mechanosensitivity113. Another possibility is that TRPV4 responds to a downstream signal initiated by a primary mechanotransducer, such as the amphipathic molecule diacylglyerol114. Cryo-EM and X-ray diffraction Nature | Vol 587 | 26 November 2020 | 571

Review structures of TRPV4 provide few clues as to what structural features may contribute to its putative gating by mechanical stimuli15.

a

Mechano-electrical transduction channel complex The first measurement of mechanically activated currents was in mechanosensory hair cells of the auditory and vestibular systems that sense vibrations induced by pressure waves at specialized hair bundle structures called stereocilia115,116. The stereocilia are arranged in a stairstep pattern of rows of similar height and are connected by filamentous tip links (Fig. 4a). When the stereocilia are deflected, the tip links are thought to transmit force to a mechanically gated ion channel, resulting in depolarization of the hair cell116. The exquisite structural complexity of the mechanosensory hair cells has impeded the decades-long search for the mechano-electrical transduction (MET) channel complex that underlies this current116. Several proteins have been identified that are essential to MET116. The tip link is a tether composed of two components: cadherin 23 is attached to the upper stereocilium and protocadherin 15 (PCDH15) is connected to a lower stereocilium where the MET channel is localized116 (Fig. 4b). Three candidate proteins are essential for MET channel currents and colocalize with PCDH15: (1) TMC1 or TMC2, (2) transmembrane inner ear (TMIE), and (3) lipoma high mobility group IC fusion partner-like (LHFPL5) (also called tetraspan membrane protein of hair cell stereocilia (TMHS))116 (Fig. 4b). Several recent studies support the contribution of TMC1 and TMC2 to the MET channel pore. First, a mutation in TMC1 that causes deafness in mice (called Beethoven (Bth)) also alters ion permeability, single channel conductance and channel blockade of MET channel currents103. Additionally, a structural model of TMC1 based on its loose homology to the TMEM16 protein family predicts a dimeric two-pore structure with a membrane-facing groove102,103. Cysteine-scanning mutagenesis has identified residues that contribute to the pore pathway on each of four pore-lining helices predicted by the homology model102. Finally, reconstitution of purified TMC1 from green sea turtles and TMC2 from budgerigars into liposomes is reported to enable the measurement of pressure-sensitive currents, providing the strongest evidence to date that TMCs are mechanically activated ion channels117. Replication of this result, especially with vertebrate channels, would be ideal given the non-conserved function of several other mechanically activated channel families. While it is likely that TMC1 contributes to the MET channel pore, it remains unclear whether it is the only molecule that underlies MET currents in all types of hair cells. For example, knockout of LHFPL5 has also been shown to affect conductance and adaptation properties of mechanically activated currents in hair cells118, and mutations in the C-terminal lipid-binding domain of TMIE also alter conductance as well as ion selectivity119. It has been difficult to obtain atomic-resolution structures of many MET channel components, possibly because their stability is dependent on the intricate structure of the hair cells16,116. However, a structural approach in which the entire multi-protein MET channel complex is reconstituted for cryo-EM single-particle analysis may be transformative for the field16. This work has been initiated with a structure of truncated PCDH15 co-expressed with LHFPL5, revealing the interactions between these components at subnanometer resolution16 (Fig. 4c). LHFPL5 is a dimeric complex of 4-TM-helix protomers; where the two protomers meet, they form a V-shaped interface that is lined by TM1 of each protomer. Each of the two PCDH15 proteins contributes a single TM pass that forms an inverted V shape that inserts into and interacts with the TM1 helices of LHFPL516 (Figs. 2e, 4c). Because the two PCDH15 protomers are connected by a stable linker domain, force along the tip link would probably pull the PCDH15 TM helices apart, transmitting force to LHFPL516.

Stereocilia

Degenerin, epithelial sodium channels and ASICs The degenerin (DEG) family of ion channels (named for the cell swelling and neurodegeneration phenotype these proteins cause) 572 | Nature | Vol 587 | 26 November 2020

b TMIE

CDH23 LHFPL5 PCDH15

TMC1/2

Cell body

c PCDH15

Out

In

LHFPL5

Fig. 4 | The MET channel complex. a, Model of a hair cell. b, MET channel complex components: TMC1 or TMC2, TMIE and LHFPL5 are localized to the stereocilia tips. Tip links are formed by dimers of PCDH15 and cadherin 23 (CDH23). c, Cartoon model of components of the MET channel complex: PCDH15 forms the lower half of the tip link (green) and LHFPL5 is a dimeric TM protein (blue) (PDB: 6C13 and 6C14).

was first discovered through a genetic screen of touch-insensitive C. elegans mutants120. MEC-4, a founding member of the DEG family, is a pore-forming subunit of the ion channel responsible for mechanosensitive currents in the gentle touch receptors of worms2. It assembles as a channel that can also include the protein MEC-102,121. Several additional C. elegans proteins are also important for mechanosensation, although they are not channel proteins. MEC-2 (homologous to stomatin-like protein 3 (STOML3) in mammals) and MEC-6 enhance mechanosensitive channel activity, probably by modifying the cholesterol content of the membrane2,121. An additional set of ECM-associated proteins, MEC-1 and MEC-9, are proposed to contribute to an extracellular tether that connects MEC-4–MEC-10 channels to the ECM2. In mammals, two

channel families, epithelial sodium channels (ENaCs) and ASICs, share homology with the C. elegans MEC-4–MEC-10 channels2. Roles for these channels in mechanosensory processes such as blood pressure sensing and nociception have been described, but there is a lack of evidence that these channels are directly activated by mechanical stimuli122,123. A recent study revealed that two glycosylated asparagines in ENaC probably serve as a tether to the ECM, building on the idea that this channel family might sense tension via a filament tether124.

Mechanistic insights With the discovery of disparate families of mechanosensitive ion channels and the subsequent structural and biophysical insights, our mechanistic understanding of mechanotransduction is developing rapidly. Two classic models for transmitting force to an ion channel remain in the spotlight: the force-from-lipids model and the force-from-tether model. We also discuss a third, hybrid model whereby modulation by both lipids and cytoskeleton can have a role.

The force-from-lipids model The force-from-lipids model involves the direct conversion of tension in the membrane to area expansion in mechanosensitive proteins, without the need for additional elements such as cytoskeleton or accessory proteins125. An ion channel is considered ‘inherently mechanosensitive’ if it can be mechanically activated after purification and reconstitution in a lipid bilayer126. MscL32,127, MscS127, TRAAK10, TREK-110, PIEZO1128 and several OSCA family members11 all meet this standard. This mechanistic conservation is unsurprising, as membrane-embedded protein domains exist in the context of the chemical and biophysical properties of the bilayer, enabling sensitivity to tension in the bilayer. At rest, all membrane-embedded proteins experience a mix of hydrophobic and steric forces exerted by the bilayer lipids, termed the ‘transbilayer pressure profile’126,129 (Fig. 2f). When the bilayer stretches, the membrane thins and the local transbilayer pressure profile changes126,129. Mechanically activated ion channels change conformation in response to hydrophobic mismatch between their membrane-facing domains and the bilayer, producing gating movements to open the pore126,129,130. Another way to conceptualize the effect of membrane tension is as an increase in the planar area of the bilayer lipids, shifting the equilibrium such that hydrophobic forces that tend to cluster lipids overcome the forces that mediate protein–lipid interactions65,131,132. In this ‘entropy-driven’ model, lipids that stabilize the closed state of the protein dissociate, resulting in a conformational change to compensate the unoccupied hydrophobic pockets131,132 (Fig. 2b). For example, in structures of the two-pore potassium channels, TRAAK and TREK-1, membrane lipids bound to a fenestration below the channel pore stabilize the closed conformation of the channel; in structures of open channels, those lipids are absent62. Lipids are predicted to occupy similar pore-adjacent fenestrations of several different channels including MscL38, MscS12,13, MSL1and OSCA1.222–24, as well as near the pore region in structures of PIEZO119,98. An alternative model suggests that bound lipids maintain their interactions with the protein rather than dissociating under tension. This ‘dragging’ model, based on observations of MscL, proposes that lipids do work on an amphipathic helix connected to a pore lining helix, straightening it as the membrane equilibrates under tension4,26 (Fig. 2a). Amphipathic helices are present in MscL26,34, MscS13, TRAAK and TREK27,62,63, PIEZOs19–21 and OSCA1.223–25, although a direct connection to a pore-lining segment is only present in the case of MscL36,37 (Fig. 2a). For PIEZOs and OSCAs, a link between the amphipathic helices and the pore is possible via a more complex network of allosteric interactions13,19,20,22–24 (Fig. 2c, d). Alternatively, the membrane-anchored loop of the beam-like domain of OSCA might substitute an amphipathic helix to serve a similar dragging function (Fig. 2d).

A different type of membrane distortion is a change in curvature32 (Fig. 2c). Channel activity of MscL, MscS, TREK-1, PIEZO1 and OSCA1.1 can all be modulated by the addition of conical lipids or amphipathic molecules that tend to bend membranes22,126. As membrane curvature can gate mechanically activated channels, the intrinsic curvature of PIEZOs is notable. A ‘dome mechanism’ has been proposed for PIEZO activation, in which its large size and curved shape enable a large increase of in-plane area as PIEZO flattens under tension18,133,134 (Fig. 2c). Bolstering this model, PIEZO1 protein reconstituted in planar membranes undergoes substantial area expansion when tension is applied to the membrane by an atomic force microscopy cantilever134. Moreover, the bending force that the curvature of PIEZO would exert on the membrane is predicted to extend past the edge of the protein into the bilayer to create a ‘membrane footprint’133. A similar membrane dome mechanism has recently been proposed for MSL156.

Tethering to ECM or cytoskeleton The alternative mechanism is one in which mechanosensitive proteins are tethered to the ECM, the cytoskeleton, or both ECM and cytoskeleton, enabling forces experienced by these cellular compartments to be transmitted to these ion channels via a connecting structure. The Drosophila channel NOMPC, with its spring-like ankyrin repeats14,28,110, and the MET channel complex, with its tip links, are clear examples of tethered channels16 (Fig. 2e). The involvement of tether molecules has also been proposed for somatosensory touch function135. Within the DEG–ENAC–ASIC superfamily, ENaC and probably MEC-4–MEC-10 channels also function via a tether to the ECM2,124, and TRPV4 and other TRP channels may also use such a tether113,114. Mechanisms of modulation There is evidence that these two models are not mutually exclusive136. For example, the conspicuously tethered channel NOMPC also responds to force-from-lipids activation4,14. A NOMPC residue that interacts with lipid head groups is, in fact, essential for mechanical activation of the channel4,14. Conversely, effects of the cytoskeleton, regulatory proteins and signalling lipids have also been observed for inherently mechanosensitive channels. PIEZO1 and TREK-1 activity can be modulated in complex ways by the presence of cytoskeletal elements9,137. For instance, the application of cytochalasin D, which disrupts the actin cytoskeleton, positively modulates PIEZO1 under some circumstances and negatively modulates it in others9. Another important modulatory factor for mechanosensitive proteins is the composition of the membrane in which they are embedded. TREK1 and TRAAK are activated by the signalling lipid PtdIns(4,5)P2138, and depletion of PtdIns(4,5)P2 and PtdIns(4)P inhibit PIEZO1 and PIEZO2139. Lipid rafts and other membrane lipid subdomains may also be important for mechanotransduction. An interesting theory suggests that mechanotransduction occurs at ‘force foci’, specific cholesteroland sphingolipid-rich membrane subdomains localized to adherens junctions and focal adhesions140. A study that followed raft-localized phospholipase D signalling after mechanical stimulation demonstrates that mechanical force is capable of disrupting cholesterol-rich rafts141. A membrane-subdomain model is compatible with tethered multi-protein assemblies such as the MET channel complex, but recent studies suggest that it may also be relevant to PIEZOs; PIEZO1 appears to localize to focal adhesions in glioma cells142. Additionally, in a subset of mechanoreceptive neurons, co-expression of PIEZO1 with cytoskeleton-associated STOML3 lowers the activation threshold of the channel by an order of magnitude143,144. STOML3 functions by binding cholesterol and forming a membrane-associated scaffold in a manner that resembles putative force foci143,144. As our ability to study membrane proteins in their native context improves, we may discover that the regulation of mechanosensitive proteins by membrane lipids is far more complex and intricate than individual channel structures have so far revealed. Nature | Vol 587 | 26 November 2020 | 573

Review Outlook The explosion of new structural models of mechanosensitive ion channels provides an excellent starting point for structure–function studies and molecular dynamics simulations to develop increasingly refined models of mechanotransduction. However, for many channel families, particularly those in vertebrates, these efforts are hindered by the absence of structures of open channels. Producing a mechanical stimulus suitable for structural studies is a substantial challenge, as high-resolution structure determination currently requires purification and isolation away from the membrane. Chemical agonists that could trap an open state have been difficult to identify for channels such as PIEZO1 and PIEZO296. For channel complexes such as the MET channel, purification and reconstitution of a complete set of proteins necessary for channel function is a formidable technical challenge16. As cryo-electron tomography and correlated light and electron microscopy improve, the ability to acquire high-resolution structures in situ might provide a new view of mechanically activated channels in native membranes. Despite the substantial recent progress in identification and characterization of mechanically activated ion channels, a variety of biological processes that depend on mechanotransduction remain poorly understood at the molecular level, and the identities of many mechanosensors remain elusive, including that of the sensor(s) for mammalian acute pain. Further characterization of the physiological roles of PIEZOs and the OSCA/TMEM63 family in animals and plants may provide some of these answers, but the search for unknown mechanosensors remains imperative. Indeed, two new families of mechanically activated ion channels, TACAN145 and Elkin146, have recently been proposed. It may also prove valuable to broaden the search for mechanosensing molecules beyond ion channels, as mechanical stimuli also affect development and growth processes that do not depend on neuronal action potentials. For example, the G-protein-coupled receptor GPR68 was recently shown to be activated by shear stress to release intracellular calcium stores, regulating blood vessel dilation147. It will be important for these initial findings on proposed mechanosensitive ion channels and G-protein-coupled receptors to be replicated by other groups in the field, and we hope that future structural and physiological studies will extend these initial observations. Finally, the identification of new families of mechanically activated channels will expand our knowledge on the role of mechanotransduction in previously unrecognized areas of physiology and disease. By using expression patterns of newly identified mechanosensors as a starting point, loss-of-function and gain-of-function mutations can be powerful genetic tools to gain in-depth insights into the in vivo consequences of sensing mechanical force. Furthermore, probing human genetic data for mutations in mechanosensors will be instrumental in defining their roles in human disease. Indeed, recent discoveries on the role of PIEZO1 in bone formation, RBC hydration and immune function in humans and in mice are examples of how genetic studies on mechanosensors can lead to mechanistic understanding of the role of pressure sensing in unexpected areas of biology and disease77,78,81,82. 1. 2. 3. 4. 5. 6. 7. 8.

Anishkin, A., Loukin, S. H., Teng, J. & Kung, C. Feeling the hidden mechanical forces in lipid bilayer is an original sense. Proc. Natl Acad. Sci. USA 111, 7898–7905 (2014). Arnadóttir, J. & Chalfie, M. Eukaryotic mechanosensitive channels. Annu. Rev. Biophys. 39, 111–137 (2010). Ranade, S. S., Syeda, R. & Patapoutian, A. Mechanically activated ion channels. Neuron 87, 1162–1179 (2015). Cox, C. D., Bavi, N. & Martinac, B. Bacterial mechanosensors. Annu. Rev. Physiol. 80, 71–93 (2018). Chalfie, M. Neurosensory mechanotransduction. Nat. Rev. Mol. Cell Biol. 10, 44–52 (2009). Douguet, D. & Honoré, E. Mammalian mechanoelectrical transduction: structure and function of force-gated ion channels. Cell 179, 340–354 (2019). Jin, P., Jan, L. Y. & Jan, Y.-N. Mechanosensitive ion channels: structural features relevant to mechanotransduction mechanisms. Annu. Rev. Neurosci. 43, 207–229 (2020). Coste, B. et al. Piezo1 and Piezo2 are essential components of distinct mechanically activated cation channels. Science 330, 55–60 (2010). This study uses an RNA interference screen to identify PIEZOs as essential components of a mechanically actived ion channel.

574 | Nature | Vol 587 | 26 November 2020

9. 10.

11.

12.

13.

14.

15. 16.

17. 18. 19. 20. 21.

22. 23. 24. 25.

26.

27. 28.

29.

30. 31. 32.

33.

34.

35. 36.

Murthy, S. E., Dubin, A. E. & Patapoutian, A. Piezos thrive under pressure: mechanically activated ion channels in health and disease. Nat. Rev. Mol. Cell Biol. 18, 771–783 (2017). Brohawn, S. G., Su, Z. & MacKinnon, R. Mechanosensitivity is mediated directly by the lipid membrane in TRAAK and TREK1 K+ channels. Proc. Natl Acad. Sci. USA 111, 3614–3619 (2014). This study demonstrates that the mechanically activated K2P channels, TRAAK and TREK1, are inherently mechanosensitive. Murthy, S. E. et al. OSCA/TMEM63 are an evolutionarily conserved family of mechanically activated ion channels. eLife 7, 1–17 (2018). Several OSCA genes and their mammalian homologues, the TMEM63 family, are shown to be inherently mechanosensitive ion channels. Rasmussen, T., Flegler, V. J., Rasmussen, A. & Böttcher, B. Structure of the mechanosensitive channel MscS embedded in the membrane bilayer. J. Mol. Biol. 431, 3081–3090 (2019). Reddy, B., Bavi, N., Lu, A., Park, Y. & Perozo, E. Molecular basis of force-from-lipids gating in the mechanosensitive channel MscS. eLife 8, e50486 (2019). These two articles present cryo-EM structures of the bacterial MscS channel in a lipidic nanodisc, substantially updating our understanding of how it is embedded within the membrane. Jin, P. et al. Electron cryo-microscopy structure of the mechanotransduction channel NOMPC. Nature 547, 118–122 (2017). This study presents the cryo-EM structure of NOMPC, revealing that the large ankyrin repeat domain is arranged with a large spring-like architecture. Deng, Z. et al. Cryo-EM and X-ray structures of TRPV4 reveal insight into ion permeation and gating mechanisms. Nat. Struct. Mol. Biol. 25, 252–260 (2018). Ge, J. et al. Structure of mouse protocadherin 15 of the stereocilia tip link in complex with LHFPL5. eLife 7, e38770 (2018). Co-expression and purification of the MET complex components PCDH15 and LHFPL5 reveal that a TM helix of the PCDH15 subunit interacts extensively with the TM helices of each LHFPL5 subunit. Noreng, S., Bharadwaj, A., Posert, R., Yoshioka, C. & Baconguis, I. Structure of the human epithelial sodium channel by cryo-electron microscopy. eLife 7, e39340 (2018). Guo, Y. R. & MacKinnon, R. Structure-based membrane dome mechanism for Piezo mechanosensitivity. eLife 6, e33660 (2017). Saotome, K. et al. Structure of the mechanically activated ion channel Piezo1. Nature 554, 481–486 (2018). Zhao, Q. et al. Structure and mechanogating mechanism of the Piezo1 channel. Nature 554, 487–492 (2018). Wang, L. et al. Structure and mechanogating of the mammalian tactile channel PIEZO2. Nature 573, 225–229 (2019). These four studies present cryo-EM structures of PIEZO1 and PIEZO2, revealing that its curved shape probably resides within the membrane, and providing a near-atomic-resolution view of several features that may be involved in gating by mechanical force. Zhang, M. et al. Structure of the mechanosensitive OSCA channels. Nat. Struct. Mol. Biol. 25, 850–858 (2018). Jojoa-Cruz, S. et al. Cryo-EM structure of the mechanically activated ion channel OSCA1.2. eLife 7, e41845 (2018). Liu, X., Wang, J. & Sun, L. Structure of the hyperosmolality-gated calcium-permeable channel OSCA1.2. Nat. Commun. 9, 5060 (2018). Maity, K. et al. Cryo-EM structure of OSCA1.2 from Oryza sativa elucidates the mechanical basis of potential membrane hyperosmolality gating. Proc. Natl Acad. Sci. USA 116, 14309–14318 (2019). These four studies present cryo-EM structures of OSCA1.2 and other OSCA family members, revealing their structural homology to TMEM16 and highlighting features that may be involved in mechanical gating. Bavi, N., Cox, C. D., Perozo, E. & Martinac, B. Toward a structural blueprint for bilayer-mediated channel mechanosensitivity. Channels 11, 91–93 (2017). This study proposed the dragging mechanism of mechanotransduction. Dong, Y. Y. et al. K2P channel gating mechanisms revealed by structures of TREK-2 and a complex with Prozac. Science 347, 1256–1259 (2015). Argudo, D., Capponi, S., Bethel, N. P. & Grabe, M. A multiscale model of mechanotransduction by the ankyrin chains of the NOMPC channel. J. Gen. Physiol. 151, 316–327 (2019). Sukharev, S. I., Martinac, B., Arshavsky, V. Y. & Kung, C. Two types of mechanosensitive channels in the Escherichia coli cell envelope: solubilization and functional reconstitution. Biophys. J. 65, 177–183 (1993). Haswell, E. S. & Meyerowitz, E. M. MscS-like proteins control plastid size and shape in Arabidopsis thaliana. Curr. Biol. 16, 1–11 (2006). Kloda, A. & Martinac, B. Common evolutionary origins of mechanosensitive ion channels in Archaea, Bacteria and cell-walled Eukarya. Archaea 1, 35–44 (2002). Perozo, E., Kloda, A., Cortes, D. M. & Martinac, B. Physical principles underlying the transduction of bilayer deformation forces during mechanosensitive channel gating. Nat. Struct. Biol. 9, 696–703 (2002). Betanzos, M., Chiang, C. S., Guy, H. R. & Sukharev, S. A large iris-like expansion of a mechanosensitive channel protein induced by membrane tension. Nat. Struct. Biol. 9, 704–710 (2002). Disulfide cross-linking between the first TM helices on adjacent subunits of MscL in the resting state and in osmotically shocked cells provides the first evidence for the area-expansion model of mechanotransduction. Chang, G., Spencer, R. H., Lee, A. T., Barclay, M. T. & Rees, D. C. Structure of the MscL homolog from Mycobacterium tuberculosis: a gated mechanosensitive ion channel. Science 282, 2220–2226 (1998). Anishkin, A. et al. On the conformation of the COOH-terminal domain of the large mechanosensitive channel MscL. J. Gen. Physiol. 121, 227–244 (2003). Iscla, I., Wray, R. & Blount, P. The dynamics of protein–protein interactions between domains of MscL at the cytoplasmic–lipid interface. Channels 6, 255–261 (2012).

37. Li, J. et al. Mechanical coupling of the multiple structural elements of the large-conductance mechanosensitive channel during expansion. Proc. Natl Acad. Sci. USA 112, 10726–10731 (2015). 38. Bavi, N. et al. The role of MscL amphipathic N terminus indicates a blueprint for bilayer-mediated gating of mechanosensitive channels. Nat. Commun. 7, 11984 (2016). 39. Wang, Y. et al. Single molecule FRET reveals pore size and opening mechanism of a mechano-sensitive ion channel. eLife 3, e01834 (2014). 40. Chiang, C. S., Anishkin, A. & Sukharev, S. Gating of the large mechanosensitive channel in situ: estimation of the spatial scale of the transition from channel population responses. Biophys. J. 86, 2846–2861 (2004). 41. Wiggins, P. & Phillips, R. Membrane–protein interactions in mechanosensitive channels. Biophys. J. 88, 880–902 (2005). 42. Bass, R. B., Strop, P., Barclay, M. & Rees, D. C. Crystal structure of Escherichia coli MscS, a voltage-modulated and mechanosensitive channel. Science 298, 1582–1587 (2002). 43. Wang, W. et al. The structure of an open form of an E. coli mechanosensitive channel at 3.45 Å resolution. Science 321, 1179–1183 (2008). 44. Anishkin, A., Kamaraju, K. & Sukharev, S. Mechanosensitive channel MscS in the open state: modeling of the transition, explicit simulations, and experimental measurements of conductance. J. Gen. Physiol. 132, 67–83 (2008). 45. Vásquez, V., Sotomayor, M., Cordero-Morales, J., Schulten, K. & Perozo, E. A structural mechanism for MscS gating in lipid bilayers. Science 321, 1210–1214 (2008). 46. Rasmussen, T. et al. Interaction of the mechanosensitive channel, MscS, with the membrane bilayer through lipid intercalation into grooves and pockets. J. Mol. Biol. 431, 3339–3352 (2019). 47. Edwards, M. D., Bartlett, W. & Booth, I. R. Pore mutations of the Escherichia coli MscS channel affect desensitization but not ionic preference. Biophys. J. 94, 3003–3013 (2008). 48. Cox, C. D. et al. Selectivity mechanism of the mechanosensitive channel MscS revealed by probing channel subconducting states. Nat. Commun. 4, 2137 (2013). 49. Rowe, I., Anishkin, A., Kamaraju, K., Yoshimura, K. & Sukharev, S. The cytoplasmic cage domain of the mechanosensitive channel MscS is a sensor of macromolecular crowding. J. Gen. Physiol. 143, 543–557 (2014). 50. Hamilton, E. S., Schlegel, A. M. & Haswell, E. S. United in diversity: mechanosensitive ion channels in plants. Annu. Rev. Plant Biol. 66, 113–137 (2015). 51. Lee, J. S., Wilson, M. E., Richardson, R. A. & Haswell, E. S. Genetic and physical interactions between the organellar mechanosensitive ion channel homologs MSL1, MSL2, and MSL3 reveal a role for inter-organellar communication in plant development. Plant Direct 3, e00124 (2019). 52. Hamilton, E. S. & Haswell, E. S. The tension-sensitive ion transport activity of MSL8 is critical for its function in pollen hydration and germination. Plant Cell Physiol. 58, 1222– 1237 (2017). 53. Basu, D., Shoots, J. M. & Haswell, E. S. Interactions between the N- and C-termini of the mechanosensitive ion channel AtMSL10 are consistent with a three-step mechanism for activation. J. Exp. Bot. 71, 4020–4032 (2020). 54. Basu, D. & Haswell, E. S. The mechanosensitive ion channel MSL10 potentiates responses to cell swelling in Arabidopsis seedlings. Curr. Biol. 30, 2716–2728.e6 (2020). 55. Li, Y. et al. Structural insights into a plant mechanosensitive ion channel MSL1. Cell Rep. 30, 4518–4527.e3 (2020). This study presents the open and closed structures of MSL1, providing insights into the gating of this plant mechanosensitive channel. 56. Deng, Z. et al. Structural mechanism for gating of a eukaryotic mechanosensitive channel of small conductance. Nat. Commun. 11, 3690 (2020). 57. Brohawn, S. G. How ion channels sense mechanical force: insights from mechanosensitive K2P channels TRAAK, TREK1, and TREK2. Ann. NY Acad. Sci. 1352, 20– 32 (2015). 58. Brohawn, S. G. et al. The mechanosensitive ion channel TRAAK is localized to the mammalian node of Ranvier. eLife 8, 713990 (2019). 59. Brohawn, S. G., Campbell, E. B. & MacKinnon, R. Domain-swapped chain connectivity and gated membrane access in a Fab-mediated crystal of the human TRAAK K+ channel. Proc. Natl Acad. Sci. USA 110, 2129–2134 (2013). 60. Schewe, M. et al. A non-canonical voltage-sensing mechanism controls gating in K2P K+ channels. Cell 164, 937–949 (2016). 61. Brohawn, S. G., del Mármol, J. & MacKinnon, R. Crystal structure of the human K2P TRAAK, a lipid- and mechano-sensitive K+ ion channel. Science 335, 436–441 (2012). 62. Brohawn, S. G., Campbell, E. B. & MacKinnon, R. Physical mechanism for gating and mechanosensitivity of the human TRAAK K+ channel. Nature 516, 126–130 (2014). 63. Lolicato, M., Riegelhaupt, P. M., Arrigoni, C., Clark, K. A. & Minor, D. L., Jr. Transmembrane helix straightening and buckling underlies activation of mechanosensitive and thermosensitive K(2P) channels. Neuron 84, 1198–1212 (2014). These two articles present crystal structures of mechanosenstive K2P channels with opposing interpretations of the activity state of the channel; Brohawn et al. observe lipid density that extends into the channel pore. 64. McClenaghan, C. et al. Polymodal activation of the TREK-2 K2P channel produces structurally distinct open states. J. Gen. Physiol. 147, 497–505 (2016). 65. Aryal, P. et al. Bilayer-mediated structural transitions control mechanosensitivity of the TREK-2 K2P channel. Structure 25, 708–718.e2 (2017). Molecular dynamics modelling of TREK-2 elucidates the gating transition upon membrane tension. 66. Clausen, M. V., Jarerattanachat, V., Carpenter, E. P., Sansom, M. S. P. & Tucker, S. J. Asymmetric mechanosensitivity in a eukaryotic ion channel. Proc. Natl Acad. Sci. USA 114, E8343–E8351 (2017). 67. Lolicato, M. et al. K2P2.1 (TREK-1)-activator complexes reveal a cryptic selectivity filter binding site. Nature 547, 364–368 (2017). 68. Murthy, S. E. et al. The mechanosensitive ion channel Piezo2 mediates sensitivity to mechanical pain in mice. Sci. Transl. Med. 10, eaat9897 (2018). 69. Szczot, M. et al. PIEZO2 mediates injury-induced tactile pain in mice and humans. Sci. Transl. Med. 10, eaat9892 (2018).

70. Zeng, W. Z. et al. PIEZOs mediate neuronal sensing of blood pressure and the baroreceptor reflex. Science 362, 464–467 (2018). 71. Nonomura, K. et al. Mechanically activated ion channel PIEZO1 is required for lymphatic valve formation. Proc. Natl Acad. Sci. USA 115, 12817–12822 (2018). 72. Choi, D. et al. Piezo1 incorporates mechanical force signals into the genetic program that governs lymphatic valve development and maintenance. JCI Insight 4, e125068 (2019). 73. Faucherre, A. et al. Piezo1 is required for outflow tract and aortic valve development. J. Mol. Cell. Cardiol. 143, 51–62 (2020). 74. Duchemin, A. L., Vignes, H. & Vermot, J. Mechanically activated piezo channels modulate outflow tract valve development through the Yap1 and Klf2–Notch signaling axis. eLife 8, e44706 (2019). 75. Kang, H. et al. Piezo1 mediates angiogenesis through activation of MT1-MMP signaling. Am. J. Physiol. Cell Physiol. 316, C92–C103 (2019). 76. He, L., Si, G., Huang, J., Samuel, A. D. T. & Perrimon, N. Mechanical regulation of stem-cell differentiation by the stretch-activated Piezo channel. Nature 555, 103–106 (2018). 77. Sun, W. et al. The mechanosensitive Piezo1 channel is required for bone formation. eLife 8, e47454 (2019). 78. Li, X. et al. Stimulation of Piezo1 by mechanical signals promotes bone anabolism. eLife 8, e49631 (2019). 79. Ellefsen, K. L. et al. Myosin-II mediated traction forces evoke localized Piezo1-dependent Ca2+ flickers. Commun. Biol. 2, 298 (2019). 80. Song, Y. et al. The mechanosensitive ion channel Piezo inhibits axon regeneration. Neuron 102, 373–389 (2019). 81. Solis, A. G. et al. Mechanosensation of cyclical force by PIEZO1 is essential for innate immunity. Nature 573, 69–74 (2019). 82. Ma, S. et al. Common PIEZO1 allele in African populations causes RBC dehydration and attenuates Plasmodium infection. Cell 173, 443–455 (2018). 83. Nguetse, C. N. et al. A common polymorphism in the druggable ion channel PIEZO1 is associated with protection from severe malaria. Proc. Natl Acad. Sci. USA 117, 9074–9081 (2020). 84. Ge, J. et al. Architecture of the mammalian mechanosensitive Piezo1 channel. Nature 527, 64–69 (2015). 85. Wu, J., Goyal, R. & Grandl, J. Localized force application reveals mechanically sensitive domains of Piezo1. Nat. Commun. 7, 12939 (2016). 86. Wu, J. et al. Inactivation of mechanically activated Piezo1 ion channels is determined by the C-terminal extracellular domain and the inner pore helix. Cell Rep. 21, 2357–2366 (2017). 87. Lewis, A. H. & Grandl, J. Inactivation kinetics and mechanical gating of Piezo1 ion channels depend on subdomains within the cap. Cell Rep. 30, 870–880 (2020). 88. Coste, B. et al. Piezo1 ion channel pore properties are dictated by C-terminal region. Nat. Commun. 6, 7223 (2015). 89. Drin, G. & Antonny, B. Amphipathic helices and membrane curvature. FEBS Lett. 584, 1840–1847 (2010). 90. Geng, J. et al. A plug-and-latch mechanism for gating the mechanosensitive Piezo channel. Neuron 106, 438–451 (2020). 91. Wang, Y. et al. A lever-like transduction pathway for long-distance chemical- and mechano-gating of the mechanosensitive Piezo1 channel. Nat. Commun. 9, 1300 (2018). 92. Taberner, F. J. et al. Structure-guided examination of the mechanogating mechanism of PIEZO2. Proc. Natl Acad. Sci. USA 116, 14260–14269 (2019). 93. Bae, C., Sachs, F. & Gottlieb, P. A. The mechanosensitive ion channel Piezo1 is inhibited by the peptide GsMTx4. Biochemistry 50, 6295–6300 (2011). 94. Alcaino, C., Knutson, K., Gottlieb, P. A., Farrugia, G. & Beyder, A. Mechanosensitive ion channel Piezo2 is inhibited by D-GsMTx4. Channels 11, 245–253 (2017). 95. Suchyna, T. M. Piezo channels and GsMTx4: two milestones in our understanding of excitatory mechanosensitive channels and their role in pathology. Prog. Biophys. Mol. Biol. 130, 244–253 (2017). 96. Syeda, R. et al. Chemical activation of the mechanotransduction channel Piezo1. eLife 4, e07369 (2015). 97. Evans, E. L. et al. Yoda1 analogue (Dooku1) which antagonizes Yoda1-evoked activation of Piezo1 and aortic relaxation. Br. J. Pharmacol. 175, 1744–1759 (2018). 98. Lacroix, J. J., Botello-Smith, W. M. & Luo, Y. Probing the gating mechanism of the mechanosensitive channel Piezo1 with the small molecule Yoda1. Nat. Commun. 9, 2029 (2018). 99. Hou, C. et al. DUF221 proteins are a family of osmosensitive calcium-permeable cation channels conserved across eukaryotes. Cell Res. 24, 632–635 (2014). 100. Yuan, F. et al. OSCA1 mediates osmotic-stress-evoked Ca2+ increases vital for osmosensing in Arabidopsis. Nature 514, 367–371 (2014). 101. Yan, H. et al. Heterozygous variants in the mechanosensitive ion channel TMEM63A result in transient hypomyelination during infancy. Am. J. Hum. Genet. 105, 996–1004 (2019). 102. Pan, B. et al. TMC1 forms the pore of mechanosensory transduction channels in vertebrate inner ear hair cells. Neuron 99, 736–753 (2018). 103. Ballesteros, A., Fenollar-Ferrer, C. & Swartz, K. J. Structural relationship between the putative hair cell mechanotransduction channel TMC1 and TMEM16 proteins. eLife 7, e38433 (2018). 104. Startek, J. B., Boonen, B., Talavera, K. & Meseguer, V. TRP channels as sensors of chemicallyinduced changes in cell membrane mechanical properties. Int. J. Mol. Sci. 20, 371 (2019). 105. Walker, R. G., Willingham, A. T. & Zuker, C. S. A. A Drosophila mechanosensory transduction channel. Science 287, 2229–2234 (2000). This study identifies the nompC gene via a genetic screen with a readout of transduction currents upon mechanical stimulus of mechanoreceptor bristles in D. melanogaster. 106. Cheng, L. E., Song, W., Looger, L. L., Jan, L. Y. & Jan, Y. N. The role of the TRP channel NompC in Drosophila larval and adult locomotion. Neuron 67, 373–380 (2010). 107. Sidi, S., Friedrich, R. W. & Nicolson, T. NompC TRP channel required for vertebrate sensory hair cell mechanotransduction. Science 301, 96–99 (2003). 108. Lee, J., Moon, S., Cha, Y. & Chung, Y. D. Drosophila TRPN(=NOMPC) channel localizes to the distal end of mechanosensory cilia. PLoS ONE 5, e11012 (2010).

Nature | Vol 587 | 26 November 2020 | 575

Review 109. Yan, Z. et al. Drosophila NOMPC is a mechanotransduction channel subunit for gentle-touch sensation. Nature 493, 221–225 (2013). 110. Sun, L. et al. Ultrastructural organization of NompC in the mechanoreceptive organelle of Drosophila campaniform mechanoreceptors. Proc. Natl Acad. Sci. USA 116, 7343–7352 (2019). 111. Lee, G. et al. Nanospring behaviour of ankyrin repeats. Nature 440, 246–249 (2006). 112. Wang, Y. et al. Push-to-open: The gating mechanism of the tethered mechanosensitive ion channel NompC. Preprint at https://doi.org/10.1101/853721 (2019). 113. Servin-Vences, M. R., Moroni, M., Lewin, G. R. & Poole, K. Direct measurement of TRPV4 and PIEZO1 activity reveals multiple mechanotransduction pathways in chondrocytes. eLife 6, e21074 (2017). 114. Nikolaev, Y. A. et al. Mammalian TRP ion channels are insensitive to membrane stretch. J. Cell Sci. 132, 238360 (2019). 115. Corey, D. P. & Hudspeth, A. J. Kinetics of the receptor current in bullfrog saccular hair cells. J. Neurosci. 3, 962–976 (1983). This study provided one of the first confirmations of the existence of a channel directly activated by mechanical stimuli, measured in hair cells from the vestibular system of a bullfrog. 116. Cunningham, C. L. & Müller, U. Molecular structure of the hair cell mechanoelectrical transduction complex. Cold Spring Harb. Perspect. Med. 9, a033167 (2019). 117. Jia, Y. et al. TMC1 and TMC2 proteins are pore-forming subunits of mechanosensitive ion channels. Neuron 105, 310–321 (2020). A study showing that TMC1 and TMC2 proteins, putative pore-forming components of the MET channel complex, form mechanically activated channels when reconstituted in liposomes. 118. Xiong, W. et al. TMHS is an integral component of the mechanotransduction machinery of cochlear hair cells. Cell 151, 1283–1295 (2012). 119. Cunningham, C. L. et al. TMIE defines pore and gating properties of the mechanotransduction channel of mammalian cochlear hair cells. Neuron 107, 126–143 (2020). 120. Driscoll, M. & Chalfie, M. The mec-4 gene is a member of a family of Caenorhabditis elegans genes that can mutate to induce neuronal degeneration. Nature 349, 588–593 (1991). In this study, mec-4, the founding member of the DEG gene family of mechanoreceptors in C. elegans, was cloned. 121. Chen, Y., Bharill, S., Isacoff, E. Y. & Chalfie, M. Subunit composition of a DEG/ENaC mechanosensory channel of Caenorhabditis elegans. Proc. Natl Acad. Sci. USA 112, 11690–11695 (2015). 122. Ben-Shahar, Y. Sensory functions for degenerin/epithelial sodium channels (DEG/ENaC). Adv. Genet. 76, 1–26 (2011). 123. Lin, S. H. et al. Evidence for the involvement of ASIC3 in sensory mechanotransduction in proprioceptors. Nat. Commun. 7, 11460 (2016). 124. Knoepp, F. et al. Shear force sensing of epithelial Na+ channel (ENaC) relies on N-glycosylated asparagines in the palm and knuckle domains of αENaC. Proc. Natl Acad. Sci. USA 117, 717–726 (2020). 125. Martinac, B., Adler, J. & Kung, C. Mechanosensitive ion channels of E. coli activated by amphipaths. Nature 348, 261–263 (1990). 126. Martinac, B. et al. Tuning ion channel mechanosensitivity by asymmetry of the transbilayer pressure profile. Biophys. Rev. 10, 1377–1384 (2018). 127. Nomura, T. et al. Differential effects of lipids and lyso-lipids on the mechanosensitivity of the mechanosensitive channels MscL and MscS. Proc. Natl Acad. Sci. USA 109, 8770–8775 (2012). 128. Syeda, R. et al. Piezo1 channels are inherently mechanosensitive. Cell Rep. 17, 1739–1746 (2016). 129. Cantor, R. S. The influence of membrane lateral pressures on simple geometric models of protein conformational equilibria. Chem. Phys. Lipids 101, 45–56 (1999). 130. Ridone, P. et al. “Force-from-lipids” gating of mechanosensitive channels modulated by PUFAs. J. Mech. Behav. Biomed. Mater. 79, 158–167 (2018). 131. Pliotas, C. et al. The role of lipids in mechanosensation. Nat. Struct. Mol. Biol. 22, 991–998 (2015). 132. Pliotas, C. & Naismith, J. H. Spectator no more, the role of the membrane in regulating ion channel function. Curr. Opin. Struct. Biol. 45, 59–66 (2017). This review discusses an entropy-based mechanism of mechanotransduction in which lipids dissociate from hydrophobic pockets, inducing conformational changes in mechanically-activated channels.

576 | Nature | Vol 587 | 26 November 2020

133. Haselwandter, C. A. & MacKinnon, R. Piezo’s membrane footprint and its contribution to mechanosensitivity. eLife 7, e41968 (2018). 134. Lin, Y.-C. C. et al. Force-induced conformational changes in PIEZO1. Nature 573, 230–234 (2019). This study uses atomic force microscopy to both induce membrane tension and measure its effects on reconstituted PIEZO1, providing evidence that PIEZO1 expands under tension. 135. Hu, J., Chiang, L. Y., Koch, M. & Lewin, G. R. Evidence for a protein tether involved in somatic touch. EMBO J. 29, 855–867 (2010). 136. Cox, C. D., Bavi, N. & Martinac, B. Biophysical principles of ion-channel-mediated mechanosensory transduction. Cell Rep. 29, 1–12 (2019). 137. Li Fraine, S., Patel, A., Duprat, F. & Sharif-Naeini, R. Dynamic regulation of TREK1 gating by polycystin 2 via a filamin A-mediated cytoskeletal mechanism. Sci. Rep. 7, 17403 (2017). 138. Lopes, C. M. B. et al. PIP2 hydrolysis underlies agonist-induced inhibition and regulates voltage gating of two-pore domain K+ channels. J. Physiol. 564, 117–129 (2005). 139. Borbiro, I., Badheka, D. & Rohacs, T. Activation of TRPV1 channels inhibits mechanosensitive Piezo channel activity by depleting membrane phosphoinositides. Sci. Signal. 8, ra15 (2015). 140. Anishkin, A. & Kung, C. Stiffened lipid platforms at molecular force foci. Proc. Natl Acad. Sci. USA 110, 4886–4892 (2013). This article proposes an innovative model for mechanotransduction in which cholesterolrich platforms, maintained by cholesterol-binding scaffold proteins and localized to focal adhesions or adherens junctions, provide specialized force-sensing domains. 141. Petersen, E. N., Chung, H. W., Nayebosadri, A. & Hansen, S. B. Kinetic disruption of lipid rafts is a mechanosensor for phospholipase D. Nat. Commun. 7, 13873 (2016). 142. Chen, X. et al. A feedforward mechanism mediated by mechanosensitive ion channel PIEZO1 and tissue mechanics promotes glioma aggression. Neuron 100, 799–815 (2018). 143. Poole, K., Herget, R., Lapatsina, L., Ngo, H. D. & Lewin, G. R. Tuning Piezo ion channels to detect molecular-scale movements relevant for fine touch. Nat. Commun. 5, 3520 (2014). 144. Qi, Y. et al. Membrane stiffening by STOML3 facilitates mechanosensation in sensory neurons. Nat. Commun. 6, 8512 (2015). 145. Beaulieu-Laroche, L. et al. TACAN is an ion channel involved in sensing mechanical pain. Cell 180, 956–967 (2020). 146. Patkunarajah, A. et al. TMEM87a/Elkin1, a component of a novel mechanoelectrical transduction pathway, modulates melanoma adhesion and migration. eLife 9, e53308 (2020). 147. Xu, J. et al. GPR68 senses flow and is essential for vascular physiology. Cell 173, 762–775 (2018). 148. Bavi, O., Vossoughi, M., Naghdabadi, R. & Jamali, Y. The combined effect of hydrophobic mismatch and bilayer local bending on the regulation of mechanosensitive ion channels. PLoS ONE 11, e0150578 (2016). Acknowledgements This work was supported by NIH grants R01 HL143297. A.P. is an investigator of the Howard Hughes Medical Institute. We thank J. Grandl, S. Murthy, S. Jojoa-Cruz and A. Gharpure for critical reading of the manuscript. Author contributions J.M.K, A.B.W. and A.P. conceptualized the content of this work. J.M.K. reviewed the literature and drafted the manuscript and figures. J.M.K., A.B.W. and A.P. discussed, wrote and edited the Review. Competing interests The authors declare no competing interests. Additional information Correspondence and requests for materials should be addressed to A.B.W. or A.P. Peer review information Nature thanks Boris Martinac and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Reprints and permissions information is available at http://www.nature.com/reprints. Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. © Springer Nature Limited 2020

Article

Experimental evidence of neutrinos produced in the CNO fusion cycle in the Sun https://doi.org/10.1038/s41586-020-2934-0

The Borexino Collaboration*

Received: 26 June 2020 Accepted: 2 October 2020 Published online: 25 November 2020 Check for updates

For most of their existence, stars are fuelled by the fusion of hydrogen into helium. Fusion proceeds via two processes that are well understood theoretically: the proton– proton (pp) chain and the carbon–nitrogen–oxygen (CNO) cycle1,2. Neutrinos that are emitted along such fusion processes in the solar core are the only direct probe of the deep interior of the Sun. A complete spectroscopic study of neutrinos from the pp chain, which produces about 99 per cent of the solar energy, has been performed previously3; however, there has been no reported experimental evidence of the CNO cycle. Here we report the direct observation, with a high statistical significance, of neutrinos produced in the CNO cycle in the Sun. This experimental evidence was obtained using the highly radiopure, large-volume, liquid-scintillator detector of Borexino, an experiment located at the underground Laboratori Nazionali del Gran Sasso in Italy. The main experimental challenge was to identify the excess signal—only a few counts per day above the background per 100 tonnes of target—that is attributed to interactions of the CNO neutrinos. Advances in the thermal stabilization of the detector over the last five years enabled us to develop a method to constrain the rate of bismuth-210 contaminating the scintillator. In the CNO cycle, the fusion of hydrogen is catalysed by carbon, nitrogen and oxygen, and so its rate—as well as the flux of emitted CNO neutrinos—depends directly on the abundance of these elements in the solar core. This result therefore paves the way towards a direct measurement of the solar metallicity using CNO neutrinos. Our findings quantify the relative contribution of CNO fusion in the Sun to be of the order of 1 per cent; however, in massive stars, this is the dominant process of energy production. This work provides experimental evidence of the primary mechanism for the stellar conversion of hydrogen into helium in the Universe.

The nuclear fusion mechanisms that are active in stars, the pp chain and the CNO cycle, are associated with the production of energy and the emission of a rich spectrum of electron-flavour neutrinos1,2 (Fig. 1, bottom). The relative importance of these two mechanisms depends mostly on stellar mass and on the abundance of elements in the core that are heavier than helium (the ‘metallicity’). For stars that are similar to the Sun but are heavier than about 1.3 solar masses4 (M☉), the energy production rate is dominated by the CNO cycle, whereas the pp chain prevails in lighter, cooler stars. The CNO cycle is thought to be the primary mechanism for the stellar conversion of hydrogen into helium in the Universe and is estimated to account for 1% of energy production in the Sun; however, the uncertainty is large because the metallicity of the Sun is poorly known. Metallicity is relevant for two reasons: first, ‘metals’ (namely, carbon, nitrogen and oxygen nuclei) act directly as catalysts in the CNO cycle; and second, they affect the plasma opacity, indirectly changing the temperature of the core and modifying the evolution of the Sun and its density profile. We note that, in the Sun, the CNO sub-cycle I (Fig. 1, top) is dominant5. The flux of CNO neutrinos scales with metal abundance in the solar core, which is itself a tracer of the initial chemical composition of the

Sun at the time of its formation. The metal abundance in the core is thought to be decoupled from the surface by a radiative zone in which no mixing occurs. CNO neutrinos are therefore a unique probe of the initial condition. The neutrinos that are produced by the solar pp chain have been extensively studied since the early 1970s, leading to the discovery of nuclear fusion reactions in the Sun and of matter-enhanced neutrino flavour conversion6–14. Recently, the Borexino experiment has reported a comprehensive study of neutrinos from the pp chain3. We report here the direct detection of neutrinos from the solar CNO cycle, providing direct evidence of the catalysed hydrogen fusion that was proposed independently by Bethe and Weizsäcker in the 1930s15,16. This result quantifies the rate of the CNO cycle in the Sun and paves the way for a solution to the long-standing ‘solar metallicity problem’2—the discrepancy between the physical properties (for example, the solar sound speed profile and the depth and composition of the external convective envelope) predicted by solar models using updated (low) metal abundances from spectroscopy (low-metallicity standard solar model, SSM-LZ)17, and those inferred from helioseismology, which favours a higher metal content (high-metallicity standard solar model,

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

Nature | Vol 587 | 26 November 2020 | 577

Article SSM-HZ). Despite detailed studies, this discrepancy remains an open problem in solar physics. Our experimental observation of CNO neutrinos confirms the overall solar picture and shows that, with future experimental improvements, a direct measurement of the metallicity of the Sun’s core could be within reach.

Reaction rate (×1034 s–1) 10–1

13C

578 | Nature | Vol 587 | 26 November 2020

Flux (cm–2 s–1 (100 keV)–1)

E+ pep

13N 15O 17F

103

16O

0.05%

7Be 7Be

109

105

(p,J) 15N

99.95% pp

17F

(p,J)

Q(15O)

(p,J)

CNO sub-cycle II

(p,α)

107

17O

(p,α)

15O

E+

CNO sub-cycle I

12C

1011

102

(p,J)

E+

Q(13N)

13N

101

14N

(p,J)

Borexino detector and data Borexino is a solar neutrino experiment, located underground at the Laboratori Nazionali del Gran Sasso in Italy, in which the cosmic muon flux is suppressed by a factor of around 106. The active core of the detector consists of approximately 280 t of liquid scintillator contained in a spherical nylon vessel with a radius of 4.25 m. Particles that interact in the scintillator emit light, which is detected by 2,212 photomultiplier tubes18. Solar neutrinos are detected by Borexino via their elastic scattering off electrons. The total number of detected photons and their arrival times are used to reconstruct the electron recoil energy and the interaction point in the detector, respectively. The energy (E) and spatial resolution (σ) of Borexino has slowly deteriorated over time owing to the steady loss of photomultiplier tubes (on average 1,238 channels are active for this analysis), with current values of σE / E ≈ 6% and σx,y,z ≈ 11 cm for 1 MeV events at the centre of the detector. The time profile of the scintillation light provides a powerful way to distinguish between different particle types (α, β− and β+) via pulse-shape discrimination methods19,20, and is essential for the selection of 210Po α decays that are used to constrain the 210Bi background, as discussed below. Despite the very large number of solar neutrinos that reach the Earth, around 6 × 1010 cm−2 s−1, their interaction rate is low—namely a few tens of counts per day (cpd) in 100 t of scintillator. Their detection is especially challenging because the signals from neutrinos cannot easily be disentangled from those of radioactive backgrounds. The success of the Borexino experiment is the result of its unprecedented radiopurity combined with the careful selection of materials21 and clean assembly protocols. This Article is based on data collected during Phase-III of the Borexino experiment, which ran from July 2016 to February 2020 and corresponds to 1,072 days of live time. The event sample is filtered by applying a set of selection criteria20 that reduce events from residual radioactive impurities, cosmic muons, cosmogenic isotopes, instrumental noise and external γ-rays. The latter are substantially suppressed by selecting events that occur within an innermost volume of the scintillator (the fiducial volume) as defined by a cut on the reconstructed radius and vertical position (r < 2.8 m and −1.8 m < z < 2.2 m). The data are analysed in the electron recoil energy interval between 320 keV and 2,640 keV. The counting rate of events that survive the selection as a function of their visible energy is shown in Fig. 2. The data distribution is understood to be the sum of solar neutrino components and of backgrounds resulting from the decays of residual radioactive contaminants in the scintillator (85Kr, 210Bi, 210Po and 40K) and of cosmogenic 11C, and from γ-rays arising from the decays of 40K, 214Bi and 208Tl in the materials external to the scintillator. These backgrounds were characterized in Phase-II of the Borexino experiment20 and their counting rates range between a few cpd and tens of cpd per 100 t, compared with the expected CNO signal of a few cpd per 100 t. The key backgrounds for this study are 11C and 210Bi. Together with solar pep neutrinos (produced by the proton–electron–proton reaction, an alternative first step of the pp chain) they represent the main obstacle in the extraction of the CNO signal, as discussed in the following section. The expected background due to the elastic scattering of 40K geo-antineutrinos22 is negligible. The yellow vertical band in Fig. 2 highlights the region of largest CNO signal-to-background ratio.

100

Q(17F)

10–2

ec13N

8B

ec15O+17F hep

101 102

103 Neutrino energy (keV)

104

Fig. 1 | CNO nuclear fusion sequences and the energy spectra of solar neutrinos. Top, the double CNO cycle in the Sun, in which sub-cycle I is dominant. The coloured arrows indicate the reaction rates integrated over the volume of the Sun. The rate of the 17O(α, p)14N reaction (dashed arrow) is lower than can be shown on the colour scale. ν, neutrino. Bottom, energy spectra of solar neutrinos from the pp chain (grey; representing pp, pep, 7Be, 8B and 3He– proton (hep) neutrinos and from the CNO cycle (in colour). The two dotted lines indicate electron capture (ec)39–41. For mono-energetic lines the flux is given in cm−2 s−1.

CNO neutrino detection and the 210Bi challenge Neutrinos from the CNO cycle have a broad energy spectrum that ranges between 0 and 1,740 keV (see Fig. 1, bottom). Consequently, the recoil energy of electrons has a rather featureless continuous distribution that extends up to 1,517 keV (Fig. 2). In this work, the three CNO neutrino components (Fig. 1) were treated as a single contribution by fixing the ratio between them according to the SSM prediction1,2. Several backgrounds contribute to the same energy interval, with a rate comparable to or larger than the signal. To disentangle all contributions, we fit the data with a procedure similar to that adopted in refs. 3,20,23 and described in Methods. The CNO analysis is affected by two additional complications: the similarity between the spectra of the CNO-neutrino recoil electron and the 210Bi β− particle, and strong correlations of these spectra with the spectrum of the pep-neutrino recoil electron. In addition, in the high-energy region of the CNO spectrum, the data are contaminated by signals from cosmogenic 11C. The muon–neutron–positron threefold-coincidence tagging technique20 for 11C is essential to enable the detection of CNO neutrinos.

200

300

400

500

600

CNO-Q pep-Q 210Bi

103 Events/(5Nh)

Nh

700

800

7Be-Q

and 8Be-Q

900

11C

External backgrounds Other backgrounds

Total fit: P value of 0.3

102

10

1 500

1,000

1,500 Energy (keV)

2,000

2,500

Fig. 2 | Spectral fit of the Borexino data. Distribution of the electron recoil energy scattered by solar neutrinos in the Borexino detector (black points) and corresponding spectral fit (magenta). CNO-neutrinos, 210Bi and pep-neutrinos are highlighted in solid red, dashed blue and dotted green, respectively, and all other components are in grey. The energy estimator Nh represents the number of photoelectrons detected by photomultipliers, normalized to 2,000 live channels. The yellow band represents the region with the largest signal-tobackground ratio for CNO-neutrinos.

The sensitivity to CNO neutrinos is low unless the 210Bi and pep-neutrino rates are sufficiently constrained in the fit24. The pep-neutrino rate is constrained to 1.4% precision24, using solar luminosity, robust assumptions on the ratio of pp-neutrino rate to pep-neutrino rate, existing solar neutrino data25,26 and the most recent oscillation parameters27. We emphasize that the luminosity of the Sun depends very weakly on the contribution of the CNO cycle, making the pep constraint essentially independent of any reasonable assumption of the CNO rate. The other main source of background in the measurement of CNO neutrinos arises from the decays of 210Bi (ref. 24), a β emitter with a short half-life (5.013 days), the decay rate of which is supported by 210 Pb through the sequence: 210

β−

Pb ⟶

22.3 yr

210

β−

α

5d

138.4 d

Bi⟶ 210Po ⟶

206

Pb.

(1)

We note that the endpoint energy of the β-decay of 210Pb is 63.5 keV, which is well below the analysis threshold (320 keV). Therefore, in order to determine the 210Bi content we must rely on measuring 210Po (ref. 28). The α particles from 210Po decay, selected event-by-event by means of pulse-shape discrimination, are ideal tracers of 210Bi, provided that the secular equilibrium in equation (1) is achieved. It is therefore crucial to understand under what conditions such an equilibrium is established. Since the beginning of the Borexino experiment in 2007, data have indicated the presence of out-of-equilibrium components of 210Po in the scintillator. A dedicated effort was therefore implemented to study these components and, by stabilizing the detector temperature, to ultimately prevent them from migrating into the fiducial volume. This upgrade enabled us to reach a sufficient equilibrium in one central sub-volume of the detector, and therefore to obtain the result that we report here. We distinguish between a scintillator (S) 210Po component (210PoS), which originates from 210Pb in the liquid and is assumed to be stable over time and in equilibrium with 210Bi, and a vessel (V) 210Po component (210PoV). The source of 210PoV for this dataset is understood to be the 210Pb that is deposited on the inner surfaces of the vessel. The daughter 210Po may detach and move into the scintillator by diffusion or by following slow convective currents. It is important to note that, as explained in detail in section ‘210Bi constraint’ below, there is no evidence of 210Pb itself leaching from these surfaces, as the rate of 210Bi observed in the scintillator has shown negligible change over several years.

The diffusion length of 210Po atoms in one half-life is considerably less than the separation between the vessel and the fiducial volume (approximately 1 m). We can therefore conclude that diffusion is negligible for both 210Po and 210Bi. However, data from the Borexino experiment show that slow convective currents—caused by temperature gradients and temperature variations—might in fact carry 210Po into the fiducial volume. This does not occur for the short-lived 210Bi, which might also detach from the vessel, because it decays before reaching the fiducial volume. Before 2016, Borexino was not equipped with detailed temperature mapping, thermal insulation or active temperature control. Convective currents were substantial, because of seasonal temperature variations and human activities affecting the temperature of the experimental hall. The large fluctuations in the activity of 210Po in the fiducial volume that were induced by these currents are shown in Fig. 3, in which the 210 Po rate at different detector positions is plotted as a function of time. It is evident that, before 2016, the 210Po counts in the fiducial volume were both high (>100 cpd per 100 t) and very unstable on timescales shorter than the 210Po half-life, because of sizeable fluid movements that prevented the separation of the scintillator Po component from the vessel Po component. In order to suppress convection, it was necessary to establish a stable vertical thermal gradient. The Borexino installation sits atop a cold floor in contact with rock that can act as an infinite thermal sink, thereby providing an opportunity to achieve such a gradient if the detector is insulated against instabilities in air temperature. Thermal insulation of the detector was completed in December 2015, and in January 2016 an active temperature control system29 was installed on top of the detector (see Methods). A residual seasonal modulation of the order of 0.3 °C over 6 months is still visible in the detector and in the rock below it, but its effect is small for the purpose of the results reported here. After this extensive stabilization effort, the 210Po rate initially decreased and reached its lowest value in a region that we named the low polonium field (LPoF), above the equator at z ≈ +80 cm. The existence of this volume, which is compatible in terms of size and location with fluid dynamics simulation30, is crucial in determining the 210Bi constraint. We note that the result we report here is stable to small variations in the shape and the location of the low polonium field. 210

Bi constraint

The amount of 210Bi in the scintillator is determined from the minimum value of the 210Po rate R(210Pomin)in the low polonium field through the relation:

R(210Pomin) = R(210Bi) + R(210PoV ),

(2)

where the 210Bi rate is equal to 210PoS according to secular equilibrium. Because 210PoV is always positive, 210Pomin yields an upper limit for the 210 Bi rate. The 210Po content is not spatially uniform within the low polonium field but exhibits a clear minimum with no sizable plateau around it. This yields a robust upper limit for the rate of 210Bi, but does not guarantee that 210PoV is actually zero. Only a spatially extended minimum of the 210Po rate would have yielded a measurement of the 210Bi rate. The minimum 210Po rate was estimated from the 210Po distribution within the low polonium field using 2D and 3D fits following two mutually compatible procedures (see Methods). The spatial position of the minimum is stable over the analysis period—it moves slowly by less than 20 cm per month—which shows that the detector is in a fluid-dynamical quasi-steady-state condition and that the 210Po rate minimum is not a statistical fluctuation. Both procedures consistently yield R(210Pomin) = 11.5 ± 1.0 cpd per 100 t . The error includes the systematic uncertainty of the fit (see Methods). Nature | Vol 587 | 26 November 2020 | 579

Article 140

15.0 50

(5)

120

(3) 14.4

20

100

80

60

210Po

14.6 30

Average temperature (°C)

Cube label

(4)

rate (cpd per 100 t)

14.8 40

40 (2)

14.2

10

20 (1) 0 24 April 2013

8 September 2014

23 January 2016

8 June 2017

23 October 2018

14.0 8 March 2020

0

Fig. 3 | Spatial and temporal distribution of 210Po activity. 210Po rate in Borexino in cpd per 100 t (rainbow colour scale) as a function of time. Small cubes of about 3 t each are ordered from the bottom (0) to the top (58) along the vertical direction (latest update: March 2020). All cubes are selected inside a sphere of radius r = 3 m. The red curve represents the average temperature in the innermost region surrounding the nylon vessel. The dashed vertical lines indicate the most important milestones of the temperature stabilization

program: (1) beginning of the ‘insulation program’; (2) turning off of the water recirculation system in the water tank; (3) first operation of the active temperature control system; (4) change of the active control set point; (5) installation and commissioning of the hall temperature control system. The white vertical bands represent different interruptions to the data acquisition due to technical issues.

The 210Bi rate can then be extrapolated over the whole fiducial volume, provided that it is uniform in the fiducial volume during the time period over which the estimation is performed. Because it is not possible to individually tag 210Bi events, the analysis is performed by selecting β-like events at energies at which the relative bismuth contribution is maximal. We find that the angular and spatial distribution of 210Bi is uniform within errors. The systematic uncertainty associated with possible spatial non-uniformity of 210Bi is conservatively estimated at 0.78 cpd per 100 t. The observed 210Bi uniformity in Phase-III of the Borexino experiment is expected as a result of the substantial fluid mixing that occurred before the thermal insulation, and agrees with 2D and 3D fluid dynamic simulations. Because of the low velocity of convection currents, the uniformity of 210Bi provides convincing evidence that 210Pb does not leach off the vessel. As a cross-check, the rate of β-like events shows the expected annual modulation of the solar neutrino rate of 3.3%—dominated by 7 Be-neutrinos—as a result of the eccentricity of the Earth’s orbit, thus proving that background β-like events are stable in time. Further details are provided in Methods. In summary, the 210Bi rate used as a constraint in the CNO-neutrino analysis is

following rates were treated as free parameters: CNO-neutrinos, 85Kr, 11 C, internal and external 40K, external 208Tl and 214Bi and 7Be-neutrinos. The pep-neutrino rate is constrained to 2.74 ± 0.04 cpd per 100 t by multiplying the standard likelihood by a symmetric Gaussian term. The upper limit to the 210Bi rate obtained from equation (3) is enforced asymmetrically by multiplying the likelihood by a half-Gaussian term—that is, leaving the 210Bi rate unconstrained between 0 and 11.5 cpd per 100 t. The reference spectral and radial probability density functions of each signal and background species that are used in the multivariate fit are obtained with a complete Monte Carlo simulation based on Geant420,31. The results of the multivariate fit for data in which the 11 C contribution has been subtracted with the threefold-coincidence technique are shown in Fig. 2. The P value of the fit is high (0.3), which demonstrates good agreement between the data and the underlying fit model. The corresponding negative log-likelihood for CNO-neutrinos, profiled over the other neutrino fluxes and background sources, is shown in Fig. 4. The best fit value is 7.2 cpd per 100 t, with an asymmetric confidence interval of −1.7 cpd per 100 t and +2.9 cpd per 100 t (68% confidence level, statistical error only), obtained from the quantile of the likelihood profile. We studied possible sources of systematic error following an approach similar to that used in refs. 3,20. By performing 2,500 fits with different fit ranges and binning values, we found that the effect of varying these fit parameters was negligible with respect to the CNO statistical uncertainty. We also considered the effect of different theoretical 210Bi shapes from refs. 32–34 and found that the CNO result is robust with respect to the selected shape32. Differences in CNO rate are included in the systematic error. We performed a detailed study of the effect of possible deviations of the energy scale and resolution from the Monte Carlo model: nonlinearity, non-uniformity and variation in the absolute magnitude of the scintillator light yield were investigated by simulating several million Monte Carlo pseudo-experiments with deformed shapes and fitting them with the regular non-deformed

R( 210Bi) ≤ 11.5 ± 1.3 cpd per 100 t,

(3)

which includes the statistical and systematic uncertainties in the determination of the 210Po minimum, and the systematic uncertainty related to the 210Bi uniformity hypothesis (added in quadrature).

Results and conclusions We performed a multivariate analysis, simultaneously fitting the energy spectra in the window between 320 keV and 2,640 keV and the radial distribution of the selected events (see Methods for details). The 580 | Nature | Vol 587 | 26 November 2020

40 Room for CNO

30

600

25 –2Δln

pep-Q 400 210Bi

200 Others 0

Fit without systematics Fit with systematics HZ-SSM 68% CI LZ-SSM 68% CI Borexino 68% CI Counting analysis

35

Events in the ROI

20

0.06 0.05 0.04 0.03

15 0.02

10

0.01

5 0 0

Counting PDF

800

2

4 6 8 10 CNO-Q rate (cpd per 100 t)

12

0.00 14

Fig. 4 | Results of the CNO counting and spectral analyses. Left, counting analysis bar chart. The height represents the number of events allowed by the data for CNO-neutrinos and backgrounds in the region of interest (ROI). On the left, the CNO signal is minimum and backgrounds are maximum; on the right, CNO is maximum and backgrounds are minimum. It is clear from this figure that the contribution of CNO neutrinos cannot be zero. Right, CNO-neutrino rate negative log-likelihood (lnL) profile obtained directly from the

multivariate fit (dashed black line) and after folding in the systematic uncertainties (black solid line). The histogram in red shows the CNO-neutrino rate obtained from the counting analysis. The blue, violet and grey vertical bands show 68% confidence intervals (CI) for the SSM-LZ (3.52 ± 0.52 cpd per 100 t) and SSM-HZ (4.92 ± 0.78 cpd per 100 t)2,24 predictions and the Borexino result (corresponding to the black solid-line log-likelihood profile), respectively.

probability density functions. The magnitude of the deformations was chosen to be within the range allowed by the available calibrations35 and by two ‘standard candles’ (210Po, 11C) present in the data. The overall contribution to the total error of all these sources is +0.6 −0.5 cpd per 100 t . Folding the systematic uncertainty over the log-likelihood profile we determine the final CNO interaction rate to be 7.2+3.0 −1.7 cpd per 100 t . This rate can be converted to a flux of CNO-neutrinos on Earth of 8 −2 −1 7.0+3.0 −2.0 × 10 cm s , assuming Mikheyev–Smirnov–Wolfenstein conversion in matter36, neutrino oscillation parameters from ref. 37 and references therein, and a density of electrons in the scintillator of (3.307 ± 0.015) × 1031 e− per 100 t. Other sources of systematic error that were investigated in the previous precision measurement of the pp chain3—such as fiducial volume, scintillator density and lifetime—were found to be negligible with respect to the large CNO statistical uncertainty. The log-likelihood profile including all the errors combined in quadrature is shown in Fig. 4. The asymmetry of the profile is a result of applying a half-Gaussian constraint on the 210Bi (see equation (3)) and causes the profile to be relatively steep on the left-hand side of the minimum. The shallow curve on the right-hand side of the profile reflects the modest sensitivity in distinguishing the spectral shapes of the 210Bi and the CNO recoil spectra. From the corresponding profile-likelihood we obtain a significance of 5.1σ for the CNO observation. Additionally, a hypothesis test based on a profile likelihood test statistics38—using 13.8 million pseudo-datasets with the same exposure as the Phase-III experiment and including systematic uncertainties (see Methods)— excludes the no-CNO signal scenario with a significance greater than 5.0σ at a 99.0% confidence level. The observed CNO rate is compatible with both SSM-HZ and SSM-LZ predictions, and as such we cannot distinguish between the two different models: the statistical compatibility for HZ is 0.5σ and for LZ is 1.3σ (Fig. 4). When combined with other solar neutrino fluxes measured by the Borexino experiment, the LZ hypothesis is disfavoured at a level of 2.1σ. We emphasize that the sensitivity to CNO neutrinos arises mainly from a small energy region between 780 keV and 885 keV (the region of interest; see yellow band in Fig. 2) at which the signal-to-background ratio is maximal24. In this region, the count rate is dominated by events from CNO and pep neutrinos, and by 210Bi decays. The remaining backgrounds contribute less than 20% (Fig. 4). A simple counting analysis confirms that the number of events in the region of interest exceeds the sum from all known backgrounds, leaving room for CNO neutrinos (Fig. 4). In this simplified approach (described in detail in Methods), we

use the 210Bi rate (as in equation (3)) and apply a symmetric Gaussian penalty while assuming an analytical description of the background model and the detector response. After accounting for statistical and systematic errors, the statistical significance of the presence in the data of CNO neutrino events from this counting analysis is around 3.5σ, which is lower than that obtained with the main analysis given the simplified nature of this approach. In conclusion, the absence of a CNO solar neutrino signal is excluded with a significance of 5.0σ. We therefore present a direct detection of CNO solar neutrinos.

Online content Any methods, additional references, Nature Research reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at https://doi.org/10.1038/s41586-020-2934-0. 1. 2. 3. 4. 5. 6. 7. 8. 9.

10. 11. 12. 13. 14. 15. 16. 17.

Bahcall, J. N. Neutrino Astrophysics (Cambridge Univ. Press, 1989). Vinyoles, N. et al. A new generation of standard solar models. Astrophys. J. 835, 202 (2017). The Borexino Collaboration. Comprehensive measurement of pp-chain solar neutrinos. Nature 562, 505–510 (2018). Salaris, M. & Cassisi, S. Evolution of Stars and Stellar Populations (John Wiley & Sons, 2005). Angulo, C. et al. A compilation of charged-particle induced thermonuclear reaction rates. Nucl. Phys. A 656, 3–183 (1999). Davis, R. Jr A Half-Century with Solar Neutrinos. Nobel Prize Lecture https://www. nobelprize.org/prizes/physics/2002/davis/lecture/ (2002). GALLEX collaboration. Solar neutrinos observed by GALLEX at Gran Sasso. Phys. Lett. B 285, 376 (1992). SAGE collaboration. Results from SAGE (The Russian–American Gallium solar neutrino experiment). Phys. Lett. B 328, 234 (1994). McDonald, A. B. The Sudbury Neutrino Observatory: Observation of Flavor Change for Solar Neutrinos. Nobel Prize Lecture https://www.nobelprize.org/prizes/physics/2015/ mcdonald/lecture/ (2015). Hirata, K. et al. Observation of 8B solar neutrinos in the Kamiokande-II detector. Phys. Rev. Lett. 63, 16 (1989). Ahmad, Q. et al. Direct evidence for neutrino flavor transformation from neutral-current interactions in the Sudbury Neutrino Observatory. Phys. Rev. Lett. 89, 011301 (2002). Araki, T. et al. Measurement of neutrino oscillation with KamLAND: evidence of spectral distortion. Phys. Rev. Lett. 94, 081801 (2005). Borexino Collaboration. Neutrinos from the primary proton–proton fusion process in the Sun. Nature 512, 383–386 (2014). Bellini, G. et al. Precision measurement of the 7Be solar neutrino interaction rate in Borexino. Phys. Rev. Lett. 107, 141302 (2011). Bethe, H. A. Energy production in stars. Phys. Rev. 55, 434–456 (1939). von Weizsäcker, C. F. Über Elementumwandlungen im Innern der Sterne I. Phys. Z. 38, 176 (1937). Serenelli, A. M., Haxton, W. C. & Peña-Garay, C. Solar models with accretion. I. application to the solar abundance problem. Astrophys. J. 743, 24 (2011).

Nature | Vol 587 | 26 November 2020 | 581

Article 18. Alimonti, G. et al. The Borexino detector at the Laboratori Nazionali del Gran Sasso. Nucl. Instrum. Methods Phys. Res. A 600, 568–593 (2009). 19. Bellini, G. et al. Final results of Borexino Phase-I on low-energy solar neutrino spectroscopy. Phys. Rev. D 89, 112007 (2014). 20. Agostini, M. et al. Simultaneous precision spectroscopy of pp, 7Be and pep solar neutrinos with Borexino Phase-II. Phys. Rev. D 100, 082004 (2019). 21. Alimonti, G. et al. Science and technology of BOREXINO: a real-time detector for low energy solar neutrinos. Astropart. Phys. 16, 205–234 (2002). 22. Agostini, M. et al. Comprehensive geoneutrino analysis with Borexino. Phys. Rev. D 101, 012009 (2020). 23. Ding, X. F. GooStats: A GPU-based framework for multi-variate analysis in particle physics. J. Instrum. 13, P12018 (2018). 24. Agostini, M. et al. Sensitivity to neutrinos from the solar CNO cycle in Borexino. Eur. Phys. J. C https://doi.org/10.1140/epjc/s10052-020-08534-2 (2020). 25. Vissani, F. Luminosity constraint and entangled solar neutrino signals. In Solar Neutrinos, Proc. 5th International Solar Neutrino Conference (eds Meyer, M. & Zuber, K.) 121–141 (World Scientific, 2019). 26. Bergström, J., Gonzalez-Garcia, M. C., Maltoni, M., Peña-Garay, C., Serenelli, A. M. & Song, N. Updated determination of the solar neutrino fluxes from solar neutrino data. J. High Energy Phys. 2016, 132 (2016). 27. Capozzi, F., Lisi, E., Marrone, A. & Palazzo, A. Global analysis of oscillation parameters. J. Phys. Conf. Ser. 1312, 012005 (2019). 28. Villante, F. L., Ianni, A., Lombardi, F., Pagliaroli, G. & Vissani, F. A step toward CNO solar neutrino detection in liquid scintillators. Phys. Lett. B 701, 336–341 (2011). 29. Bravo-Berguño, D. et al. The Borexino Thermal Monitoring & Management System and simulations of the fluid-dynamics of the Borexino detector under asymmetrical, changing boundary conditions. Nucl. Instrum. Methods Phys. Res. A 885, 38–53 (2018). 30. Di Marcello, V. et al. Fluid-dynamics and transport of 210Po in the scintillator Borexino detector: a numerical analysis. Nucl. Instrum. Methods Phys. Res. A 964, 163801 (2020). 31. Agostini, M. et al. The Monte Carlo simulation of the Borexino detector. Astropart. Phys. 97, 136–159 (2018). 32. Daniel, H. Das β-spektrum des RaE. Nucl. Phys. 31, 293–307 (1962). 33. Grau Carles, A. & Grau Malonda, A. Precision measurement of the RaE shape factor. Nucl. Phys. A 596, 83–90 (1996). 34. Alekseev, I. E. et al. Precision measurement of 210Bi β-spectrum. Preprint at https://arxiv. org/abs/2005.08481 (2020). 35. Back, H. et al. Borexino calibrations: hardware, methods, and results. J. Instrum. 7, P10018 (2012). 36. de Holanda, P. C., Liao, W. & Smirnov, A. Yu. Toward precision measurements in solar neutrinos. Nucl. Phys. B 702, 307–332 (2004). 37. Capozzi, F., Lisi, E., Marrone, A. & Palazzo, A. Current unknowns in the three neutrino framework. Prog. Part. Nucl. Phys. 102, 48–72 (2018). 38. Cowan, G., Cranmer, K., Gross, E. & Vitells, O. Asymptotic formulae for likelihood-based tests of new physics. Eur. Phys. J. C 71, 1554 (2011). 39. Bahcall, J. N. Line versus continuum solar neutrinos. Phys. Rev. D 41, 2964–2966 (1990). 40. Stonehill, L. C., Formaggio, J. A. & Robertson, R. G. H. Solar neutrinos from CNO electron capture. Phys. Rev. C 69, 015801 (2004). 41. Villante, F. L. ecCNO solar neutrinos: a challenge for gigantic ultra-pure liquid scintillator detectors. Phys. Lett. B 742, 279–284 (2015). Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

582 | Nature | Vol 587 | 26 November 2020

© The Author(s), under exclusive licence to Springer Nature Limited 2020 The Borexino Collaboration M. Agostini1,2, K. Altenmüller2, S. Appel2, V. Atroshchenko3, Z. Bagdasarian4,27, D. Basilico5, G. Bellini5, J. Benziger6, R. Biondi7, D. Bravo5,28, B. Caccianiga5, F. Calaprice8, A. Caminata9, P. Cavalcante10,29, A. Chepurnov11, D. D’Angelo5, S. Davini9, A. Derbin12, A. Di Giacinto7, V. Di Marcello7, X. F. Ding8, A. Di Ludovico8, L. Di Noto9, I. Drachnev12, A. Formozov5,13, D. Franco14, C. Galbiati8,15, C. Ghiano7, M. Giammarchi5, A. Goretti8,29, A. S. Göttel4,16, M. Gromov11,13, D. Guffanti17, Aldo Ianni7, Andrea Ianni8, A. Jany18, D. Jeschke2, V. Kobychev19, G. Korga20,21, S. Kumaran4,16, M. Laubenstein7, E. Litvinovich3,22, P. Lombardi5, I. Lomskaya12, L. Ludhova4,16, G. Lukyanchenko3, L. Lukyanchenko3, I. Machulin3,22, J. Martyn17, E. Meroni5, M. Meyer23, L. Miramonti5, M. Misiaszek18, V. Muratova12, B. Neumair2, M. Nieslony17, R. Nugmanov3,22, L. Oberauer2, V. Orekhov17, F. Ortica24, M. Pallavicini9, L. Papp2, L. Pelicci5, Ö. Penek4,16, L. Pietrofaccia8, N. Pilipenko12, A. Pocar25, G. Raikov3, M. T. Ranalli7, G. Ranucci5 ✉, A. Razeto7, A. Re5, M. Redchuk4,16, A. Romani24, N. Rossi7, S. Schönert2, D. Semenov12, G. Settanta4, M. Skorokhvatov3,22, A. Singhal4,16, O. Smirnov13, A. Sotnikov13, Y. Suvorov3,7,30, R. Tartaglia7, G. Testera9, J. Thurn23, E. Unzhakov12, F. L. Villante7,26, A. Vishneva13, R. B. Vogelaar10, F. von Feilitzsch2, M. Wojcik18, M. Wurm17, S. Zavatarelli9, K. Zuber23 & G. Zuzel18 Department of Physics and Astronomy, University College London, London, UK. Physik-Department E15, Technische Universität München, Garching, Germany. 3National Research Centre Kurchatov Institute, Moscow, Russia. 4Institut für Kernphysik, Forschungszentrum Jülich, Jülich, Germany. 5Dipartimento di Fisica, Università degli Studi and INFN, Milan, Italy. 6Chemical Engineering Department, Princeton University, Princeton, NJ, USA. 7INFN Laboratori Nazionali del Gran Sasso, Assergi, Italy. 8Physics Department, Princeton University, Princeton, NJ, USA. 9Dipartimento di Fisica, Università degli Studi and INFN, Genoa, Italy. 10Physics Department, Virginia Polytechnic Institute and State University, Blacksburg, VA, USA. 11Lomonosov Moscow State University Skobeltsyn Institute of Nuclear Physics, Moscow, Russia. 12St Petersburg Nuclear Physics Institute, NRC Kurchatov Institute, Gatchina, Russia. 13Joint Institute for Nuclear Research, Dubna, Russia. 14Astroparticule et Cosmologie, Université de Paris, CNRS, Paris, France. 15Gran Sasso Science Institute, L’Aquila, Italy. 16III. Physikalisches Institut B, RWTH Aachen University, Aachen, Germany. 17 Institute of Physics and Excellence Cluster PRISMA+, Johannes Gutenberg-Universität Mainz, Mainz, Germany. 18M. Smoluchowski Institute of Physics, Jagiellonian University, Krakow, Poland. 19Institute for Nuclear Research of NAS Ukraine, Kyiv, Ukraine. 20Department of Physics, School of Engineering, Physical and Mathematical Sciences, Royal Holloway, University of London, Egham, UK. 21Institute of Nuclear Research (Atomki), Debrecen, Hungary. 22National Research Nuclear University MEPhI (Moscow Engineering Physics Institute), Moscow, Russia. 23Department of Physics, Technische Universität Dresden, Dresden, Germany. 24Dipartimento di Chimica, Biologia e Biotecnologie, Università degli Studi e INFN, Perugia, Italy. 25Amherst Center for Fundamental Interactions and Physics Department, University of Massachusetts, Amherst, MA, USA. 26Dipartimento di Scienze Fisiche e Chimiche, Università dell’Aquila, L’Aquila, Italy. 27Present address: Department of Physics, University of California, Berkeley, Berkeley, CA, USA. 28Present address: Departamento de Física Teórica, Universidad Autónoma de Madrid, Madrid, Spain. 29 Present address: INFN Laboratori Nazionali del Gran Sasso, Assergi, Italy. 30Present address: Dipartimento di Fisica, Università degli Studi Federico II e INFN, Naples, Italy. ✉e-mail: [email protected] 1

2

Methods Experimental setup and neutrino detection technique The Borexino detector18 was designed and built to achieve the utmost radiopurity at its core. It is made of an unsegmented stainless steel sphere (SSS) mounted within a large water tank. The SSS contains the organic liquid and supports the photomultiplier tubes (PMTs), while the water shields the SSS against external radiation and is the active medium of a Cherenkov muon tagger. A schematic is shown in Extended Data Fig. 1. Within this SSS, two thin (125 μm) nylon vessels separate the volume in three shells of radii 4.25 m, 5.50 m and 6.85 m, the latter being the radius of the SSS itself. The inner nylon vessel, concentric to the SSS, contains a solution of pseudocumene as solvent and 2,5-diphenyloxazole (PPO) as fluor dissolved at a concentration of about 1.5 g l−1. The second and the third shells are filled with a buffer liquid comprised of a solution of dimethylphthalate (DMP) in pseudocumene. The purpose of this double buffer is to shield the inner vessel against γ radiation emitted by contaminants present in the PMTs and the steel, while the outer nylon vessel prevents the diffusion of emanated radon into the inner vessel. The total amount of liquid within the SSS is approximately 1,300 t, of which about 280 t are the active liquid scintillator. The inner vessel scintillator density is slightly smaller than that of the buffer liquid, yielding an upward buoyant force. The inner vessel is therefore anchored to the bottom of the SSS through thin high-molecular-weight polyethylene cords, thus minimising the amount of material close to the scintillator and keeping the inner vessel in stable mechanical equilibrium. The SSS is equipped with nominally 2,212 8-inch (20.3 cm) PMTs that collect scintillation light emitted when a charged particle, either produced by neutrino interactions or by radioactivity, releases energy in the scintillator. Most of the PMTs (1,800) are equipped with light concentrators (Winston cones) for an effective optical coverage of 30%. Scintillation light is detected at approximately 500 photoelectrons per MeV of electron equivalent of deposited energy (normalized to 2,000 PMTs). In organic liquid scintillators, the light yield per unit of deposited energy is affected by ionization quenching42. Alpha particles, characterized by higher ionization rates along their path, experience more quenching compared to electrons and thus produce less scintillation light. The distribution of photon arrival times on PMTs allows the reconstruction of the location of the energy deposit by means of time-of-flight triangulation and the determination of the particle type by exploiting the pulse shape20. The very nature of the scintillation emission makes it impossible to distinguish the signal emitted by electrons scattered by neutrinos from that produced by electrons emitted in nuclear β-decays or Compton-scattered by γ-rays. Therefore, the radioactive background must be kept at or below the level of the expected signal rate, which for the total solar neutrino spectrum is of the order of a few events per tonne per day and, in the case of CNO neutrinos, two orders of magnitude smaller. Taking into account that typical materials (air, water, metals) are normally contaminated with radioactive impurities at the level of 10,000 or even 100,000 decays per tonne per second, this requirement is indeed a formidable challenge. The scintillator procurement procedure was conceived to select an organic hydrocarbon with a very low 14C (β−, with the total energy released (the Q-value) being Q = 156 keV) content. Carbon-14 is cosmogenically activated in atmospheric carbon and an irreducible radioactive contaminant in organic hydrocarbons. The scintillator was delivered to the Gran Sasso laboratory in special tanks following procedures conceived to avoid contamination and to minimize the exposure to cosmic rays, which also produce other long-living isotopes. Once underground, it was purified following various steps in plants specifically developed over more than 10 years for this purpose

and installed close to the detector. The purification during the initial scintillator fill in 2007 was done mainly by distillation and counterflow sparging using low-argon-krypton nitrogen. A dedicated purification campaign in 2010–2011 processed the scintillator through several cycles of ultra-pure water extraction. These purification techniques are described in refs. 20,43,44. After this effort, the extreme purity of the scintillator and the careful selection of the material surrounding it (nylon, plastic supports of the nylon vessels, steel and PMT glass in particular), and the use of carefully selected components (valves, pumps, fittings, etc.) together with special care during detector construction and installation, yielded unprecedented low values of radioactive contaminants in the active scintillator. In addition, through the selection of a fiducial volume, the residual external gamma ray background (from the interval vessel nylon, the SSS and the PMTs) is further substantially reduced. All results from the Borexino experiment can be directly attributed to this unprecedented radiopurity. The water tank is itself equipped with 208 PMTs to detect Cherenkov light emitted by muons crossing the water. The capability to detect muons and to reconstruct their tracks through the scintillator was crucial to identify and tag cosmogenic contaminants (that is, short-lived nuclei produced by muon spallation with scintillator components45,46), especially the 11C background. Muon tagging enables Borexino to also efficiently detect cosmogenic neutrons47, which occasionally are produced with high multiplicity, another crucial ingredient in 11C tagging.

Thermal insulation system and control The thermal stability of the Borexino detector is required to avoid undesired background variations due to the mixing of the scintillator inside the inner vessel. This mixing is caused by convective currents induced by temperature changes due to human activities in the underground hall and to seasonal effects. A substantial upgrade of the detector in this respect was carried out. Between May and December 2015, 900 m2 of thermal insulation was installed on the outside of the Borexino water tank. In addition, the system used to recirculate water inside the water tank was stopped in July 2015 to contribute to the inner detector thermal stability and allow its fluid to vertically stratify. The thermal insulation consists of two layers: an outer 10-cm layer of Ultimate Tech Roll 2.0 mineral wool (thermal conductivity at 10 °C of 0.033 W m−1 K−1) and an inner 10-cm layer of Ultimate Protect wired Mat 4.0 mineral wool reinforced with Al foil (65 g cm−2) with glass grid on one side (thermal conductivity at 10 °C of 0.030 W m−1 K−1). The thermal insulation material is anchored to the water tank with 20-m-long nails on a metal plate attached to the tank (5 nails per m2). In addition, an active temperature control system (ATCS) was completed in January 2016. Extended Data Fig. 2 shows the Borexino water tank wrapped in thermal insulation. A system of 66 probes with 0.07 °C resolution, the position of which is shown in Extended Data Fig. 3, monitors the temperature of Borexino. They are arranged as follows: 14 protruding 0.5 m radially inward into the SSS (ReB probes), in operation since October 2014, measure the temperature of the outer part of the buffer liquid; 14 mounted 0.5 m radially outward from the SSS (ReW probes), in operation since April 2015, measure the temperature of the water; 20 installed between the insulation layer and the external surface of the water tank (WT probes), in operation since May 2015; 4 located inside a pit underneath the Borexino water tank, in operation since October 2015; 14 on the Borexino detector water tank dome, installed in early 2016. Since 2016 the average temperature of the floor underneath the detector in contact with the rock is 7.5 °C, whereas at the top of the detector it is 15.8 °C. This temperature difference corresponds to a naturally driven gradient ΔT/Δz > 0 ≈ 0.5 °C m−1. Ensuring this gradient does not decrease is the key to reducing convective currents and scintillator mixing, and consequently to stabilizing the 210Po background for the CNO analysis.

Article Out of the last 14 probes, three are part of the ATCS. The ATCS consists of a water-based system made with copper tube coils installed on the upper part of the dome of the detector. The coils are in contact with the water tank steel, with the addition of an Al layer to enhance the thermal coupling. A 3-kW electric heater, a circulation pump, a temperature controller and an expansion tank are connected to the coils. The ATCS trims the natural thermal gradient and is essential to eliminate convection motion. The outer detector head tank (a 70-l vessel connected with the 1,346 m3 volume of the SSS) is used as a sensitive detector thermometer. After installation of the thermal insulation system the head tank had to be refilled with 289 kg of pseudocumene because of the overall cooling of the detector and corresponding shrinkage. Calibration established the sensitivity of this thermometer to be of the order of 10−2 °C per 100 mm change of fluid height. The deployment of both the thermal insulation and the temperature control systems were quickly effective in stabilizing the inner detector temperature. As of 2016 the heat loss due to the thermal insulation system was equal to 247 W. However, changes of the experimental hall temperature induced residual variations in the top buffer probes of the order of 0.3 °C over 6 months. To further reduce these effects, an active system to control the seasonal changes in the air temperature entering the experimental hall and surrounding the Borexino water tank was designed and installed in 2019. It consists of a 70-kW electrical heater installed inside the inlet air duct, which has a capacity of 12,000 m3 h−1 (in normal conditions). The heater is deployed just a few metres before the main door to the hall. The temperature control is based on a master/slave architecture with a master PID (proportional–integral–derivative) controller that acts on a second slave PID controller. Probes deployed around the water tank monitor the temperature of the air. After commissioning, a set point temperature for the master PID of 14.5 °C is chosen. This system controls the temperature of the inlet air within approximately 0.05 °C. The thermal insulation, active temperature control of the detector, and control of the air temperature in the hall have enabled remarkable temperature stability of the detector. Extended Data Fig. 4 shows the temperature time profile read by all probes since 2016. A stable temperature gradient was clearly established, as required to avoid mixing of the scintillator.

The low polonium field and its properties After the completion of the thermal insulation (Phase-III), the 210Bi background activity is measured from the 210Po activity assuming secular equilibrium of the mass number A = 210 chain. The measured 210 Po rate is the sum of two contributions: a scintillator 210Po component supported by the 210Pb in the liquid (210PoS), which we assume to be stable in time and equal to the intrinsic rate of 210Bi in the scintillator, and a vessel component (210PoV). The latter has a 3D diffusive-like structure as a result of polonium detaching from the inner vessel and migrating into the fiducial volume. The origin of this component is the 210Pb contamination of the vessel. The 210Po migration process is driven by residual convective currents. A rough estimation of the migration length λmig, obtained by fitting the spatial distribution of 210Po, is found to range between 50 and 100 cm, which corresponds to a migration coefficient Dmig = (1.0 ± 0.4) × 10−9 m2 s−1 (where we have used the relation λ mig = DmigτPo with the 210Po lifetime, τPo = 199.7 days). This value is slightly lower than the diffusion coefficient Ddiff ≈ 1.5 × 10−9 m2 s−1 (corresponding to a diffusion length λdiff ≈ 20 cm), predicted by the Stokes–Einstein formula48 and observed for heavy atoms in hydrocarbons49. We attribute this difference to the presence of residual convective motions in Phase-III. These motions are localized in small regions and create a diffusive-like structure with an effective migration length λmig ≳ λdiff. The α particles from 210Po decays are selected event-by-event with a highly efficient α/β pulse shape discrimination neural network

method based on a multi-layer perceptron (MLP)50. The resulting three-dimensional 210Po activity distribution, named the low polonium field (LPoF), exhibits an effective migration profile with an almost stable minimum located above the detector equator (see 3D shape in Extended Data Fig. 5, and dark blue regions in Extended Data Fig. 6, top). The qualitative shape and approximate position of the LPoF is reproduced by fluid dynamical numerical simulations reported in ref. 30. Assuming azimuthal symmetry around the detector z-axis, confirmed by 3D analysis, the 210Po minimum activity is determined by fitting LPoF with a 2D paraboloidal function:

d2 R(210Po) = d(ρ 2 )dz  ρ 2 (z − z0) 2  [R(210Pomin)ϵEϵMLP + Rβ ] × 1 + 2 +  , 2 b   a

(4)

where ρ2 = x2 + y2, a and b are the paraboloid axes, z0 is the position of the minimum along the z axis, ϵE and ϵMLP are the efficiency of energy and MLP cuts used to select α particles from 210Po decays, and Rβ is the residual rate of β events after the selection of α particles. The fit is initially performed in data bins of 2 months, but compatible results are obtained using the bins of 1 month. Extended Data Fig. 6, top shows the result of the z0 minimum position as a function of time. The minimum slowly moves along the z direction by less than 20 cm per month. In order to perform a better estimation of the 210Po minimum, we sum up all the time bins after aligning the 3D distributions with respect to z0. Possible intrinsic biases, due to the minimum determination in different time intervals, have been minimized by blindly aligning the data from each time bin according to the z0 inferred from the previous time interval. The distribution of 210Po events after applying this procedure is shown in Extended Data Fig. 6, bottom, in which the LPoF structure is clearly visible. The final fit is then performed on 20 t of this aligned dataset containing about 5,000 210Po events. From this fit we extract the 210Po minimum. This value might still have a small contribution from the vessel component (equation (2), that is, R(210Bi) ≤ R(210Pomin)). Therefore this method provides only an upper limit for the 210Bi rate. A companion analysis was performed using a 3D paraboloidal function. The 2D and 3D fits were performed with a standard binned likelihood and a Bayesian approach using non-informative priors. In particular, the latter was implemented with MultiNest51–53, a nested sampling algorithm. In addition, because the shape of the LPoF might show more complexity along the z-axis than a simple paraboloidal shape, a Bayesian framework was also used to perform the fit with a cubic spline along the z axis. Splines are piecewise polynomials connected by knots. The number of knots defines the complexity of the curve. To prevent overfitting, a Bayesian factor analysis was used to decide on the most appropriate number of knots for the dataset. Although it was found that the splines were, in general, a better fit to the data (Bayes factor >102), the final result is compatible with the simpler model within statistical uncertainties. This result has been further cross-checked by fitting the 210Po distribution along different angular directions with a family of analytical functions found as a solution of the Fick diffusion equation54 for the migration of decaying 210Po. Possible biases have been quantified by testing the fit model on simulated LPoF patterns based on numerical fluid dynamical simulations. They were found to be negligible for our purpose.

Spatial uniformity and time stability of 210Bi The 210Bi independent constraint inferred from the LPoF can be extended over the whole fiducial volume if, and only if, the 210Bi itself is uniform in space. Observation of the time stability of the 210Bi rate, not strictly required if the time periods of the LPoF and main analyses

are the same, can additionally cross-check the overall robustness of the dataset. We have evidence that at the beginning of Borexino Phase-II, after the purification campaign performed from 2010 to mid-2011, the 210Bi was not uniform: the cleanest part of the scintillator was concentrated on the top, partially out of the fiducial volume. In fact, the purification was performed in loop, taking the scintillator out from the bottom, purifying it and re-inserting it from the top. For this reason, at the beginning of Phase-II the apparent 210Bi rate was higher and slowly decreased over time as mixing was taking place, thanks to the strong pre-insulation convective currents. This decreasing trend stopped in early 2016, suggesting that the mixing had completed. Numerical fluid dynamical simulations, performed using the velocity field obtained from 210Po movements during the pre-insulation time as input, confirm this hypothesis. A more conservative approach, which uses heuristic arguments based on the effective migration of ions as measured from LPoF, suggests that 210 Bi at the beginning of Phase-III (mid-2016) must be uniform at least within a volume scale of about 20 m3. This argument is also verified by means of fluid dynamics numerical simulations. All the a priori arguments and qualitative studies described above are confirmed a posteriori by looking at the β event rate in optimized energy windows in which the 210Bi signal-to-background ratio is maximal. The observed non-uniformity is then conservatively assigned only to 210Bi, contributing about 15% to the overall rate in the selected energy window. In order to test the spatial uniformity of the 210Bi rate in the fiducial volume and to associate a systematic uncertainty to its possible non-uniformity, we split the spatial distribution into radial and angular components. Extended Data Fig. 7 (top) shows the angular power spectrum of observed β events (black points). The dark pink and pink bands are the allowed 1σ and 2σ, regions respectively, obtained from 104 Monte Carlo simulations of uniformly distributed events. The analysis is performed with the HEALPix55 software package, available, for example, for cosmic microwave background analysis. Extended Data Fig. 7 (bottom) shows the linear fit to the r3 distribution of the β events, expected to be flat for uniform spatial distribution, from which we determine the allowed residual non-uniformity along the radial direction. All these studies show no evidence for a sizeable non-uniformity of the β-like event distribution inside the fiducial volume. In particular, the rate measured in the LPoF is fully consistent with that measured in the total fiducial volume. This evidence further supports a very small systematic uncertainty on the 210Bi independent constraint. Combining in quadrature the uncertainties from the radial (0.52 cpd per 100 t) and angular (0.59 cpd per 100 t) components, we obtain a systematic error associated with the 210Bi spatial uniformity of 0.78 cpd per 100 t. Finally, we checked the stability of the 210Bi rate over time by applying two methods on the observed rate of β events in the optimized energy windows: first, we studied the range of possible polynomial distortions; and second, we performed a Lomb–Scargle spectral decomposition (see ref. 56 and references therein). We found no evidence of any relevant time variation apart from the expected annual modulation due to solar neutrinos (7Be-neutrinos contribute more than 60% to the β rate in the selected energy windows). Actually, the fact that we are capable of seeing the tiny 3.3% sinusoidal variation induced by the eccentricity of the Earth’s orbit around the Sun is in itself further proof of the excellent stability of the 210Bi rate over time. In particular, by studying the time dependence of the β-like events in the optimized window, the uncertainty in the 210Bi rate change is 0.18 cpd per 100 t, which is indeed negligible as compared with the global error quoted in equation (3). We note that, even after complete mixing, the true 210Bi rate is not perfectly constant in time, as it must follow the decay rate of the parent 210 Pb (τ = 32.7 years). This effect is not detectable over the approximately

3-year time period of our analysis, but for substantially longer periods it could be used for better constraining the 210Bi by fitting its long-lived temporal trend.

Details of the CNO analysis The analysis presented in this work is based on the data collected from June 2016 to February 2020 (Borexino Phase-III) and is performed in a fiducial volume defined as r < 2.8 m and −1.8 m < z < 2.2 m (r and z being the reconstructed radial and vertical position, respectively). The total exposure of this dataset corresponds to 1,072 d × 71.3 t. In Borexino, the energy of each event is given by the number of collected photoelectrons, whereas its position is determined by the photon arrival times at the PMTs. The energy and spatial resolution in Borexino has slowly deteriorated over time due to the steady loss of PMTs (the average number of active channels in Phase-III is 1,238) and is currently σE / E ≈ 6% and σx,y,z ≈ 11 cm for 1 MeV events in the centre of the detector. Events are selected by a sequence of cuts, which are specifically designed to veto muons and cosmogenic isotopes, to remove 214Bi –214Po fast coincidence events from the 238U chain, electronic noise, and external background events. The fraction of neutrino events lost by this selection criteria is measured with calibration data to be of the order of 0.1% and is therefore negligible. More details on data selection can be found in ref. 20. The main backgrounds surviving the cuts and affecting the CNO analysis are: 210Bi and 210Po in secular equilibrium with 210Pb which, as discussed above, have a rate in Borexino Phase-III of ≤11.5 ± 1.3 cpd per 100 t ; 210PoV from the vessel; 85Kr (β, Q = 687 keV); 40K (β and γ, Q = 1,460 keV); 11C (β+, Q = 960 keV; τ = 30 min), which is continuously produced by cosmic muons crossing the scintillator; and γ-rays emitted by 214Bi, 208 Tl and 40K from materials external to the scintillator (buffer liquid, PMTs, stainless steel sphere, etc.). CNO neutrinos are disentangled from residual backgrounds using a multivariate analysis, which includes the energy and radial distributions of the events surviving the selection. Data are split into two complementary datasets: the threefold-coincidence (TFC)-subtracted spectrum, in which 11C is selectively filtered out using the muon–neutron–positron threefold coincidence algorithm19,57 and the TFC-tagged spectrum, enriched in 11C. The TFC is a space and time coincidence vetoing the 11C β+ decay events, by tagging the spallation muon and the neutron capture from the reactions: μ + 12C → 11C + n and n + p → d + γ. The reference shapes— that is, the probability density functions (PDFs) for signal and backgrounds used in the fit—are obtained through a complete Geant4-based Monte Carlo code31, which simulates all physics processes occurring in the scintillator, including energy deposition, photon emission, propagation, and detection, generation and processing of the electronic signal. The simulation takes into account the evolution in time of the detector response and produces data that are reconstructed and selected following the same pipeline of real data. The relevant input parameters of the simulation—mainly related to the optical properties of the scintillator and of the surrounding materials—have been initially obtained through small-scale laboratory tests and subsequently fine-tuned on calibration data, reaching an agreement at the sub-per-cent level35. Data are then fitted as the sum of signal and background PDFs: the weights of this sum (the energy integral of the rates with zero threshold of each component in Borexino) are the only free parameters of the fit. The details of the multivariate fit tool, used also to perform other solar neutrino analysis in Borexino, are described thoroughly in refs. 3,20. Contrary to the previous comprehensive pp chain analysis, the fit is performed between 320 and 2,640 keV, thus excluding the contribution of 14C decays and its pile-up. This choice is motivated by the loss of energy and position resolutions due to the decreased number of active channels in Phase-III, which affects mainly the low-energy region. In addition to the energy shape, other information is exploited to help the fit to disentangle the signal from background: the 11C β+ events are

Article tagged by TFC, and contributions from the external backgrounds (208Tl, 214 Bi and 40K) are further constrained due to their radial distribution. In order to enhance the sensitivity to CNO neutrinos, the pep neutrino rate is constrained to the value 2.74 ± 0.04 cpd per 100 t derived from a global fit25,26 to solar neutrino data and imposing the pp/pep ratio and the solar luminosity constraint, considering the Mikheyev–Smirnov– Wolfenstein matter effect on the neutrino propagation, as well as the errors on the neutrino oscillation parameters. As discussed in the main text, the spectral fit has little capability to disentangle events due to CNO neutrino interactions and 210Bi decay. Therefore, we use the results of the independent analysis on the 210Po distribution in the LPoF to set an upper limit to the 210Bi rate of 11.5 ± 1.3 cpd per 100 t. The results of the simultaneous multivariate fit are given in Extended Data Fig. 8, showing the TFC-subtracted and TFC-tagged energy spectra, and in Extended Data Fig. 9, demonstrating the fit of the radial distribution. The fit is performed in the energy estimator Nh (defined as the sum of all photons triggering a PMT, normalized to 2,000 active PMTs) and the results are reported also in keV. The P value of the fit is 0.3, demonstrating fair agreement between data and the underlying fit model. A non-zero CNO neutrino rate is clearly better suited to the fit, as shown in the log-likelihood profile of Fig. 4 (dashed black curve). Many sources of possible systematic errors have been considered. The systematic error associated with the fit procedure was studied by performing 2,500 fits with slightly altered conditions (different fit ranges and binning), and was found to be negligible with respect to the statistical uncertainty. Because the multivariate analysis relies critically on the simulated PDFs of signal and backgrounds, any mismatch between the realistic and simulated energy shapes can alter the result of the fit and bias the significance on the CNO neutrinos. In order to study the effect of these possible mismatches, we simulated more than a million pseudo-datasets with the same exposure as Phase-III, injecting deformations in the signal and background shapes, following ref. 58. Each dataset is then fitted with the standard non-deformed PDFs. The study was performed injecting different values of CNO, including the one obtained by our best fit. We studied the effect of the following sources of deformations: (1) Energy response function: inaccuracies in the energy scale (at the level of ~0.23%) and in the description of non-uniformity and nonlinearity of the response (at the level of ~0.28% and ~0.4%, respectively). The size of the applied deformations has been chosen in the range allowed by calibration data and by data from specific internal backgrounds (11C and 210Po) taken as reference ‘standard candles’. (2) Deformations of the 11C spectral shape induced by cuts to remove noise events, not fully taken into account by the Monte Carlo PDFs (at the level of 2.3%). (3) Spectral shape of 210Bi: we studied the systematic error associated with the shape of the forbidden β-decay of 210Bi simulating data with alternative spectra (found in refs. 33,34) with respect to the default spectrum32. Differences in the shapes may be as large as 18%. From this Monte Carlo study we evaluate the CNO systematic error due to a mismatch between real and simulated PDFs to be +0.6 −0.5 cpd per 100 t. This uncertainty is deduced by comparing the CNO output distributions from toy Monte Carlo PDFs with and without injecting systematic distortions as described above. In order to evaluate the significance (space) of our result in rejecting the no-CNO hypothesis, we performed a frequentist hypothesis test using a profile likelihood test, with statistics q defined (following ref. 38) as:

q = − 2log

L(CNO = 0) , L(CNO)

(5)

where L(CNO = 0) and L(CNO) is the maximum likelihood obtained by keeping the CNO rate fixed to zero or free, respectively. Extended

Data Fig. 10 shows the q distribution obtained from 13.8 million pseudo-datasets simulated with deformed PDFs (see above) and no-CNO injected (q0, grey curve). In the same plot, the theoretical q0 distribution in the case of no PDF deformation is shown (blue curve). The result of data obtained from the fit is the black line (qdata = 30.05). The plot in Extended Data Fig. 10 enables us to reject the CNO = 0 hypothesis with a significance better than 5.0σ at 99.0% confidence level59. This construction is consistent with the significance evaluation of 5.1σ, reported in the main text, by means of the quantiles of the profile likelihood folded with the systematic uncertainty. In Extended Data Fig. 10, we also provide as reference the q distribution (red) obtained with 1 million pseudo-datasets, including systematic deformations and injected CNO rate equal to 7.2 cpd per 100 t—that is, our best fit value. A cross-check of the main analysis has been performed with an almost independent method—counting analysis—in which we simply count events in an optimized energy window (region of interest, ROI) and subtract the contributions due to known backgrounds in order to reveal the CNO signal. This method is simpler, albeit less powerful, with respect to the multivariate fit and is less prone to possible correlations between different species. However, whereas the multivariate analysis implicitly checks the validity of the background model by the goodness of the fit, the counting analysis relies completely on the assumption that there are no unknown backgrounds that contribute to the ROI. The counting analysis is based on a different energy estimator than the multivariate analysis ( Npe, the total charge of all hits, normalized to 2,000 active channels) and relies on a different response function (analytically derived, instead of Monte Carlo-based) to determine the percentage of events for each of the signal and background species that falls inside the ROI. The chosen ROI, 780–885 keV, is obtained optimizing the CNO signal-to-background ratio. An advantage of this method is that, in the ROI, some of the backgrounds that affect the multivariate analysis (such as 85Kr and 210Po) are not present or contribute less than 2% (for example, external backgrounds). The count rate is dominated by CNO, pep and 210Bi (80%), with smaller contributions from 7Be neutrinos and residual 11C (18%). The rate of pep neutrinos and 210 Bi are constrained to the same values used in the multivariate fit. Note that whereas in the spectral fit the 210Bi rate is left free to vary between 0 up to 11.5 ± 1.3 cpd per 100 t (the upper limit determined in the LPoF analysis), the counting analysis conservatively constrains it to the maximum value with a Gaussian error of 1.3 cpd per 100 t. The 7 Be neutrino rate is sampled uniformly between the low-metallicity (43.7 ± 2.5 cpd per 100 t) and the high-metallicity (47.9 ± 2.8 cpd per 100 t) values predicted by the Standard Solar Model17 with 1σ error, whereas the 11C rate is obtained from the average Borexino Phase-II results with an additional conservative error of 10% derived from uncertainties on the energy scale (quenching of the 1 MeV annihilation γ-rays). The CNO rate is obtained by subtracting all background contributions defined above and by propagating the uncertainties by randomly sampling their rates from Gaussian distributions with proper widths. Note that the uncertainty related to the energy response (which affects the percentage of the spectrum of each component falling in the ROI) also contributes to the total error associated with the count rate of each species. The CNO rate obtained with this method is demonstrated by the red histogram in Fig. 4. The mean value and width of the distribution are 5.6 ± 1.6 cpd per 100 t, confirming the presence of CNO at the 3.5σ level. The counting analysis shows that the core of the sensitivity to CNO neutrinos in Borexino mainly comes, as expected, from a narrow energy region in which the contributions from CNO, pep and 210Bi are dominant over the residual backgrounds, as discussed in ref. 24. Conversely, the multivariate fit effectively exploits additional information contained in the data with a substantial enhancement of the significance of the CNO solar neutrino.

Data availability The datasets generated during the current study are freely available from the repository https://bxopen.lngs.infn.it/. Additional information is available from the Borexino Collaboration spokesperson ([email protected]) upon reasonable request. 42. Birks, J. B. The Theory and Practice of Scintillation Counting (Pergamon, 1964). 43. Benziger, J. et al. The scintillator purification system for the Borexino solar neutrino detector. Nucl. Instrum. Methods Phys. Res. A 587, 277–291 (2008). 44. Alimonti, G. et al. The liquid handling systems for the Borexino solar neutrino detector. Nucl. Instrum. Methods Phys. Res. A 609, 58–78 (2009). 45. Bellini, G. et al. Cosmic-muon flux and annual modulation in Borexino at 3800 m water-equivalent depth. J. Cosmol. Astropart. Phys. 2012, 015 (2012). 46. Bellini, G. et al. Cosmogenic backgrounds in Borexino at 3800 m water-equivalent depth. J. Cosmol. Astropart. Phys. 2013, 049 (2013). 47. Bellini, G. et al. Muon and cosmogenic neutron detection in Borexino. J. Instrum. 6, P05005 (2011). 48. Cruickshank Miller, C. The Stokes–Einstein law for diffusion in solution. Proc. R. Soc. Lond. A 106, 724–729 (1924). 49. Wójcik, M., Wlazlo, W., Zuzel, G. & Heusser, G. Radon diffusion through polymer membranes used in the solar neutrino experiment Borexino. Nucl. Instrum. Methods Phys. Res. A 449, 158–171 (2000). 50. Hoecker, A., Speckmayer, P., Stelzer, J., Therhaag, J., von Toerne, H. & Voss, E. TMVA toolkit for multivariate data analysis. Preprint at https://arxiv.org/abs/physics/0703039 (2007). 51. Feroz, F., Hobson, M. P., Cameron, E. & Pettitt, A. N. Importance nested sampling and the MultiNest algorithm. Open J. Astrophys. 2, 10 (2019). 52. Feroz, F., Hobson, M. P. & Bridges, M. MultiNest: an efficient and robust Bayesian inference tool for cosmology and particle physics. Mon. Not. R. Astron. Soc. 398, 1601–1614 (2009). 53. Feroz, F. & Hobson, M. P. Multimodal nested sampling: an efficient and robust alternative to Markov Chain Monte Carlo methods for astronomical data analyses. Mon. Not. R. Astron. Soc. 384, 449–463 (2008). 54. Fick, A. Ueber Diffusion. Ann. Phys. 170, 59–86 (1855). 55. Gorski, K. M., Wandelt, B. D., Hansen, F. K., Hivon, E. & Banday, A. J. The HEALPix Primer. Preprint at https://arxiv.org/abs/astro-ph/9905275 (1999). 56. Agostini, M. et al. Seasonal modulation of the 7Be solar neutrino rate in Borexino. Astropart. Phys. 92, 21–29 (2017).

57. Bellini, G. et al. First evidence of pep solar neutrinos by direct detection in Borexino. Phys. Rev. Lett. 108, 051302 (2012). 58. Cousins, R. D. & Highland, V. L. Incorporating systematic uncertainties into an upper limit. Nucl. Instrum. Methods Phys. Res. A 320, 331–335 (1992). 59. Brown, L. D., Cai, T. T. & Das Gupta, A. Interval estimation for a binomial proportion. Stat. Sci. 16, 101–133 (2001). Acknowledgements We acknowledge the hospitality and support of the Laboratori Nazionali del Gran Sasso (Italy). The Borexino program is made possible by funding from Istituto Nazionale di Fisica Nucleare (INFN) (Italy), National Science Foundation (NSF) (USA), Deutsche Forschungsgemeinschaft (DFG) and Helmholtz-Gemeinschaft (HGF) (Germany), Russian Foundation for Basic Research (RFBR) (grant numbers 16-29-13014ofi-m, 17-02-00305A and 19-02-00097A), Russian Science Foundation (RSF) (grant number 17-12-01009) and Ministry of Science and Higher Education of the Russian Federation (contract number 075-15-2020-778) (Russia), and Narodowe Centrum Nauki (NCN) (grant number UMO 2017/26/M/ST2/00915) (Poland). We acknowledge the computing services of Bologna INFN-CNAF data centre and U-Lite Computing Center and Network Service at LNGS (Italy), and the computing time granted through JARA on the supercomputer JURECA at Forschungszentrum Jülich (Germany). This research was supported in part by PLGrid Infrastructure (Poland). Author contributions The Borexino detector was designed, constructed and commissioned by the Borexino Collaboration over the span of more than 30 years. The Borexino Collaboration sets the science goals. Scintillator purification and handling, material radiopurity assay, source calibration campaigns, photomultiplier tube and electronics operations, signal processing and data acquisition, Monte Carlo simulations of the detector, and data analyses were performed by Borexino members, who also discussed and approved the scientific results. This Article was prepared by a subgroup of authors that was appointed by the Collaboration and was subjected to an internal collaboration-wide review process. All authors reviewed and approved the final version of the manuscript. Competing interests The authors declare no competing interests. Additional information Supplementary information is available for this paper at https://doi.org/10.1038/s41586-0202934-0. Correspondence and requests for materials should be addressed to G.R. Peer review information Nature thanks Marc Pinsonneault, Gabriel Orebi Gann and David Wark for their contribution to the peer review of this work. Peer reviewer reports are available. Reprints and permissions information is available at http://www.nature.com/reprints.

Article

Extended Data Fig. 1 | The Borexino detector. Schematic view of the structure of the Borexino apparatus. From inside to outside: the liquid scintillator, the buffer liquid, the stainless steel sphere with the photomultipliers, and the water tank.

Extended Data Fig. 2 | The Borexino detector after the thermal stabilization. The Borexino water tank after completion of the thermal insulation and deployment of the active temperature control system.

Article

Extended Data Fig. 3 | Temperature probes of the Borexino detector. Distribution of temperature probes around and inside the Borexino detector. For simplicity, the probes on the water tank (WT) dome and in the pit below the detector are not shown.

Extended Data Fig. 4 | Temperature evolution over time in the Borexino detector. Graph depicting the temperature as a function of time in different volumes of the Borexino detector. The vertical dashed lines show the beginning of the thermal insulation installation (1), the turning off of the water

loop inside the water tank (2), the completing of the thermal insulation installation (3), the activation of the temperature control system on the dome of the water tank (4), the set-point change (5) and the activation of the air control system in experimental hall C (6).

Article

Extended Data Fig. 5 | The low polonium field in the Borexino scintillator. Three-dimensional view of the 210Po activity inside the entire nylon vessel (see colour code). The innermost blue region contains the LPoF (black grid). The white grid is the software-defined fiducial volume. a.u., arbitrary units.

Extended Data Fig. 6 | Analysis of the low polonium field. Top, the rate of 210 Po in cylinders of 3-m radius and 10-cm height located along the z axis from −2 m to 2 m, as a function of time with 1-month binning. The dashed lines indicate the z coordinate of the fiducial volume. The markers show the positions of the centre of the LPoF obtained with two fit methods: paraboloid (red) and spline (white). Both fit methods follow the dark-blue minimum of the

210 Po activity well. The structure visible in mid-2019 is due to a local instability produced by a tuning of the active temperature control system. This transient has no effect on the final result. Bottom, distribution of 210Po events after the blind alignment of data using the z0 from the paraboloidal fit (red markers in the top graph). The red solid lines indicate the paraboloidal fit within 20 t with equation (4).

Article

Extended Data Fig. 7 | Angular and radial uniformity of the β events in the optimized energy window. Top, angular power spectrum as a function of the multipole moment l of observed β events (black points) compared with 104 uniformly distributed events from Monte Carlo simulations at 1σ (dark pink) and 2σ (pink) confidence levels (C.L.). Data are compatible with a uniform distribution within the uncertainty of 0.59 cpd per 100 t. Inset, angular

distribution of the β events. Bottom, normalized radial distribution of β events r/r0 (black points), where r0 = 2.5 m is the radius of the sphere surrounding the analysis fiducial volume. The linear fit of the data (red solid line) is shown along with the 1σ (yellow) and 2σ (green) confidence level bands. The data are compatible with a uniform distribution within 0.52 cpd per 100 t.

3

Events / 5Nh

10

400

500

Nh

600

700

800

900

200

CNO- 210

external bkgs

11

Total fit: p-value = 0.3

other bkgs

C

4 3 2

Residuals

2 1

0 1

2

1500

2000

600

700

800

900

10

4 3

Energy [keV]

Nh

102

1

1000

500

103

1

500

400

Be- and B-

pep-

10

3 4

300

8

7

Bi

102

Residuals

300

Events / 5Nh

200

2500

Extended Data Fig. 8 | Energy distributions from a multivariate fit of the Borexino data. Full multivariate fit results for the TFC-subtracted (left) and the TFC-tagged (right) energy spectra with corresponding residuals. In both graphs the magenta lines represent the resulting fit function, the red line is the

1 0 1

2 3 4

500

1000

1500

Energy [keV]

2000

2500

CNO neutrino electron recoil spectrum, the green dotted line is the pep neutrino electron recoil spectrum, the dashed blue line is the 210Bi β spectrum, and in grey we report the remaining background (bkgs) contributions.

Counts / 0.10 m

Article Uniform Component External Component Best Fit - 2/NDF = 21 / 30

102

10

Residuals

1

3 2 1 0 1 2 3 0

0.5

1

Extended Data Fig. 9 | Radial distribution from a multivariate fit of the Borexino data. Radial distribution of events in the multivariate fit. The red line is the resulting fit, the green line represents the internal uniform contribution

1.5

r [m]

2

2.5

3

and the blue line shows the non-uniform contribution from the external background. NDF, the number of degrees of freedom in the fit.

Extended Data Fig. 10 | Frequentist hypothesis test for the CNO observation. Distribution of the test statistics q (equation (5) from Monte Carlo pseudo-datasets). The grey distribution q0 is obtained with no CNO simulated data and includes the systematic uncertainty. The black vertical line represents qdata = 30.05. The corresponding P value of q0 with respect to qdata

gives the significance of the CNO discovery (>5.0σ at 99% confidence level). For comparison, in blue is the q0 without the systematics. The red histogram represents the expected test statistics distribution for an injected CNO rate equal to 7.2 cpd per 100 t—that is, our best fit value.

Article

Observing the emergence of a quantum phase transition shell by shell https://doi.org/10.1038/s41586-020-2936-y Received: 24 April 2020

Luca Bayha1,7 ✉, Marvin Holten1,7 ✉, Ralf Klemt1, Keerthan Subramanian1, Johannes Bjerlin2,3,4, Stephanie M. Reimann2, Georg M. Bruun5,6, Philipp M. Preiss1 & Selim Jochim1

Accepted: 3 September 2020 Published online: 25 November 2020 Check for updates

Many-body physics describes phenomena that cannot be understood by looking only at the constituents of a system1. Striking examples are broken symmetry, phase transitions and collective excitations2. To understand how such collective behaviour emerges as a system is gradually assembled from individual particles has been a goal in atomic, nuclear and solid-state physics for decades3–6. Here we observe the few-body precursor of a quantum phase transition from a normal to a superfluid phase. The transition is signalled by the softening of the mode associated with amplitude vibrations of the order parameter, usually referred to as a Higgs mode7. We achieve fine control over ultracold fermions confined to two-dimensional harmonic potentials and prepare closed-shell configurations of 2, 6 and 12 fermionic atoms in the ground state with high fidelity. Spectroscopy is then performed on our mesoscopic system while tuning the pair energy from zero to a value larger than the shell spacing. Using full atom counting statistics, we find the lowest resonance to consist of coherently excited pairs only. The distinct non-monotonic interaction dependence of this many-body excitation, combined with comparison with numerical calculations allows us to identify it as the precursor of the Higgs mode. Our atomic simulator provides a way to study the emergence of collective phenomena and the thermodynamic limit, particle by particle.

A key element of our understanding of nature is that macroscopic systems are characterized by the presence of phase transitions and collective modes, which cannot be extrapolated from the two-body solution1. Although these effects in principle exist only in the thermodynamic limit, they are sometimes observed in surprisingly small systems. For instance, atomic nuclei consisting of only around 50 particles exhibit a collective mode spectrum consistent with a superfluid3,6. In liquid helium droplets, superfluidity has been found to set in for similar particle numbers4. Ultracold atoms offer the exciting possibility of studying the onset of many-body physics in systems with full tunability of interactions, particle number and single-particle spectra8. The emergence of a Fermi sea was observed in a one-dimensional trap in ref. 5. Two- and three-dimensional systems promise even richer physics such as quantum phase transitions and symmetry breaking, as well as degenerate energy levels and single-particle spectra akin to the shell structure of atoms and nuclei. In this work, we observe the few-body precursor of a quantum phase transition. The measurement relies on our experimental breakthrough in the preparation of a tunable number of Fermionic atoms in the ground state of a two-dimensional (2D) harmonic potential. We study the interplay of the shell structure and Pauli blocking with the attractive interactions for closed-shell configurations. The competition of the gapped single-particle spectrum—given by the confinement—with the interactions gives rise to particular excitations exhibiting a non-trivial

dependence on the attraction. We study this dependence and demonstrate that the modes consist of coherent excitations of particle pairs. Combined with a careful comparison with numerical calculations, this allows us to identify these excitations as few-body precursors of a Higgs mode associated with a quantum phase transition to a superfluid of Cooper pairs9 (see Methods). Comparing measured spectra for different atom numbers allows us to observe the approach towards the thermodynamic limit. In the many-body limit, a Higgs mode has been observed in cold-atom, superconducting and ferromagnetic systems10–17. Our results pioneer the study of emergent quantum phase transitions and the associated Higgs mode starting from a few-body system.

Experimental setup We perform our experiments with a balanced mixture of two hyperfine states of 6Li confined in a trap created by the superposition of an optical tweezer (OT) and a single layer of an optical lattice (see Fig. 1a). The radial trapping frequencies of fr ≈ 1,000 Hz are much smaller than the axial frequency fz ≈ 6,800 Hz. Hence, for low temperatures and only a few occupied shells, the sample is in the quasi-2D regime and the dynamics along the third direction is frozen out. We prepare the ground state of up to 12 atoms in this trap by applying a novel spilling technique for quasi-2D systems based on the method in ref. 18. A magnetic field

Physikalisches Institut der Universität Heidelberg, Heidelberg, Germany. 2Mathematical Physics and NanoLund, LTH, Lund University, Lund, Sweden. 3The Niels Bohr Institute, University of Copenhagen, Copenhagen, Denmark. 4Department of Physics and Astronomy, University of Southern California, Los Angeles, CA, USA. 5Center for Complex Quantum Systems, Department of Physics and Astronomy, Aarhus University, Aarhus, Denmark. 6Shenzhen Institute for Quantum Science and Engineering and Department of Physics, Southern University of Science and Technology, Shenzhen, China. 7These authors contributed equally: Luca Bayha, Marvin Holten. ✉e-mail: [email protected]; [email protected] 1

Nature | Vol 587 | 26 November 2020 | 583

m = –2

m=0 m = +2

m = –1

m = +1

2 m=0

1

E2 = 3hfr

E1 = 2hfr

( 2n+| m| +1) hf r E=

m = –2

m=0 m = +2

m = –1

E0 = 1hfr

m = +1

m=0

c

100 7

0 2

3 4 5 6 7 8 9 Optical trap depth, V (arbitrary units)

Fig. 1 | Deterministic preparation of two-dimensional closed-shell configurations. a, Sketch of the experimental setup. The Fermionic atoms are trapped in a single layer of an attractive optical lattice (horizontal disk) providing a tight confinement that freezes out motion along the z direction. A superimposed tightly focused optical tweezer (vertical cone) provides radial harmonic confinement. b, Mean trapped atom number as a function of the final depth of the optical tweezer. The stable atom numbers of 2, 6 and 12 correspond to the closed-shell configurations of the two-dimensional harmonic oscillator. The third plateau is observed slightly below the expected value owing to current experimental limitations that sometimes result in an incompletely filled third shell. The insets show sketches of the corresponding atom distribution. The standard error of the mean is indicated by the width of the band connecting the points. Each data point is the average of 84 or 85 measurements. c, Standard deviation of the detected atom number. At trap depths where the atom number reaches a plateau fluctuations are suppressed indicating the deterministic preparation of closed shells of atoms.

gradient is applied and the power of the OT lowered such that only the lowest states inside the trap remain bound. This removes atoms initially occupying higher energy trap levels and the minimal power of the OT determines the initialized atom number. The degeneracy of the 2D harmonic oscillator is k + 1 for the kth energy level, resulting in the emergence of shells. The lowest three closed-shell configurations, that is, where all states up to some energy are occupied and all other states are empty, contain 1, 3 and 6 fermions per spin state, respectively. Working with two hyperfine states, we expect closed-shell configurations of 2, 6 and 12 atoms. In the experiment, we observe plateaus for these ‘magic’ numbers as a function of the OT depth (Fig. 1b). The increased stability of the closed shells is also visible in the atom number fluctuations, which are strongly suppressed for closed-shell configurations (Fig. 1c). An optimized sequence gives preparation fidelities of 97 ± 2%, 93 ± 3% and 76 ± 2%, for 2, 6 and 12 atoms respectively. The deterministic loading of the ground state in two-dimensional systems is one of the main results of our work and opens up prospects for quantum simulation with ultracold atoms. More details on the experimental procedure are provided in the Methods. The gapped closed-shell configurations are an interesting starting point at which to introduce interactions: whereas a gapless Fermi gas at zero temperature undergoes a phase transition from normal to superfluid at any attraction, a gap in the single-particle spectrum gives rise to a quantum phase transition from a normal to a superfluid phase at a certain critical interaction strength19,20. For the 2D harmonic oscillator, this means that for partly filled (open) shells the particles in this system will pair for arbitrarily weak attraction, whereas for completely filled (closed) shells and weak attraction the system is dominated by the energy gap to empty shells and pairing is suppressed. Rich physics thus arises from the competition of interactions with the single-particle shell structure7,9,21–23. Experimentally, we tune and control the interactions using a Feshbach resonance24. Because the experiment is performed in a quasi-2D geometry there exists a two-body bound state 584 | Nature | Vol 587 | 26 November 2020

90 80

6

70 60

5

50 4 3 2

40 30 20

B = 850G EB /hfr = 0.33 1,400

1,500

Probability (%)

c

b

| +1) hf r ( 2n+| m E=

10

2fr

r

a

14 12 10 8 6 4 2 0

Atom number

b z

Standard deviation, V

a

Mean atom number, N

Article

1,600 1,700 1,800 1,900 Excitation frequency (Hz)

2,000

2,100

0

Fig. 2 | Excitation spectrum for 6 particles. a, b, Modulating the interaction strength at a frequency close to twice the radial trap frequency fr leads to excitations of pairs (a) and single particles (b). We start with 6 atoms in the ground state, modulate the interaction at different frequencies and count the number of atoms remaining in the lowest two shells. The probability of detecting N remaining atoms as a function of the modulation frequency is shown in c. The experiment is repeated 45 times for each drive frequency. We observe the lowest resonance at 1,890 Hz below twice the trap frequency, indicating mode softening. Remarkably, only the probability of detecting 4 atoms is increased, showing that this resonance consists of a coherent superposition of pair excitations. The second resonance at 2,060 Hz lies above twice the trap frequency consistent with a mean-field estimate. It is composed of single-particle excitations, as the probabilities for detecting both 4 and 5 atoms are enhanced.

for any attractive contact interaction. The binding energy EB of the pair uniquely characterizes the interaction strength25. We explore an interaction range where the binding energy EB is much smaller than the axial confinement and the effective interactions can be treated as quasi-2D (see Methods).

Excitation spectra We utilize many-body spectroscopy to probe the effect of interactions on closed-shell configurations. The system is excited by modulating the axial confinement at frequencies far below the bandgap of the lattice, which modulates only the effective two-dimensional interaction strength between the different hyperfine states26 (see Methods). This interaction perturbation couples strongly to collective excitations driven by pairing correlations9. After modulation, all atoms excited to higher states are removed by a second spilling procedure and we count the remaining atoms. Repeating this procedure for different excitation frequencies gives the spectrum shown in Fig. 2c. The full counting statistics not only contains the resonance positions, but also reveals the difference between pair and single-particle excitations. The spectrum shows two resonances, where the probability of detecting 6 atoms is greatly reduced. The higher resonance at 2,060 Hz lies above twice the noninteracting trap frequency 2fr ≈ 2,000 Hz. This is consistent with an attractive mean-field potential from the atom cloud, which increases the effective trapping frequency. The mean-field

2.2

2.0

2.2 0.5

2.0

1.9

1.0

2.0

1.8

1.8

0

1.8 1.6

1.6

1.4

Singleparticle excitations

1.2 0.4

0.8

1.2 EB/hfr

1.6

2.0

0

0.4

0.8 EB/hfr

1.2

1.6

0.0

N=6 N = 12 0.5

1.0 EB/hfr

1.5

y

0

1.4

d y-bo Man

1.5 1.4

Pair excitations

1.6 Excitation energy

1.7

(PN – 2/PN)scaled

N = 12

2.0

od e

N=6

2.1

Eexc/hfr

c

b 2.2

Hig gs m

a

Attraction

Fig. 3 | Excitation spectra as function of interaction strength for 6 and 12 particles. a, b, We show the probability (the colour scale refers to all panels) of exciting a pair of particles PN−2 divided by the probability of retaining the initial particle number PN when modulating at an excitation frequency fexc, which is linked to the excitation energy by Eexc = hfexc. The system is excited by modulating the interaction strength for a fixed time for all interactions except for the spectra shown for EB = 0, where the system is excited by modulating the radial trapping potential and where we show the normalized excitation probability 1 − PN . We note that the drive parameters are outside the linear response regime and the coupling strength cannot be extracted from the respective peak amplitudes (see Methods). To increase visibility, the amplitude of the pair excitation mode is scaled to unity at each interaction strength. The

resonance position for the lowest three excitations shown in c are extracted by Gaussian fits to the respective spectra. The error bars extracted from the fit are smaller than the symbol size. Comparing the data for 6 atoms (a) and 12 atoms (b), we observe a deepening of the minimum of the pair excitation mode with particle number, as expected for the transition to the many-body limit. Moreover, the position of the minimum shifts to smaller interaction strength for more particles. The pair character of the lower excitations becomes evident in the full counting statistics for a and b shown in Extended Data Figs. 4 and 5. The inset of c sketches the approach of the pair excitation (blue) towards the many-body limit (black), upon increasing the particle number as discussed in ref. 9. We scale the attraction by the critical attraction to compare to the thermodynamic limit.

energy is proportional to the atom density, thus predicting a larger interaction shift for the ground state than for the more dilute excited state. This higher resonance consists of single-particle excitations two shells up in energy. For the chosen drive strength and time there is a substantial probability of exciting the system more than once. This explains the enhanced probabilities of detecting 4 and 5 atoms. Mean-field theory, however, completely fails to explain the lower resonance at 1,890 Hz, below twice the noninteracting trap frequency. Furthermore, the observed atom number distribution is strikingly different. Here only the probability of detecting 4 atoms is enhanced, whereas all other probabilities are flat. Thus, at this frequency it is only possible to excite a single pair of atoms and not individual atoms or two pairs. Both the energy and the atom number distribution are clear signatures of the collective nature of this excitation arising from the competition between the single-particle gap and the attractive interactions. The spectrum is obtained for a binding energy of EB = 0.33hfr, which is smaller than the single-particle gap and pairing is suppressed for closed shells owing to Pauli blocking. However, by exciting a coherent superposition of particle pairs from the completely filled shell the remaining atoms can enhance their overlap by occupying the now empty states and thereby gain pairing energy. The excited particles form a pair in the otherwise empty shell and have a lower energy compared to two noninteracting particles in the same shell. Thus, the pair excitation lies below twice the trap frequency. Next, we investigate the competition between pairing and the shell structure in more detail by tuning their relative strength using a Feshbach resonance. The spectrum for different binding energies shown in Fig. 3a allows us to track the evolution of the different excitations discussed above. The branch highest in energy shows a monotonous increase of frequency with interaction, as expected, from the increasing mean-field shift. Remarkably, the lower two branches show a non-monotonic behaviour. As we shall discuss in detail below, they correspond to coherent excitations of pairs with angular momentum 0 and ±2ħ. For small interactions the energy of these excitations decreases with increasing attraction. This is due to the increasing gain in binding

energy and the larger pair correlations in the excited state. This picture breaks down above an interaction strength of EB ≈ 1.1hfr, where the lower mode energies start to increase with the attraction. In this regime the binding energy is comparable to the radial trap frequency and pairing becomes important also for the closed-shell ground state. Here, it is energetically favourable to have an admixture of higher harmonic oscillator levels to form a pair. Consequently, the ground state has notable pairing correlations and its energy decreases faster than that of the excited states. We identify the position of the minimal excitation gap with the critical interaction strength. To study the scaling of the spectrum towards the many-body limit, we fill one more shell in our trap, working with 12 particles. The corresponding excitation spectrum is shown in Fig. 3b. Qualitatively, the spectra for N = 12 and 6 show the same features. For the larger system the number of states that are shifted upwards in energy above 2hfr increases, rendering it impossible to resolve a single well defined excitation peak. Importantly, the minima of the pair excitation branches below 2hfr deepen and move to smaller interaction strengths for larger particle numbers, as is evident from the resonance positions shown in Fig. 3c.

Many-body picture Crucially, the qualitative behaviour of this spectrum and its evolution from the few- to the many-body limit can be understood from theory7,9. In the thermodynamic limit, a closed-shell system undergoes a quantum phase transition from a normal to a superfluid phase with increasing attraction7,27. As a generic feature of quantum phase transitions2, this gives rise to a collective mode that goes soft, that is, the excitation gap closes at the transition point. In the case at hand, the lowest collective mode corresponds to the coherent excitations of time-reversed pairs across the gap. The energy cost of these excitations vanishes at the transition point, indicating that the system is spontaneously forming Cooper pairs. From a broken-symmetry perspective, the mode corresponds to amplitude vibrations in the order parameter (pairing strength) around its average value, which is zero in the normal phase Nature | Vol 587 | 26 November 2020 | 585

Article Fit to PN = 6

PN = 6

1.0

PN = 4

PN = 4 + P N = 6

pair in the excited shell at a Rabi rate of 8.0 ± 0.1 Hz (see Fig. 4). The 1/e decay rate (where e is Euler’s number) of the oscillation of 4.5 ± 0.5 Hz gives a quantitative upper limit on the lifetime of the pair excitation mode, exceeding the transition frequency of 1,480 Hz by a factor of more than 300. The long lifetime of the excited state can be attributed to the discrete level spectrum of our trap7. No decay channels to single-particle excitations are available that conserve the energy of the isolated system.

PN = 5

Probability

0.8

0.6

0.4

Outlook

0.2

0

0

50

100

150

200

250

Modulation time (ms) Fig. 4 | Coherent driving of the lower pair excitation mode. The interactions are modulated at the resonance frequency of the lower pair excitation mode for variable times. The probabilities for 4 and 6 particles coherently oscillate out of phase, revealing the pair character of the excitation. Coupling to other states is negligible, as seen from the almost constant value of PN = 4 + PN = 6 and the fact that PN = 4 and PN = 6 converge to the same value. Thus, the ground state plus the lower Higgs mode precursor can be described as a coherently driven two-state system. We fit PN = 6 with an exponentially damped Rabi oscillation. The data are taken for EB = 0.57hfr. The error bars represent the standard error of the mean. For each modulation time the measurement is repeated between 177 and 181 times.

In conclusion, we have shown that systems consisting of only a few particles exhibit precursors of a quantum phase transition to a superfluid phase with an associated Higgs mode present in the thermodynamic limit. In addition to the emergence of pairing, the degree of control achieved over this mesoscopic system will allow us to study thermalization in isolated quantum systems28 and fermionic superfluidity at its fundamental level. As a next step, we will go beyond the excitation spectrum studied here and investigate the emergence of pair correlations across the precursor of the normal-to-superfluid transition directly in momentum space. Another interesting question to be explored concerns the emergence of Cooper pairs and Goldstone modes with increasing system size. Since Goldstone modes are driven by phase fluctuations of the order parameter, they exist when the phase can be defined on a length scale much smaller than the system size, or equivalently, when the superfluid gap is much larger than the trap level spacing27.

Online content and non-zero in the superfluid phase. In the superfluid phase, this mode is referred to as the Higgs mode. The pair excitation modes we observe in the experiment are the few-body precursors of the Higgs mode9. Owing to the finite particle number, the phase transition is broadened to a crossover and the gap does not close completely. However, the lowest excitation corresponding to zero angular momentum retains the non-monotonic dependence on interactions and the pair correlation character. Adding more particles to the system decreases the minimal gap, consistent with an eventual complete gap closure in the many-body limit. When the Fermi energy increases, the relative importance of the single-particle gap decreases, and the minimal gap moves towards smaller binding energies. Both the softening of the mode and the shift to smaller critical binding energies when approaching the many-body limit are clearly visible when going from two to three closed shells. This interpretation is explicitly confirmed by comparing to a numerical diagonalization of the microscopic Hamiltonian (see Methods). The higher non-monotonic branch in fact corresponds to two nearly degenerate modes consisting of coherent excitations of pairs with angular momentum ±2ħ. They are precursors of Higgs modes of a superfluid with higher-angular-momentum Cooper pairs. Although modulating the interaction strength does not add angular momentum, these modes are visible owing to a slight breaking of the circular symmetry of the trap (see Methods).

Any methods, additional references, Nature Research reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at https://doi.org/10.1038/s41586-020-2936-y. 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12. 13. 14. 15.

Coherent drive At all interaction strengths shown in Fig. 3, the Higgs mode precursor is a well defined excitation: The linewidth is of the order of 10 Hz and is thus much smaller than the excitation energy. To probe the stability of the excited state we drive the 6-particle system for variable times at the frequency of the lower pair excitation mode. We observe oscillations between the probabilities of detecting 4 and 6 particles in the lowest two shells, indicating the coherent formation and destruction of a 586 | Nature | Vol 587 | 26 November 2020

16. 17. 18. 19. 20.

Anderson, P. W. More is different. Science 177, 393–396 (1972). Sachdev, S. Quantum Phase Transitions 2nd edn (Cambridge Univ. Press, 2011). Bohr, A. & Mottelson, B. R. Nuclear Structure Vols. I, II (Benjamin, 1975). Grebenev, S., Toennies, J. P. & Vilesov, A. F. Superfluidity within a small helium-4 cluster: the microscopic Andronikashvili experiment. Science 279, 2083–2086 (1998). Wenz, A. N. et al. From few to many: observing the formation of a Fermi sea one atom at a time. Science 342, 457–460 (2013). Launey, K. D. Emergent Phenomena in Atomic Nuclei from Large-Scale Modeling (World Scientific, 2017). Bruun, G. M. Long-lived Higgs mode in a two-dimensional confined Fermi system. Phys. Rev. A 90, 023621 (2014). Bloch, I., Dalibard, J. & Zwerger, W. Many-body physics with ultracold gases. Rev. Mod. Phys. 80, 885–964 (2008). Bjerlin, J., Reimann, S. M. & Bruun, G. M. Few-body precursor of the Higgs mode in a Fermi gas. Phys. Rev. Lett. 116, 155302 (2016). Sooryakumar, R. & Klein, M. V. Raman scattering by superconducting-gap excitations and their coupling to charge-density waves. Phys. Rev. Lett. 45, 660–662 (1980). Rüegg, C. et al. Quantum magnets under pressure: controlling elementary excitations in TlCuCl3. Phys. Rev. Lett. 100, 205701 (2008). Bissbort, U. et al. Detecting the amplitude mode of strongly interacting lattice bosons by Bragg scattering. Phys. Rev. Lett. 106, 205303 (2011). Endres, M. et al. The ‘Higgs’ amplitude mode at the two-dimensional superfluid/Mott insulator transition. Nature 487, 454–458 (2012). Matsunaga, R. et al. Higgs amplitude mode in the BCS superconductors Nb1−xTixN induced by terahertz pulse excitation. Phys. Rev. Lett. 111, 057002 (2013). Léonard, J., Morales, A., Zupancic, P., Donner, T. & Esslinger, T. Monitoring and manipulating Higgs and Goldstone modes in a supersolid quantum gas. Science 358, 1415–1418 (2017). Katsumi, K. et al. Higgs mode in the d-wave superconductor Bi2Sr2CaCu2O8+x driven by an intense terahertz pulse. Phys. Rev. Lett. 120, 117001 (2018). Behrle, A. et al. Higgs mode in a strongly interacting fermionic superfluid. Nat. Phys. 14, 781–785 (2018). Serwane, F. et al. Deterministic preparation of a tunable few-fermion system. Science 332, 336–338 (2011). Kohmoto, M. & Takada, Y. Superconductivity from an insulator. J. Phys. Soc. Jpn 59, 1541–1544 (1990). Nozières, P. & Pistolesi, F. From semiconductors to superconductors: a simple model for pseudogaps. Eur. Phys. J. B 10, 649–662 (1999).

21. Heiselberg, H. & Mottelson, B. Shell structure and pairing for interacting fermions in a trap. Phys. Rev. Lett. 88, 190401 (2002). 22. Bruun, G. M. Low-energy monopole modes of a trapped atomic Fermi gas. Phys. Rev. Lett. 89, 263002 (2002). 23. Rontani, M., Eriksson, G., Åberg, S. & Reimann, S. M. On the renormalization of contact interactions for the configuration-interaction method in two-dimensions. J. Phys. At. Mol. Opt. Phys. 50, 065301 (2017). 24. Zürn, G. et al. Precise characterization of 6Li Feshbach resonances using trap-sideband-resolved RF spectroscopy of weakly bound molecules. Phys. Rev. Lett. 110, 135301 (2013). 25. Randeria, M., Duan, J. M. & Shieh, L. Y. Superconductivity in a two-dimensional Fermi gas: evolution from Cooper pairing to Bose condensation. Phys. Rev. B 41, 327–343 (1990).

26. Idziaszek, Z. & Calarco, T. Analytical solutions for the dynamics of two trapped interacting ultracold atoms. Phys. Rev. A 74, 022712 (2006). 27. Bruun, G. M. & Mottelson, B. R. Low energy collective modes of a superfluid trapped atomic Fermi gas. Phys. Rev. Lett. 87, 270403 (2001). 28. D’Alessio, L., Kafri, Y., Polkovnikov, A. & Rigol, M. From quantum chaos and eigenstate thermalization to statistical mechanics and thermodynamics. Adv. Phys. 65, 239–362 (2016). Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. © The Author(s), under exclusive licence to Springer Nature Limited 2020

Nature | Vol 587 | 26 November 2020 | 587

Article Methods Experimental sequence The experimental sequence starts by transferring a gas of 6Li atoms from a magneto-optical trap into a red-detuned crossed-beam optical dipole trap. Here, we make use of radio-frequency pulse sequences to prepare a balanced mixture of the two hyperfine states |1⟩ and |3⟩ of the ground state of 6Li. We label the hyperfine states according to their energy from lowest to highest, |1⟩ to |6⟩. After a first evaporative cooling stage in the crossed-beam optical dipole trap we transfer approximately 1,000 atoms into a tightly focused OT. In the OT, quantum degeneracy is reached by spilling around 20 atoms close to the ground state of the approximately harmonic confinement by the procedure described in ref. 18. Subsequently, we begin the crossover to a quasi-2D system (see Extended Data Fig. 1). This is achieved with an adiabatic transfer of the atoms from the effective confinement of the OT alone with fr: fz ≈ 5:1 to fr: fz ≈ 1:7 in a combined potential of OT and a single layer of a one-dimensional optical lattice. To this end, we lower the radial trap frequency fr of the OT from approximately 20 kHz to 1 kHz. This is done by ramping the pattern displayed on a spatial light modulator in 20 ms, which changes the aperture of the optical setup creating the OT. This changes the waist of the OT from about 1 μm to about 5 μm. The transfer is performed at a magnetic field of 750 G in order to have sizeable coupling between the different states. The axial confinement is solely defined by the optical lattice with fz ≈ 6.8 kHz. The measurements with 6 atoms were performed at a final radial trap frequency of fr = 1,001 Hz and the 12-atom data were taken at fr = 992 Hz. In the combined trap we create closed-shell configurations of the quasi-2D harmonic oscillator, by applying a magnetic gradient of approximately 70 G cm−1 in the axial direction and reducing the power of the OT such that only the lowest one to three energy shells remain bound. The spilling procedure is performed at 750 G, where the interaction energy is sufficiently small that one recovers the noninteracting shell structure. After preparation, we increase the OT power back up until we recover the trap frequencies and aspect ratio discussed in the main text. The measurements of the excitation spectrum are performed by a sinusoidal modulation of the power of either the OT or of the optical lattice with frequency fexc for t = 400 ms. To detect excitations, we implement a final spilling stage after which we transfer all the remaining trapped atoms from the OT back into the magneto-optical trap. In the magneto-optical trap we are able to determine the total atom number in both spin states with a fidelity exceeding 99%. The latter is achieved by integrating the total fluorescence signal of the magneto-optical trap for one second on a charge-coupled device (CCD) camera. Excitations of the system show up as a reduced atom number compared to the closed-shell ground states of 2, 6 or 12 atoms. The modulation parameters are chosen to optimize the signal of the pair excitation mode. Therefore, the recorded spectra do not allow us to extract quantitatively the excitation strength of the different modes directly. The ground state plus the isolated pair excitation mode are well described by a two-level system for which one would not expect a linear response. Nevertheless, we note that qualitatively the higher angular momentum pair excitation mode has a reduced strength for the 12-particle system, indicating a much weaker coupling to the ground state. Whether this is a result of systematic experimental effects, such as a drift of the anisotropy on a level below 1%, or is due to a decreased coupling for larger systems, is an interesting question for further experiments. Experimental parameters To compute the binding energy EB, we use the exact analytical solution of the Schrödinger equation for two ultracold atoms confined in an axially symmetric harmonic potential provided by ref. 26. The solution

depends on three parameters: the three-dimensional s-wave scattering length a3D and the trap frequencies in the radial (fr) and axial (fz) directions. The scattering length is set via the magnetic offset field B by making use of the Feshbach resonance of the |1⟩–|3⟩ mixture at B0 = 690 G (ref. 24). To verify the accuracy of the calculated value for the binding energy EB, we compare the analytical result for the excitation energy of the two-body problem to the spectrum measured for two atoms (see Extended Data Fig. 2). The trap frequencies are determined using the same sequence as explained above. The only difference is that the system is excited by modulating the confinement at a magnetic field of B = 568 G, that is, at the zero crossing of the scattering length. In the noninteracting system the lowest monopole excitation is at f = 2fr(2fz), for radial (axial) excitations. This allows us to extract the frequencies by modulating the harmonic confinement in the respective directions. The measurements for fr for 6 and 12 atoms are shown at EB = 0 in Fig. 3a, b. For the axial direction we find fz = 6,803 ± 2 Hz.

Different modulation schemes As discussed above, we use two different schemes to drive excitations above the closed-shell ground state. All the data shown in the main text, except the spectra taken at EB = 0, are recorded by modulating the optical lattice. We modulate the depth of the axial confinement at frequencies well below its bandgap, in order not to create excitations along this tightly confined direction. The wavefunction adiabatically follows the potential change and is compressed periodically. The effective 2D interaction is obtained by integrating out the wavefunction along the third direction. Thus this effectively modulates only the two-dimensional binding energy. The strength of the modulation corresponds to a change of EB by approximately 2%. For reference, we compare the modulation of the radial trapping potential with the modulation of interactions at 300 G. We find that both schemes lead to different relative transition probabilities for the Higgs and the other excited modes (see Extended Data Fig. 3). The locations of the respective excitations in the spectrum remain unaffected. The qualitative result that a modulation of interaction strength leads to an increased transition matrix element of the Higgs mode precursor is consistent with the few-body calculation by ref. 9. Consequently, fz modulation was applied for all the data shown in the main text, except for the spectra taken at EB = 0. Anisotropy and anharmonicity The small size of our OT results in a finite anharmonicity of the trapping potential. The transition frequency from the lowest shell two shells up is about 10% larger than the transition frequency from the second shell to the fourth shell. The anharmonicity extracted from the noninteracting spectra matches our expectations, owing to the finite size of the optical tweezer. Its waist of around 5 μm is of the same magnitude as the harmonic oscillator length lho = ħ /ωm  ≈ 1.3 μm and the atoms probe the non-harmonic parts of the trap. In addition, the trap shows a slight anisotropy (ωx – ωy)/(ωx + ωy) of approximately 2%. These corrections should not affect the qualitative behaviour of the measured spectra for the interacting system. However, they might quantitatively change the coupling strengths to the different modes and the exact shape of the Higgs mode precursor. We neglect the influence of the anharmonicity for calculating the binding energy. A comparison of the calculated and measured excitation energies for two particles, shown in Extended Data Fig. 2, confirms that this is only a small effect, as expected. Numerical modelling We model the experiment using a trapping potential of the form V = V2D(x , y) +

1 mω 2z z 2 . 2

(1)

Here, V2D describes the potential in the x–y plane, which is provided by the Gaussian profile of the OT so that  y2   −γx 2 + γ  /(l 2hoA) A    ,  V2D(x , y) = ħωr × 1 − e   2  

(2)

with the (lowest order) harmonic trapping frequency ωr. The parameter γ controls the ratio of the trap frequencies in the x- and y-directions and hence the anisotropy. The parameter A is the depth of the trap (in units of ħωr) and determines the anharmonicity. Pure contact interactions in three dimensions are represented by a term g3Dδ(rk − rl ). We assume ωz ≫ ωr so that that the fermions reside in the lowest harmonic oscillator state along the z-direction. Consequentially, the z-direction can be integrated out, yielding a quasi-2D model with an effective coupling strength ∼ g (ref. 8). Since the integral depends on the wavefunction, ∼ g increases with the confinement in the z-direction. It follows that modulating ωz will modulate the effective 2D coupling strength, which is an efficient way to excite the Higgs mode precursor9. We end up with the effective 2D N-particle Hamiltonian N  2  −ħ 2 ∇i + V2D(ri ) + g~ H^2D = ∑   i =1   2m

∑ δ(rk − rl),

(3)

kl

2

2

where ri = (xi, yi) is the coordinate of particle i, ∇i2 = ∂ xi + ∂ yi, and k and l denote fermions in the two different hyperfine (pseudo spin) states. Only fermions with different spins interact for a contact potential. We solve the Hamiltonian numerically by using a single-particle basis of 2D isotropic harmonic oscillator states |m, n⟩ with energies (2n + |m| + 1)ħωr. They are determined by principal and angular momentum quantum numbers n = 0, 1, 2… and m = 0, ± 1, ± 2… respectively. Following ref. 9, we employ a two-parameter cutoff scheme in which we define a single-particle subspace by the highest allowed total single-particle energy E cut sp , and then define the relevant many-body subspace in terms of the maximum many-body energy. For circular symmetry (γ = 1) the many-body eigenstates split into subspaces with different total angular momentum Lz. In this case, it is sufficient to consider the Lz = 0 subspace only, since a modulation of the trapping frequency ωz does not impart angular momentum to the ground state. However, the experimental trap is slightly anisotropic (γ ≈ 0.99) and as a result a finite coupling between subspaces with different Lz has to be considered. To accelerate the convergence of our calculations for an approximately circular trap, we neglect negative angular momentum states in our basis and include a subspace of total angular momenta Lz = 0, ħ and 2ħ only. This enables us to use a higher energy cutoff by exploiting the (approximate) symmetry between the positive and negative Lz states. Care has to be taken when treating the δ(r) interaction in equation (3), which causes an ultraviolet divergence. This divergence can be regularized by expressing the coupling constant ∼ g and the cut-off in terms of the two-body binding energy EB (refs. 9,23). We note that our regularization procedure accounts for the inherent logarithmic energy dependence of the 2D contact interaction. The parameters of the 2D potential (equation (2)), used in the numerical calculations, are obtained by comparing to the lowest monopole excitation observed experimentally with six noninteracting particles. This yields γ ≈ 0.99 and A = 20 for the anisotropy and anharmonicity, respectively. The combination of a slightly broken circular symmetry with the high degree of collectivity when passing through a few-body precursor of a quantum critical point makes the modelling by a full numerical diagonalization of the many-particle Hamiltonian a highly non-trivial task. The complexity of configuration-interaction diagonalization methods often makes it challenging to reach full convergence in terms of the basis set used, which we also found to be the case in our present simulations, where we made partial use of ref. 29. We employed a parallel full diagonalization

using implicitly restarted Arnoldi routines for sparse matrices of bases with up to about ten million states. Trends going towards larger numbers of basis states were carefully analysed. In this way, we were able to reach consistent results for the N = 6 particle system, whereas reasonably converged numerical calculations for a N = 12 particle system in a realistic trap are currently beyond our reach. We emphasize, however, that the qualitative features of the spectra, including the presence of three Higgs-like modes with a non-monotonic behaviour are reassuringly robust with increasing basis-set size. In general, we found that the essential features of the Higgs modes, such as the minimum in energy and the strong coupling via interaction modulation, become more pronounced with increasing basis set size.

Comparing experimental and numerical spectra Extended Data Fig. 6a shows the calculated as well as experimentally observed spectrum for the N = 6 particle system, with N = 20 and γ = 0.99. The numerical calculcation includes states up to E cut sp = 10ħωr and up to a many-body energy of 28ħωr. We see that there is reasonable agreement between theory and experiment and that all qualitative features in the spectrum are recovered by the calculations. In particular, the existence of two non-monotonic Higgs branches is confirmed by the calculations. The lowest branch connects smoothly to the Lz = 0 Higgs mode for the isotropic case, whereas the two higher modes connect smoothly to the Lz = ±2ħ Higgs modes when the anisotropy goes to zero (γ = 1). The latter modes have a higher energy because they describe Cooper pairing with finite angular momentum, and are almost degenerate owing to the small anisotropy. Since our numerics include only positive angular momentum states, only one of the Lz =  ± 2ħ Higgs modes is visible in the calculated spectrum. In the experimental data they appear as a single resonance owing to the small energy splitting for small anisotropy. Extended Data Fig. 6b shows the spectrum weighted with the matrix element E Γint = |⟨G| ∑ δ(rk − rl )|E ⟩|2 , k,l

(4)

which gives the coupling between the ground state |G⟩ and the excited state |E⟩ when the interaction strength is modulated in the experiment. The slight breaking of the circular symmetry leads to the coupling of the ground state to all three Higgs modes. This plot also highlights why the manifold of states around the energy ħωr is not observed experimentally: These states correspond to exciting one fermion one shell up, which changes the angular momentum by ±ħ. The trap anisotropy leads solely to quadrupole excitations, where the angular momentum changes by ΔLz = ±2ħ. As a result, these modes are not visible in the experimental spectrum.

Limitations of the model We note that the agreement between the theoretical and experimental spectra becomes worse in the region where the binding energy is substantially larger than the critical binding energy. There are two main reasons that explain this behaviour. First, this region corresponds to the few-body precursor of the BEC regime of dimers and the large binding energy requires a cut-off beyond what is numerically feasible. Second, the modelled potential only approximates the actual experimental confinement. The fitting of the potential parameters is performed only for a small set of experimental values, all within the lowest few harmonic shells. This allows for a qualitative simulation of the low-lying excitation spectra observed in the few-body experiments but we did not perform a systematic fit including all observed modes. This is because the experimental trap is not precisely described by equation (2): there are deviations away from the Gaussian profile especially close to the continuum. More importantly, the experimental aspect ratio ωz/ωr ≃ 6.8 implies that three-dimensional effects become important at higher energies. In this regime, effects such as excitations to higher axial states or the confinement-induced effective range30 have to be considered. Including such effects would enable us to achieve a more quantitative agreement

Article but is at the same time computationally very challenging. In addition, this would require very precise control and knowledge of the experimental potential at high energies beyond what is achieved here. We estimated that, in the regime relevant for this work, these effects do not change the results qualitatively and therefore do not include them in our model. In conclusion, the overall agreement between the numerical calculations and the experiment confirms the physical interpretation of the data. In particular, we are indeed observing a few-body precursor of a quantum phase transition with the associated emergence of a Higgs mode. We note that obtaining an even better quantitative agreement requires a more accurate determination of the shape of the trap, inclusion of three-dimensional physics for higher energies, and the use of substantially larger computational resources. All qualitative features of the low energy spectrum, however, were found to be insensitive towards these effects. We finally note that simpler theoretical approaches are not able to capture the non-monotonous behaviour of the pair excitation mode. Mean-field theory would predict only modes above 2ωr. Second-order perturbation theory would capture the initial decrease of the excitation frequency of the pair mode. It would however fail to describe the non-monotonic behaviour at larger interaction strengths, as this is caused by non-perturbative pairing correlations.

Data availability The data that support the findings of this study are available from the corresponding authors upon reasonable request. Source data are provided with this paper.

29. Cremon, J. Quantum Few-Body Physics with the Configuration Interaction Approach: Method Development and Application to Physical Systems. Ph.D. thesis, Lund University (2010). 30. Hu, H., Mulkerin, B. C., Toniolo, U., He, L. & Liu, X.-J. Reduced quantum anomaly in a quasi-two-dimensional fermi superfluid: significance of the confinement-induced effective range of interactions. Phys. Rev. Lett. 122, 070401 (2019).

Acknowledgements The experimental work has been supported by the ERC consolidator grant 725636, the Heidelberg Center for Quantum Dynamics, the DFG Collaborative Research Centre SFB 1225 (ISOQUANT) and the European Union’s Horizon 2020 research and innovation programme under grant agreement number 817482 PASQuanS. K.S. acknowledges support by the Landesgraduiertenförderung Baden-Württemberg. P.M.P. acknowledges funding from the Daimler and Benz Foundation. S.M.R. and J.B. acknowledge financial support by the Swedish Research Council, the Knut and Alice Wallenberg Foundation and NanoLund. G.M.B. acknowledges financial support from the Independent Research Fund Denmark—Natural Sciences via grant number DFF-8021-00233B and the Danish National Research Foundation through the Center of Excellence “CCQ” (grant agreement number DNRF156). Author contributions L.B. and M.H. contributed equally to this work. L.B., M.H. and K.S. performed the measurements and analysed the data. J.B., S.M.R. and G.M.B. developed the theoretical framework. J.B. performed the numerical calculations. P.M.P. and S.J. supervised the experimental part of the project. All authors contributed to the discussion of the results and the writing of the manuscript. Competing interests The authors declare no competing interests. Additional information Supplementary information is available for this paper at https://doi.org/10.1038/s41586-0202936-y. Correspondence and requests for materials should be addressed to L.B. or M.H. Peer review information Nature thanks Hui Hu and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Peer reviewer reports are available. Reprints and permissions information is available at http://www.nature.com/reprints.

Preparation 1D Spill 1D to 2D

20

Excitation

2D Spill

6/12

Modulate

Detection 2D Spill

2D to 1D Counting

6/12

Microtrap Power

2D Trap Power

Tweezer Size

0

Time

Extended Data Fig. 1 | Experimental protocol. The sequence can be separated into three parts. First, several evaporation and spilling stages are combined with a transfer from a quasi-1D to a quasi-2D trap geometry. This is needed to prepare closed-shell ground state configurations of up to 12 atoms. Second, we excite the system at some defined frequency fexc and magnetic offset field B using a sinusoidal modulation of either the radial or axial confinement. Third, detection is implemented by spilling to the ground state a second time and a transfer of all remaining atoms to the magneto-optical trap, where we count them.

Article 2.5

Measurement Calculation

2.4

Eexc / hfr

2.3

2.2

2.1

2.0 700

750

800

Magnetic Field, B [Gauss]

850

Extended Data Fig. 2 | Excitation spectrum for two particles. We define the two-body excitation energy Eexc as the energy difference between the ground state and the lowest monopole excitation of the two atom system. It is measured using the same modulation scheme as for the Higgs mode. The system is initialized with one filled shell, that is, two particles. The analytical solution of the two-body problem (solid line) shows good agreement with the measurement (blue points). The systematic uncertainty of around 2% on the measured radial and axial trap frequencies that enters into the analytical solution is indicated by the grey error band. Residual systematic deviations can be explained by the trap anharmonicity. For the measurement (blue points) error bars are extracted from the fit to the spectrum and are smaller than the data points.

5.5

Mean Atom Number

5.0 4.5 4.0 3.5 3.0 2.5

Modulation of fr

Mean Atom Number

5.4 5

4.6 4.2 3.8 1400

Modulation of EB 1500

1600

1700

1800

Excitation Frequency [Hz]

1900

2000

2100

Extended Data Fig. 3 | Comparison of different modulation schemes. A modulation of the radial trap frequency leads to similar transition probabilities for the pair excitation mode and the higher excited states (top). In contrast, a modulation of the axial confinement effectively only modulates the interaction strength and couples predominantly to the pair excitation mode. Excitations to higher states are suppressed by this modulation scheme (bottom). This qualitative observation agrees with the coupling elements that were predicted in ref. 9. The two modulation amplitudes have been chosen such that they lead to a similar response of the pair excitation mode. The data are taken for EB = 0.09hfr. For this measurement the radial trap frequency was 2fr = 1,660 Hz. Error bars show the standard error of the mean. Each data point is the average of at least 24 measurements.

Article 1.0

a

0.4

b

0.9

2.1 2.0

0.3

0.7

1.8

0.2 0.6

1.7

0.2

0.5

1.6

Excitation Energy / hfr

0.4

0.3

0.8

1.9

1.5

0.1

0.1

0.4

N=6

1.4

2.2

0.5

c

d

0.3 0.2

N=5

e

0.0 0.1

N=4

f

0.0 0.1

2.1 2.0

Probability of Occurrence, PN

2.2

1.9 0.1

1.8 1.7 1.6 1.5

N=3

1.4 0.4

0.8

1.2

1.6

2.0

0.0

N=2 0.4

0.8

1.2

EB / hfr

Extended Data Fig. 4 | Probabilities of different atom numbers remaining in the lowest two shells for the N = 6 initial state. a–f, The probabilities for different remaining atom numbers after modulating the 6-atom ground state with a defined frequency given on the y axis and subsequent removal of excited atoms. All possible excitations manifest themselves by a reduced probability of

1.6

2.0

0.0

N=1 0.4

0.8

1.2

1.6

2.0

0.0

remaining in the ground state of 6 atoms (a). We find that the lowest excitation, or Higgs mode, mostly consists of excitations to four atoms (c), while the higher excited peaks are predominantly generated by the loss of a single atom (b). For each setting the experiment is repeated between 42 and 47 times.

a

0.8

2.2

b

c

0.7 2.0

0.5

0.3

0.4

0.6

1.8

0.5

0.3

0.2

0.4 0.2

0.3

Excitation Energy / hfr

1.4

0.1

0.2

1.2

N = 12

d

N = 11

0.0 0.4

2.2

0.1

0.1

N = 10

0.0

e

0.0

f 0.3

2.0

Probability of Occurrence, PN

1.6

0.3

1.8

0.1

0.2 0.2

1.6

1.4

0.1

0.1

1.2

N=9 0.4

0.8

1.2

1.6

0.0

N=8 0.4

0.8

EB / hfr

Extended Data Fig. 5 | Probabilities of different atom numbers remaining in the lowest three shells for the N = 12 initial state. a–f, The probabilities for different remaining atom numbers after modulating the 12-atom ground state with a defined frequency given on the y axis and subsequent removal of excited atoms. All possible excitations manifest themselves by a reduced probability of

1.2

1.6

0.0

N=7 0.4

0.8

1.2

1.6

0.0

remaining in the ground state of 12 atoms (a). We find that lowest excitation, or Higgs mode, mainly consists of excitations to ten atoms (c), while the higher excited peaks are predominantly generated by the loss of even more atoms (d–f). For each setting the experiment is repeated between 19 and 63 times.

Article a

0.2

b

Excitation Energy / hfr

Scaled Transition Matrix Element

2.5

2.0

0.1

1.5

Numerics Experiment 1.0

0

0.5

1

EB / hfr

1.5

0 2

Extended Data Fig. 6 | Numerically calculated excitation spectrum for 6 particles. a, The level spectrum obtained by exact diagonalization with parameters A = 20 and γ = 0.99 for the potential as well as the experimental results. The calculation includes states up to E cut sp = 10ħω r and up to a many-body energy of 28ħωr. For comparison the experimental data are shown by green diamonds. Values and errors bars are obtained as in Fig. 3c. b, The numerically

0

0.5

1.0

EB / hfr

1.5

2.0

calculated excitation spectrum for a modulation of the interaction strength. As in the experiment we observe that this modulation couples to two non-monotonous modes. We note that the calculations performed for b employ a smaller cut-off (E cut sp = 6ħω r and a maximal many-body energy of 24ħωr) than in a owing to the computational demand in calculating the matrix element (equation (4)).

Article

Spin-enhanced nanodiamond biosensing for ultrasensitive diagnostics https://doi.org/10.1038/s41586-020-2917-1 Received: 24 May 2019 Accepted: 16 September 2020 Published online: 25 November 2020 Check for updates

Benjamin S. Miller1,2 ✉, Léonard Bezinge1, Harriet D. Gliddon1, Da Huang1, Gavin Dold1,3, Eleanor R. Gray1, Judith Heaney4, Peter J. Dobson5, Eleni Nastouli6, John J. L. Morton1,3 & Rachel A. McKendry1,2 ✉

The quantum spin properties of nitrogen-vacancy defects in diamond enable diverse applications in quantum computing and communications1. However, fluorescent nanodiamonds also have attractive properties for in vitro biosensing, including brightness2, low cost3 and selective manipulation of their emission4. Nanoparticle-based biosensors are essential for the early detection of disease, but they often lack the required sensitivity. Here we investigate fluorescent nanodiamonds as an ultrasensitive label for in vitro diagnostics, using a microwave field to modulate emission intensity5 and frequency-domain analysis6 to separate the signal from background autofluorescence7, which typically limits sensitivity. Focusing on the widely used, low-cost lateral flow format as an exemplar, we achieve a detection limit of 8.2 × 10−19 molar for a biotin–avidin model, 105 times more sensitive than that obtained using gold nanoparticles. Single-copy detection of HIV-1 RNA can be achieved with the addition of a 10-minute isothermal amplification step, and is further demonstrated using a clinical plasma sample with an extraction step. This ultrasensitive quantum diagnostics platform is applicable to numerous diagnostic test formats and diseases, and has the potential to transform early diagnosis of disease for the benefit of patients and populations.

Fluorescent nanodiamonds (FNDs) that contain nitrogen-vacancy (NV) centres—defects that have optical transitions within the bandgap—have received considerable attention as a spin system for use as a qubit in quantum computing and communication, and for quantum sensing1,4,8,9. Such applications arise from the ability of the negative NV (NV−) centre spin state to be optically initialized and read out, while being manipulated using direct current and microwave magnetic fields. FNDs also have attractive fluorescent properties—including a high quantum yield and a lack of blinking and photobleaching—as well as high stability, low cytotoxicity2,10, the availability of surface groups for bio-functionalization11, and ease of mass manufacture, such as by the milling of high-pressure, high-temperature diamond3,12. The sensing applications of NV centres4 include magnetic field quantification13–15, temperature sensing16,17 and biological labelling2,18, the latter of which includes cellular imaging19, drug delivery20 and contrast enhancement in magnetic resonance imaging21. A key advantage of NV− centres is that, unlike neutral (NV0) centres, their fluorescence can be selectively modulated by spin manipulation4, which enables signal separation for imaging in high-background environments. This property has been used to improve the contrast for imaging by modulating the fluorescence with microwaves5,22, magnetic fields23,24 or near-infrared light25. Here we investigate the use of FNDs for in vitro diagnostics. Communicable diseases represent an enormous global health challenge, and disproportionately affect poorer populations with limited

access to healthcare26. At the end of 2015, there were 36.9 million people living with HIV worldwide, of whom 9.4 million (25%) were unaware of their HIV status27. Early diagnosis is crucial for effective treatment and prevention and benefits patients and populations. For example, patients in the UK that started antiretroviral therapy for HIV after a late diagnosis had a reduction in life expectancy of more than 12 years compared with those that started treatment after an earlier diagnosis28. The earliest marker of HIV is viral RNA, which is detectable 7 days before antigen and 16 days before antibodies29. Point-of-care nucleic acid testing therefore offers the potential for diagnosis at an earlier stage than either existing laboratory-based nucleic acid tests or point-of-care protein tests. Rapid point-of-care tests have transformed access to disease testing in various community settings, including clinics, pharmacies and the home30. Among the most frequently used tests worldwide are paper microfluidic lateral flow assays (LFAs), of which 276 million were sold in 2017 for malaria alone31. LFAs satisfy many of the REASSURED criteria32 for diagnostics; however, despite their widespread use, they are still limited by inadequate sensitivity to the low levels of biomarkers present in early disease. Fluorescent markers can be highly sensitive, but their practical application is limited by background fluorescence from the sample, the substrate or the readout technique. In the case of nitrocellulose substrates used in LFAs, there is a considerable background

1 London Centre for Nanotechnology, University College London, London, UK. 2Division of Medicine, University College London, London, UK. 3Department of Electronic and Electrical Engineering, University College London, London, UK. 4Advanced Pathogens Diagnostic Unit, University College London Hospitals, London, UK. 5The Queens College, University of Oxford, Oxford, UK. 6Department of Virology, University College London Hospitals, London, UK. ✉e-mail: [email protected]; [email protected]

588 | Nature | Vol 587 | 26 November 2020

a

b

Fluorescent nanodiamond

55

Fluorescent nanodiamond

0-

nm

ex

cit

Resonator

at

ion

ion

iss

nm

em

5-

67

dsDNA

NV defect Streptavidin

Lateral flow strip

Resonator

Amplitude-modulated microwave electromagnetic field

Nitrocellulose

Fig. 1 | Schematic illustration of the use of FNDs in LFAs. a, Illustration of the concept of using FNDs in an LFA. The binding of DNA modifications causes FNDs to be immobilized at the test line in a sandwich-format LFA. The inset shows the atomic structure of an NV− centre, the origin of FND fluorescence. An omega-shaped stripline resonator applies a microwave field over the LFA, modulating the fluorescence intensity. b, Schematic showing more detail of the principle. FNDs are immobilized at the test line in a sandwich structure in

the presence of double-stranded DNA (dsDNA) amplicons. Exciting at 550 nm (green) produces fluorescence emission centred at 675 nm (red), imaged with a camera. An amplitude-modulated microwave field, applied by the resonator, selectively modulates the fluorescence of the immobilized FNDs at a set frequency. This enables specific separation of the FND fluorescence from background fluorescence in the frequency domain, to improve the signal-to-noise ratio.

autofluorescence7, which inherently limits sensitivity. Various methods have been reported to reduce this effect, such as membrane modification to reduce background fluorescence33, exciting in the near-infrared range and using upconverting nanoparticles34, and time-gated detection using long-persistent phosphors35 to separate out the shorter-lived background fluorescence. The use of these methods has led to improvements in sensitivity of approximately 10-fold compared with gold nanoparticles, but they are limited by relatively low brightness. Here we show the use of FNDs as a fluorescent label in an LFA format, demonstrating their first—to our knowledge—use for in vitro diagnostics. This strategy takes advantage of the high brightness and selective modulation of FNDs, and the use of a narrowband resonator enables the low-power generation of microwave-frequency electromagnetic fields—suitable for a point-of-care device—that can efficiently separate the signal from the background in the frequency domain by lock-in6 detection. After characterization, functionalization and optimization, we applied FND-based LFAs first to a model system and then to a molecular HIV-1 RNA assay in order to demonstrate their clinical utility.

coupled to an omega-shaped stripline resonator that provides a uniform peak field over the measurement area (the area of the resonator). The fluorescence of the FNDs, and its response to the microwave field, was measured on the nitrocellulose paper substrate. To investigate the fluorescence intensity as a function of microwave frequency on paper, a wideband resonator was used to perform continuous-wave electron spin resonance spectroscopy, as shown in Extended Data Fig. 1a–c. A plot of FND fluorescence over a wide frequency range is shown in Fig. 2b, showing two peaks at ∆E = 2.87 GHz and ∆E* = 1.43 GHz, corresponding to the triplet level splitting in the ground and excited states, respectively. Figure 2c shows a narrowband resonator (characterized in Extended Data Fig. 1d–f) that is designed to have a resonant frequency at 2.87 GHz with quality factor Q = 100; this induced a reduction in measured fluorescence of around 3–6% (Extended Data Fig. 1f), which varied linearly with the microwave input power in dBm (see Extended Data Fig. 1g, h). The excitation and emission spectra of FNDs are shown in Fig. 2d. The presence of NV− centres is indicated by the zero-phonon line at around 640 nm. Using an amplitude-modulated microwave field to specifically vary the FND fluorescence at a fixed frequency enables the application of a computational lock-in algorithm6 (shown schematically in Extended Data Fig. 2a) to selectively extract signals at the reference frequency. This lock-in analysis, shown in ref. 5, separates the periodic FND fluorescence from the non-periodic background fluorescence (the autofluorescence of nitrocellulose), thus improving sensitivity. The fluorescence modulation is shown in Fig. 2e, f. Figure 2e shows pixel variation over time: the test line, on which FNDs are immobilized, has a high variance compared to the background, which does not modulate and has low variance. The time series is shown in Fig. 2f (top), in which a square-wave 4 Hz amplitude-modulated microwave field modulates the fluorescence intensity. Application of the lock-in algorithm over a small frequency range gave the plot in Fig. 2f (bottom)—an absolute sinc function, the Fourier transform of a square pulse. The maximum response is shown when the reference frequency matches the modulation frequency at 4 Hz. The optimization of modulation frequency, sampling frequency, exposure time and measurement time are shown in Extended Data Fig. 2b–e. Microwave generation was miniaturized using a voltage-controlled oscillator, amplifier, and

Microwave modulation of FND emission on paper An illustration of the use of FNDs in LFAs is shown in Fig. 1. FNDs can be used as nanoparticle labels on nitrocellulose strips, on which they undergo a multiple-step binding assay with little user input to bind at the test line in the presence of the target nucleic acids. After immobilization, the fluorescence of the FNDs can be modulated at a fixed frequency using a microwave field, enabling them to be specifically detected and quantified. Figure 2a shows an energy-level diagram of the NV− centre, which is the origin of FND fluorescence. The triplet ground state is optically driven into an excited triplet state, which then undergoes radiative decay back to the ground states. Throughout the process, the electron spin state (ms = 0, ±1) is conserved; however, the ms =  ±1 excited-state levels can decay into a metastable ‘dark’ state with a corresponding reduction in fluorescence4. Resonant microwave radiation drives electron spin population from the ms = 0 to the ms =  ±1 levels, reducing the intensity of the fluorescence. The microwave field was produced by a voltage-controlled oscillator connected to an antenna, capacitively

Nature | Vol 587 | 26 November 2020 | 589

Article NV– centre

b

ms ±1

99.5

ΔE*

0

Singlet ‘dark’ states

ΔE* = 1.43 GHz

99.0 98.5 98.0 97.5

±1

ΔE

1.0

0

d

1 mm

1.5

2.0

2.5

3.0

Frequency (GHz)

1.0

Normalized intensity (a.u.)

ΔE = 2.87 GHz

97.0

Excitation Microwave off Microwave on

0.8

e

Pixel intensity variation (counts) –1,500 –1,000 –500 0

f

Intensity (%)

Triplet ground states

100

Amplitude (%)

Triplet excited states

c

100.0

Intensity (%)

a

2

98 96

0

0.5 Time (s)

0.6 0.4 0.2 0 450

500

550

600

650

700

750

800

850

Wavelength (nm)

1.0

1 0 3.0

3.5

4.0

4.5

5.0

Reference frequency (Hz)

Fig. 2 | Microwave modulation of FNDs on paper. a, Energy-level diagram of an NV− centre. Splitting in the ground and excited states is labelled as ∆E and ∆E*, respectively. b, The variation in FND fluorescence under microwave fields of different frequency, showing dips at energies ∆E and ∆E*. Some peak splitting is observed around the centre frequency due to a magnetic field of approximately 3 G generated by the resonator (measured with a Gaussmeter). c, A microscope image of the omega-shaped stripline resonator used to produce a uniform field at 2.87 GHz. d, Excitation and emission spectra of

FNDs. The green shaded area shows the filtered excitation light used. The area of the emission spectrum is reduced under the application of the microwave field. e, The pixel variation at the test line (with immobilized FNDs) of an LFA strip under an amplitude-modulated microwave field. f, Top, the variation of mean fluorescence intensity over time under the application of the same amplitude-modulated field. Bottom, applying a lock-in algorithm over a range of frequencies gives a sinc function with a peak at the modulation frequency. a.u., arbitrary units.

a custom power and timing circuit that measured 65 mm × 38 mm (Extended Data Fig. 2f, g).

shown schematically in Fig. 3a. The high-affinity interaction between biotin and streptavidin, as well as the flow rate and the high binding capacity of nitrocellulose, ensures that the residency time of the FNDs at the test line is much longer than the binding time of biotinylated FNDs to the streptavidin (Extended Data Fig. 5a, b). This implies that all FNDs bind at the test line, making it ideal for benchmarking the best-case sensitivity and comparing with other nanomaterials. The LODs were quantified for FNDs with three different particle-core diameters: 120, 200 and 600 nm. The resulting fluorescence signals from the LFA test line were analysed using lock-in analysis and conventional intensity analysis, in which the difference in fluorescence intensity between the test line and the background is measured. The results were compared with those of gold nanoparticles, which are frequently used in LFAs40. The signal-to-blank ratios (SBRs) were plotted against the concentration of 600-nm FNDs (Fig. 3b). Each dilution series was fitted to a simple linear regression, and the LOD was defined as the intersection of the lower 95% confidence interval of the linear fit with the upper 95% confidence interval of the blanks41. Figure 3c shows images of the test lines as well as time series of the fluorescence modulation at each FND concentration, demonstrating that signal modulation can be measured well below the concentration at which there is a visible test line. LODs were found to be 200 aM, 46 aM and 820 zM for particles of 120, 200 and 600 nm in diameter, respectively (Extended Data Fig. 5d). The best LODs were found with the larger particles, because the lock-in amplitude scales with the fluorescence modulation intensity, which in turn scales with the number of NV− centres. The number of NV− centres per particle scales with the volume, so the LOD should scale with the

Fundamental limits using a biotin–avidin model After fluorescence characterization and optimization of the modulation parameters, FNDs were functionalized with biomolecules for incorporation into LFAs. We used FNDs with a hydrophilic polyglycerol layer in order to reduce non-specific binding to the nitrocellulose36 (Extended Data Fig. 3a), a key factor that limits LFA sensitivity. Three sizes of FND–polyglycerol (characterized by dynamic light scattering, as shown in Extended Data Fig. 3b) were functionalized with antibodies via the activation of the alcohol groups of polyglycerol with disuccinimidyl carbonate37, as shown in Extended Data Fig. 3c. Characterization by scanning electron microscopy, dynamic light scattering and Fourier transform infrared spectroscopy (Extended Data Fig. 3d–i) showed successful conjugation, with minimal aggregation upon functionalization and increases in size that were consistent with the size of the conjugants38,39. The number of active binding sites on the surface of 600-nm-diameter FNDs was subsequently quantified using quantitative polymerase chain reaction (qPCR; Methods, Extended Data Fig. 4). The measured value of 4,300 active binding sites per FND is consistent with geometric calculations of the number of antibodies that could bind. The fundamental limit of detection (LOD) of FND-based LFAs was investigated using a model biotin–avidin interaction. A serial dilution of a bovine serum albumin (BSA)–biotin–functionalized FND suspension was run on LFA strips, on which the FNDs bound directly to a printed poly-streptavidin test line (rather than in a sandwich formation) as 590 | Nature | Vol 587 | 26 November 2020

b

104

Signal-to-blank ratio

a

103

FND lock-in FND intensity analysis Gold nanoparticles

FND–BSA–Biotin

Streptavidin

Single-copy detection of HIV-1 RNA

102 820 zM

84 aM

81 fM

101 100

Paper

10–20

c

10–18 10–16 10–14 FND concentration (M)

10–12

FND concentration 1.8 fM

70 aM

110 zM

2.8 aM

0M

Fluorescence images

Signal modulation (pixel variation) ×10

0

1

0

1

×100

0

1 Time (s)

×100

0

1

×100

0

1

Fig. 3 | Characterizing the fundamental limit of detection using biotin–avidin binding of FNDs on LFAs. a, Schematic of the assay. FNDs functionalized with BSA–biotin were run on streptavidin-printed LFA strips, on which they bound directly to the test line. The arrow shows the flow direction. b, The SBRs of a dilution series of 600-nm FNDs were measured by both lock-in analysis and conventional intensity analysis, and compared to that of 50-nm gold nanoparticles. LODs were 820 zM, 83 aM and 81 fM, respectively. Lock-in analysis provided a 100-fold improvement over conventional intensity analysis, and a 98,000-fold improvement over gold nanoparticles. Dots show the means and error bars show s.d. of n = 3 technical replicates and n = 3 measurement replicates for each sample. c, An illustration of the comparison between lock-in and conventional analysis. Top, example fluorescence images at various concentrations; bottom, intensity–time plots, showing that a periodic signal is still evident after the test line is no longer visible in the images.

cube of the diameter. Additionally, surface effects reduce the fluorescence of NV− centres close to the surface, so a larger volume-to-surface ratio should increase fluorescence intensity. LODs using 600-nm FNDs were 820 zM and 83 aM using lock-in and conventional analysis, respectively, indicating that lock-in analysis yields a 620-fold improvement in the signal-to-background ratio and a 100-fold improvement in the LOD. This increases to an 810-fold improvement in signal-to-background ratio using 120-nm FNDs, giving a 380-fold improvement in the LOD. This fundamental LOD of 820 zM corresponds to 0.5 particles per μl, or just 27 particles in a 55-μl sample. For comparison, the same experiment was performed with 50-nm gold nanoparticle labels, which are frequently used in LFAs40 owing to their ease of functionalization and strong light absorption. The 600-nm FNDs were five orders of magnitude more sensitive than gold nanoparticles. The size of gold nanoparticles that can be used on LFAs is also limited by the broadening of the plasmonic peak, whereas FNDs at larger sizes become brighter. Owing to the low numbers of particles detected, the LODs of biological assays are expected to be limited by non-specific binding and equilibrium considerations, rather than by the fundamental sensitivity of FNDs.

This platform was then applied to a proof-of-concept assay for the detection of DNA amplicons. The assay is based on a reverse transcriptase–recombinase polymerase amplification (RT–RPA) reaction for the detection of HIV-1 RNA, which is performed with modified primers to form a sandwich structure on the nitrocellulose (Fig. 4a). After assay optimization (Methods, Extended Data Figs. 6–8), LFAs were performed with serial dilutions of RT–RPA products using FNDs of three particle sizes (120, 200 and 600 nm). The initial aim was to determine the sensitivity of the detection system, rather than the amplification step, so the amplicon concentration used was measured post-amplification. The resulting plots of SBR against amplicon concentration are shown in Fig. 4b, and fitted to the Langmuir adsorption isotherm model (equation (6), Methods). The LODs were measured as 9.0, 7.5 and 3.7 fM for 120, 200 and 600-nm FNDs, respectively. The LOD of 3.7 fM, achieved with 600-nm FNDs, corresponds to 2,200 copies per μl, or 1.1 × 105 copies in total (190 zmol of DNA). A model ‘amplicon’ (described in Methods and characterized in Extended Data Fig. 9a, b) was used to compare the 600-nm FNDs with 40-nm gold nanoparticles. The resulting LODs, plotted in Extended Data Fig. 9c, show that FNDs provide an approximately 7,500-fold improvement over 40-nm gold nanoparticles. The approximately 13-fold reduction in improvement over gold nanoparticles compared to that found for the biotin–avidin model is due to non-specific binding. The blanks in the DNA assay have the same FND concentration as the positives, whereas in the biotin–avidin assay they are ‘true blanks’ (running buffer only). The resulting small lock-in amplitude in the blanks is around 13-fold higher than a ‘true blank’ signal (noise), showing no significant difference from the blanks from the biotin–avidin assay multiplied by this factor of 13 (two-tailed t-test, P = 0.33). Achieving this level of sensitivity from FND labelling means that a short amplification step before addition of the sample to the LFA could lead to single-copy detection, with typical amplification factors42 for isothermal RPA of 104 in 10 min. This was subsequently demonstrated by performing 10 min (37 °C) RT–RPA reactions on serial dilutions of HIV-1 transcript RNA, before adding a 6× running buffer solution to the purified products and running on LFAs as previously. The resulting SBRs are plotted against RNA input copy number in Fig. 4c, showing a LOD of 1 copy. Positive results were achieved down to a single RNA copy. Statistical analysis of the lock-in amplitudes (analysis of variance) is shown in Extended Data Fig. 10a–c. Owing to the 10-min amplification time, all concentrations of at least 1 copy reach the saturation signal, so a qualitative yes/no result is given. The variation of SBR with amplification time is shown in Extended Data Fig. 10d, which shows the results of single-copy reactions run for different times. A detectable signal was observed after a 7-min amplification time. The sensitivity of the FNDs conveys improved LODs in shorter amplification times compared to previous work with RPA using gold nanoparticles43–46. In addition, as a proof-of-concept, a clinical sample (University College London Hospitals (UCLH) clinical standard, 4 × 104 copies per μl) was successfully detected. This involved the addition of an RNA-extraction step (Fig. 4d), which would need to be adapted for point-of-care testing. RPA has been shown to be relatively robust to complex samples, but this remains a major challenge for the field of nucleic acid testing47. The positive clinical standard had a mean SBR of around 19 compared to the negative plasma control. In order to demonstrate the suitability of this system for rapid detection of early disease, a small proof-of-concept experiment was performed using a seroconversion panel of thirteen samples taken over six weeks, spanning the initial stages of an HIV-1 infection. Extended Data Fig. 10e shows that RNA is detected as early as when using the PCR with reverse transcription (RT–PCR) gold standard, giving positive results for six out of seven RT–PCR-positive samples and zero out of six RT–PCR-negative samples. However, this is a preliminary study, Nature | Vol 587 | 26 November 2020 | 591

Article a

Reverse transcription

ssRNA extraction

Virus in patient sample

FND–anti-DIG

10-min RPA followed by purification

DNA primer binds to viral ssRNA

Primer binds to complementary DNA

RPA amplicons

Streptavidin

DIG–dsDNA–biotin

Paper

12

d 25

10

20

200 nm 600 nm

9.0 fM 7.5 fM

3.7 fM

101

*

Signal-to-blank ratio

102 Signal-to-blank ratio

c

120 nm

*

*

Signal-to-blank ratio

b

8 6 4

15

10

5

2 100 10–16 10–14 10–12 DNA amplicon concentration (M)

10–10

0

10–2

10–1 100 RNA copies

101

102

0

Water control

Negative plasma

Clinical standard

Fig. 4 | Single-copy detection of HIV-1 RNA on LFAs using RT–RPA and FNDs. a, A schematic of the assay. Digoxigenin (DIG) and biotin-modified primers were used in a RT–RPA reaction to produce labelled amplicons, which bind to anti-DIG-functionalized FNDs and streptavidin-printed test lines on the LFA strips, forming a sandwich structure in the presence of amplicons. b, Dilution series of amplicons were run on LFAs for three different FND sizes (120, 200 and 600 nm). Serial dilutions were plotted (dots showing means with error bars showing s.d., n = 3–9 technical replicates, n = 3 measurement replicates), and fitted to the Langmuir adsorption model in equation (6) (Methods). Limits of detection for 120, 200 and 600-nm FNDs were 9.0, 7.5 and 3.7 fM respectively. Asterisks mark the lowest concentrations for each particle size that are significantly different from the blanks at the 95% confidence level, calculated

by ANOVA (P = 10 −7, 10 −5 and 0.03 from Tukey HSD post hoc test for 120, 200 and 600-nm FNDs, respectively). c, Serial dilutions of HIV-1 RNA copies were amplified with RT–RPA (10 min), purified and run on LFAs with 600-nm FNDs. The RNA concentration was plotted against the SBR (data are mean ± s.d., crosses show individual measurements), with n = 4 experimental replicates and n = 3 measurement replicates for each sample. Single-copy sensitivity was achieved (P = 10 −8 from ANOVA from Tukey HSD post hoc test; see Extended Data Fig. 10b, c). d, The system was applied to a proof-of-concept positive clinical sample (UCLH clinical standard) and negative human plasma control, giving a mean SBR of around 19 and a P value (comparison between the negative and positive clinical samples) of 8 × 10 −13 with a t value of −19.3 from an unpaired one-tailed t-test.

and further optimization with clinical samples and a larger study will be required to precisely ascertain the clinical sensitivity.

shown in Supplementary Table 2. The system was also demonstrated with HIV-1-positive and -negative clinical samples with the addition of an RNA-extraction step. The remaining challenges in the translation of this exemplar RNA detection assay towards a rapid point-of-care test that meets the REASSURED criteria32 are summarized in Supplementary Table 3. The incorporation of the amplification step on the LFA strip45 is a major challenge, along with sample processing and RNA extraction in resource-limited settings47, and removal of the wash step. However, the sensitivity of this transduction technique means that there is leeway for sensitivity reductions while retaining clinical relevance: we have demonstrated single-copy detection with a 10-min RT-RPA step, up to 50-fold greater sensitivity than the World Health Organization viral suppression threshold49 of 1,000 copies per ml. This technique is also easily translatable to other assays, for example to amplification methods using modified primers—including existing PCR assays—by changing only the primers; to direct detection by the hybridization of complementary modified probe sequences to a molecular target; or to protein detection in a sandwich assay using modified antibodies. In order to demonstrate this, detection of the HIV-1 capsid protein using FNDs on paper was evaluated (Extended Data Fig. 11), giving a LOD of 120 fM. FNDs are also applicable to a range of other in vitro diagnostic test formats. In addition, owing to the long fluorescence lifetimes of

Conclusions Here we demonstrate the use of FNDs as an ultrasensitive fluorescent label for in vitro diagnostic assays, using microwave-based spin manipulation to increase the signal-to-background ratio and therefore the sensitivity. The system was demonstrated in an LFA format with two assays. Using a biotin–avidin model, a fundamental LOD of 0.5 particles per μl was measured—five orders of magnitude more sensitive than gold nanoparticles—with the caveat of increased cost due to the need for a fluorescence reader (see Supplementary Table 1), but the advantage of enabling the capture of quantitative data (as compared to visual interpretation). By applying FNDs to a sandwich assay for oligonucleotide detection, single-copy sensitivity was achieved for the detection of RNA, using 600-nm FNDs and with a 10-min RT–RPA step. The sensitivity of the FND detection system (LOD of 2,200 copies per μl with RT–RPA amplicons, measured post-amplification) means that a short amplification time is possible while achieving higher sensitivity than has been previously demonstrated with other nanomaterials45,48, making the test more suitable for point-of-care applications. A comparison with other fluorescence-based amplicon detection on LFAs is 592 | Nature | Vol 587 | 26 November 2020

NV centres50 compared to nitrocellulose7, time-gated fluorescence measurements could be used to further improve the sensitivity of FND-based LFAs. The low power consumption (0.25 W microwave power), optical readout and potential portability of this technique render it suitable for ultrasensitive diagnosis and monitoring in low-resource settings, using a portable fluorescence reader or smartphone-based device including microwave modulation. The nature of lock-in readout makes it robust to background light, minimising sensitivity losses when moving from a microscope to such a portable device. FNDs on paper microfluidics offer a sensitive, robust labelling and readout method for in vitro disease diagnostics.

Online content Any methods, additional references, Nature Research reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at https://doi.org/10.1038/s41586-020-2917-1. 1. 2. 3. 4.

5. 6. 7.

8. 9. 10.

11. 12. 13. 14. 15.

16. 17.

18. 19. 20.

Childress, L. & Hanson, R. Diamond NV centers for quantum computing and quantum networks. MRS Bull. 38, 134–138 (2013). Mochalin, V. N., Shenderova, O., Ho, D. & Gogotsi, Y. The properties and applications of nanodiamonds. Nat. Nanotechnol. 7, 11–23 (2012). Boudou, J.-P. et al. High yield fabrication of fluorescent nanodiamonds. Nanotechnology 20, 235602 (2009). Schirhagl, R., Chang, K., Loretz, M. & Degen, C. L. Nitrogen-vacancy centers in diamond: nanoscale sensors for physics and biology. Annu. Rev. Phys. Chem. 65, 83–105 (2014). Igarashi, R. et al. Real-time background-free selective imaging of fluorescent nanodiamonds in vivo. Nano Lett. 12, 5726–5732 (2012). Leis, J., Martin, P. & Buttsworth, D. Simplified digital lock-in amplifier algorithm. Electron. Lett. 48, 259 (2012). Shah, K. G. & Yager, P. Wavelengths and lifetimes of paper autofluorescence: a simple substrate screening process to enhance the sensitivity of fluorescence-based assays in paper. Anal. Chem. 89, 12023–12029 (2017). Childress, L. et al. Coherent dynamics of coupled electron and nuclear spin qubits in diamond. Science 314, 281–285 (2006). Chang, H.-C., Hsiao, W. W.-W. & Su, M.-C. Fluorescent Nanodiamonds Ch. 11 (Wiley, 2018). Yu, S. J., Kang, M. W., Chang, H. C., Chen, K. M. & Yu, Y. C. Bright fluorescent nanodiamonds: no photobleaching and low cytotoxicity. J. Am. Chem. Soc. 127, 17604–17605 (2005). Shenderova, O. A. & McGuire, G. E. Science and engineering of nanodiamond particle surfaces for biological applications (review). Biointerphases 10, 030802 (2015). Chang, Y. R. et al. Mass production and dynamic imaging of fluorescent nanodiamonds. Nat. Nanotechnol. 3, 284–288 (2008). Maze, J. R. et al. Nanoscale magnetic sensing with an individual electronic spin in diamond. Nature 455, 644–647 (2008). Balasubramanian, G. et al. Nanoscale imaging magnetometry with diamond spins under ambient conditions. Nature 455, 648–651 (2008). Tetienne, J. P. et al. Magnetic-field-dependent photodynamics of single NV defects in diamond: An application to qualitative all-optical magnetic imaging. New J. Phys. 14, 103033 (2012). Acosta, V. M. et al. Temperature dependence of the nitrogen-vacancy magnetic resonance in diamond. Phys. Rev. Lett. 104, 070801 (2010). Hsiao, W. W. W., Hui, Y. Y., Tsai, P. C. & Chang, H. C. Fluorescent nanodiamond: a versatile tool for long-term cell tracking, super-resolution imaging, and nanoscale temperature sensing. Acc. Chem. Res. 49, 400–407 (2016). Vaijayanthimala, V. & Chang, H.-C. Functionalized fluorescent nanodiamonds for biomedical applications. Nanomedicine 4, 47–55 (2009). Fu, C.-C. et al. Characterization and application of single fluorescent nanodiamonds as cellular biomarkers. Proc. Natl Acad. Sci. USA 104, 727–732 (2007). Chang, B. M. et al. Highly fluorescent nanodiamonds protein-functionalized for cell labeling and targeting. Adv. Funct. Mater. 23, 5737–5745 (2013).

21. Waddington, D. E. et al. Nanodiamond-enhanced MRI via in situ hyperpolarization. Nat. Commun. 8, 15118 (2017). 22. Hegyi, A. & Yablonovitch, E. Molecular imaging by optically detected electron spin resonance of nitrogen-vacancies in nanodiamonds. Nano Lett. 13, 1173–1178 (2013). 23. Sarkar, S. K. et al. Wide-field in vivo background free imaging by selective magnetic modulation of nanodiamond fluorescence. Biomed. Opt. Express 5, 1190 (2014). 24. Chapman, R. & Plakhoitnik, T. Background-free imaging of luminescent nanodiamonds using external magnetic field for contrast enhancement. Opt. Lett. 38, 1847 (2013). 25. Doronina-Amitonova, L., Fedotov, I. & Zheltikov, A. Ultrahigh-contrast imaging by temporally modulated stimulated emission depletion. Opt. Lett. 40, 725 (2015). 26. Bhutta, Z. A., Sommerfeld, J., Lassi, Z. S., Salam, R. A. & Das, J. K. Global burden, distribution, and interventions for infectious diseases of poverty. Infect. Dis. Poverty 3, 21 (2014). 27. Global HIV & AIDS statistics — 2018 fact sheet. https://www.unaids.org/en/resources/ fact-sheet (UNAIDS, 2018). 28. May, M. et al. Impact of late diagnosis and treatment on life expectancy in people with HIV-1: UK Collaborative HIV Cohort (UK CHIC) Study. Br. Med. J. 343, d6016 (2011). 29. Gray, E. R. et al. p24 revisited: a landscape review of antigen detection for early HIV diagnosis. AIDS 32, 2089–2102 (2018). 30. Price, C. P. Point of care testing. Br. Med. J. 322, 1285–1288 (2001). 31. World Malaria Report. https://www.who.int/malaria/publications/world-malaria-report2018/en/ (WHO, 2018). 32. Land, K. J., Boeras, D. I., Chen, X. S., Ramsay, A. R. & Peeling, R. W. REASSURED diagnostics to inform disease control strategies, strengthen health systems and improve patient outcomes. Nat. Microbiol. 4, 46–54 (2019). 33. Walter, J. G. et al. Protein microarrays: reduced autofluorescence and improved LOD. Eng. Life Sci. 10, 103–108 (2010). 34. Kim, J. et al. Rapid and background-free detection of avian influenza virus in opaque sample using NIR-to-NIR upconversion nanoparticle-based lateral flow immunoassay platform. Biosens. Bioelectron. 112, 209–215 (2018). 35. Paterson, A. S. et al. A low-cost smartphone-based platform for highly sensitive point-of-care testing with persistent luminescent phosphors. Lab Chip 17, 1051–1059 (2017). 36. Boudou, J. P., David, M. O., Joshi, V., Eidi, H. & Curmi, P. A. Hyperbranched polyglycerol modified fluorescent nanodiamond for biomedical research. Diamond Relat. Mater. 38, 131–138 (2013). 37. Hermanson, G. T. Zero-length crosslinkers. In Bioconjugate Techniques 3rd edn (ed. Hermanson, G. T.) Ch. 4 (Academic, 2013). 38. González Flecha, F. L. & Levi, V. Determination of the molecular size of BSA by fluorescence anisotropy. Biochem. Mol. Biol. Educ. 31, 319–322 (2003). 39. Reth, M. Matching cellular dimensions with molecular sizes. Nat. Immunol. 14, 765–767 (2013). 40. Ngom, B., Guo, Y., Wang, X. & Bi, D. Development and application of lateral flow test strip technology for detection of infectious agents and chemical contaminants: a review. Anal. Bioanal. Chem. 397, 1113–1135 (2010). 41. Armbruster, D. A. & Pry, T. Limit of blank, limit of detection and limit of quantitation. Clin. Biochem. Rev. 29, S49–S52 (2008). 42. Daher, R. K., Stewart, G., Boissinot, M. & Bergeron, M. G. Recombinase polymerase amplification for diagnostic applications. Clin. Chem. 62, 947–958 (2016). 43. Lillis, L. et al. Cross-subtype detection of HIV-1 using reverse transcription and recombinase polymerase amplification. J. Virol. Methods 230, 28–35 (2016). 44. Crannell, Z. A., Rohrman, B. & Richards-Kortum, R. Equipment-free incubation of recombinase polymerase amplification reactions using body heat. PLoS ONE 9, e112146 (2014). 45. Rohrman, B. A. & Richards-Kortum, R. R. A paper and plastic device for performing recombinase polymerase amplification of HIV DNA. Lab Chip 12, 3082 (2012). 46. Boyle, D. S., Lehman, D. A. & Lillis, L. Rapid detection of HIV-1 proviral DNA for early infant diagnosis using rapid detection of HIV-1 proviral DNA for early infant diagnosis. MBio 4, e00135-13 (2013). 47. Dineva, M. A., Mahilum-Tapay, L. & Lee, H. Sample preparation: a challenge in the development of point-of-care nucleic acid-based assays for resource-limited settings. Analyst 132, 1193 (2007). 48. Jauset-Rubio, M. et al. Ultrasensitive, rapid and inexpensive detection of DNA using paper based lateral flow assay. Sci. Rep. 6, 37732 (2016). 49. Phillips, A. et al. Sustainable HIV treatment in Africa through viral-load-informed differentiated care. Nature 528, S68–S76 (2015). 50. Kuo, Y., Hsu, T.-Y., Wu, Y.-C., Hsu, J.-H. & Chang, H.-C. Fluorescence lifetime imaging microscopy of nanodiamonds in vivo. In Proc. Advances in Photonics of Quantum Computing, Memory, and Communication VI Vol. 8635, 863503 (SPIE, 2013). Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. © The Author(s), under exclusive licence to Springer Nature Limited 2020

Nature | Vol 587 | 26 November 2020 | 593

Article Methods Resonator design CST Studio Suite 2015 (Dassault Systems) was used to create a 3D model of the resonator design, solving Maxwell’s equations over a sweep of microwave frequencies to determine reflected and absorbed power. The design was based on copper patterned on a printed circuit board, using Rogers 4003C substrate for low dielectric loss at microwave frequencies. The top side had an interdigitated capacitor and a capacitor-inductor omega-shaped loop, and the bottom had a ground plane. The dimensions of these components were varied iteratively to maximize the absorption at 2.87 GHz and ensure an impedance of 50 Ω for coupling to the frequency generator. The final design was exported as a 2D CAD file. Preparation of functionalized FNDs Polyglycerol (PG)-functionalized FNDs were conjugated to antibodies using disuccinimidyl carbonate (DSC) as shown in Extended Data Fig. 3c. DSC activates hydroxyl surface groups to form succinimidyl carbonates, which can then react with antibodies to form stable carbamate or urethane bonds37. In a typical synthesis, 100 μl FND–PG (1 mg ml−1, Adámas Nanotechnologies, high brightness 120 nm core + 20 nm PG FND NDNV140nmHiPG2mg) were centrifuged at 21,130g for 7.5 min to condense the particles into a pellet. The supernatant was then removed and the FNDs were resuspended in anhydrous N,N-dimethylformamide (DMF, 99.8%, Sigma-Aldrich). After resuspension in DMF, the colloidal solution was sonicated for 5 min in an ultrasonic bath. The washing and sonication steps were repeated three times to remove water. After the last centrifugation, the particles were resuspended in 100 μl of a 50 mg ml−1 solution of DSC (≥95%, Sigma-Aldrich) in DMF and placed in a thermoshaker for 3.5 h at 300 rpm at room temperature. Excess reagents were removed by three cycles of centrifugation and resuspension in DMF (as described above). After the third centrifugation, the particles were resuspended in 100 μl deionized water. Depending on the desired surface functionalization, 13.7 μl of anti-DIG antibodies (1 mg ml−1, Abcam, ab76907) or 6.8 μl BSA–biotin (2 mg ml−1 in deionized water, Sigma-Aldrich) were added to the activated FNDs. The mixture was placed in a thermoshaker overnight for 15 h at 300 rpm at room temperature. The remaining succinimidyl carbonates were quenched by adding of 10 μl of Tris-HCl pH 7.5 (1 M, Thermo Fisher Scientific). After 30 min, the unbound reagents were removed by three cycles of centrifugation and resuspension in deionized water (100 μl) and stored in 100 μl of PBS with 0.1 wt% BSA. After functionalization, the FND concentrations were quantified by fluorescence intensity, as this remains unchanged during the functionalization reactions: the fluorescence originates from the atomic structure of the FNDs, so is unaffected by surface modifications. This was carried out by performing a serial dilution of the FND stock solution (of know mass concentration, cP, of 1 mg ml−1 based on manufacturers specifications) and using a spectrophotometer to measure the fluorescence compared to the functionalized-FND solution. A linear regression was fitted to the fluorescence intensity of the serial dilution of the stock FND solution against FND concentration and interpolated to calculate the mass concentration of the functionalized particles. Examples of this for the three different particle sizes are shown in Extended Data Fig. 5c. This was converted to molar concentration (CP) using the diameter (d), density of diamond (ρ), and Avogadro constant (NA), shown in equation (1): −1

CP(M) =

cP(mg ml ) 3

d (nm3 )

×

103 π 6

−1

× ρ(mg nm−3 ) × NA(mol )

(1)

Characterization of nanoparticles. Excitation spectra of the FNDs were acquired with a fluorescence microplate reader (SpectraMax i3,

Molecular Devices LLC) and served as a reference to estimate the final FND concentration by comparison of the fluorescence intensity with the stock solution. Emission spectra were recorded with a spectrometer (SPM-002, Photon Control) with a 500-nm LED light source (pE-4000, CoolLED). FTIR spectroscopy was performed by conjugating particles as described above, and storing in deionized water (maximum of 2 days), before centrifuging at 21,130g to condense the particles into a pellet and removing as much supernatant as possible to form a paste. This paste was pipetted onto the spectrometer (Bruker, Alpha). Three measurements of each sample were taken using 16 reads per measurement. Dynamic light scattering data and zeta potentials were measured with a Zetasizer (Zeta Sizer Nanoseries, Malvern Instruments) using a 150-fold dilution of the FNDs. The resulting number plots were fitted to the skewed exponential in equation (2) to find the peak diameter.

 −(x − μ) 2  exp erfc 2  2σ  N (x ) = 2π σ

(

−α (x − μ) 2σ

)

(2)

where N is the number fraction, x is the diameter, μ is the mean of the diameter distribution, σ is the standard deviation and α is the skew parameter.

Quantification of antibody binding sites on FND surface In order to quantify the number of active antibody binding sites on the surface, an assay similar to PCR–ELISA (enzyme-linked immunosorbent assay) and that described in ref. 51 was developed. FNDs (300 μl) were functionalized with anti-DIG antibodies, as described in the section ‘Preparation of functionalized FNDs’, except the final suspension was in DNase/RNase-free distilled water (Thermo Fisher UltraPure) rather than storage buffer, and the particles were concentrated fivefold (to 5 pM, 60 μl). The suspension was subsequently split in half for a positive sample and a negative control, and 6 μl of a 6× running buffer solution was added to each, to a final concentration of 1× running buffer (5% milk + 0.05% Empigen in water). A large excess of a DIG-modified DNA sequence (0.9 μM final concentration) was added to the positive sample, and the same excess of the same DNA sequence but with no DIG modification was added to the negative control. A short DNA sequence (82 bp) was used to avoid the bound DNA blocking available sites on the FND surface. The DIG–DNA was left to bind to the FND–antibody for 2 h. After binding, each solution was diluted to 400 μl in DNase/ RNase-free distilled water before centrifuging at 376g for 2 min and removing the supernatant. This washing was repeated four times to remove excess DNA, with the final suspension in 150 μl 100 μg ml−1 salmon sperm DNA solution (Thermo Fisher UltraPure). qPCR was then performed on the final suspensions. The template, primers and probe sequences are listed in Extended Data Fig. 8d (assay taken from ref. 52). The master mix was the TaqMan Fast Virus 1-Step Master Mix (Thermo Fisher) with primers at 300 nM and the probe at 150 nM, and 4 μl of sample in a total volume of 15 μl. The standard was constructed from serial dilutions of the pHRSIN-CSGW plasmid53. The qPCR was performed by an Applied Biosystems 7500 Real-Time PCR System (Thermo Fisher), and the copy numbers quantified by the 7500 software (v.2.0.6). The FND concentrations in the final suspensions were measured as described in the section ‘Preparation of functionalized FNDs’. Dividing the DNA copy number by the FND number gave the number of active binding sites per FND. Target amplification by recombinase polymerase amplification RNA template generation. The template was designed using an alignment of 2,929 clinical isolates of HIV-1 from the Los Alamos HIV Sequence Database54 to identify conserved areas. The alignment was mapped to the HXB2 (K03455) HIV-1 reference genome using Geneious Software (v.10.0.6) and a highly conserved region of 229 bp (1573–1801 bp from HXB2) selected to design five forward and five reverse primers

to be tested in RPA primer selection. Starting from a R9BAL∆Env plasmid (a gift from G. Towers, University College London), DNA template was produced by polymerase chain reaction amplification of the 1,503 bp template sequence using the Phusion High-Fidelity PCR Kit (New England Biolabs). Primer sequences used are shown in Extended Data Fig. 8d. The thermocycling was performed at 98 °C for 30 s, then 30 cycles of: 98 °C for 10 s, 65 °C for 20 s, 72 °C for 25 s, and a final extension of 72 °C for 10 min. The DNA was then transcribed to RNA using the MEGAscript T7 Transcription Kit (Invitrogen) and purified using MEGAclear Transcription Clean-Up Kit (Invitrogen), following the manufacturer’s instructions. The concentration of RNA template was measured via Qubit RNA HS assay kit (Invitrogen) with the Qubit 4 Fluorometer. RT–RPA reaction (amplicon serial dilution). RT–RPA assay was performed on a 1.5 kb HIV-1 in vitro transcribed RNA template. Optimization of the assay is shown in Extended Data Fig. 8. RT–RPA of the template was performed using TwistAmp Exo Reverse Transcription Kit (TwistDx), following the manufacturer’s instructions. The reaction time was 30 min at 37 °C shaking at 200 rpm in an incubator (New Brunswick Innova 42). Nucleic acid sequences are listed in Extended Data Fig. 8d, including a fluorescent probe. During amplification, exonuclease cuts the tetrahydrofuran, releasing the fluorescent tag (FAM) from the quencher, producing a quantitative signal. The resulting RPA products were incubated with RNase A (QIAGEN) for 2 h, before purification of amplified template to remove primers and fragments of RNA using QIAquick PCR Purification Kit (QIAGEN), following the manufacturer’s instructions. Quantification by measuring absorption at 260 nm is confounded by RNA contamination, so double-stranded (ds)DNA quantification was performed using a Quant-iT PicoGreen dsDNA Assay Kit (Invitrogen), following the manufacturer’s instructions. Fluorescence measurements were taken with an ultraviolet–visible spectrophotometer (Molecular Devices, SpectraMax i3). RT–RPA reaction (final assay with amplification). RT–RPA of the template was performed using TwistAmp Basic Kit (TwistDx). The master mix, containing 480 nM of forward and reverse primers (for sequences see Extended Data Fig. 8d, Integrated DNA Technologies), 1× rehydration buffer (TwistDx), reverse transcriptase (M-MLV Reverse Transcriptase, Invitrogen) and nuclease-free water (Invitrogen), was prepared in a tube. For each RPA reaction, 2 μl of target HIV-1 RNA template was added to 45.5 μl of master mix and a freeze-dried RPA pellet. The reaction was started by adding 2.5 μl of magnesium acetate to each reaction, giving a final reaction volume of 50 μl. The RT–RPA reactions proceeded at 37 °C in a thermal incubator for 10 min. The RT–RPA products were purified by QIAquick PCR Purification Kit (QIAGEN) and resuspended in a final volume of 50 μl elution buffer for each reaction. RT-RPA reaction (UCLH clinical standards). RNA from the UCLH HIV-1 viral load positive and negative standards (personal communication, gift from P. Grant, UCLH) was extracted from a 140 μl sample using the QIAamp Viral RNA Mini Kit (Qiagen) essentially according to the manufacturer’s instructions, except that elution was in 60 μl water. Ten microlitres of extracted RNA in water was used for each RT–RPA reaction. RT–RPA reaction (seroconversion panel). RNA from an HIV-1 seroconversion panel (thirteen samples: ZeptoMetrix Corporation, Panel Donor No. 73698) was extracted from a 140-μl sample using the QIAamp Viral RNA Mini Kit (Qiagen) essentially according to the manufacturer’s instructions, except that elution was in 60 μl water. Two microlitres of extracted RNA in water was used for each RT–RPA reaction. The RT– RPA reactions proceeded at 37 °C in a thermal incubator for 10.5 min. The RT–RPA products were purified by QIAquick PCR Purification Kit

(Qiagen) and resuspended in a final volume of 50 μl elution buffer for each reaction.

Lateral flow assay The following assays all use LFA strips with a poly-streptavidin test line, blocked by a proprietary polyvinylpyrrolidone-sucrose method (Mologic). The strips were 5 mm wide with the test line positioned 7 mm from the bottom of the strip. A major challenge in developing sensitive LFAs is non-specific binding. To this end, sweeps of running buffers and washing buffers were performed to identify the combination that gave the best SBR (see Extended Data Fig. 6). This gave rise to a reduction in non-specific binding to the strip, reducing the blanks, and increasing the signal in turn. The optimum buffers in this study were found to be non-fat milk 5 wt% + 0.05 vol% Empigen in deionized water (running buffer) and 0.2 wt% BSA with 0.2 vol% Tween 20 in acetate buffer 10 mM pH 5 (washing buffer). Having chosen running and washing buffers, the background was further reduced by optimising the concentration of FNDs, as shown in Extended Data Fig. 7a, b. LFA strips were run with a dilution series of FND concentration. A positive test (500 pM of DNA) and a negative control (deionized water) were run at each FND concentration. The fitted relationships between positive and negative lock-in amplitude signals and FND concentration were used, along with modelling of equilibrium binding, depending on antigen and FND concentration. This enabled the estimation of the LODs and dynamic ranges at each FND concentration, as explained in Extended Data Fig. 7c, d and Supplementary Information section 2, leading to the selection of the FND concentration. The dynamic range is limited by the total number of FNDs at the top end and the non-specific signal in the negative at the bottom end. The chosen concentration gave a per-strip FND cost of less than 0.02 cents (USD) (4.8 ng of FNDs per strip). The total cost of consumables per test and estimated costs of a strip reader are shown in Supplementary Table 1. The LFAs were performed by pipetting the solutions to be run into wells of a 96-well plate, then dipping the strips into the wells. All LFAs were performed at room temperature. Purified single-strand (ss)DNA concentrations were measured by absorption using the Nanodrop OneC (Thermo Scientific). Assay with FND–BSA–biotin. BSA-biotin-functionalized FNDs were diluted in running buffer to the particle concentrations shown in Extended Data Fig. 5d. Then, 55 μl of this suspension was run on each LFA strip. Assay for model RT–RPA products. The initial optimization and benchmarking was performed using a model ssDNA RT–RPA ‘amplicon’ (a short ssDNA strand with digoxigenin and biotin modifications at opposite ends), before moving to real RT–RPA amplicons for the final assay. A comparison of real RT–RPA amplicons with the model ssDNA ‘amplicon’ is shown in Extended Data Fig. 9a, validating its use for optimization, with similar dissociation constant (KD) values and dynamic ranges, although more variation in the blanks with real amplicons gives a higher LOD. A Monte Carlo simulation of the variances of the clinical sample lock-in amplitudes that can be explained by FND size distribution gives a value of around 8–9% of the total variance (Extended Data Fig. 9d). A further approximately 0.1–2% of variance is explained by periodic drift in the modulation amplitude (Extended Data Fig. 9e), and frequency noise contributes negligible variation (Extended Data Fig. 9f), indicating that the majority is from other factors, such as strip-to-strip inconsistency. This strip-to-strip variation is more evident with larger FNDs, which could be because they are close to the minimum pore size of the nitrocellulose. LODs for the three FND diameters using the model ssDNA ‘amplicon’ is shown in Extended Data Fig. 9b.

Article A single strand of DNA (26 bp), functionalized with digoxigenin at the 3′ end and biotin at the 5′ end (Integrated DNA Technologies, 5′ biotin-GTCCGAGCGTACGACGAACGGTCGCT-digoxigenin 3′) was used as a model for RT–RPA amplicons produced with biotin and digoxigenin functionalized primers. These model ssDNA strands were diluted in running buffer and 50 μl of this solution was mixed with 5 μl of anti-digoxigenin antibody-functionalized FND suspension (1,400, 170 and 3 fM in PBS for 120, 200 and 600 nm diameters, respectively). After 10 min at room temperature, these solutions were run on LFA strips. After all the solution was run (approximately 10 min), the strips were transferred to wells of a 96-well plate with 75 μl of washing buffer (around 12 min). Assay for real RT–RPA products (amplicon serial dilution). After purification and quantification of amplicons, the assay was run and washed identically to the model RPA products with FND concentrations of 2,600, 120 and 4 fM for 120, 200 and 600 nm diameters, respectively. RT–RPA used a digoxigenin-modified forward primer and a biotin-modified reverse primer. The RT–RPA products, therefore, consist of dsDNA (181 bp), each copy including a digoxigenin molecule at one end and a biotin molecule at the other. These modifications bind to anti-digoxigenin-functionalized FNDs and the poly-streptavidin test line on the nitrocellulose paper, respectively, forming a sandwich structure and immobilising FNDs in the presence of amplicons, as shown in Fig. 4a. Final assay for RNA quantification with RT-RPA. After purification, 10 μl of 6× running buffer (30 wt% non-fat milk with 0.3 vol% Empigen in deionized water) was added to the 50 μl RT–RPA product. Then, 5 μl of anti-digoxigenin antibody-functionalized FND suspension was added before running the strips as already described. For the lowest positive sample (average of 1.26 copies), there is a 71% chance of having at least one copy, based on the Poisson distribution. This gives a 26% chance of all four experimental replicates having at least one copy, using the binomial distribution, and a 42% chance for three of the four replicates. For the next dilution (average of 0.13 copies), these probabilities decrease to 0.019% and 0.60%. These probabilities are consistent with the results in Fig. 4. Fluorescence Modulation and Imaging. The paper strips were imaged using a fluorescence microscope (Olympus BX51) with a 550 nm green LED as excitation light source (pE-4000, CoolLED), with a filter cube containing an excitation filter (500 nm bandpass, 49 nm bandwidth, Semrock), a dichroic mirror (596 nm edge, Semrock) and emission filter (593 nm long-pass, Semrock). A 20×/0.4 BD objective (LMPlanFl, Olympus) was used. Images were recorded with a high-speed camera (ORCA-Flash4.0 V3, Hamamatsu) using HCImage Live software (Hamamatsu). All strips were measured when dry to eliminate any possible variation due to drying during measuring. Extended Data Fig. 12 shows the detection on wet strips and the effect of drying on the lock-in amplitude of the FND signal. This experiment was performed by running positive and negative LFAs with the model ‘amplicon’ as above, then fixing each strip to the microscope stage directly after completing the wash step. A 15-s lock-in measurement at an exposure time of 20 ms was taken every 1 min. The light source was only on during measurement to prevent it speeding up drying. One of the negative controls was measured for less than 55 min (35 min), so its mean was used after this time in Extended Data Fig. 12. There is a small loss in sensitivity on wet strips (around 1.4−1.9 times), corresponding to a necessary increase in isothermal amplification time of less than 1 min. A microwave field was generated by a voltage controlled oscillator (VCO, Mini-Circuits, ZX95-3360+) and a low noise amplifier (Mini-Circuits, ZX60-33LN+) connected to the resonator circuit board

(Minitron, Rogers 4003C 0.8 mm substrate and 1 ozft−2 (300 g m−2) copper weight). The resonator was attached to the microscope stage. The tuning voltage of the VCO was set to maximize the decrease in fluorescence. Modulation of the signal was achieved by modulating the input voltage of the VCO with an on-chip reference frequency generator at 4Hz using a 32.768kHz crystal oscillator (DS32KHZ, Farnell) with a 14-stage frequency divider (CD4060BM, Farnell). Circuit board design was performed using EAGLE (Autodesk). A sweep of modulation frequencies was performed using this VCO and amplifier, using a microcontroller (Arduino Nano 3.0) to generate the different modulation frequencies. The power dependence of the decrease in fluorescence was recorded using a benchtop microwave generator (HM8135, Rohde & Schwartz Hameg) and a low noise amplifier (Mini-Circuits, ZRL-3500+). A broad sweep of microwave frequencies was measured with a radiofrequency signal generator (WindFreak Technologies, SynthUSBII).

Computation lock-in and LOD The fluorescence signal was modulated with a set modulation frequency (Fm) and the amplitude of the resulting signal was computed with a computational lock-in algorithm. Images were recorded with the high-speed camera (ORCA-Flash4.0 V3, Hamamatsu) at a sampling frequency Fs. Each frame was averaged to get a mean pixel value at each time point t0 = 0 to tL = L/Fs, where L was the total number of frames. A moving average low-pass filter with a span width of 1.5 × Fs/Fm was applied to the fluorescence time series. The filtered signal, Vin was multiplied by two reference signals: in-phase (sin(2πFmt)) and π/2 out-of-phase (cos(2πFmt)) to obtain Vx and Vy, respectively: Vx = Vin × sin(2πFmt)

(3)

V y = Vin × cos(2πFmt)

(4)

The d.c. components of these two signals, X and Y, were calculated by finding the mean of Vx and Vy, respectively, and enabled the evaluation of the magnitude R of the lock-in amplitude at the frequency Fm according to:

R = X2 + Y 2

(5)

Where there was no FND saturation (BSA–biotin assays), the LOD was computed by fitting the lock-in amplitude, as a function of concentration, c, to a linear regression. Where there was saturation, (all assays except BSA–biotin assays), a Langmuir isotherm was fitted:

SBR = k0 + k1 ×

[T ] KD + [T ]

(6)

where k0 is SBR of the negative control, k1 is a scaling constant representing the SBR at target saturation, [T] is the amplicon concentration and KD is the equilibrium dissociation constant. Fitting was performed in MATLAB using the fitlm and nlinfit functions for linear and Langmuir fits, respectively, weighting the fit by the variance at each concentration. The LOD was defined as the intersection of the lower 95% confidence bound of the fit with the upper 95% confidence bound of the blank measurements41.

Data availability The datasets generated during and/or analysed during the current study, and the computer code used are available from the corresponding author on reasonable request, in line with the requirements of UCL and the funder (EPSRC policy framework on research data).

51. Kim, E. Y. et al. A real-time PCR-based method for determining the surface coverage of thiol-capped oligonucleotides bound onto gold nanoparticles. Nucleic Acids Res. 34, e54 (2006). 52. Besnier, C., Takeuchi, Y. & Towers, G. Restriction of lentivirus in monkeys. Proc. Natl Acad. Sci. USA 99, 11920–11925 (2002). 53. Bainbridge, J. W. et al. In vivo gene transfer to the mouse eye using an HIV-based lentiviral vector; efficient long-term transduction of cornealendothelium and retinal pigment epithelium. Gene Ther. 8, 1665–1668 (2001). 54. Foley, B. et al. HIV Sequence Compendium 2017. LA-UR-18-25673 (Los Alamos National Laboratory, 2018). 55. Kong, J. & Yu, S. Fourier transform infrared spectroscopic analysis of protein secondary structures. Acta Biochim. Biophys. Sin. (Shanghai) 39, 549–559 (2007). 56. Zadeh, J. N. et al. NUPACK: Analysis and design of nucleic acid systems. J. Comput. Chem. 32, 170–173 (2011). 57. SantaLucia, J. A unified view of polymer, dumbbell, and oligonucleotide DNA nearest-neighbor thermodynamics. Proc. Natl Acad. Sci. USA 95, 1460–1465 (1998). 58. Laitinen, M. P. & Vuento, M. Affinity immunosensor for milk progesterone: Identification of critical parameters. Biosens. Bioelectron. 11, 1207–1214 (1996). Acknowledgements We thank M. Schormans for help with circuit design, M. Thomas for assistance with dynamic light scattering measurements and M. Towner for assistance with FTIR measurements. This work was funded by the i-sense EPSRC IRC in Early Warning Sensing Systems for Infectious Diseases (EP/K031953/1); i-sense EPSRC IRC in Agile Early Warning Sensing Systems in Infectious Diseases and Antimicrobial Resistance (EP/R00529X/1); the National Institute for Health Research University College London Hospitals Biomedical Research Centre; Royal Society Wolfson Research Merit Award to R.A.M. (WM130111); LCN

Departmental Studentship to B.S.M.; EPSRC Centre for Doctoral Training in Delivering Quantum Technologies to G.D. (EP/L015242/1); H2020 European Research Council Local quantum operations achieved through the motion of spins to J.J.L.M. (771493); and the UCLH NHS Foundation Trust to J.H. and E.N. Author contributions B.S.M. and R.A.M. conceived the research and led the study; P.J.D. advised on nanodiamonds and J.J.L.M. on microwave modulation. B.S.M. demonstrated the initial proof-of-concept; B.S.M. and L.B. designed and optimized the lock-in analysis, functionalization and LFA design; B.S.M., L.B. and D.H. performed all the FND LFA experiments; H.D.G. designed, optimized and performed RT-RPA assays including primer design and template generation; D.H. adapted and performed RT-RPA assays and purification; J.J.L.M. and G.D. designed the microwave delivery including resonators; E.R.G. performed clinical RNA extraction, and advised on virology including primer design; J.H. performed qPCR on the seroconversion panel; E.N. provided clinical expertise; B.S.M. and E.R.G. designed and performed binding-site quantification experiments; B.S.M., L.B. and R.A.M. drafted the manuscript; and all authors reviewed and revised the manuscript. Competing interests B.S.M., L.B., G.D., P.J.D., J.J.L.M. and R.A.M. are inventors on the UK patent application number 1814532.6 filed by University College London Business. Additional information Supplementary information is available for this paper at https://doi.org/10.1038/s41586-0202917-1. Correspondence and requests for materials should be addressed to B.S.M. or R.A.M. Peer review information Nature thanks Takuya Segawa and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Reprints and permissions information is available at http://www.nature.com/reprints.

Article

Extended Data Fig. 1 | Optimization of microwave modulation. a–f, A linear resonator was designed to have a wideband response over the range 1–4 GHz, and an omega narrowband resonator was designed to have a stronger, narrower resonance at 2.87 GHz with quality factor Q = 100. The schematic printed circuit board layouts for the two resonators are shown in a and d, respectively. The resulting simulated fields are shown in b and e, respectively. The reflected power (S11) is plotted against frequency in c and f. The narrowband resonator shows 5–6 orders of magnitude greater absorption than the wideband resonator at 2.87 GHz, indicating resonant coupling giving strong absorption. Panel f also shows the corresponding FND intensity dip. g, Emission spectra of FNDs acted on by a 2.87

GHz microwave field. The powers listed in decibel-milliwatts are the output power of the microwave generator (before the 17 dB amplifier). h, Each spectrum is integrated over the whole wavelength range to give a total intensity, which is plotted against preamplifier power. This shows a linear relationship between fluorescence intensity and microwave power (in dBm) above a threshold power, and up to 7 dBm, where the amplifier reaches its 1 dB compression power. At this point, the fluorescence starts to increase again owing to a loss in the quality of the sinusoid leading to power lost in harmonics. Dots show means and error bars show the s.d., with n = 3 measurement repeats.

Extended Data Fig. 2 | See next page for caption.

Article Extended Data Fig. 2 | Optimization of lock-in analysis. a, Schematic of the computational lock-in algorithm used to extract the microwave modulated FND signal from the background. The input signal is high-pass filtered using a moving average filter to remove low-frequency drift. It is subsequently multiplied by cosine and sine functions with frequency Fm, and the resulting signals are lowpass filtered to generate the in phase and quadrature components, respectively, of the vector representation of the signal. The magnitude of this vector is calculated to remove the effect of phase, giving the output magnitude. b, The variation of lock-in amplitude with modulation rate (Fm) at various sampling rates (Fs). A single strip with very high intensity was modulated at Fm values between 1–450 Hz, and sampled at various Fs values between 3.89–996 Hz. The resulting plot shows that lock-in amplitude is independent of Fs when Fs > 2Fm. c, d, The relationships between lock-in amplitude, exposure time (Te) and modulation frequency (Fm). An identical LFA strip was measured with exposure times between 10–50 ms, using the maximum possible Fs for each Te, and Fm values between 1–15 Hz. d shows Fm against lock-in amplitude at various exposure times. It is shown that the lock-in amplitude has its maximum at around 5 Hz for all frequencies, and reduces when Fm is close to Fs/2, its maximum possible value. This is evident in the raw signal plots in c for each Fm at a fixed exposure time of 30 ms. As Fm approaches Fs/2, the sampling effects obscure the square wave, decreasing lock-in amplitude. For maximum lock-in amplitude, the highest

possible Te should be used. Here, we are limited to 50 ms by the background autofluorescence of the nitrocellulose, which saturates the camera above this value. A corresponding Fm of 4 Hz was chosen as it is in the optimal range and is a power of 2, so can be achieved by simply dividing the temperature compensation crystal oscillator (TCXO) frequency. e, The variation of lock-in amplitude with total measurement time at Fm = 4 Hz and Fs = 20 Hz for five different concentrations of FNDs and a negative control, immobilized with a biotin–avidin interaction. The positive amplitudes stabilize quickly, reaching 5% of their 15 s value in 3.9 s for positive results. The negative results take longer to stabilize, reaching 5% of their 15 s value in 13 s. A measurement time of 15 s (300 frames) was used for subsequent measurements. f, Schematic circuit design of temperature compensation crystal oscillator (TCXO)-based modulated microwave source. It is powered by a 5 V source which powers a TCXO, which outputs a 32.768 kHz square wave. This is converted to a 4 Hz signal by a 4060 counter chip. This square wave controls two transistors which deliver 12 V stepped up power (d.c. converter) to the microwave VCO. The bias voltage is regulated from 12 V to 8.15 V by a voltage regulator. The VCO microwave output is amplified by the MW amplifier and transmitted to the omega resonator. g, Printed circuit board layout of the prototype (65 mm × 38 mm). Outputs for the microwave amplifier and microwave VCO are at the top right and bottom right, respectively. A photo of the printed circuit board with a £1 coin (GBP) for scale is shown below.

Extended Data Fig. 3 | See next page for caption.

Article Extended Data Fig. 3 | FND characterization and functionalization. a, Comparison of the non-specific binding of various commercial FNDs with various surface functionalizations on LFAs. The lock-in amplitude at the test line was measured to quantify non-specific binding. The LFAs were also pre-blocked with a polyvinylpyrrolidone-sucrose solution (proprietary method, Mologic). The lowest non-specific binding was from the PG-functionalized particles (FND–PG), as the PG adds a hydrophilic layer. b, Dynamic light scattering of three different FND particle core diameters: 120, 200 and 600 nm. c, A schematic of antibody functionalization of FND–PG. DSC activates hydroxyl surface groups to form succinimidyl carbonates, which can then react with antibodies to form stable carbamate or urethane bonds. d–f, Scanning electron microscope images of FNDs with particle core diameters of 120, 200 and 600 nm, respectively. g, Dynamic light scattering was also used to measure the size and aggregated fraction after functionalization of 120 nm FND–PG before and after functionalization with BSA-biotin or antibodies. Dots show the means of n = 3

measurement replicates. Fitting the number plots to skew exponentials (equation (3), plotted as lines) gave peak particle hydrodynamic diameters of 106, 121 and 128 nm. h, The fitted peak diameters are plotted with error bars denoting their 95% confidence intervals (n = 3 measurement replicates), showing no significant difference between the bio-functionalized diameters (FND–biotin, FND–antibody), but both are significantly different from the prefunctionalization diameter (FND–PG): *P ≤ 0.05; **P ≤ 0.01, ANOVA with Tukey HSD post hoc test. i, FTIR spectroscopy of FND–PG and antibody-functionalized FND–PG. Lines show means of n = 6 measurement replicates for FND samples and n = 2 measurement replicates for the blank. C–O and C–H peaks, indicative of the PG layer, can be seen in both FND–PG and FND–PG–antibody at around 1,100 cm−1 and at around 2,900 cm−1, respectively. The FND–PG-antibody spectrum displays additional peaks at around 1,640 cm−1 and at around 1,540 cm−1, suggesting protein amide I and amide II bonds, respectively55, showing that protein functionalization was successful.

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

Article Extended Data Fig. 4 | Quantification of the number of available binding sites per FND. a, Initially, binding constants of the anti-DIG antibody binding to DIG were measured using interferometry. Full experimental details are shown in Supplementary Information section 1. Binding at different concentrations was measured and the resulting curves were fitted to exponentials. To find the equilibrium dissociation constant (KD), equilibrium binding values, B, were plotted here against concentration, C. A Langmuir adsorption isotherm was

(

a×C

)

fitted B ∞ = K + C giving a KD value of 5.1 × 10−10 M. b, In order to find the on- and D off-rates, kon and koff, the observed reaction rates, kobs, at each concentration were plotted and fitted to the linear relationship: kobs = koff + C × kon. The resulting fitted values are kon = 1.6 × 105 M−1 s−1 and koff = 9.1 × 10−5 s−1. c, A schematic of the assay to quantify the number of available binding sites per FND. After functionalization of FNDs with anti-DIG antibodies, an approximately 50-fold excess of DIG-modified DNA was added and left to bind for 2 h. The negative DNA control used the same sequence, but with no DIG modification to compensate for non-specific binding and adequate washing. After multiple washes by centrifugation to remove the excess DNA, the remaining DNA (bound to FNDs) was quantified by qPCR. See Extended Data Fig. 8d for template, primer and probe sequences, and Methods for full experimental details. d, A kinetic binding simulation was performed to verify that all available sites would be occupied after 2 h with the above excess. The graph shows the fraction of sites on the FNDs which are occupied, with this approximately 50-fold excess, over a range of KD, kon and koff values. The red cross

in circle marks the location of the anti-DIG antibody used in this paper (using the values measured in a and b), indicating that more than 99.9% of available sites will be occupied after 2 h. This means that quantifying the DNA gives a true measure of available binding sites. e, Amplification plot showing the normalized fluorescence intensity against the number of cycles. A standard curve of each decade from 40 copies to 4 × 108 copies is plotted, along with the sample and negative control FND samples described above. The negative diluent controls are also plotted along with the Cq threshold. The lines show means and shaded areas show the s.d. of repeats (n = 3 technical replicates for standard curve, and n = 6 for samples). f, The resulting Cq values are plotted against copy number per reaction. Dots show means and error bars show s.d. (n = 3 technical replicates for standard curve and n = 6 for samples). The standard curve was fitted to a logarithmic curve (Cq = −3.2log10 copies + 39), enabling calculation of the number of copies in the DIG–DNA sample and negative DNA control. Dividing by the particle concentration (measured as shown in Extended Data Fig. 5c) and subtracting the negative DNA control value gives the number of available binding sites per particle as 4,300 sites. This is within what is geometrically plausible, giving an area per antibody of at least 200 nm2 (assuming at least 1 paratope available of at least 75% of the bound antibodies). The corresponding calculated values for 120 and 200 nm particles are 172 and 477 available binding sites per FND respectively, assuming the same loading density.

Extended Data Fig. 5 | Lateral flow and FND benchmarking. a, Measurement of flow rate of lateral flow strips. During wetting, the flow follows the Washburn 1 equation, where V ≈ t 2 (inset), and during fully-wetted flow, Darcy’s law for capillary flow is followed (V ≈ t), with a constant flow rate of 6.9 μl min−1. b, Using a one-to-one receptor–ligand binding approximation, the binding of biotinylated FNDs to streptavidin was modelled kinetically, indicating that all the FNDs bind with a residency time of more than about 10−3 s. Here, the residency time is measured as 4 s, using the flow rate from a, so all the FNDs should bind. c, An example of the measurement of FND concentration. FND fluorescence is unaffected by surface chemistry, so is used to quantify concentration. A serial dilution of FND suspensions from a known stock concentration was performed (dots showing means with error bars showing s.d., n = 6 measurement replicates). This was then fitted with a linear regression (lines) to find a relationship between fluorescence intensity and concentration. After each FND functionalization,

the fluorescence intensities of the final suspensions were measured, and the linear fit was used to estimate concentration (crosses). d, Fundamental LODs for different sized FNDs on LFAs, using a model biotin–avidin interaction. Suspensions (55 μl) of BSA–biotin-functionalized FNDs were run at different concentrations on poly-streptavidin strips. Concentrations were chosen to span the dynamic range of the camera, limited by over-exposure, as seen with the top concentration of 200 and 600 nm FNDs. Dots show means and error bars show s.d. (n = 3 technical replicates, n = 3 measurement replicates). Each series is fitted to a simple linear regression, shown as the solid line, with 95% confidence intervals shown shaded. LODs for 120, 200 and 600 nm diameter FNDs are 200 aM, 46 aM, and 820 zM respectively, defined by the intersection of the lower 95% confidence intervals of the linear fit with the upper 95% confidence intervals of the blanks for each particle size.

Article

Extended Data Fig. 6 | Assay optimization by buffer selection. Sensitivity is limited by the non-specific binding of FNDs at the LFA test line. LFA strip blocking, running buffer and washing step are, therefore, key factors in improving LOD. In this section 120 nm FNDs were used for optimization. a, Signal-to-background comparison for the FNDs in different running buffers. There is no wash step. Error bars show s.d. (n = 3 measurement replicates). Milk was selected as the basis for the running buffer. b, Subsequently, a sweep of different surfactants was performed (n = 1 measurement replicate). The best signal-to-background ratio came from adding 0.05 vol% Empigen, showing a significant increase in the signal-to-background. There is no wash step. c–e, The best running buffer was then used for a washing buffer pH sweep (n = 1 measurement replicate) (c). All washing buffers were run at a volume of 75

μl, chosen because preliminary experiments showed it to be a good compromise between assay time and washing success. Although results were similar, pH 5 gave the best signal-to-background ratio, so acetate buffer at 10 mM pH 5 was used as the basis for a second washing buffer sweep, shown in (d), testing a number of detergents and adding casein at 0.2 wt% as a blocking protein (n = 1 measurement replicate). As a final test, the three best running buffers were tested, each with the three best washing conditions, displayed as a grid in (e). Each square is the average of three measurements (n = 3 measurement replicates). The results were consistent with previous sweeps, the combination of the best running buffer and best washing buffer giving the best signal-to-background. Milk and protein percentages are by weight and detergent percentages are by volume.

Extended Data Fig. 7 | Optimization of FND concentration. The background was reduced by optimising the particle concentration, shown here for 120-nm FNDs. a, A positive LFA strip (500 pM of ssDNA) and a negative control (deionized water) were run at varying FND concentrations between 3.88 fM and 496 fM, plotted against FND concentration, and fitted to simple linear regressions. The dots show means and error bars show the s.d. of repeat measurements (n = 3 measurement replicates). Linear regressions are shown by solid lines, and shaded areas show the 95% confidence intervals of the fits. b, Signal-to-background ratio, found by dividing the fitted linear regressions in a, is plotted against FND concentration. At higher concentrations, where the gradient term of the linear regression dominates, the positive and negative lock-in values tend to a constant separation on the log–log plot, so the signal-to-background ratio tends to a constant value of around 27. At low

concentrations, the positive and negative curves converge as the negative lock-in amplitude levels off at the noise threshold, and the signal-tobackground ratio tends to 1. c, The fitted linear regressions in a were used, along with the antibody equilibrium dissociation constant measured in Extended Data Fig. 4, to estimate the variation of lock-in amplitude with analyte concentration at different FND concentrations. The principles and equations are described in full in Supplementary Information section 2. The LOD for each FND concentration is defined as the intersection of this plot with the value of the blank plus two times the 95% confidence interval at that value, assuming a low concentration positive would have a similar confidence interval. d, The estimated LODs and dynamic ranges from c, plotted against FND concentration, to determine the optimum.

Article

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

Extended Data Fig. 8 | Primer optimization. a, List of forward primers (F1–F5) and reverse primers (R1–R5) tested for the initial primer screen. b, An initial primer screen was performed to achieve the highest amplification efficiency (n = 3 technical replicates) using the TwistAmp Exo Reverse Transcription Kit (TwistDx). The yield of each primer combination was measured by the fluorescence of the exo probe with a fluorescence microplate reader (SpectraMax i3, Molecular Devices LLC). Primers F5 and R3 gave the highest yield, although all the yields were above 63% of this value. c, Interactions between forward primers and reverse primers to predict the minimum free energy structures for the ten primer combinations that gave the largest yield of RPA product in the primer screen. The table shows the results of simulations in NUPACK 56, using an input of 10 μM for each oligonucleotide. The minimum free

energy secondary structures are the most energetically favourable secondary structures that can be assumed for oligonucleotides of a given primary sequence, calculated using the nearest-neighbour method57. Primers F1 and R4 were selected for future work since the energetics of their hybridization are much less favourable than that of F3 and R5, yet they still gave a high RPA yield in the primer screen (93% of the highest yield pair). d, A list of oligonucleotides used for PCR, RPA and qPCR assays. The PCR reverse primer included a T7 promoter for RNA transcription (underlined) and a spacer (bold). e, Gel electrophoresis of 1,503 bp template sequence produced by PCR using a 1% agarose gel. f, Gel electrophoresis of 181 bp double-stranded RT–RPA products using a 1% agarose gel.

Article

Extended Data Fig. 9 | See next page for caption.

Extended Data Fig. 9 | Comparison of LODs of model ssDNA with real RPA amplicons and gold nanoparticles. a, The dilution series of the real RPA amplicons and the model ssDNA ‘amplicons’ were plotted against concentration for 600 nm FNDs (dots showing means with error bars showing s.d., n = 3–9 technical replicates, n = 3 measurement replicates) with their respective linear fits (solid lines with 95% confidence intervals of the fit shown shaded). The curves are similar, with fitted KD values of 29 and 22 fM for model and real amplicons, respectively, and similar dynamic ranges. The real amplicons showed increased variation in the blanks, leading to a higher blank cutoff giving a higher LOD, and slightly reduced signal-to-blank ratio. b, The dilution series of model ssDNA ‘amplicons’ were plotted against concentration for 120, 200 and 600 nm FNDs (dots showing means with error bars showing s.d., n = 3 technical replicates, n = 3 measurement replicates) with their respective linear fits (solid lines with 95% confidence intervals of the fit shown shaded). The LODs are 3.7, 3.6 and 0.8 fM respectively. c, Comparison of 600 nm FNDs with 40 nm gold nanoparticles on LFAs, often used in LFAs owing to a good compromise between stability (and therefore ease of functionalization), and sensitivity58. Serial dilutions are plotted (dots showing means with error

bars showing s.d., n = 3 technical replicates, n = 3 measurement replicates for the FNDs; and dots with error bars showing the s.d. across the test line, n = 1 technical replicate, n = 1 measurement replicate for the gold nanoparticles). LODs are calculated as previously, giving 800 aM and 6.0 pM, respectively. d, e, A Monte Carlo simulation of the signal variation that can be explained by the FND size distribution (from DLS measurements in Extended Data Fig. 3b) was performed (n = 200,000). The violin plots (d) show the normalized simulated random variation in lock-in amplitudes due to the 600-nm FND size distribution in the clinical sample assays in Fig. 4d (negative plasma control and clinical standard). The experimental data are overlaid, showing that FND size distribution explains approximately 8–9% of the total experimental signal variance. Full details of the simulation are given in Supplementary Information section 3. A further approximately 0.1–2% of the variance is explained by periodic drift in modulation amplitude, shown over 45 min in e, normalized to the mean. f, A plot of the variation in lock-in amplitude due to small changes in the modulation frequency, Fm. The variance of the frequency is 3 × 10 −8% over the same period, giving negligible differences in lock-in amplitude.

Article

Extended Data Fig. 10 | See next page for caption.

Extended Data Fig. 10 | Further analysis of RT-RPA samples. a, ANOVA analysis was performed on the measured lock-in amplitudes of the FND LFAs, giving a P value of 7.4 × 10 −29 and F value of 95.6, with 71 total degrees of freedom. Box plots of the data groups are shown (grouped by RNA concentration). The horizontal red lines represent the medians, the horizontal blue lines represent the 25th and 75th percentiles and the notches represent the 95% confidence intervals of the medians. The black dashed lines represent the range for each group. b, A graphical comparison of the means of the groups (grouped by RNA concentration). The circles represent the means, and the horizontal lines represent the comparison intervals of the means from Tukey HSD post hoc test (overlap of these intervals denotes statistical similarity). The negative control, highlighted in blue, is shown to be not significantly different from the 10 −2 and 10 −1 RNA copy number samples (P values >0.999, shown in grey), but it is significantly different from the 1, 101 and 102 RNA copy number samples (P values ≈10 −8, shown in red). c, A table of ANOVA P values. The P value for the null hypothesis that the difference between the means of the two groups is zero. d, Comparison of amplification time for a low copy number RT–RPA sample (average of 1.26 RNA copies). Multiple RPA reactions were run and

stopped after different times, before adding to FND LFAs, as described in Methods. A negative control is shown for comparison, and the dashed line represents the upper 95% confidence interval of the negative control. Dots show the mean of n = 3 measurement replicates, crosses show the individual measurements, and error bars represent the s.d. e, Early disease detection using FND LFAs was demonstrated by a seroconversion panel (ZeptoMetrix Corporation, Panel Donor No. 73698), taken from a single donor over a period of six weeks spanning the early stages of an HIV-1 infection. The thirteen samples of the panel were measured on FND LFAs (n = 1–2 experimental replicates, n = 3 measurement replicates). The measured values are plotted along with positive and negative non-amplification controls. They are colourcoded for RT–PCR results, and labelled with sample numbers, dates, and copy numbers in brackets. The blank cutoff is defined as the upper 95% interval of the negative control. The results show that the RNA was detectable on FND LFAs as early as RT–PCR, and six out of seven RT–PCR-positive samples were detected on FND LFAs, while six out of six RT–PCR-negative samples were negative.

Article

Extended Data Fig. 11 | Detection of HIV-1 capsid protein on using 600 nm FNDs. A serial dilution of the capsid protein was detected on streptavidin-modified LFAs using a sandwich of a biotinylated capture nanobody and antibody-modified FNDs. The results are plotted (n = 3–4 experimental replicates, n = 3 measurement replicates), normalized to the blanks for each sample set with dots showing means and error bars showing s.d. These data were then fitted to a Langmuir curve (equation (6), shown as a line with shaded area denoting the 95% confidence interval of the fit). This gives a LOD of 120 fM, and a lowest concentration that is significantly different from the blank (at the 95% confidence level) of 3 pM, marked with *. Full experimental details are shown in Supplementary Information section 4.

Extended Data Fig. 12 | Effect of lateral flow test strip drying on lock-in amplitude of FND assay. a, Positive and negative lateral flow test strips were measured over time after running was complete (time = 0), showing a small increase in the positive strip lock-in amplitude as the strip dries (the initial lock-in amplitude is around 70% of the final value); however, no increase is seen

in the negative control. The shaded areas show the standard deviation between repeats (n = 3 technical replicates). b, The resulting signal-to-blank ratio variation over time. The shaded areas show the standard deviation between repeats (n = 3 technical replicates), showing that the effect of drying is quite small compared to strip-to-strip variation.

Article

Lanthanide-doped inorganic nanoparticles turn molecular triplet excitons bright https://doi.org/10.1038/s41586-020-2932-2 Received: 24 April 2019 Accepted: 23 September 2020

Sanyang Han1,15, Renren Deng1,2,15 ✉, Qifei Gu1, Limeng Ni1, Uyen Huynh1, Jiangbin Zhang1,3,4, Zhigao Yi5, Baodan Zhao1,6, Hiroyuki Tamura7, Anton Pershin8, Hui Xu9, Zhiyuan Huang10, Shahab Ahmad11, Mojtaba Abdi-Jalebi1,12, Aditya Sadhanala1, Ming Lee Tang10, Artem Bakulin3, David Beljonne8, Xiaogang Liu5,13,14 ✉ & Akshay Rao1 ✉

Published online: 25 November 2020 Check for updates

The generation, control and transfer of triplet excitons in molecular and hybrid systems is of great interest owing to their long lifetime and diffusion length in both solid-state and solution phase systems, and to their applications in light emission1, optoelectronics2,3, photon frequency conversion4,5 and photocatalysis6,7. Molecular triplet excitons (bound electron–hole pairs) are ‘dark states’ because of the forbidden nature of the direct optical transition between the spin-zero ground state and the spin-one triplet levels8. Hence, triplet dynamics are conventionally controlled through heavy-metal-based spin–orbit coupling9–11 or tuning of the singlet–triplet energy splitting12,13 via molecular design. Both these methods place constraints on the range of properties that can be modified and the molecular structures that can be used. Here we demonstrate that it is possible to control triplet dynamics by coupling organic molecules to lanthanide-doped inorganic insulating nanoparticles. This allows the classically forbidden transitions from the ground-state singlet to excited-state triplets to gain oscillator strength, enabling triplets to be directly generated on molecules via photon absorption. Photogenerated singlet excitons can be converted to triplet excitons on sub-10-picosecond timescales with unity efficiency by intersystem crossing. Triplet exciton states of the molecules can undergo energy transfer to the lanthanide ions with unity efficiency, which allows us to achieve luminescent harvesting of the dark triplet excitons. Furthermore, we demonstrate that the triplet excitons generated in the lanthanide nanoparticle–molecule hybrid systems by near-infrared photoexcitation can undergo efficient upconversion via a lanthanide–triplet excitation fusion process: this process enables endothermic upconversion and allows efficient upconversion from near-infrared to visible frequencies in the solid state. These results provide a new way to control triplet excitons, which is essential for many fields of optoelectronic and biomedical research.

Figure 1a shows a schematic of the lanthanide (Ln)-doped nanoparticles (nanocrystals of NaLnF4) and the structures of some of the model molecules used in our study (rubrene and tetracene derivatives) along with their triplet energies. Unlike semiconductor quantum dots, these lanthanide-doped nanocrystals are insulators and their optoelectronic properties are governed solely by the lanthanide ions. We begin by preparing blended films of rubrene and NaGdF4 nanocrystals by drop-casting (Supplementary Figs. 1 and 2).

Figure 1b shows the absorption spectra, measured by photothermal deflection spectroscopy (PDS), of a NaGdF4–rubrene blend film, a pristine rubrene film and a pure NaGdF4 film. Apart from the typical absorption features associated with the S0 → Sn (ground-state singlet to excited-state singlets) transitions in rubrene, we observe new absorption features between 700 nm and 1,100 nm in the NaGdF4–rubrene blend film. By contrast, the pristine rubrene film and the NaGdF4-only film have no absorptions in the same region. The spectra reveal an

Cavendish Laboratory, University of Cambridge, Cambridge, UK. 2Institute for Composites Science Innovation, School of Materials Science and Engineering, Zhejiang University, Hangzhou, China. 3Department of Chemistry, Imperial College London, London, UK. 4College of Advanced Interdisciplinary Studies, National University of Defense Technology, Changsha, China. 5 Department of Chemistry, National University of Singapore, Singapore, Singapore. 6State Key Laboratory of Modern Optical Instrumentation, College of Optical Science and Engineering, International Research Center for Advanced Photonics, Zhejiang University, Hangzhou, China. 7Department of Chemical System Engineering, The University of Tokyo, Tokyo, Japan. 8Laboratory for Chemistry of Novel Materials, University of Mons, Mons, Belgium. 9School of Chemistry and Material Science, Heilongjiang University, Harbin, China. 10Department of Chemistry, University of California, Riverside, Riverside, CA, USA. 11Advanced Energy Materials Group, Department of Physics, Indian Institute of Technology Jodhpur, Jodhpur, India. 12Institute for Materials Discovery, University College London, London, UK. 13The N.1 Institute for Health, National University of Singapore, Singapore, Singapore. 14Joint School of National University of Singapore and Tianjin University, International Campus of Tianjin University, Fuzhou, China. 15These authors contributed equally: Sanyang Han, Renren Deng. ✉e-mail: [email protected]; [email protected]; [email protected] 1

594 | Nature | Vol 587 | 26 November 2020

a

Tetracene: R = 5-CT: R =

H

COOH

R CPT: R = Rubrene ET = 1.14 eV

COOH

CPPT: R =

Tetracene derivatives ET = 1.25 eV

COOH

NaLnF4

b

NaGdF4 only NaLuF4 + rubrene NaYF4 + rubrene

c 4

NaGdF4 + 5-CT NaErF4 + 5-CT NaLuF4 + 5-CT

NaYF4 + 5-CT 5-CT only

d

0.6

0–1

0–0

0–2

0.4

S 0 → Tn

0.2

700 800 900 1,000 1,100

0.0 500 600 700 800 900 1,000 1,100 Wavelength (nm)

3

Absorbance (10–2 )

0.8

2 1 0 750

NaGdF4 + 5-CT NaGdF4 + CPT

NaGdF4 + CPPT NaGdF4 only

4

1.0 Absorbance (10–2 )

Normalized absorbance

b

NaGdF4 + rubrene Calculated S0 → Tn Rubrene only

2

0 800

850 900 950 1,000 1,050 Wavelength (nm)

750

800

850 900 950 1,000 1,050 Wavelength (nm)

Fig. 1 | Lanthanide-nanocrystal-coupled triplet excitation. a, Schematic illustration of a lanthanide-doped nanocrystal (NaLnF4) and the organic molecules in our study. Note that ET refers to the triplet energy of the molecules. b, Comparison of the PDS spectra of films of NaGdF4–rubrene, NaLuF4–rubrene, NaYF4–rubrene with pristine rubrene and with NaGdF4 only. Only the system in which the lanthanide has unpaired spin (that is, the NaGdF4– rubrene films) shows a broadband absorption from 700 nm to 1,100 nm; the inset shows that this absorption matches the calculated direct transition from

the ground singlet state (S0) to the lowest triplet state (T1). c, Comparison of the absorption spectra of 5-CT coupled with various lanthanide nanoparticles, including NaGdF4, NaErF4, NaLuF4 and NaYF4. Enhanced NIR absorption, related to the direct excitation of triplets, is only observed in the presence of lanthanide ions with unpaired spins (Gd3+ and Er3+) d, Comparison of absorption spectra of NaGdF4 nanocrystals coupled with tetracene derivatives (5-CT, CPT, CPPT). The enhanced NIR absorption decreases with increasing spacing between the lanthanide nanoparticle and the core of the molecule.

approximately 200-fold enhancement in the near-infrared (NIR) absorbance of the NaGdF4–rubrene blend compared with the pristine rubrene film. To understand this observation, we performed density functional theory and multireference second-order Møller–Plesset perturbation theory calculations (see Supplementary Information section 2). We found that the experimentally measured absorption in the NIR region (800—1,100 nm) matches well with the calculated absorption for the S0 → T1 transition of an isolated rubrene molecule (Fig. 1b, inset). Note that the theoretical prediction for the Sv0 → T1v (v–v = 0–0, where v represents a vibronic state) transition shows a much higher intensity. We attributed this suppression of the 0–0 band to the Herzberg–Teller mechanism, as described previously14. The extra absorption in the NaGdF4–rubrene blend film is thus assigned to the S0 → Tn transition of rubrene, implying that the usually dark S0 → Tn transition has become bright in the blended system. Similarly, the enhanced dark S0  →  Tn transition feature was also observed in tetracene-derivative-based blends (Supplementary Fig. 3). One explanation for the enhanced S0 → Tn absorption could be related to the spin–orbit coupling associated with the presence of heavy atoms (atomic number Z = 64 for Gd). To test this hypothesis, we prepared blended films of rubrene with different types of lanthanide-doped nanoparticles, including NaGdF4, NaYF4 and NaLuF4. Gd3+ has seven unpaired 4f electrons15, whereas Y3+ (Z = 39) and Lu3+ (Z = 71) have zero spin momentum (Supplementary Table 1). The absorption of the S0 → Tn transition was observed only in the blends with non-zero spin, while no features could be observed for the Y3+- and Lu3+-based blends despite the higher atomic mass of Lu3+ (Fig. 1b). These results suggest that the enhanced S0 → Tn transition is not due to the heavy-atom-induced spin–orbit coupling but related to the spins of unpaired 4f electrons of lanthanide ions.

To further investigate the nature of the coupling between organic molecules and lanthanide nanocrystals, we prepared lanthanide nanoparticles modified with a series of carboxylic acid-functionalized tetracene derivatives, 5-carboxylic acid tetracene (5-CT), 4-(tetracen-5-yl)benzoic acid (CPT) and 4′-(tetracen-5-yl)[1,1′-biphenyl]-4-carboxylic acid (CPPT), as shown in Fig. 1a. These molecules can selectively bind to surface cations of the nanocrystals through their carboxylic groups. The different spacer groups allow us to control the distance between lanthanide ions and the tetracene core, where the triplet excitons will be localized. We studied blended films of 5-CT with different NaLnF4 (Ln = Gd3+, Er3+, Y3+ or Lu3+) nanocrystals, as shown in Fig. 1c. The absorption spectra showed that all the lanthanides with unpaired 4f electrons (Gd3+ and Er3+) give rise to an enhanced NIR absorption of 5-CT molecules, whereas the lanthanides without unpaired electrons (Y3+ and Lu3+) have a negligible effect. Figure 1d compares the effects of longer spacer groups. As we move from 5CT to CPT and CPPT, it was observed that the absorption in the NIR spectral region decreases. This result confirms that the coupling between lanthanides and tetracene molecules is very sensitive to the distance between them. On the basis of these observations, we propose that the proximity of the organic molecules to the lanthanide ions with unpaired spins permits photon absorption to directly generate triplet states, S0 → Tn. We return to the nature of this interaction later. We next explore the effect of the lanthanide ions on the excited states of the organic semiconductors. We chose 9-[3-carboxyl-4-(diph enylphosphinoyl)phenyl]-9H-carbazole (CPPOA) as a model molecule owing to its high-lying triplet state (which will be important for the discussion of energy transfer below). We prepared colloidal solutions of NaYF4@NaLnF4 core–shell nanoparticles with surface-bound Nature | Vol 587 | 26 November 2020 | 595

Article a

b 1.0

WS = 12.9 ns

WS = 82.3 ps 1

O

P

C

T1

0.4 0.2

N

NaYF4@NaLnF4

Normalized ΔT/T

WT = 18.4 ns 1

WT = 96.8 ps 1

0.0

1.0

Fast ISC

0.6

CPPOA

d

S1

0.8 Normalized ΔT/T

O

O

c

1

CPPOA-NaYF4@NaGdF4 10 ps

100 ps

1 ns Time

Pristine CPPOA 10 ns

100 ns

S0

WEu = 9.3 ps WTb = 92.4 ps

0.8

WGd = 96.8 ps

0.6

WY = 17.4 ns

WYb = 2.2 ns WLu = 12.3 ns

0.4 WTb,decay = 7.66 ns 0.2 WEu,decay = 883.4 ps 0.0 10 ps

100 ps

1 ns

10 ns Time

100 ns

1 μs

10 μs

100 μs

Fig. 2 | Ultrafast intersystem crossing in organic molecules coupled to lanthanide-doped nanoparticles. a, Schematic illustration of a NaYF4@ NaLnF4 core–shell nanoparticle modified with CPPOA. b, Extracted kinetics showing the singlet (S1) decay and triplet (T 1) growth of a solution containing pristine CPPOA molecules and of a solution of CPPOA-modified NaYF4@ NaGdF4 nanoparticles. The singlet lifetime decreases from 12.9 ns in the pristine CPPOA to 82.3 ps in CPPOA-modified NaYF4@NaGdF4, indicating greatly enhanced intersystem crossing (ISC). Note that ΔT/T refers to the fractional differential transmission signal of the probe in transient absorption spectra. c, The interaction between the lanthanides and the molecules accelerates the ISC from the singlet to triplet exciton states of the molecule.

d, Kinetics of triplet generation and decay in CPPOA molecules attached to different types of core–shell nanoparticles. The compositions of the core– shell nanoparticles are NaYF4@NaEuF4, NaYF4@NaGdF4, NaYF4@NaTbF4, NaYF4@NaYbF4, NaYF4@NaLuF4 and NaYF4@NaYF4. The key at top left gives the fitted values of τLn for the increasing part of each curve. For lanthanides with unpaired 4f electrons (Tb3+, Eu3+, Gd3+ and Yb3+) enhanced ISC is seen, whereas those with no unpaired spins (Y3+ and Lu3+) show no obvious enhancement in ISC. An enhanced triplet decay (τTb,decay = 7.66 ns and τEu,decay = 883 ps) of CPPOA on the nanoparticles containing Tb3+ or Eu3+ suggests triplet energy transfer to the lanthanide ions.

CPPOA molecules (Fig. 2a) and studied the dynamics of photoexcitations using pump–probe spectroscopy. The S1 state of CPPOA on the NaYF4@NaGdF4 nanoparticle was found to decay with a time constant of 82 ps (Fig. 2b), concomitant with the rise time of the triplet excitons (T1, 97 ps). This indicates that the photogenerated singlet of CPPOA attached to the NaYF4@NaGdF4 nanoparticles undergoes rapid intersystem crossing (ISC). By contrast, the singlet of the pristine CPPOA shows a decay time of 12.9 ns with a concomitant triplet rise over 18.4 ns. Thus, the presence of the Gd3+-based nanoparticles increases the rate of the ISC by three orders of magnitude (Fig. 2c). To investigate this further, we attached CPPOA to a series of NaYF4@NaLnF4 core–shell nanoparticles with different lanthanide ions in the shell and measured the triplet generation rate. As shown in Fig. 2d, we observe an enhanced ISC rate for nanoparticles with unpaired 4f electrons (Tb3+, Eu3+, Gd3+ and Yb3+), but not for those with no unpaired spins (Y3+ and Lu3+) (Supplementary Tables 1–3). The same trend was observed in sub-gap absorption enhancement for CPPOA-capped nanoparticles, analogous to the results in Fig. 1 (Supplementary Figs. 9 and 10a). For Eu3+-doped nanoparticles, we measured a triplet rise time of 9.3 ps, which is 1,978 times faster than that of the pristine CPPOA molecules. The ISC efficiency is estimated to be 99.4%, based on the singlet lifetime quenching. Thus, in addition to turning S0 → Tn transitions bright, the interaction between

the CPPOA and the unpaired spins on the lanthanide nanoparticles also yields highly efficient ISC (ref. 16; see Supplementary Figs. 11–32 for full details). It can be seen in Fig. 2d that the fast rise of the T1 state for Tb3+- and Eu3+-containing nanoparticles is accompanied by a quick decay of that state (7.66 ns for Tb3+ and 883 ps for Eu3+). This decay is caused by the energy transfer of the T1 state from CPPOA to the 5D1/5D0 and 5D4 levels of Eu3+ and Tb3+, respectively (Fig. 3a). On the basis of the quenching of the triplet lifetime, the calculated quantum efficiency of triplet energy transfer from CPPOA to lanthanide nanoparticles exceeds 99%. This near-quantitative triplet energy transfer gives rise to bright luminescence from the lanthanide ions on excitation of the coupled systems at 365 nm (Fig. 3b and Supplementary Figs. 10b, 33–35). These results show that molecular triplet excitons can be efficiently transferred to lanthanide-doped nanoparticles, allowing the luminescent harvesting of normally dark triplet excitons. This luminescent harvesting of triplet excitons is not restricted to triplets generated by ISC on the surface of the nanoparticles but is also relevant for triplets generated by other processes such as singlet fission. Tetracene and rubrene are both well-known singlet fission materials, in which the photogenerated singlet excitons rapidly and efficiently convert to a pair of triplet excitons13,17. Figure 3c shows data for blend films of tetracene and rubrene with NaGdF4:Yb (50 mol%)

596 | Nature | Vol 587 | 26 November 2020

b

T1 T TE

S0 Organic

c Rubrene

Eu3+

Tb3+

450 Ln3+

550

650

Wavelength (nm)

Normalized intensity

ISC st Fa r SF o

S1

Normalized intensity

a

Tetracene Yb3+ → 2F7/2

2F 5/2

Sn → S0

500

700

900

1,100

Wavelength (nm)

Fig. 3 | Triplet energy transfer from molecules to nanoparticles. a, Simplified energy diagram showing triplet energy transfer (TET) from the molecular triplet state to lanthanide emitters (Ln3+) following a fast intersystem crossing (ISC) or singlet fission (SF) process. b, Photoluminescence spectra and corresponding luminescence photographs of colloidal solutions containing CPPOA-modified NaYF4@NaEuF4 nanoparticles (top) and CPPOA-modified NaYF4@NaTbF4 nanoparticles (bottom) under excitation at

365 nm. c, Photoluminescence spectra of NaGdF4:Yb–tetracene blend (green curve) and NaGdF4:Yb–rubrene blend (red curve) films excited at 405 nm. Luminescence arises from the transfer of triplet excitons generated via singlet fission to the 2F5/2 → 2F7/2 transition of Yb3+. The process is inefficient in rubrene owing to its triplet energy being lower than the 2F5/2 → 2F7/2 transition of Yb3+ (1.14 eV versus 1.25 eV).

nanoparticles. These nanoparticles feature an energy gap of 1.25 eV between the lowest excited state (2F5/2) and the ground state (2F7/2) of Yb3+, which lies near the triplet energy of tetracene (1.25 eV; ref. 17) and above that of rubrene (1.14 eV; ref. 18). When a NaGdF4:Yb–tetracene blend film was excited at 405 nm, we recorded a strong quenching of the characteristic visible emission from tetracene in favour of a Yb3+ emission located at 950–1,100 nm (Fig. 3c and Supplementary Fig. 36). We observed that the emission of Yb3+ was strongly quenched when the blend film was exposed to air (Supplementary Fig. 37), suggesting that the Yb3+ emission arises from triplet-mediated energy transfer from tetracene to Yb3+. Magnetic-field-dependent photoluminescence measurements showed an increase in emission from the tetracene while the emission from Yb3+ decreased with increasing magnetic field19,20, confirming the transfer of triplets generated by the singlet fission process to Yb3+ (Supplementary Figs. 44–46). By contrast, almost no quenching of the visible and no NIR emission was observed in the rubrene blend owing to inefficient triplet energy transfer to Yb3+. Figure 4a shows the photoluminescence spectra of the NaGdF4:Yb– rubrene and NaGdF4:Yb–tetracene blends on 980-nm excitation. Spectra corresponding to the singlet emission from both tetracene and rubrene were obtained, consistent with the upconversion of absorbed energy. This upconverted emission is visible even under ambient light conditions (Supplementary Fig. 38) and is found to have a quadratic dependence on the excitation power at low excitation fluence, followed by a slope change from 2 to 1 at higher excitation density (Supplementary Fig. 39). Considering the broadband absorption of the NaGdF4:Yb–rubrene blended film in the NIR wavelength region (Supplementary Fig. 40), we carried out excitation in a spectral range from 850 nm to 1,020 nm and observed upconverted emission at all excitation wavelengths (Supplementary Figs. 41, 42). These results suggest that it is the interaction between the molecular triplet exciton and the lanthanide, rather than the conventional triplet–triplet annihilation (TTA) process21–23, that mediates the upconversion process in organic molecules after triplet transfer from photoexcited Yb3+-doped nanoparticles under NIR irradiation (Fig. 4b). To further investigate the upconversion mechanism, we prepared a series of blends with varying concentration ratios of NaGdF4:Yb and rubrene or tetracene. The main emission peak shifted from 540 nm (2.29 eV) to 480 nm (2.58 eV) for the NaGdF4:Yb–tetracene blends when the concentration ratio of tetracene to nanoparticle was changed from 10:1 to 1:100 (see Fig. 4c and Supplementary Fig. 43 for rubrene blends). Interestingly, an emission characteristic of a single tetracene molecule24 was

obtained when the concentration of tetracene was diluted to 1 molecule per 100 nanoparticles. This upconverted emission is recorded at room temperature with moderate excitation density (106 W cm−2). Conventional TTA upconversion through bimolecular triplet–triplet states does not enable the single molecular emission of tetracene because the emission would be shifted to lower energies due to excitonic coupling, as happens when we increase the tetracene concentration in the blend films. This again suggests that the upconverted emission is produced by a different process. Magnetic-field-dependent photoluminescence studies show no change in upconverted emission under applied magnetic field (Supplementary Fig. 47), confirming that the upconversion process is not mediated by the TTA (see Supplementary Information section 9 and Supplementary Figs. 48, 49 for full details). Given that the singlet energy of tetracene (for a single molecule, 2.62 eV; Supplementary Figs. 48, 50) is higher than the total energy contained in two excitation photons (1.25 eV × 2 = 2.50 eV), this new mechanism enables us to obtain endothermic upconversion. To investigate this endothermic upconversion process further, we performed temperature-dependent upconversion measurements. The upconversion emission intensity for the 1:100 film gradually increased as the temperature was reduced to 80 K, presumably due to the suppression of non-radiative loss channels at low temperatures (Fig. 4d). However, further lowering the temperature resulted in a decrease in upconversion emission (Supplementary Figs. 51, 52). We note that the downshifting luminescence of the same sample under 405-nm excitation increases monotonically as the temperature drops, and the integrated emission intensity at 20 K is 12 times stronger than that measured at room temperature (Supplementary Fig. 53). At room temperature, we measured an internal photoluminescence quantum yield (PLQY) of more than 1% for the NaGdF4:Yb–rubrene blend film with moderate excitation of >16 W cm−2. A maximum PLQY value of 1.9% ± 0.5% was reached at an irradiance power density of 75 W cm−2 (Fig. 4e). We note that the singlet PLQY of rubrene in the blend under 405-nm excitation at room temperature was measured to be 20% ± 2.1%. This suggests a maximum singlet yield of about 10% per absorbed NIR photon for NaGdF4:Yb–rubrene. A maximum PLQY value of 16.2% ± 3.4% was attained at 10 K (Supplementary Fig. 54), due to reduction of the non-radiative energy loss pathways at low temperatures. Given the fact that two lower-energy photons are converted to Nature | Vol 587 | 26 November 2020 | 597

Article c

d

Rubrene

1:100

700 S1

b LTF

T1

5/2

10:1 Tetracene: Tetracene : YbNCs nanoparticles of NaGdF4:Yb 1:100 1:100 1:10 1:10 10:1 10:1

12 10 8

×3.2

6 4 0

e 2.0

50

100 150 200 Temperature (K)

S0 Yb3+

Rubrene/tetracene

300

1.5 18 1.0

12 6 00

7/2

250

T = 298 K NaGdF4:Yb-rubrene

0.5 2F

Tetracene:nanoparticles Tetracene 1:100 : YbNCs = of NaGdF4:Yb = 1:100

15%

2

PLQY (%)

500 550 600 650 Wavelength (nm)

1:10

Normalized intensity

450

2F

Integrated PL (103 )

Normalized intensity

Tetracene

PLQY (%)

a

450

500

550 600 650 Wavelength (nm)

700

0.0

0

100 200 T (K)

20 40 60 Excitation power density (W cm–2)

300 80

Fig. 4 | Lanthanide–triplet excitation fusion upconversion in nanoparticle– molecule blends. a, Photoluminescence spectra of NaGdF4:Yb–tetracene blend (50:1) and NaGdF4:Yb–rubrene blend (1:10) films excited at 980 nm (about 40 W cm−2), showing upconverted emission arising from the singlet state of the organics. b, Proposed lanthanide–triplet excitation fusion (LTF) upconversion process. c, Top, scheme represents the isolated molecule (left) and molecular conjugates (right) when tetracene couples with the NaGdF4:Yb nanoparticles. Note that grey and blue spheres in the nanoparticle refer to Gd3+ and Yb3+ ions, respectively. Bottom, upconversion spectra and corresponding emission photographs (middle) of NaGdF4:Yb–tetracene blend films with varying tetracene-to-nanocrystal ratios (1:100, 1:10 and 10:1). d, Plot of the

temperature-dependent upconversion emission (integrated from 475 nm to 650 nm) of the NaGdF4:Yb–tetracene blend films (tetracene: NaGdF4:Yb = 1:100). Note that PL refers to photoluminescence. Dashed arrows suggests that upconversion emission showed a 3.2-fold increase as the temperature decreased from 300 K to 80 K and then a 15% decrease when the temperature was further reduced to 10 K. e, The internal photoluminescence quantum yield (PLQY) of the NaGdF4:Yb–rubrene blend measured as a function of excitation power density (excitation at 980 nm). Inset, temperature-dependent quantum yield of the same sample under excitation of 980 nm for a power density of 76 W cm−2.

one higher-energy photon during the upconversion process, our system has thus converted about 32% of the absorbed photons. In comparison with the conventional lanthanide25–28- or TTA5,22-based upconversion, a key feature of the lanthanide–triplet excitation fusion approach demonstrated here is that the excitation energy can be directly amassed in both organic and inorganic components without the need for a sensitization step. Therefore, energy loss during the sensitization process can be effectively reduced to zero29. In addition, lanthanide-doped nanoparticles have no absorption at higher energies, thereby eliminating the problem of reabsorption associated with quantum dot/molecule systems30,31. Furthermore, owing to the nature of the spin states, normal spin statistical limitations that apply to the conventional TTA do not apply to the lanthanide–triplet upconversion. We have demonstrated that it is possible to control and manipulate triplet exciton dynamics by coupling conventional molecular systems to the unpaired spins of lanthanide ions doped in inorganic nanoparticles. Further experimental and theoretical work is required to understand the nature of the coupling in these systems. Our results open up new avenues for triplet sensitization, photocatalysis, optoelectronics, sensing, and photon frequency conversion driven by optically bright triplet excitons.

1.

Online content Any methods, additional references, Nature Research reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at https://doi.org/10.1038/s41586-020-2932-2. 598 | Nature | Vol 587 | 26 November 2020

2. 3. 4.

5.

6. 7.

8. 9.

10. 11.

12. 13. 14.

15. 16.

Bolton, O., Lee, K., Kim, H. J., Lin, K. Y. & Kim, J. Activating efficient phosphorescence from purely organic materials by crystal design. Nat. Chem. 3, 205–210 (2011); correction 3, 415 (2011). Baldo, M. A. et al. Highly efficient phosphorescent emission from organic electroluminescent devices. Nature 395, 151–154 (1998). Uoyama, H., Goushi, K., Shizu, K., Nomura, H. & Adachi, C. Highly efficient organic light-emitting diodes from delayed fluorescence. Nature 492, 234–238 (2012). Yanai, N. & Kimizuka, N. New triplet sensitization routes for photon upconversion: thermally activated delayed fluorescence molecules, inorganic nanocrystals, and singlet-to-triplet absorption. Acc. Chem. Res. 50, 2487–2495 (2017). Schulze, T. F. & Schmidt, T. W. Photochemical upconversion: present status and prospects for its application to solar energy conversion. Energy Environ. Sci. 8, 103–125 (2015). Ravetz, B. D. et al. Photoredox catalysis using infrared light via triplet fusion upconversion. Nature 565, 343–346 (2019); correction 570, E24 (2019). Mongin, C., Garakyaraghi, S., Razgoniaeva, N., Zamkov, M. & Castellano, F. N. Direct observation of triplet energy transfer from semiconductor nanocrystals. Science 351, 369–372 (2016). Köhler, A. & Bassler, H. Triplet states in organic semiconductors. Mater. Sci. Eng. Rep. 66, 71–109 (2009). Lamansky, S. et al. Highly phosphorescent bis-cyclometalated iridium complexes: synthesis, photophysical characterization, and use in organic light emitting diodes. J. Am. Chem. Soc. 123, 4304–4312 (2001). Bünzli, J.-C. G. On the design of highly luminescent lanthanide complexes. Coord. Chem. Rev. 293–294, 19–47 (2015). Klink, S. I., Kerizer, H. & van Veggel, F. C. J. M. Transition metal complexes as photosensitizers for near-infrared lanthanide luminescence. Angew. Chem. Int. Ed. 39, 4319–4321 (2000). Penfold, T. J., Gindensperger, E., Daniel, C. & Marian, C. M. Spin-vibronic mechanism for intersystem crossing. Chem. Rev. 118, 6975–7025 (2018). Smith, M. B. & Michl, J. Singlet fission. Chem. Rev. 110, 6891–6936 (2010). Wykes, M., Parambil, R., Beljonne, D. & Gierschner, J. Vibronic coupling in molecular crystals: a Franck-Condon Herzberg-Teller model of H-aggregate fluorescence based on quantum chemical cluster calculations. J. Chem. Phys. 143, 114116 (2015). Auzel, F. Upconversion and anti-Stokes processes with f and d ions in solids. Chem. Rev. 104, 139–174 (2004). Tobita, S., Arakawa, M. & Tanaka, I. The paramagnetic metal effect on the ligand localized S1→T1 intersystem crossing in the rare-earth-metal complexes with methyl salicylate. J. Phys. Chem. 89, 5649–5654 (1985).

17. 18. 19. 20. 21. 22.

23. 24.

25.

Tiberghien, A. & Delacote, G. Evaluation of the crystalline tetracene triplet Davydov splitting. Chem. Phys. Lett. 8, 88–90 (1971). Tao, S. et al. Optical pump-probe spectroscopy of photocarriers in rubrene single crystals. Phys. Rev. B 83, 075204 (2011). Thompson, N. J. et al. Energy harvesting of non-emissive triplet excitons in tetracene by emissive PbS nanocrystals. Nat. Mater. 13, 1039–1043 (2014). Tabachnyk, M. et al. Resonant energy transfer of triplet excitons from pentacene to PbSe nanocrystals. Nat. Mater. 13, 1033–1038 (2014). Wu, M. et al. Solid-state infrared-to-visible upconversion sensitized by colloidal nanocrystals. Nat. Photon. 10, 31–34 (2016). Zhao, J., Ji, S. & Guo, H. Triplet–triplet annihilation based upconversion: from triplet sensitizers and triplet acceptors to upconversion quantum yields. RSC Adv. 1, 937–950 (2011). Singh-Rachford, T. N. & Castellano, F. N. Photon upconversion based on sensitized triplet– triplet annihilation. Coord. Chem. Rev. 254, 2560–2573 (2010). Burdett, J. J., Müller, A. M., Gosztola, D. & Bardeen, C. J. Excited state dynamics in solid and monomeric tetracene: the roles of superradiance and exciton fission. J. Chem. Phys. 133, 144506 (2010). Zhao, J. et al. Single-nanocrystal sensitivity achieved by enhanced upconversion luminescence. Nat. Nanotechnol. 8, 729–734 (2013).

26. Haase, M. & Schäfer, H. Upconverting nanoparticles. Angew. Chem. Int. Ed. 50, 5808– 5829 (2011). 27. Garfield, D. J. et al. Enrichment of molecular antenna triplets amplifies upconverting nanoparticle emission. Nat. Photon. 12, 402–407 (2018). 28. Pollnau, M., Gamelin, D. R., Lüthi, S. R., Güdel, H. U. & Hehlen, M. P. Power dependence of upconversion luminescence in lanthanide and transition-metal-ion systems. Phys. Rev. B 61, 3337–3346 (2000). 29. Nienhaus, L., Wu, M., Bulović, V., Baldo, M. A. & Bawendi, M. G. Using lead chalcogenide nanocrystals as spin mixers: a perspective on near-infrared-to-visible upconversion. Dalton Trans. 47, 8509–8516 (2018). 30. Nienhaus, L. et al. Speed limit for triplet-exciton transfer in solid-state PbS nanocrystal-sensitized photon upconversion. ACS Nano 11, 7848–7857 (2017). 31. Huang, Z. et al. Hybrid molecule−nanocrystal photon upconversion across the visible and near-infrared. Nano Lett. 15, 5552–5557 (2015). Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. © The Author(s), under exclusive licence to Springer Nature Limited 2020

Nature | Vol 587 | 26 November 2020 | 599

Article Methods Materials Gd(CH3CO2)3·xH2O (99.9%), Y(CH3CO2)3·xH2O (99.9%), Yb(CH3CO2)3· 4H2O (99.9%), Tb(CH3CO2)3·xH2O (99.9%), Eu(CH3CO2)3·xH2O (99.9%), Lu(CH3CO2)3·xH2O(99.9%),NaOH(>98%),NH4F(99%),1-octadecene(90%), oleic acid (90%), rubrene (99.99%), tetracene (99.99%) and all anhydrous solvents were purchased from Sigma-Aldrich. If not stated otherwise, all chemicals were used as received without further purification. Synthesis of 5-nm NaLnF4 nanocrystals The lanthanide nanocrystals were synthesized according to a well-documented co-precipitation method32. In a typical experiment for synthesizing 5-nm NaGdF4:Yb (50 mol%) nanocrystals, a water solution (2 ml) containing Gd(CH3CO2)3 (0.2 mmol) and Yb(CH3CO2)3 (0.2 mmol) was mixed with oleic acid (3.5 ml) and 1-octadecene (10.5 ml) in a 50-ml flask, followed by heating at 150 °C for 2 h. Thereafter, the reactant was cooled to 50 °C, and a methanol solution (6 ml) containing NH4F (1.36 mmol) and NaOH (1 mmol) was added. The mixed solution was stirred for 30 min. The reaction temperature was then raised to 100 °C to remove the methanol from the reaction solution. After that, the reactant was heated at 270 °C under a nitrogen atmosphere for 1 h, followed by cooling to room temperature. The resulting nanocrystals were extracted through repeated precipitation with a mixture of ethanol and methanol, collected by centrifugation at 4,000 rpm for 5 min, and re-dispersed in 4 ml of hexane. Pure NaGdF4 nanocrystals were synthesized with the same procedure with the addition of the corresponding Ln(CH3CO2)3 solution. Synthesis of NaYF4 core nanoparticles In a typical procedure32, an aqueous solution of Y(CH3CO2)3·xH2O (2 ml, 0.2 M) was mixed with 3 ml of oleic acid (OA) in a 50-ml flask. The mixture was heated at 150 °C in an oil bath for 30 min. Then 7 ml of 1-octadecene (ODE) was added to the flask. The mixture was cooled to 50 °C after 30 min. After that, a methanol solution (6 ml) containing NH4F (1.6 mmol) and NaOH (1 mmol) was added to the core precursor and stirred continuously for 30 min. After the removal of the low boiling point solvent, the temperature was increased to 290 °C under an argon atmosphere. After 2 h, the mixture was cooled down and washed with ethanol several times. The product was re-dispersed in 4 ml of cyclohexane. Synthesis of core–shell nanoparticles NaLnF4 shell precursor was then prepared by adding an aqueous solution of Ln(CH3CO2)3·xH2O (1 ml, 0.2 M, Ln = Y, Gd, Eu, Tb, Yb, Lu) to a mixture of OA (3 ml) and ODE (7 ml). The mixture was heated at 150 °C in an oil bath for 1 h. After cooling to 80 °C, NaYF4 core nanoparticles in 4 ml of cyclohexane were added, and the resulting mixture was kept at 80 °C for 30 min. Subsequently, a methanol solution of NH4F (0.8 mmol) and NaOH (0.5 mmol) was added under magnetic stirring and kept for 30 min at 50 °C. After that, the temperature was increased to 100 °C to evaporate the low boiling point solvents. The mixture was finally heated at 290 °C under an argon atmosphere for 2 h. After cooling to room temperature, the product nanoparticles were precipitated, washed several times with ethanol, and re-dispersed in 4 ml of cyclohexane for further use. Preparation of ligand-free nanocrystals The as-prepared OA-capped nanoparticles were precipitated from the hexane solution by adding acetone and were then redispersed in an acetone solution containing HCl (0.1 M). The solution was ultrasonicated for 20 min to remove the oleate ligands on the surface. After the reaction, the nanocrystals were collected by centrifugation. The precipitants were washed with acetone/methanol several times and finally redispersed in methanol33.

Nanoparticle–molecule film fabrication Samples were fabricated on 15-mm round glass substrates. In a typical experiment, the glass substrates were first cleaned by sequential sonication in isopropanol and acetone, followed by treatment with oxygen plasma for 10 min. The substrates were then transferred to a nitrogen glovebox. An anhydrous chloroform solution of tetracene or rubrene was mixed with the methanol solution of nanoparticles in the glovebox. The resulting nanoparticle–molecule mixed solution was drop-cast onto the glass substrates to form blend films. The as-prepared sample films were covered with a 0.13-mm thin glass slide and encapsulated with epoxy glue in the glovebox before exposure to air. Sample characterization Transmission electron microscopy (TEM) measurements were carried out on a JEOL-2010F transmission electron microscope ( JEOL) operating at an acceleration voltage of 200 kV. Scanning electron microscopy (SEM) images were recorded using a Leo Gemini 1530 VP SEM with an acceleration voltage of 3 kV. Ultraviolet-visible diffuse reflection spectra were recorded on a Lambda 750 spectrophotometer equipped with an integrating sphere to collect all of the diffuse reflection from the samples. Photothermal deflection spectroscopy This is a highly sensitive surface-averaged absorption measurement technique. For the measurements, a monochromatic pump light beam (produced by a combination of a Light Support MKII 100 W xenon arc source and a CVI DK240 monochromator) is shone on the sample (a thin film on a quartz substrate) perpendicular to the plane of the sample, which on absorption produces a thermal gradient near the sample surface via non-radiative relaxation induced heating. This results in a refractive index gradient in the area surrounding the sample surface. This refractive index gradient is further enhanced by immersing the sample in a deflection medium comprising an inert liquid (FC-72 Fluorinert, 3M Co.) that has a high refractive index change per unit change in temperature. A fixed wavelength CW transverse laser probe beam, produced using a Qioptiq 670-nm wavelength fibre-coupled diode laser with temperature stabilizer for reduced beam pointing noise, was passed through the thermal gradient in front of the sample to produce a deflection proportional to the absorbed light at that particular wavelength. The signal is detected by a differentially amplified quadrant photodiode and a Stanford Research SR830 lock-in amplifier combination. Scanning through different wavelengths gives complete absorption spectra. Steady-state photoluminescence Photoluminescence spectra were measured by exciting the solid film using a diode laser (BrixX976 NB, 980 nm, for upconversion; LDM405.100.CWA.L, 405 nm, for downshifting) with a laser spot size of about 1 mm. The spectra were recorded using a spectrometer (Andor, Shamrock SR-303i) integrated with a CCD detector (Andor, DU420A-BVF). For upconversion spectral measurements, a 900-nm short-pass filter was placed in front of the spectrometer to cut off the scattering from the laser. The magnetic field dependent photoluminescence measurements were carried out using a spectrometer (Andor, Shamrock SR-303i) integrated with a CCD detector (Andor, DU420A-BVF) with the samples placed in the centre of an electromagnet (GMW, Model 3470). Transient photoluminescence spectroscopy Time-resolved photoluminescence measurements were obtained with a customized phosphorescence lifetime spectrometer (Edinburgh, FSP920-C). A nanosecond optical parametric oscillator (OPO) pumped by a 3.8-ns-pulsed Nd:YAG laser (Ekspla, NT352) was used as the excitation source. The emission from the samples was collected at an angle of 90° to the excitation beam using a pair of lenses.

Quantum yield measurements The quantum yield was measured using an integrating sphere method34. Samples were placed in an integrating sphere (Labsphere, 150 mm, internally coated with barium sulfate). A continuous-wave diode laser (BrixX976 NB, 980 nm) was used to excite the samples. The emission from the samples in the integrating sphere was collected by a spectrometer (Andor, Shamrock SR-303i) through an optical fibre. The signal was recorded by a CCD detector (Andor, DU420A-BVF). To avoid saturation of the detector by NIR signals, an NIR neutral density filter (Thorlabs, NENIR40B) was used to reduce the signal from the laser. The quantum yield of upconversion emission was calculated by measuring the number of photons emitted versus the number of photons absorbed. In our quantum yield measurements, three experiments were carried out in an integrating sphere, and the total light intensity collected at the spectrometer was measured: (a) laser excitation with no sample; (b) laser excitation and sample emission with direct illumination of the sample; and (c) laser excitation and sample emission with indirect illumination of the sample. By integrating the excitation and emission signals, the upconversion efficiency was obtained as follows: the number of photons absorbed equals LaA, where La is the excitation intensity in experiment (a), and A = 1 − (Lc/Lb), where Lb and Lc are the excitation intensities in experiments (b) and (c), respectively. The number of photons emitted equals Pc − [(1 − A)Pb], where Pb and Pc are the emission intensities in experiments (b) and (c) respectively. Thus, the quantum yield is given by PLQY = {Pc − [(1 − A)Pb]}/(LaA). This method accounts for all photons absorbed by direct excitation, indirect excitation via scattering in the integration sphere, and sample emission. Details of the quantum yield measurement and calculation can be found elsewhere34. Transient absorption spectroscopy The samples were excited by a pump pulse and then probed at different delayed times using a broadband probe pulse. Transient absorption spectra were recorded over short time delays (500 fs to 6 ns) with a probe pulse covering 500–850 nm and 750–1,600 nm, and over long (1 ns to 1 ms) time delays with a probe pulse covering 350–750 nm and 850–1,020 nm. The short-time transient absorption (ps-TA) measurements were performed with a commercial transient absorption spectrometer (HELIOS, Ultrafast Systems). Part of the output from a regenerative Ti:sapphire amplifier system (Spectra Physics, Solstice, generating ultrafast pulses at about 790 nm) was used to pump a TOPAS-Prime (light conversion) to generate tunable pump pulses (355– 2,600 nm). Another part of the output from this Ti:sapphire system was introduced to a YAG crystal to generate a broadband probe pulse (800–1,550 nm). The probe light is delayed using a computer-controlled piezoelectric translation stage, and a sequence of probe pulses with and without the pump is generated using a chopper wheel on the pump beam. The pump and probe pulses were focused onto a spot of area about 0.5 mm2. The time resolution of the laser pulse was about 200 fs. In long-time transient absorption (ns-TA) measurements, an electronically controlled delay was employed. A separate frequency-doubled Q-switched Nd:YVO4 laser (AOTYVO-25QSPX, Advanced Optical Technologies) is used to generate pump pulses with a temporal breadth below 1 ns at 530 nm. The pump and probe beams overlap on the sample adjacent to a reference probe beam. This reference is used to account for any shot-to-shot variation in transmission. The sample is held in a 1-mm quartz cuvette, mounted in a holder. The beams are focused into an imaging spectrometer (Andor, Shamrock SR 303i) and detected using a pair of linear image sensors (Hamamatsu, G11608) driven and read out at the full laser repetition rate by a custom-built board from Stresing Entwicklungsburo. In all measurements, every second pump shot is omitted, either electronically for long-time measurements or using a mechanical chopper for short-time measurements. The fractional differential transmission (ΔT/T) of the probe is calculated for each data point once 1,000 shots are collected.

In pump–probe experiments as described above, the differential transmission (ΔT/T) signal in the transient absorption spectra refers to features that define excited states. To identify different components from the transient absorption data, a genetic algorithm analysis was also used to distinguish different spectral species and the corresponding kinetics. In the pump–probe technique, a short light pulse (the ‘pump’) excites the sample, and the other pulse (the ‘probe’), which is broad in energy but short in time, interrogates the same spot after a time delay. The transmitted light from the probe is compared with and without the pump light and resolved by both spectral wavelength and delay time. If there is a change in the spectra of the probe because of bleaching of the ground-state transitions (‘ground-state bleach’), stimulated emission, or excited-state absorption from one excited state to another, these will manifest as a change in the transmittance of the probe, ΔT. We recorded the signal normalized by the ground-state transmittance, ΔT/T, to facilitate comparison across experimental configurations. We excited CPPOA at 355 nm to create singlet excitons (S1) on the molecules and subsequently probed the evolution of the spectral features as a function of time. We note that the lanthanide ions have no transient absorption features and thus the entire response arises from the excited-state features of CPPOA.

Data availability The data underlying all figures in the main text and Supplementary Information are publicly available from the University of Cambridge repository at https://doi.org/10.17863/CAM.59063. 32. Wang, F., Deng, R. & Liu, X. Preparation of core-shell NaGdF4 nanoparticles doped with luminescent lanthanide ions to be used as upconversion-based probes. Nat. Protoc. 9, 1634–1644 (2014). 33. Bogdan, N., Vetrone, F., Ozin, G. A. & Capobianco, J. A. Synthesis of ligand-free colloidally stable water dispersible brightly luminescent lanthanide-doped upconverting nanoparticles. Nano Lett. 11, 835–840 (2011). 34. de Mello, J. C., Wittmann, H. F. & Friend, R. H. An improved experimental determination of external photoluminescence quantum efficiency. Adv. Mater. 9, 230–232 (1997). Acknowledgements This project received funding from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (grant agreement no. 758826) and from Marie Skłodowska-Curie grant agreements nos 797619 (TET-Lanthanide project), 748042 (MILORD project) and 646176 (EXTMOS project). We acknowledge support from the Engineering and Physical Sciences Research Council (EPSRC) and the Winton Programme for the Physics of Sustainability, the Singapore Ministry of Education (grant MOE2017-T2-2-110), the Singapore Agency for Science, Technology and Research (grant A1883c0011), and the National Research Foundation, Prime Minister’s Office, Singapore under the NRF Investigatorship programme (award no. NRF-NRFI05-2019-0003). R.D. acknowledges support from the National Natural Science Foundation of China (grant 51872256) and the Zhejiang Provincial Natural Science Foundation of China (grant LR19B010002). Computational resources were provided by the Consortium des Équipements de Calcul Intensif (CÉCI), funded by the Fonds de la Recherche Scientifique de Belgique (FRS-FNRS) under grant no. 2.5020.11, and by the Tier-1 supercomputer of the Fédération Wallonie-Bruxelles, which is infrastructure funded by the Walloon Region under grant agreement no. 1117545. L.N. acknowledges support from the Jardine Foundation. J.Z. thanks the China Scholarship Council for a PhD scholarship (no. 201503170255). S.A. acknowledges financial support from DST-UKIERI (DST/INT/UK/P-167/2017) and SERB-ECRA (ECR/2018/002056). Author contributions S.H., R.D. and A.R. designed the experiments. S.H., R.D., Z.Y. and B.Z. performed nanocrystal synthesis and film preparation. S.H., R.D., L.N., U.H., A.S. and S.A. carried out spectroscopic measurements. S.H., Q.G. and J.Z. contributed to transient absorption experiments and data analysis under the supervision of A.B. and A.R. H.T., A.P. and D.B. carried out theoretical calculations. H.X. prepared organic molecules. Z.H. prepared tetracene derivatives under the supervision of M.L.T. M.A.-J. and A.S. performed PDS measurements. S.H., R.D., X.L. and A.R. wrote the manuscript. X.L. and A.R. supervised the project. All authors discussed the results and commented on the manuscript. Competing interests The authors declare no competing interests. Additional information Supplementary information is available for this paper at https://doi.org/10.1038/s41586-0202932-2. Correspondence and requests for materials should be addressed to R.D., X.L. or A.R. Peer review information Nature thanks Jiajia Zhou and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Reprints and permissions information is available at http://www.nature.com/reprints.

Article

Antarctic ice dynamics amplified by Northern Hemisphere sea-level forcing https://doi.org/10.1038/s41586-020-2916-2

Natalya Gomez1 ✉, Michael E. Weber2, Peter U. Clark3,4, Jerry X. Mitrovica5 & Holly K. Han1

Received: 20 September 2019 Accepted: 16 September 2020 Published online: 25 November 2020 Check for updates

Sea-level rise due to ice loss in the Northern Hemisphere in response to insolation and greenhouse gas forcing is thought to have caused grounding-line retreat of marine-based sectors of the Antarctic Ice Sheet (AIS)1–3. Such interhemispheric sea-level forcing may explain the synchronous evolution of global ice sheets over ice-age cycles. Recent studies that indicate that the AIS experienced substantial millennial-scale variability during and after the last deglaciation4–7 (roughly 20,000 to 9,000 years ago) provide further evidence of this sea-level forcing. However, global sea-level change as a result of mass loss from ice sheets is strongly nonuniform, owing to gravitational, deformational and Earth rotational effects8, suggesting that the response of AIS grounding lines to Northern Hemisphere sea-level forcing is more complicated than previously modelled1,2,6. Here, using an ice-sheet model coupled to a global sea-level model, we show that AIS dynamics are amplified by Northern Hemisphere sea-level forcing. As a result of this interhemispheric interaction, a large or rapid Northern Hemisphere sea-level forcing enhances grounding-line advance and associated mass gain of the AIS during glaciation, and grounding-line retreat and mass loss during deglaciation. Relative to models without these interactions, the inclusion of Northern Hemisphere sea-level forcing in our model increases the volume of the AIS during the Last Glacial Maximum (about 26,000 to 20,000 years ago), triggers an earlier retreat of the grounding line and leads to millennial-scale variability throughout the last deglaciation. These findings are consistent with geologic reconstructions of the extent of the AIS during the Last Glacial Maximum and subsequent ice-sheet retreat, and with relative sea-level change in Antarctica3–7,9,10.

There are several mechanisms to explain near-synchronous interhemispheric climate changes on orbital timescales despite opposite insolation forcing11,12. However, synchronous changes in surface climate cannot explain synchronous changes in the Northern Hemisphere and Antarctic ice sheets3,13, because such changes would have induced opposing ice-sheet surface-mass-balance responses, with warming climate over the AIS leading to a more positive surface mass balance3. In the absence of surface melting, mechanisms that affect the primary controls on AIS mass balance (basal melting of buttressing ice shelves and ice discharge across grounding lines of marine-based sectors) are therefore required for ice-sheet synchronization. Studies have shown that an increase in subsurface warming from changes in ocean circulation contributes to AIS deglaciation4,5,14,15. Marine-based sectors of ice sheets are also vulnerable to sea-level change at their grounding lines: A local fall in sea level may slow or stabilize grounding-line retreat, or initiate or enhance grounding-line advance, whereas a rise in sea level may slow grounding-line advance, or initiate or enhance grounding-line retreat16,17. Previous work suggests that sea-level rise from deglaciation of Northern Hemisphere ice sheets triggered the retreat of the grounding lines

of the AIS and thus synchronized ice-sheet variability globally1,2,18,19. Well-dated geologic records of AIS fluctuations support synchronization on orbital timescales3,13 and identify links between periods of sea-level rise and millennial-scale AIS variability during the last deglaciation (20,000–9,000 years ago; 20–9 ka). For example, deep-sea sediments from Scotia Sea’s Iceberg Alley record eight discrete episodes of increased flux of iceberg-rafted debris originating from the AIS during the last deglaciation4. Three of these AIS discharge (AID) events occurred at the same time as well-documented periods of sea-level rise, suggesting a possible link: AID7 corresponds to the onset of deglacial sea-level rise about 19.5–19 ka3,20, AID6 corresponds to Meltwater Pulse 1A (MWP-1A) around 14.5 ka21, and AID2 corresponds to an acceleration of sea-level rise during the early Holocene, starting roughly 11.5 ka22,23. Additional evidence for this dynamic ice-sheet behaviour comes from isotopic records from a horizontal ice core in the Patriot Hills of the Weddell Sea Embayment6, which suggest that the ice surface in that region lowered by at least 600 m during AID6 and around AID2. Finally, marine records from the Ross Sea indicate a step-wise retreat of the grounding line of the West Antarctic Ice Sheet, coincident with AID events 6 and 2 (ref. 7). The iceberg-rafted debris record since AID1 (about 10.4–9 ka)

Department of Earth and Planetary Sciences, McGill University, Montreal, Quebec, Canada. 2Department of Geochemistry and Petrology, Institute for Geosciences, University of Bonn, Bonn, Germany. 3College of Earth, Ocean, and Atmospheric Sciences, Oregon State University, Corvallis, OR, USA. 4School of Geography and Environmental Sciences, University of Ulster, Coleraine, UK. 5Department of Earth and Planetary Sciences, Harvard University, Cambridge, MA, USA. ✉e-mail: [email protected] 1

600 | Nature | Vol 587 | 26 November 2020

0.5 1.0 1.5 (Dynamic AIS model)

2.0

130

Ice loss since 21 ka (km)

(ICE5G)

140

120 110 100 90 80

2.5

70

d 100

0

–100 –200 –300

Change in sea level since 21 ka (m)

b

100

0

–100 –200 –300

suggests that there has been a substantial reduction in the amplitude of iceberg-flux variability over the past 8 kyr24 and global-mean sea level has experienced only decimetre-scale changes over the past 6 kyr25. Current uncertainties in these far-field sea-level reconstructions and in the age model for Antarctic iceberg-rafted debris cores4 preclude determination of the relative phasing between the AID events and the

Iceberg-rafted-debris flux (grains yr–1 cm–3)

c

138.1

MWP-1A

4

110.5

MWP-1B

1

ICE5G 82.8 ICE6GC ANU 55.2 GLAC1D_9927 GLAC1D_9894 27.6

0 3.1

0 85.6

3

Fixed North Hemisphere ice 82.8

2.9

80.1

2.8

77.3

2.7

74.6

2.6

71.8

3 2

AID6

0.15

AID2

LGM in the Weddell Sea

0.10

AID1

0.05

0 40

35

30

25

20 Time (ka)

15

10

5

Antarctic ice volume (m.s.l.e.)

Antarctic ice volume (107 km3)

b

5

Fig. 1 | Contributions to deglacial sea-level changes in Antarctica. a, Changes in grounded ice thickness since 21 ka in the coupled ice-sheet–sea-level simulation, based on the ICE5G27 ice history in the Northern Hemisphere and the dynamic ice-sheet model in Antarctica. b, Predicted total relative sea-level change since 21 ka in Antarctica. c, Component of sea-level change in b associated with ice-cover changes in the Northern Hemisphere in the ICE5G27 ice history, computed from a simulation with only the sea-level model. d, Difference between b and c, showing changes in relative sea level associated with Antarctic ice-cover changes.

sea-level change associated with Northern Hemisphere ice melting, but modelling can provide insights into the mechanisms that led to these observed changes. Here we investigate possible interhemispheric ice-sheet coupling through sea-level change over the past 40,000 years and assess its effect on the evolution and behaviour of the AIS. We model the evolution

Northern Hemisphere ice volume (m.s.l.e.)

Northern Hemisphere ice volume (107 km3)

a

Sea-level change since 21 ka related to Northern Hemisphere ice-cover change (m)

c 0

Sea-level change since 21 ka related to Antarctic ice-cover change (m)

a

Fig. 2 | Timing of Northern Hemisphere sea-level forcing and its influence on Antarctic ice volume changes. a, Prescribed changes in Northern Hemisphere ice volume from the ICE5G27 (solid black line), ICE6GC31 (dashed black line) and ANU30 (solid cyan line) ice histories. Two composite ice histories, in which ice cover over North America and Greenland in ICE5G has been replaced by GLAC1D29 regional ice histories (blue solid and dashed lines), are also shown. Blue vertical bands indicate the timings of MWP-1A 21 and MWP-1B22. b, Changes in Antarctic ice volume predicted in simulations with fixed (red line) and evolving (blue, cyan and black lines; ice histories as in a) Northern Hemisphere ice. Green vertical bands indicate periods during MWP-1A and the early Holocene in the model simulations that are examined in the text and other figures. c, Iceberg-rafted-debris flux from sediment cores in Iceberg Alley4 (Methods, Extended Data Fig. 9). Vertical red bands indicate the timing of AID events 1, 2 and 6; horizontal blue bands indicate the extent of the LGM in the Weddell Sea3. m.s.l.e. indicates metres sea-level equivalent.

0

Nature | Vol 587 | 26 November 2020 | 601

Article Antarctic ice flux (107 km3 yr–1)

a

0.20 Fixed Northern Hemisphere ice Evolving Northern Hemisphere ice

0.15 0.10 0.05 0 –0.05 15

14

12

11

10

14.5–13.5 ka

9

8 7 Time (ka)

6

d

5

4

3

2

1

0

11.5–9.0 ka

Evolving Northern Hemisphere ice

b

13

e

Fixed Northern Hemisphere ice

c

150 m in previously glaciated regions in Antarctica, while the bedrock deepens in the interior of Antarctica and sea level rises by 150 m in the surrounding ocean (Fig. 1b). Mass loss from Northern Hemisphere ice sheets contributes a sea-level rise that increases from 80 m in East Antarctica to 130 m in West Antarctica (Fig. 1c), in agreement with previous work3. This sea-level gradient is driven by a shift of the Earth’s rotation axis towards North America in the Northern Hemisphere, where most ice is being lost (Fig. 1a), and towards East Antarctica in the Southern Hemisphere, driving lower-than-average sea-level rise in these regions and higher-than-average sea-level rise in the opposing quadrants of the Earth’s surface (which include West Antarctica and Eurasia). By contrast, mass loss from the AIS drives a sea-level fall of up to 300 m in previously glaciated regions of Antarctica, owing to gravitationally driven lowering of the sea surface and viscoelastic uplift of the solid Earth under the areas of ice mass loss (Fig. 1d). The sea-level fall associated with local AIS loss thus dominates the total Antarctic signal in these areas. However, sea-level rise due to the much larger Northern Hemisphere ice-mass loss substantially decreases the geographic spread and magnitude of sea-level fall at Antarctic grounding lines over the last deglaciation (compare Figs. 1b and d).

AIS response to Northern Hemisphere sea-level forcing

0

–100 –200 –300 –400 –500 –600 Change in Antarctic ice thickness (m)

Fig. 3 | Enhanced Antarctic ice loss during MWP-1A and the early Holocene. a, Rate of change of Antarctic ice volume, including grounded and floating ice, calculated as a 100-year running mean, predicted from simulations including (black line) and excluding (red line) Northern Hemisphere ice-cover changes. The black line and shading represent the mean and standard deviation of predictions generated with the five ice histories described in the text. b–e, Change in Antarctic ice thickness during MWP-1A (14.5–13.5 ka; b, c) and during the early Holocene, starting from the time of MWP-1B (11.5–9 ka; d, e). Panels b and d are based on simulations that include ice-mass flux from the Northern Hemisphere from the ICE5G ice history. These simulations predict Antarctic ice loss equivalent to global-mean sea-level rise of 1.14 m and 1.95 m for MWP-1A (b) and the early Holocene (d), respectively. Panels c and e are based on simulations in which Northern Hemisphere ice sheets remain fixed throughout the simulation. The grey–blue and black lines indicate the grounding-line position at the start and end of the time interval, respectively.

of the AIS and global sea-level changes using the Pennsylvania State University three-dimensional ice-sheet model coupled to a gravitationally self-consistent global sea-level model that includes viscoelastic deformation of the solid Earth, rotational feedbacks and migrating shorelines26 (see Methods).

Deglacial sea-level change in Antarctica Figure 1a shows ice loss since 21 ka from the AIS, as calculated in a coupled ice-sheet–sea-level model simulation, and from Northern Hemisphere ice sheets, derived from the ICE5G ice history27. The AIS simulation is based on parameters identified in a large ensemble analysis as best fitting a range of palaeo and modern data constraints28, and is characterized by a global-mean sea-level-equivalent mass loss of 5 m from the AIS, with 107 m from the Northern Hemisphere in ICE5G (see Methods and Extended Data Fig. 1 for results of simulations with a larger AIS contribution to global-mean sea level). Peak sea-level fall is predicted to reach 602 | Nature | Vol 587 | 26 November 2020

We quantify how sea-level changes associated with Northern Hemisphere ice-sheet variations (Figs. 1c, 2a) influenced the dynamics of the AIS leading up to the Last Glacial Maximum (LGM; around 26–20 ka13) and during the last deglaciation. We compare AIS mass changes predicted from: (1) model simulations that include the evolution of the Northern Hemisphere ice sheets prescribed from five ice-history reconstructions27,29–31; and (2) a simulation in which the Northern Hemisphere ice sheets remain fixed in their initial configuration at 40 ka and do not contribute to sea-level changes in Antarctica over the simulation (Fig. 2b, Extended Data Fig. 8, Methods). In the simulations, Northern Hemisphere ice growth leading up to the LGM contributes a sea-level fall beginning around 30 ka (Fig. 2a), which drives additional AIS growth at about 28 ka (Fig. 2b). This growth occurs primarily in the Antarctic Peninsula and Weddell Sea regions (Extended Data Figs. 2c, 3, at 20 ka; Extended Data Fig. 4), and is consistent with evidence for when ice reached its LGM extent in these regions3. During the last deglaciation, sea-level rise from Northern Hemisphere ice-sheet retreat (Fig. 2a) greatly enhances the magnitude and rate of AIS mass loss since 15 ka (Fig. 2b, Extended Data Figs. 2, 3). By contrast, the simulation with no Northern Hemisphere sea-level forcing is characterized by net AIS growth over much of the same period (Fig. 2b). In particular, with fixed Northern Hemisphere ice, extensive grounded ice remains in the Weddell Sea and, to a lesser extent, the Ross Sea regions until the present day (Extended Data Figs. 2, 3), whereas including a Northern Hemisphere sea-level forcing causes these regions to completely deglaciate, reaching a modern Antarctic ice volume close to the observed volume (2.69 × 107 km3; ref. 32). Differences in AIS evolution are greatest in the Weddell Sea region (Extended Data Fig. 4), where the largest sea-level forcing is predicted from Northern Hemisphere ice-mass loss (Fig. 1c). We have highlighted the sensitivity of the ice sheet to this geographic variability in Extended Data Fig. 7.

MWP-1A and early Holocene ice loss Our simulations suggest that an increase in AIS mass loss since 15 ka, driven by Northern Hemisphere sea-level forcing, contributed to MWP-1A (around 14.5 ka) and support an Antarctic source for early-Holocene acceleration in sea-level rise33 (Fig. 2). Specifically, the AIS simulations that include a Northern Hemisphere sea-level forcing during the deglaciation show distinct corresponding periods of rapid mass loss (Figs. 2b, 3a) during and after these two episodes of rapid sea-level rise. This behaviour may therefore explain the large

4.0 3.5 3.0

10 ka

9 ka

2.0 1.5 1.0

Ice thickness (km)

2.5

c

0.5 0 Ice shelf Site S Site 1

1.0

AID2

Site S, ICE5G Site S, ANU Site S, fixed Northern Hemisphere ice

AID1

0.25

0.8

0.20

0.6

0.15 0.4

0.10

0.2

0.05

Iceberg-rafted-debris flux (grains yr−1 cm−3)

b

11 ka

Ice thickness above present-day (km)

12 ka

0

0 200

Sea-level change (m)

a

Site 1, cosmogenic data Site 1, ICE5G Site 1, fixed Northern Hemisphere ice

Site S, ICE5G, global-mean Site S, ANU, global-mean

150 100 50 0 –50 20

18

16

14

12

10

8

6

4

2

0

Time (ka)

Fig. 4 | Agreement of predicted sea-level and ice-cover changes with geological records in the Ross Sea sector. a, Predicted Antarctic ice cover in the simulation that includes Northern Hemisphere ice from ICE5G27 at snapshots in time as the Ross Sea region deglaciates. b, Predicted ice thickness (left axis) above the present-day thickness at site S (red squares in a) and site 1 (blue squares in a; ref. 35), for simulations in which Northern Hemisphere ice cover is evolving according to ICE5G27 or ANU30 ice histories (black and cyan lines for site S, solid blue line for site 1) or fixed (red line for site S, dashed blue line for site 1). Error bars show cosmogenic exposure age data with 2σ

uncertainty from ref. 35 at site 1. The grey time series is the recorded flux of Antarctic iceberg-rafted debris from Fig. 2c (right axis); red vertical bands indicate AID events 1 and 2. c, Predictions of relative sea-level change at site S for evolving (solid black and cyan lines) and fixed (solid red line) Northern Hemisphere ice cover. Dashed black and cyan lines indicate the contribution to global-mean sea-level change from Northern Hemisphere ice-cover prescribed in ICE5G27 and ANU30 ice histories, respectively. Black markers with 2σ error bars show two-way (circles) or lower-bound (triangles) relative sea-level constraints9.

increases in iceberg-rafted-debris flux in Iceberg Alley (AID events 6, 2 and 1; ref. 4; Fig. 2c) and mass loss from the Weddell Sea6 and Ross Sea7 regions at these times. Alternative physical processes (such as ocean and atmospheric forcing) must be sought to explain the other AIS discharge events during the last deglaciation4. In the case of MWP-1A, the AIS experiences more extensive mass loss (Figs. 2b, 3a, Extended Data Fig. 5) in the simulations in which Northern Hemisphere ice sheets evolve than in the simulation in which they are fixed and AIS mass loss is driven only by climate forcing on the ice sheet14 (Fig. 3a–c). The net volume of ice lost in the latter simulation over this period is 2.5–3 times less than predicted in the former across the range of evolving Northern Hemisphere ice histories we consider. In the case of the early Holocene, including Northern Hemisphere sea-level forcing increases the rate of AIS mass loss by up to a factor of about 4, starting around 11.5 ka (Fig. 3a)—the time of MWP-1B suggested by the far-field Barbados sea-level record23 and AID2 (ref. 4). This mass loss continues throughout AID1 until 9.5–9 ka (Figs. 2, 3, Extended data Fig. 5), with substantial grounding-line retreat in the Ross Sea and Weddell Sea regions (compare Fig. 3d and Fig. 3e). The amplified AIS response during the early Holocene occurs regardless of whether there is an acceleration in Northern Hemisphere ice loss during that time (compare Extended Data Fig. 5a with ICE5G27 and Extended Data Fig. 5b, c with ICE6G31 and ANU30). It is also consistent with the hypothesis of a substantial or dominant Antarctic source for acceleration in global-mean sea-level rise31,33 during this period (Extended Data Fig. 6). We have confirmed that the general behaviour evident in Figs. 2 and 3 holds for coupled model simulations using a range of Earth and ice-model parameters (Extended Data Figs. 1, 8). The largest difference is found in simulations with less basal sliding. These simulations result in a larger LGM ice sheet (Extended Data Fig. 1), with AIS growth in the

first half of the simulations being less sensitive to Northern Hemisphere ice growth and the associated sea-level fall than simulations with more basal sliding. The concordance between the two simulations with less basal sliding is probably because the AIS margin nearly reaches its maximum possible extent at the continental shelf edge in both cases.

Comparison to geologic records We next consider local ice-sheet and sea-level changes in the Ross Sea region where there are reconstructions of relative sea level9,34 (Southern Scott Coast; site S in Fig. 4a), grounding-line migration7 and changes in ice-surface elevation (site 1 in Fig. 4; other sites in the Ross Sea and Weddell Sea regions are shown in Extended Data Figs. 10, 11)35 that provide a test of the local response to Northern Hemisphere sea-level forcing. When this forcing is included, deglaciation of the region is predicted to begin in the early Holocene (Fig. 4a), with regional ice thinning occurring from 11 ka to 8 ka (blue, black and cyan curves show thinning at sites S and 1 in Fig. 4b). This model result is consistent with observations of ice-surface lowering at site 1 (ref. 35; Fig. 4b, error bars; nearby sites 3–5, which are just outside the region of substantial ice thinning, are discussed in Extended Data Fig. 11), a grounding-line retreat of at least 200 km in the Ross Sea7 and a peak iceberg-rafted-debris flux during AID events 2 and 1 (ref. 4). By contrast, this retreat takes place from 10 ka to 6 ka at sites S and 1 when Northern Hemisphere sea-level forcing is not included in the simulation (Fig. 4b), which is inconsistent with a relatively low observed iceberg-rafted-debris flux since about 8.5 ka24 (Fig. 4b, grey line). At site S, local sea-level change during the deglacial phase is initially dominated by Northern-Hemisphere-driven sea-level rise. However, as local ice loss begins (about 11 ka), the associated gravitational and Nature | Vol 587 | 26 November 2020 | 603

Article deformational effects dominate the local sea-level change (Fig. 4c). When the Northern-Hemisphere-driven sea-level forcing is excluded (Fig. 4c), relative sea-level fall is predicted to begin later (around 9 ka), coincident with local ice loss for this simulation (Fig. 4b); there is almost no change in relative sea level before 10 ka. The oldest relative sea-level indicators from this area (about 6.5 ka; Fig. 4c) are more consistent with the lower relative sea level predicted by the simulations that include Northern Hemisphere sea-level forcing. However, older indicators are needed to clearly corroborate this, especially given the uncertainty in the local viscoelastic structure of the Earth in this region.

Summary We conclude that geographically variable sea-level changes around Antarctica, driven by Northern Hemisphere ice-sheet changes, strongly modulated the growth and decay of the AIS. Northern Hemisphere ice growth leading up to the LGM causes local sea-level fall and further AIS growth in our simulations, yielding a higher peak AIS volume at the LGM than without this forcing. Conversely, in our model, Northern Hemisphere ice loss during the last deglaciation produces a sea-level rise of 80–130 m in Antarctica, which drives earlier, greater and more rapid AIS retreat that is in better agreement with geologic evidence than predictions that omit this forcing. The simulations indicate that the Weddell Sea region of the AIS was subject to the largest sea-level changes driven by Northern Hemisphere ice changes, suggesting that ice-mass changes in this region were particularly sensitive to this far-field sea-level forcing. Finally, simulations with Northern Hemisphere sea-level forcing predict increases in AIS mass flux during MWP-1A and the early Holocene, consistent with multiple lines of geologic evidence for AIS mass loss at these times.

Online content

9.

10. 11. 12. 13. 14.

15. 16. 17. 18. 19.

20.

21. 22.

23.

24.

25. 26. 27.

Any methods, additional references, Nature Research reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at https://doi.org/10.1038/s41586-020-2916-2.

28.

1.

31.

2.

3. 4. 5. 6. 7.

8.

Denton, G. H. & Hughes, T. J. Milankovitch theory of ice ages: hypothesis of ice-sheet linkage between regional insolation and global climate. Quat. Res. 20, 125–144 (1983). Huybrechts, P. J. Sea-level changes at the LGM from ice-dynamic reconstructions of the Greenland and Antarctic ice sheets during the glacial cycles. Quat. Sci. Rev. 21, 203–231 (2002). Weber, M. E. et al. Interhemispheric ice-sheet synchronicity during the Last Glacial Maximum. Science 334, 1265–1269 (2011). Weber, M. E. et al. Millennial-scale variability in Antarctic ice-sheet discharge during the last deglaciation. Nature 510, 134–138 (2014). Golledge, N. R. et al. Antarctic contribution to meltwater pulse 1A from reduced Southern Ocean overturning. Nat. Commun. 5, 5107 (2014). Fogwill, C. J. et al. Antarctic ice sheet discharge driven by atmosphere–ocean feedbacks at the Last Glacial Termination. Sci. Rep. 7, 39979 (2017). Bart, P. J., DeCesare, M., Rosenheim, B. E., Majewski, W. & McGlannan, A. A centuries-long delay between a paleo-ice-shelf collapse and grounding-line retreat in the Whales Deep Basin, eastern Ross Sea, Antarctica. Sci. Rep. 8, 12392 (2018). Milne, G. A. & Mitrovica, J. X. Searching for eustasy in deglacial sea-level histories. Quat. Sci. Rev. 27, 2292–2302 (2008).

604 | Nature | Vol 587 | 26 November 2020

29.

30.

32. 33.

34. 35.

Hall, B. L. & Denton, G. H. New relative sea-level curves for the southern Scott Coast, Antarctica: evidence for Holocene deglaciation of the western Ross Sea. J. Quat. Sci. 14, 641–650 (1999). Fogwill, C. J. et al. Southern Ocean carbon sink enhanced by sea-ice feedbacks at the Antarctic Cold Reversal. Nat. Geosci. 13, 489–497 (2020). Kawamura, K. et al. Northern Hemisphere forcing of climatic cycles in Antarctica over the past 360,000 years. Nature 448, 912–916 (2007). Huybers, P. & Denton, G. Antarctic temperature at orbital timescales controlled by local summer duration. Nat. Geosci. 1, 787–792 (2008). Clark, P. U. et al. The last glacial maximum. Science 325, 710–714 (2009). Pollard, D., Chang, W., Haran, M., Applegate, P. & DeConto, R. Large ensemble modeling of the last deglacial retreat of the West Antarctic Ice Sheet: comparison of simple and advanced statistical techniques. Geosci. Model Dev. 9, 1697–1723 (2016). Clark, P. U. et al. Oceanic forcing of penultimate deglacial and last interglacial sea-level rise. Nature 577, 660–664 (2020). Weertman, J. Stability of the junction of an ice sheet and an ice shelf. J. Glaciol. 13, 3–11 (1974). Schoof, C. Ice sheet grounding line dynamics: steady states, stability, and hysteresis. J. Geophys. Res. 112, F03S28 (2007). Denton, G. H., Hughes, T. J. & Karlén, W. Global ice-sheet system interlocked by sea level. Quat. Res. 26, 3–26 (1986). Tigchelaar, M., Timmermann, A., Friedrich, T., Heinemann, M. & Pollard, D. Nonlinear response of the Antarctic Ice Sheet to late Quaternary sea level and climate forcing. Cryosphere 13, 2615–2631 (2019). Yokoyama, Y., Lambeck, K., De Deckker, P., Johnston, P. & Fifield, L. K. Timing of the Last Glacial Maximum from observed sea-level minima. Nature 406, 713–716 (2000); corrigendum 412, 99 (2001). Deschamps, P. et al. Ice-sheet collapse and sea-level rise at the Bølling warming 14,600 years ago. Nature 483, 559–564 (2012). Bard, E., Hamelin, B. & Delanghe-Sabatier, D. Deglacial meltwater pulse 1B and Younger Dryas sea levels revisited with boreholes at Tahiti. Science 327, 1235–1237 (2010). Abdul, N. A., Mortlock, R. A., Wright, J. D. & Fairbanks, R. G. Younger Dryas sea level and meltwater pulse 1B recorded in Barbados reef crest coral Acropora palmata. Paleoceanography 31, 330–344 (2016). Bakker, P., Clark, P. U., Golledge, N. R., Schmittner, A. & Weber, M. E. Centennial-scale Holocene climate variations amplified by Antarctic Ice Sheet discharge. Nature 541, 72–76 (2017). Hallmann, N. et al. Ice volume and climate changes from a 6000 year sea-level record in French Polynesia. Nat. Commun. 9, 285 (2018). Gomez, N., Pollard, D. & Mitrovica, J. X. A 3-D coupled ice sheet – sea level model applied to Antarctica through the last 40 ky. Earth Planet. Sci. Lett. 384, 88–99 (2013). Peltier, W. R. Global glacial isostasy and the surface of the ice-age Earth: the ICE-5G (VM2) model and GRACE. Annu. Rev. Earth Planet. Sci. 32, 111–149 (2004). Pollard, D., Gomez, N. & DeConto, R. M. Variations of the Antarctic Ice Sheet in a coupled ice sheet-Earth-sea level model: sensitivity to viscoelastic Earth properties. J. Geophys. Res. Earth Surf. 122, 2124–2138 (2017). Tarasov, L., Dyke, A. S., Neal, R. M. & Peltier, W. R. A data-calibrated distribution of deglacial chronologies for the North American ice complex from glaciological modeling. Earth Planet. Sci. Lett. 315–316, 30–40 (2012). Lambeck, K., Rouby, H., Purcell, A., Sun, Y. & Sambridge, M. Sea level and global ice volumes from the Last Glacial Maximum to the Holocene. Proc. Natl Acad. Sci. USA 111, 15296–15303 (2014). Peltier, W. R., Argus, D. F. & Drummond, R. Space geodesy constrains ice age terminal deglaciation: the global ICE-6G_C (VM5a) model. J. Geophys. Res. Solid Earth 120, 450–487 (2015). Fretwell, P. et al. Bedmap2: improved ice bed, surface and thickness datasets for Antarctica. Cryosphere 7, 375–393 (2013). Bard, E., Hamelin, B., Deschamps, P. & Camoin, G. Comment on “Younger Dryas sea level and meltwater pulse 1B recorded in Barbados reefal crest coral Acropora palmata” by N. A. Abdul et al. Paleoceanography 31, 1603–1608 (2016). Briggs, R. D. & Tarasov, L. How to evaluate model-derived deglaciation chronologies: a case study using Antarctica. Quat. Sci. Rev. 63, 109–127 (2013). Jones, R., Whitehouse, P., Bentley, M., Small, D. & Dalton, A. Impact of glacial isostatic adjustment on cosmogenic surface-exposure dating. Quat. Sci. Rev. 212, 206–212 (2019).

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. © The Author(s), under exclusive licence to Springer Nature Limited 2020

Methods Coupled ice-sheet–sea-level modelling The evolution of the AIS and global sea-level changes were modelled using the coupled ice-sheet–sea-level model developed and described in detail in ref. 26 and applied in refs. 28,36. The model consists of the Pennsylvania State University (PSU) 3D ice-sheet model37 coupled to a gravitationally self-consistent global sea-level model that includes viscoelastic deformation of the solid Earth, rotational feedbacks onto sea level and migrating shorelines38,39. PSU 3D is a finite-difference, ice-sheet–ice-shelf model that adopts hybrid combinations of the scaled shallow-ice and shallow-shelf equations40,41 to treat ice dynamics, and includes grounding-line migration through a parameterization of flux across the grounding line17. This grounding-line treatment performs reasonably well in comparison with higher-order ice-sheet models42 and facilitates the computational feasibility of glacial–interglacial timescale simulations. Basal sliding in the model is treated with a standard Weertman-type sliding law and basal sliding coefficients are determined through inverse fitting to ice thickness under regions with modern grounded ice43. In modern oceanic regions, where basal sliding is relatively unconstrained, the coefficient is set to 10−5 m yr−1 Pa−2 in simulations presented in the main text and to 10−6 m yr−1 Pa−2 in additional simulations summarized in Extended Data Fig. 1, representing the range identified28 to best fit a suite of palaeo ice-sheet and sea-level constraints. Alternative basal sliding laws have been proposed44,45, but model treatments remain unverified by observations and should be explored further in future work. Other ice model parameters such as the calving coefficient and ocean melt factor are similarly set to best-fitting values identified in refs. 14,28. Here and in these references, atmospheric climate forcing was applied by perturbing modern climatology (ALBMAP46) to mimic past conditions according to a deep-sea δ18O stack47. Sub-ice-shelf melt rates are determined through a parameterization that depends on subsurface oceanic temperatures from ref. 48, with sensitivity inferred from the aforementioned large ensemble model–data comparisons28. We note that the best-fitting simulations in these studies produce a relatively small contribution from Antarctica to sea level over the last deglaciation. We therefore explore simulations with a larger excess ice volume at the LGM and greater ice loss through the deglaciation in Extended Data Fig. 1. Northern Hemisphere (NH) ice-cover variations were prescribed in the global sea-level model using five different ice histories: three global ice histories (ICE5G27, ICE6GC31 and the ANU model30) and two histories with GLAC1D reconstructions29 over North America and Greenland and ICE5G elsewhere. These ice-history reconstructions are widely used, cover the whole time period under consideration and are constrained by glacial isostatic adjustment modelling and a suite of sea-level and ice-cover records. The elastic and density structure of the solid Earth in the sea-level model are prescribed from the preliminary reference Earth model (PREM)49. Two different models of the viscosity structure of the Earth’s mantle are adopted in the simulations. Figs. 1–4 show results using the ‘LVZ’ model, which is representative of structure beneath the West Antarctic. This model is characterized by a 50-km-thick lithosphere, a low-viscosity zone of 1019 Pa s extending from the base of the lithosphere to a depth of 200 km, and a viscosity of 2 × 1020 Pa s in the upper mantle and 3 × 1021 Pa s in the lower mantle. The LVZ model was adopted in refs. 28,36. Additional simulations in Extended Data Fig. 1 adopt a viscosity structure that falls within a range of models that best fit a suite of global observations related to glacial isostatic adjustment30,50. The model, labelled ‘HV’, is characterized by a lithospheric thickness of 120 km and upper and lower mantle viscosities of 5 × 1020 Pa s and 5 × 1021 Pa s, respectively. The ice-sheet model is run on a polar stereographic projection with a 20-km grid resolution, while sea-level calculations are performed up

to spherical harmonic degree 512. To couple the models, the ice-sheet model first computes changes in Antarctic ice thickness over a 200-year ‘coupling interval’ (sensitivity tests described in ref. 26 show that this choice is sufficiently short for ice-age simulations). These AIS changes are then combined with NH ice cover—which is either fixed at its configuration at 40 ka throughout the run for ‘fixed NH ice’ simulations or evolves according to the chosen ice history in ‘evolving NH ice’ simulations (Extended Data Fig. 8)—over this interval. The combination is then used as input to the sea-level model to compute the associated global changes in sea level. The predicted changes in sea level, which are equivalent to the negative of changes in topography or bedrock elevation, are passed back to the ice-sheet model and used to update bedrock topography in Antarctica. The ice-sheet model then proceeds forward across another coupling interval and the process repeats over the full 40-kyr simulation. Initial conditions of the ice sheet at 40 ka are provided by a longer, full-glacial-cycle run of the ice-sheet model, along with bedrock deformation given by a simpler, elastic-lithosphere relaxed-asthenosphere (ELRA) model51. Global topography and bedrock elevation in Antarctica at the start of the simulation (40 ka) are initially unknown. They are determined through an iterative procedure, whereby the predicted modern topography at the end of a 40-kyr simulation is compared to observed topography (ETOPO252 globally and Bedmap232 in Antarctica), and the difference between the two is used to correct the initial topography at 40 ka in the next iteration. The process is repeated four times, which guarantees sufficient convergence of predicted and observed modern topography.

Iceberg Alley sites and iceberg-rafted debris (IBRD) record Sample-based investigations have concentrated on deep-sea cores retrieved in the Scotia Sea’s Iceberg Alley during Marion Dufresne II cruise 160 in March 2007. Sites MD07-3133 (57° 26′ S, 43° 27′ W; 3,101 m water depth; 32.8 m long) and MD07-3134 (59° 25′ S, 41° 28′ W; 3,663 m water depth; 58.2 m long) originate from the northern end of Dove Basin and Pirie Bank, respectively. The age models of sites MD07-3133 and MD07-3134 are based on distinct dust–climate couplings between Southern Ocean sediment and the Antarctic EPICA Dronning Maud Land (EDML) ice core53 on the EDML1 age model54. The EDML1 age model appears more consistent with local ash-layer correlations than the later AICC 2012 age model, which relies on interhemispheric methane correlation55,56. Comparison between age models (Extended Data Fig. 9c) shows older ages for the AICC 2012 age scale. Differences are minimal for the mid to late Holocene. At the time of MWP-1B, the difference is roughly 150 years, whereas it is 350 years during MWP-1A and about 500 years at the LGM. This means that for AIS discharge (AID) event 2 the shift is very small and the event aligns well with MWP-1B regardless of the age model. AID event 6 would extend from about 14.3 ka to 15.2 ka in AICC 2012, a range that still encompasses MWP-1A (14.65–14.3 ka), especially within the uncertainties of the ice-core age models, which increase from a few centuries for the time of MWP-1B up to a millennium for the LGM55,56. Therefore, the correlations we make here and the conclusions we draw hold regardless of the age model applied. Magnetic susceptibility and Ca and Fe records measured through X-ray fluorescence can be used to study coherent and synchronous changes in dust deposition across much of the Southern Ocean and the AIS across the last deglaciation4, and on longer, glacial–interglacial timescales57,58. IBRD counting was conducted every 1 cm on X-radiographs taken from 1-cm-thick slices cut out from the centre of each core segment and exposed to an X-ray system. The transitions from low to high and high to low IBRD contents form the basis of our AID event classification. The counting interval of 1 cm translates to 8–17-year resolution for AID events 1–7, depending on the time interval and core. The IBRD data presented here correspond to a stack of sites MD07-3133 and MD07-3134, to obtain a regional rather than local record for 20–0 ka. This stack is combined with previous data for the periods

Article 27–7 ka4 and 8–0 ka24. This new IBRD stack, the EDML1 and AICC 2012 age models and uncertainty calculations are shown in Extended Data Fig. 9. Three tables containing the IBRD data on both age scales, the age-scale tie points and the uncertainty calculations are available from the PANGAEA data server (see ‘Data availability’).

Comparison to cosmogenic exposure age data In Fig. 4 and Extended Data Figs. 10, 11, we compare modelled ice-loss history in the Ross Sea and Weddell Sea regions to records of ice thinning from cosmogenic exposure age data35 (also discussed in ref. 59). Records at site 1 in the Ross Sea (Fig. 4) and at sites 11–15 in the Weddell Sea region (Extended Data Fig. 10) are consistent with the earlier deglaciation predicted in simulations that include a NH sea-level forcing. While exposure age data at sites 3–5 (Extended Data Fig. 11a) appear to be more consistent with a later deglaciation, these sites are just outside the region of substantial ice loss in the simulation (Extended Data Fig. 11d, e). The sites are geographically close to site S, but they exhibit substantially less thinning than the rest of the Ross Sea region (compare ice loss at red and blue squares in Extended Data Fig. 11e). On the other hand, thinning rates and timing at sites 1 and S are comparable (Extended Data Fig. 11, Fig. 4b), and more representative of regional-scale ice loss that occurs in the model during the main part of the deglaciation. Taken into consideration with the Ross Sea grounding-line record7, it is possible that major deglaciation ends in this sector of the Ross Sea by around 8 ka, and the thinning rates observed at sites 3–5 indicate smaller-magnitude, late-Holocene ice changes. Higher-resolution ice-sheet modelling would be needed to investigate this issue further, which is infeasible in the long-timescale coupled model simulations described here.

Data availability The datasets generated for this publication are available in the PANGAEA database (https://doi.org/10.1594/PANGAEA.919498) and as source data for Extended Data Fig. 9. The modelling results are available in the OSF database (https://osf.io/g5ur2/?view_only=8acbf1e38c 184d9c8f09811c8bbef036). Source data are provided with this paper.

Code availability The coupled ice-sheet–sea-level model used is reported in refs. 26,28; the PSU 3D ice-sheet model is reported in ref. 37. Ice-sheet and sea-level models are available on request from the authors of the references listed. 36. Gomez, N., Pollard, D. & Holland, D. Sea-level feedback lowers projections of future Antarctic Ice Sheet mass loss. Nat. Commun. 6, 8798 (2015). 37. Pollard, D. & DeConto, R. M. Description of a hybrid ice sheet-shelf model, and application to Antarctica. Geosci. Model Dev. 5, 1273–1295 (2012). 38. Kendall, R. A., Mitrovica, J. X. & Milne, G. A. On post-glacial sea level – II. Numerical formulation and comparative results on spherically symmetric models. Geophys. J. Int. 161, 679–706 (2005). 39. Gomez, N., Mitrovica, J. X., Tamisiea, M. E. & Clark, P. U. A new projection of sea level change in response to collapse of marine sectors of the Antarctic Ice Sheet. Geophys. J. Int. 180, 623–634 (2010).

40. MacAyeal, D. R. Large-scale ice flow over a viscous basal sediment: theory and application to ice stream B, Antarctica. J. Geophys. Res. Solid Earth 94, 4071–4087 (1989). 41. Andrews, J. T. & Mahaffy, M. A. W. Growth rate of the Laurentide Ice Sheet and sea level lowering (with emphasis on the 115 000 BP sea level low). Quat. Res., 6, 167–183 (1976). 42. Pattyn, F. et al. Grounding-line migration in plan-view marine ice-sheet models: results of the ice2sea MISMIP3d intercomparison. J. Glaciol. 59, 410–422 (2013). 43. Pollard, D. & DeConto, R. M. A simple inverse method for the distribution of basal sliding coefficients under ice sheets, applied to Antarctica. Cryosphere 6, 953–971 (2012). 44. Pattyn, F. Sea-level response to melting of Antarctic ice shelves on multi-centennial time scales with the fast elementary thermomechanical ice sheet model (f.ETISh v1.0). Cryosphere 11, 1851–1878 (2017). 45. Tsai, V. C., Stewart, A. L. & Thompson, A. F. Marine ice-sheet profiles and stability under Coulomb basal conditions. J. Glaciol. 61, 205–215 (2015). 46. Le Brocq, A. M., Payne, A. J. & Vieli, A. An improved Antarctic dataset for high resolution numerical ice sheet models (ALBMAP v1). Earth Syst. Sci. Data 2, 247–260 (2010). 47. Lisiecki, L. E. & Raymo, M. E. A Pliocene-Pleistocene stack of 57 globally distributed benthic δ18O records. Paleoceanography 20, PA1003 (2005); correction 20, PA2007 (2005). 48. Liu, Z. et al. Transient simulation of last deglaciation with a new mechanism for Bolling-Allerod warming. Science 325, 310–314 (2009). 49. Dziewonski, A. M. & Anderson, D. L. Preliminary reference Earth model. Phys. Earth Planet. Inter. 25, 297–356 (1981). 50. Mitrovica, J. X. & Forte, A. M. A new inference of mantle viscosity based upon joint inversion of convection and glacial isostatic adjustment data. Earth Planet. Sci. Lett. 225, 177–189 (2004). 51. Huybrechts, P. & de Wolde, J. The dynamic response of the Greenland and Antarctic ice sheets to multiple-century climatic warming. J. Clim. 12, 2169–2188 (1999). 52. National Geophysical Data Center. 2-minute Gridded Global Relief Data (ETOPO2) v2 (NOAA, 2006); https://data.nodc.noaa.gov/cgi-bin/iso?id=gov.noaa.ngdc.mgg.dem:301. 53. Weber, M. E. et al. Dust transport from Patagonia to Antarctica – a new stratigraphic approach from the Scotia Sea and its implications for the last glacial cycle. Quat. Sci. Rev. 36, 177–188 (2012). 54. Ruth, U. et al. “EDML1”: a chronology for the EPICA deep ice core from Dronning Maud Land, Antarctica, over the last 150 000 years. Clim. Past 3, 475–484 (2007). 55. Veres, D. et al. The Antarctic ice core chronology (AICC2012): an optimized multi-parameter and multi-site dating approach for the last 120 thousand years. Clim. Past 9, 1733–1748 (2013). 56. Bazin, L. et al. An optimized multi-proxy, multi-site Antarctic ice and gas orbital chronology (AICC2012): 120–800 ka. Clim. Past 9, 1715–1731 (2013). 57. Lamy, F. et al. Increased dust deposition in the Pacific Southern Ocean during glacial periods. Science 343, 403–407 (2014). 58. Martínez-García, A. et al. Iron fertilization of the subantarctic ocean during the last ice age. Science 343, 1347–1350 (2014). 59. Small, D., Bentley, M. J., Jones, R. S., Pittard, M. L. & Whitehouse, P. L. Antarctic ice sheet palaeo-thinning rates from vertical transects of cosmogenic exposure ages. Quat. Sci. Rev. 206, 65–80 (2019). Acknowledgements N.G. and H.K.H. were supported by the Natural Sciences and Engineering Research Council (NSERC), the Canada Research Chair’s programme and the Canadian Foundation for Innovation, M.E.W. by the Deutsche Forschungsgemeinschaft (DFG; grant numbers We2039/8-1 and We 2039/17-1), and J.X.M. by NASA grant NNX17AE17G and Harvard University. We thank G. Tseng for assistance with exploratory research that informed this study, and D. Pollard for insight on and use of the PSU ice-sheet model. Author contributions N.G. contributed the numerical modelling and analysis; H.K.H. prepared model input; M.E.W. contributed iceberg-rafted debris records and, together with P.U.C. and J.X.M., other published data and related discussion. All authors contributed to developing the idea and to writing and refining the manuscript. Competing interests The authors declare no competing interests. Additional information Correspondence and requests for materials should be addressed to N.G. Peer review information Nature thanks Frank Pattyn and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Reprints and permissions information is available at http://www.nature.com/reprints.

3.4 3.3

Antarctic ice volume (107 km3 )

3.2

Low viscosity mantle, thin lithosphere evolving NH, b=10 -5 m a -1 Pa-2

MWP WP A 1A

evolving NH, b=10 -6 m a -1 Pa-2 -5

fixed NH, b=10 m a

-1

Pa

100

-2

fixed NH, , b=10 -6 m a -1 Pa-2 NH gmsl contribution ICE5G

W MWP B 1B

80

60

3.1 40

3

20

2.9 2.8

0

N.Hemisphere gmsl contribution (m)

a

2.7

35

30

b

25

20

time (ka)

15

10

5

0

High viscosity mantle, thick lithosphere

3.4

AID6

3.3

100

AID1 ID2 AID2

80

Antarctic ice volume (107 km3 )

3.2

60

3.1 40

3

20

2.9 2.8

0

N.Hemisphere gmsl contribution (m)

2.6 40

2.7 2.6 40

35

30

25

20

time (ka)

Extended Data Fig. 1 | Sensitivity of results to ice and Earth model parameters. a, Changes in Antarctic ice volume predicted in simulations with evolving (solid lines) and fixed (dashed lines) Northern Hemisphere ice mass, and adopting the LVZ Earth model (Methods). Blue lines are identical to those in Fig. 2b, corresponding to a basal sliding coefficient of b = 10 −5 m yr−1 Pa−2; red lines correspond to a basal sliding coefficient of b = 10 −6 m yr−1 Pa−2. The black

15

10

5

0

dotted line shows changes in Northern Hemisphere ice volume (right axis; in metres of global-mean sea level (gmsl) equivalent) prescribed in the ICE5G27 ice history. b, As in a, but adopting the HV Earth model (Methods). Blue (a) and red (b) vertical bands represent the timing of MWP and AID events, as in Fig. 2a, c, respectively.

Article 30 ka

20 ka

10 ka

0 ka

4000 3500

a. varying NH

3000 2500 2000 1500 1000 500 0

b. fixed NH (m)

Ice shelf -500 500 400 300 200 100 0 -100

c. a-b

-200 -300 -400 -500

(m)

Extended Data Fig. 2 | Evolution of Antarctic ice cover with and without Northern Hemisphere sea-level forcing. a, b, Thickness of grounded ice (in metres) and extent of ice shelves, at 30 ka, 20 ka, 10 ka and the present day, predicted from simulations that include variations in the Northern Hemisphere ice sheets represented by the ICE5G27 ice history (a) and from simulations in which ice cover in the Northern Hemisphere remains fixed (b). Black lines show

the grounding lines. c, The difference in grounded ice thickness between simulations in a and b, representing the effect of sea-level changes associated with Northern Hemisphere ice sheets on the evolution of the AIS. Green and black lines represent the positions of the grounding lines with (a) and without (b) the Northern Hemisphere sea-level forcing included.

19 ka

18 ka

17 ka

16 ka

15 ka

14 ka

13 ka

12 ka

11 ka

10 ka 500 400 300 200 100

9 ka

8 ka

7 ka

6 ka

5 ka

0 -100 -200 -300 -400 -500 (m)

4 ka

3 ka

2 ka

Extended Data Fig. 3 | Influence of Northern Hemisphere sea-level forcing on Antarctic ice cover during the deglaciation. The colour scale indicates the difference in the thickness (in metres) of grounded ice, at the indicated times, between simulations that include variations in the Northern Hemisphere ice sheets from ICE5G27 ice history and in which ice cover in the

1 ka

0 ka

Northern Hemisphere remains fixed throughout the simulation. Differences are displayed as in Fig. 3c, but every 1 kyr for the past 19 kyr. Green and black lines represent the positions of the grounding lines with and without the Northern Hemisphere sea-level forcing included, respectively.

Article a Weddell Sea Sector ICE5G fixed NH

0.15

c

0

7

3

Ice volume (10 km )

0.2

-200

0.1

Weddell Sea

0.05

b

0 40

-600 35

30

0.2

25

20

15

10

5

-800

0

time (ka)

-1000

Ross Sea Sector

-1200 -1400

0.15

7

3

Ice volume (10 km )

-400

Ross Sea

0.1

-1600 -1800 -2000

0.05

(m) 0 40

35

30

25

20

15

10

5

0

time (ka) Extended Data Fig. 4 | Antarctic ice-volume changes in the Ross Sea and Weddell Sea sectors. a, Changes in ice volume in the Weddell Sea sector predicted in simulations with fixed (red) and evolving (black) Northern Hemisphere ice from the ICE5G27 ice history. b, As in a, but for the

Ross Sea sector. c, Blue lines outline the areas included in the calculations in a and b; colour scale indicates the change in ice thickness (in metres) from 20 ka to the present day in the simulations that include Northern Hemisphere ice-cover changes from ICE5G27.

Antarctic ice flux

0.2

a

N.H.ICE5G fixed N.H. ice

0.1 0 15 0.2

14

13

12

11

10

9

8

7

6

5

4

3

b

2

1

0

N.H.ICE6GC

0.1 0

7

3

Ice flux (10 km / year)

15 0.2

14

13

12

11

10

9

8

7

6

5

4

3

2

c

1

0

N.H.ANU

0.1 0 15 0.2

14

13

12

11

10

9

8

7

6

5

4

d

3

2

1

0

N.H.GLAC1D_9927

0.1 0 15 0.2

14

13

12

11

10

9

8

7

6

5

4

e

3

2

1

0

N.H.GLAC1D_9894

0.1 0 15

14

13

12

11

10

9

8

7

6

5

4

3

2

1

0

time (ka) Extended Data Fig. 5 | Influence of Northern Hemisphere sea-level forcing on the rate of Antarctic ice loss. a–e, Rate of change of Antarctic ice volume, including grounded and floating ice, calculated with a 100-year running mean, predicted from simulations including (black) and excluding (red) Northern

Hemisphere ice-cover changes, using the ice histories indicated in the legend (Methods). The mean and standard deviation of these five panels are shown in Fig. 3a.

Article a

MWP-1A

1.4

1.2

1.0

0.8

0.6

0.4

0.2

0

Barbados value (pink) = 1.22, Tahiti value (green) = 1.15

b

Early Holocene

1.4

1.2

1.0

0.8

0.6

0.4

0.2

0

Barbados value (pink) = 1.21, Tahiti value (green) = 1.17 Extended Data Fig. 6 | Patterns of sea-level change for Antarctic ice loss during MWP-1A and the early Holocene. a, b, Predicted sea-level change, normalized by the global-mean sea-level-equivalent associated with Antarctic ice loss during MWP-1A (a) and the early Holocene (including MWP-1B; b). Calculations are associated with simulations that include Northern

Hemisphere ice cover changes given by ICE5G27. The patterns of sea-level change and the global mean sea-level equivalent used in the normalization are calculated over the time windows indicated by the green vertical bands in Fig. 2b. Green and magenta asterisks indicate the locations of the far-field relative sea-level records in Tahiti and Barbados.

14.5 -13.5 ka

11.5 - 9.0 ka

a

d

b

e

(m)

c

f 3.07

3

full coupling eustatic

3.06

2.98 2.96

Ice volume (107 km 3)

Ice volume (107 km 3)

3.05

2.94

3.04

2.92

3.03 3.02

2.9

2.88

3.01

2.86

3

2.84

2.99

2.82

2.98 -14.5

-14

time (ka)

-13.5

-13

Extended Data Fig. 7 | Sensitivity of the Weddell Sea sector to geographic variability in sea-level forcing. a, Same as Fig. 3b, but zoomed in on the Weddell Sea region, where geographically variable sea-level changes associated with Northern Hemisphere ice loss are largest (Fig. 1c). The colour scale shows the change in ice thickness predicted from a simulation adopting the ICE5G27 ice history in the Northern Hemisphere, which includes geographically variable sea-level changes associated with gravitational, deformational and Earth rotational effects activated by ice-cover changes globally, during MWP-1A (14.5–13.5 ka). Grey and black lines indicate the grounding-line position at the start and end of the time interval, respectively. b, The difference between a and the same calculation but adopting the

2.8 -11.5

-11

-10.5

-10

time (ka)

-9.5

-9

simulation with globally uniform sea-level change from the Northern Hemisphere. The black line is as in a; the blue line indicates the grounding-line position at the end of the time interval for the uniform sea-level simulation. c, Antarctic ice-volume variations from simulations with geographically variable (black) and uniform (red) sea-level changes associated with Northern Hemisphere ice loss over the MWP-1A interval. d–f, As in a–c, but for the early Holocene interval (11.5–9 ka). In this case, d is the same as Fig. 3d, but zoomed in on the Weddell Sea region. The uniform sea-level change is calculated relative to modern topography and scaled such that the total contribution to global sea-level change from the Northern Hemisphere over the last deglaciation (since 21 ka) is 95.5 m, in agreement with ref. 27.

Article a Antarctic ice volume

0.305 0.3

Ice volume (107 km3 )

0.295 0.29

0.285 0.28 MWP 1A

0.275

MWP 1B

0.27 0.265 40

35

30

b

25

20

time (ka)

15

10

0

Global mean sea level equivalent

4

fixed NH fixed NH at modern ICE5G ICE6GC ANU GLAC1D_9927 GLAC1D_9894

3

Antarctic ice loss in GMSLE (m)

5

2 1

AID 6

0

AID 1 AID 2

-1 -2 -3 -4 -5 40

35

30

25

Extended Data Fig. 8 | Predicted Antarctic ice-volume changes and global-mean sea-level contributions. a, Changes in AIS volume predicted in a simulation with Northern Hemisphere ice cover fixed at the 40 ka configuration within ICE5G27 (solid red line) and in simulations with evolving Northern Hemisphere ice adopting the ICE5G27 (solid black line), ICE6GC31 (dashed black line) and ANU30 (cyan line) ice histories, as well as two composite ice histories in which ice cover over North America and Greenland in ICE5G has been replaced by regional GLAC1D29 models (blue lines). The dashed red line represents a simulation in which the Northern Hemisphere ice sheets are fixed at the modern configuration rather than at the 40 ka configuration throughout the simulation. In this case, marine-based sectors of the AIS start on even

20

time (ka)

15

10

5

0

shallower bedrock, and hence the predicted ice-sheet growth is larger at the LGM, while the ice loss during the deglaciation occurs later and is of even smaller magnitude than in the original simulation. Note that this is not a realistic starting configuration. b, As in a, but expressed as a global-mean sea-level-equivalent (GMSLE) relative to the modern state. This is calculated by taking the ice above floatation thickness in Antarctica relative to the palaeo bedrock topography at each time step in the model, and dividing by the area of the modern ocean. Note that a and b are not directly proportional because as the bedrock topography in Antarctica evolves the volume of ice above floatation in marine sectors also changes. Blue (a) and red (b) vertical bands represent the timing of MWP and AID events, as in Fig. 2a, c, respectively.

17

16

15

14

13

12

11

10

9

8

0.5

a 0.4 0.3 0.9

0.2

0.8

Uncertainty AICC 2012 (kyr)

Age difference AICC 2012 EDML1 (kyr)

18

0.7

0.1

0.6

0

0.5

b

0.4 0.3 0.2

MWP-1A

MWP-1B

0.1

Uncertainty AID 7

AID 6

AID 5

AID 4

AID 3

AID 2

AID 1

0.14

AICC 2012 age scale EDML1/EDC3 age scale

IBRD flux (grains/yr/cm3)

0.12 0.10 0.08

c

0.06 0.04 0.02 0

18

17

16

15

14

13 Age (ka)

Extended Data Fig. 9 | Age model comparison and uncertainty for IBRD flux record from Iceberg Alley. a, Age difference between the AICC 2012 55,56 and EDML1 54 age models. b, Age uncertainty in the AICC 2012 age model. c, IBRD flux time series adopting the AICC 2012 (black line, as in Figs. 2c, 4b) and EDML1/EDC3 (blue dotted line) age scales. The IBRD stack is composed of records from sites MD07-3133 and MD07-3134. It is presented here for 20–0 ka

12

11

10

9

8

and was combined with previous data for 27–7 ka4 and 8–0 ka24. Vertical brown bars indicate AID events 1–74 on the AICC 2012 age scale. Blue vertical bars indicate MWP-1A 21 and MWP-1B22. Horizontal black error bars show propagated uncertainties for the upper and lower bounds of each AID event for errors in tie-point correlation to EDML4 and uncertainties of the AICC 2012 age model..

Article a sites 11-13

800

with NH no NH cosmo data at site 11 cosmo data at site 12 cosmo data at site 13

ice thickness above modern (m)

700 600 500 400 300

c

12 ka

200 100 0

14

12

10

8

6

4

2

0

time (ka)

14,15

11-13

b

1

sites 14,15

800

ice thickness above modern (m)

S

sim. with ICE5G in NH sim. with fixed NH ice cosmo data at site 14 cosmo data at site 15

700 600

3-5

500 (m)

400 300 200 100 0

14

12

10

8

6

4

2

0

time (ka)

Extended Data Fig. 10 | Comparison of predicted and observed ice-thickness changes in the Weddell Sea region. a, b, Predicted (lines) and observed (error bars) ice thickness (in metres) above the modern thickness at sites 11–13 (a) and 14, 15 (b) from ref. 35. Predictions are from simulations in which Northern Hemisphere ice cover is evolving according to ICE5G27 (black

lines) or is fixed (blue lines). Error bars show cosmogenic exposure age data with 2σ uncertainty from ref. 35. c, Map of predicted ice thickness at 12 ka, in the simulation with ICE5G27. The locations of the relevant sites in the Weddell Sea and Ross Sea (see Extended Data Fig. 11) regions are indicated. See Methods for further discussion of these results.

a

1000

AID 1

AID 2

Cosmo site 1 Cosmo sites 3-5 Scott Coast site1 site3 site4 site5

900

ice thickness above modern (m)

800 700 600 500 400 300 200 100 0

15

14

b

d

13

12

11

. ..

10 9 time (ka)

c

12 ka

. ..

12-10 ka

Extended Data Fig. 11 | Comparison of predicted and observed ice-thickness changes in the Ross Sea region. a, Predicted (lines) and observed (2σ error bars) ice thickness (in metres) above the modern thickness at Scott Coast site S (black) and sites 1 (red) and 3–5 (shades of blue) from ref. 35. The locations of the sites are indicated in b–e. Predictions are from simulations in which Northern Hemisphere ice cover is evolving according to ICE5G27 (solid

8

7

6

5

4

. ..

12-11 ka

1

S

Ice shelf

e

3-5

. ..

12-9 ka

lines) or is fixed (dashed lines). Observations are cosmogenic exposure age data from ref. 35. Red vertical bands represent the timing of AID events 1 and 2, as in Fig. 2c. b, Map of predicted ice thickness 12 ka in the Ross Sea, in the simulation with evolving Northern Hemisphere ice. c–e, The difference in ice thickness between 12 ka (b) and 11 ka (c), 10 ka (d) and 9 ka (e). See Methods for further discussion of these results.

Article

Sensory pollutants alter bird phenology and fitness across a continent https://doi.org/10.1038/s41586-020-2903-7 Received: 30 January 2019 Accepted: 12 August 2020 Published online: 11 November 2020 Check for updates

Masayuki Senzaki1,2,14, Jesse R. Barber3, Jennifer N. Phillips1,4, Neil H. Carter5, Caren B. Cooper6,7, Mark A. Ditmer5, Kurt M. Fristrup8, Christopher J. W. McClure3,9, Daniel J. Mennitt10, Luke P. Tyrrell11, Jelena Vukomanovic12,13, Ashley A. Wilson1 & Clinton D. Francis1,14 ✉

Expansion of anthropogenic noise and night lighting across our planet1,2 is of increasing conservation concern3–6. Despite growing knowledge of physiological and behavioural responses to these stimuli from single-species and local-scale studies, whether these pollutants affect fitness is less clear, as is how and why species vary in their sensitivity to these anthropic stressors. Here we leverage a large citizen science dataset paired with high-resolution noise and light data from across the contiguous United States to assess how these stimuli affect reproductive success in 142 bird species. We find responses to both sensory pollutants linked to the functional traits and habitat affiliations of species. For example, overall nest success was negatively correlated with noise among birds in closed environments. Species-specific changes in reproductive timing and hatching success in response to noise exposure were explained by vocalization frequency, nesting location and diet. Additionally, increased light-gathering ability of species’ eyes was associated with stronger advancements in reproductive timing in response to light exposure, potentially creating phenological mismatches7. Unexpectedly, better light-gathering ability was linked to reduced clutch failure and increased overall nest success in response to light exposure, raising important questions about how responses to sensory pollutants counteract or exacerbate responses to other aspects of global change, such as climate warming. These findings demonstrate that anthropogenic noise and light can substantially affect breeding bird phenology and fitness, and underscore the need to consider sensory pollutants alongside traditional dimensions of the environment that typically inform biodiversity conservation.

Anthropogenic noise and light pollution are increasing even faster than the human population1,2. Laboratory work and small-scale field studies suggest that both pollutants can affect animal behaviour and physiology by altering sensory performance and perceptions of environments6. Anthropogenic noise impairs the perception of auditory signals, altering communication, orientation, foraging and vigilance behaviours4,8,9. Analogously, anthropic night lighting modifies activities and interactions mediated by vision10,11 and alters circadian rhythms, which are tightly controlled by photoperiod5. Despite growing evidence documenting behavioural responses to these globally pervasive sensory pollutants, fitness consequences of these stimuli are known only from a few species at the local scale12–15. Thus, there is a clear need to understand whether fitness consequences of exposure to these stimuli are widespread, whether responses to these stimuli vary across species and, if so, why.

To fill this knowledge gap, we investigated the macroecological consequences of both noise and light pollution across a continent (Fig. 1, Extended Data Figs. 1 and 2) and determined whether species-specific responses to each stimulus can be linked to functional traits or environmental contexts. We used 58,506 nest records from 142 species collected throughout the contiguous United States between 2000 and 2014 by citizen science volunteers through the NestWatch programme of the Cornell Laboratory of Ornithology (Methods, Supplementary Table 1). We combined this dataset with high-resolution geospatial data for anthropogenic noise and night lighting (Supplementary Table 2) to examine how these variables influence first egg-laying date (clutch initiation), clutch size, partial hatching success (in which one or more eggs fail to hatch), clutch failure (nest failure at egg stage) and nest success (fledging of one or more young from nest) using linear and generalized linear mixed-effect models with spatially explicit correlation

Biological Sciences, California Polytechnic State University, San Luis Obispo, CA, USA. 2Faculty of Environmental Earth Science, Hokkaido University, Sapporo, Japan. 3Biological Sciences, Boise State University, Boise, ID, USA. 4Department of Science and Mathematics, Texas A&M San Antonio, San Antonio, TX, USA. 5School for Environment and Sustainability, University of Michigan, Ann Arbor, MI, USA. 6Department of Forestry and Environmental Resources, North Carolina State University, Raleigh, NC, USA. 7Leadership in Public Science, North Carolina State University, Raleigh, NC, USA. 8National Park Service Natural Sounds and Night Skies Division, Fort Collins, CO, USA. 9The Peregrine Fund, Boise, ID, USA. 10Exponent, Denver, CO, USA. 11 Department of Biological Sciences, State University of New York Plattsburgh, Plattsburgh, NY, USA. 12Department of Parks, Recreation and Tourism Management, North Carolina State University, Raleigh, NC, USA. 13Center for Geospatial Analytics, North Carolina State University, Raleigh, NC, USA. 14These authors contributed equally: Masayuki Senzaki, Clinton D. Francis. ✉e-mail: [email protected] 1

Nature | Vol 587 | 26 November 2020 | 605

Article a

dB 60 50 40 30 20 10 0

Light

Noise

b

Timing

Clutch size

Clutch failure

Partial hatch

Success

Fig. 1 | Anthropogenic noise and night lighting are widespread and affect a variety of species. a, Colours denote median sound energy (L 50 = sound levels exceed the value 50% of the measurement period) in A-weighted decibels (dB) from anthropogenic sources. Points denote nest locations that were successful (black) or unsuccessful (red). Sound level data from United States National Park Service Data Store. No claim to original US Government works. b, Sample

of species for which reproduction is negatively influenced by exposure to noise or light pollution. Top row, noise-affected species: left to right, Northern cardinal, oak titmouse, barn swallow, Eastern bluebird, purple martin. Bottom row, light-affected species: white-breasted nuthatch, Carolina wren, house sparrow, house finch, violet-green swallow. Photograph credits: Carolina wren, public domain; all others, David Keeling.

structure. We accounted for the influence of latitude on responses, because it is a well-known macroecological proxy for environmental drivers of life history variation in avian reproduction16, and controlled for potential differences between coastal and interior populations. Additionally, we separated the effects of noise and light from other metrics reflective of human activity and urbanization in general by including human population density and proportion of anthropogenic impervious surface in our models (Supplementary Table 2). We scaled all continuous predictors to enable direct comparisons and ensured that our models did not suffer from problems of multicollinearity (Methods). We examined the influence of these variables on nesting metrics of all species and of those that nest in open environments, such as grasslands and wetlands (hereafter, open-habitat species), and in forests (hereafter, closed-habitat species). These two environments contrast strongly in vegetation structure that could provide microhabitat refugia from anthropogenic stimuli, affording the opportunity to test whether responses to noise and light pollution are stronger in open environments where exposure to these stimuli may be most severe17,18. We followed these analyses with species-specific models for 27 species that were represented by at least 100 nests in the dataset (Supplementary Table 3) and used phylogenetically informed models to determine whether individual species’ responses (± s.e.) to noise or light are linked to traits hypothesized to predict sensitivities to these stimuli (Table 1, Methods). All-species models of the five responses revealed weak associations with noise, light and other anthropogenic predictors (Supplementary Table 4). However, open-habitat birds in the brightest conditions were estimated to begin laying eggs, on average, a month earlier than those in

the darkest areas (nests = 4,251, β = –4.73, 95% CI –8.21, –1.25), although the confidence interval in brighter conditions was also wide (Fig. 2, Supplementary Table 5). Species in closed habitats exhibited the same apparent trend, advancing laying by approximately 18 d over a smaller range of light exposure; however, the confidence in the effect was lower (nests = 5,076, β = –2.30, 95% CI –5.05, 0.44) (Fig. 2, Supplementary

606 | Nature | Vol 587 | 26 November 2020

Table 1 | Hypotheses and predicted relationships between traits and responses to noise and light exposure Variable

Hypothesis

Stressor | predicted effect

Vocal frequency

Birds with higher-frequency vocalizations should experience less interference from noise

Noise | +

Light-gathering ability

Greater light-gathering ability (that is, better vision in low light) will correlate with sensitivity to light exposure

Light | −

Cavity

Birds that nest in cavities will be less sensitive to noise and light than open-cup-nesting species because cavities may reduce exposure to these stimuli

Noise | + Light | +

Diet

Birds with animal-based diets will be more sensitive to noise than those with plant-based diets; all birds, regardless of diet, will benefit from light through extended foraging time

Noise | − Light | +

b

Scaled noise exceedance 0 2

4

140

120

6

5

100

4

80 0

2.5 5.0 7.5 Scaled light levels

10.0

0

2 4 Scaled light levels

6

d

c 0.5

0.9

0.4

0.8 Nest success

Clutch failure

–2 7

160

Clutch size

Clutch initiation date (1 January = 1)

a

0.3 0.2 0.1

0.7 0.6 0.5

0

0.4 –2

0 2 Scaled noise exceedance

4

–2

0 2 Scaled noise exceedance

4

Fig. 2 | Responses to light and noise by birds in open and closed habitats. a, Light exposure resulted in advancements in clutch initiation for birds in both open (blue solid line; n = 4,251, β = –4.73, 95% CI –8.21, –1.25) and closed (orange dashed line, n = 5,076, β = –2.30, 95% CI –5.05, 0.44) environments. b, Birds in closed environments also experienced increased and decreased clutch sizes with light (orange dashed line, bottom-axis, β = 0.06, 95% CI 0.00, 0.13) and noise exposure (red solid line, top axis, β = –0.06, 95% CI –0.14, 0.01), respectively (n = 5,076). c, d, Closed-habitat birds also experienced increased clutch failure (n = 5,076, β = 0.17, 95% CI –0.02, 0.37) (c) and lower nest success (d) (n = 4,980, β = –0.19, 95% CI –0.37, –0.003) with noise exposure. Analyses based on spatially explicit linear mixed-effect models (a) and generalized linear mixed-effect models (b–d). Marginal effects and 95% CIs (denoted by dotted lines) are shown.

Table 6). In terms of clutch size, closed-habitat birds produced clutches that were approximately 16% larger in well-lit compared to dark areas (nests = 5,076, β = 0.06, 95% CI 0.00, 0.13) (Fig. 2, Supplementary Table 6), but there was no apparent influence of light on clutch size among open-habitat birds. There is no simple explanation for these habitat-associated contrasts. Dense vegetation should substantially suppress light from skyglow for birds in closed environments. However, relative to open-habitat birds, those in closed environments tend to have eye geometries that improve low-light vision (β = 0.04, 95% CI 0.00, 0.09, λ = 0) (Supplementary Table 8). Thus, closed-habitat birds may take advantage of even low light levels to extend foraging time to support larger clutches and broods. Closed-habitat, but not open-habitat, birds also tended to experience a decline in clutch size with noise exposure (nests = 5,076, β = –0.06, 95% CI –0.14, 0.01) (Fig. 2, Supplementary Table 6) such that those in the loudest areas laid clutches that were on average 0.64 eggs (12%) smaller than those in the quietest conditions. Birds that inhabit areas with dense vegetation vocalize at lower frequencies than those in more open areas19. Because lower-frequency vocalizations are more susceptible to energetic masking from anthropogenic noise20, and masking can negatively influence female sexual receptivity and maternal investment in clutch size21, the decline in clutch size may be an outcome of masking. Elevated noise also tended to increase the probability of clutch failure for closed-habitat birds (nests = 5,076, β = 0.17, 95% CI

–0.02, 0.37) (Fig. 2, Supplementary Table 6). Most importantly, nest success among closed-habitat, but not open-habitat, birds declined with noise exposure (nests = 4,980, β = –0.19, 95% CI –0.37, –0.003) (Fig. 2, Supplementary Table 6). These results provide multi-species evidence on a continental scale that supports the negative influence of noise exposure on reproductive success previously documented for only a few species12–14. Models of responses by individual species revealed widespread but heterogeneous effects of noise and light (Extended Data Figs. 3–8). Half of the species experienced changes in nesting phenology or reproduction due to both stimuli, and 19 of 27 species experienced strong responses to noise or light (Extended Data Fig. 3, Supplementary Table 7). In general, average noise and light pollution exposures for each species exhibited positive covariance (n = 27, rho = 0.830, P 0; lighter dashed lines throughout reflect relationships with no phylogenetic structure.

Our continental-scale analyses document complex effects of sensory pollution on avian reproduction even while controlling for other potentially influential natural and anthropogenic macroecological variables. We found that closed-habitat birds show more responses to these stimuli than birds affiliated with open environments, and our species-level analyses revealed many important links between responses to noise and light and functionally relevant traits. Specifically, variation in acoustic spectra used by birds to communicate and variation in light-gathering ability of the avian eye were important predictors of responses to noise and to light exposure, respectively. The advancement in reproductive phenology due to light exposure among birds with better light-gathering ability was expected. However, improvements in overall nest success with light exposure for species with better low-light vision were unforeseen, prompting new questions about how responses to sensory stimuli interact with or counteract responses to other forms of global change, such as the warming climate. Because we detected underlying variation in exposure to noise and light among species (Extended Data Fig. 3), determining whether the reported multi-species responses can be attributed to species sorting, general responses across species30 or both will be important. Future work should also prioritize evaluating how these and other functional traits and contexts are related to sensitivities to these stimuli across more animal species. Such an approach could be both useful and practical for forecasting responses among poorly studied species. Finally, because these sensory pollutants are pervasive1,2, and our results point to widespread responses to these stimuli, mitigating sensory pollution may be a powerful tool for habitat restoration and improved ecological resilience.

Online content Any methods, additional references, Nature Research reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at https://doi.org/10.1038/s41586-020-2903-7.

1. 2. 3. 4.

an increased phenological mismatch between food availability and brood rearing as temperatures rise. Given our finding that sensory pollutants can influence phenology and fitness outcomes for multiple species, understanding how altered sensory environments influence our interpretation of biological responses to climate change is a critical frontier. For example, there is widespread consensus that climate change is advancing reproductive phenology in temperate birds26, especially as insectivores track changes in their arthropod prey27. However, if responses to climate change are collected within light-polluted areas, documented shifts in clutch initiation that are attributed to temperature are confounded with the effects of light exposure and thus likely overestimated. Similarly, delays in the onset of breeding due to noise exposure could offset responses to climate change and result in underestimation. Much of our existing knowledge of phenological responses to climate change comes from studies in North America and Europe27–29. Yet because noise and light are pervasive in both regions (especially near cities)1,2, re-evaluation of documented phenological responses to climate change with explicit consideration of influential environmental sensory gradients could paint a much more nuanced picture of how organisms are coping with all aspects of global change. 608 | Nature | Vol 587 | 26 November 2020

5. 6. 7. 8.

9. 10.

11.

12. 13.

14.

15.

Buxton, R. T. et al. Noise pollution is pervasive in U.S. protected areas. Science 356, 531–533 (2017). Kyba, C. C. M. et al. Artificially lit surface of Earth at night increasing in radiance and extent. Sci. Adv. 3, e1701528 (2017). Barber, J. R., Crooks, K. R. & Fristrup, K. M. The costs of chronic noise exposure for terrestrial organisms. Trends Ecol. Evol. 25, 180–189 (2010). Swaddle, J. P. et al. A framework to assess evolutionary responses to anthropogenic light and sound. Trends Ecol. Evol. 30, 550–560 (2015). Gaston, K. J., Davies, T. W., Nedelec, S. L. & Holt, L. A. Impacts of artificial light at night on biological timings. Annu. Rev. Ecol. Evol. Syst. 48, 49–68 (2017). Dominoni, D. M. et al. Why conservation biology can benefit from sensory ecology. Nat. Ecol. Evol. 4, 502–511 (2020). Visser, M. E. & Gienapp, P. Evolutionary and demographic consequences of phenological mismatches. Nat. Ecol. Evol. 3, 879–885 (2019). Francis, C. D. & Barber, J. R. A framework for understanding noise impacts on wildlife: an urgent conservation priority. Front. Ecol. Environ. 11, 305–313 (2013). Shannon, G. et al. A synthesis of two decades of research documenting the effects of noise on wildlife. Biol. Rev. Camb. Philos. Soc. 91, 982–1005 (2016). van Langevelde, F., Ettema, J. A., Donners, M., WallisDeVries, M. F. & Groenendijk, D. Effect of spectral composition of artificial light on the attraction of moths. Biol. Conserv. 144, 2274–2281 (2011). Hale, J. D., Fairbrass, A. J., Matthews, T. J., Davies, G. & Sadler, J. P. The ecological impact of city lighting scenarios: exploring gap crossing thresholds for urban bats. Glob. Chang. Biol. 21, 2467–2478 (2015). Halfwerk, W., Holleman, L. J. M., Lessells, C. M. & Slabbekoorn, H. Negative impact of traffic noise on avian reproductive success. J. Appl. Ecol. 48, 210–219 (2011). Kight, C. R., Saha, M. S. & Swaddle, J. P. Anthropogenic noise is associated with reductions in the productivity of breeding Eastern Bluebirds (Sialia sialis). Ecol. Appl. 22, 1989–1996 (2012). Injaian, A. S., Poon, L. Y. & Patricelli, G. L. Effects of experimental anthropogenic noise on avian settlement patterns and reproductive success. Behav. Ecol. 29, 1181–1189 (2018). Kempenaers, B., Borgström, P., Loës, P., Schlicht, E. & Valcu, M. Artificial night lighting affects dawn song, extra-pair siring success, and lay date in songbirds. Curr. Biol. 20, 1735–1739 (2010).

16. Cooper, C. B., Hochachka, W. M., Butcher, G. & Dhondt, A. A. Seasonal and latitudinal trends in clutch size: thermal constraints during laying and incubation. Ecology 86, 2018–2031 (2005). 17. Van Renterghem, T., Botteldooren, D. & Verheyen, K. Road traffic noise shielding by vegetation belts of limited depth. J. Sound Vibrat. 331, 2404–2425 (2012). 18. Luginbuhl, C. B. et al. From the ground up II: sky glow and near-ground artificial light propagation in Flagstaff, Arizona. Publ. Astron. Soc. Pacif. 121, 204–212 (2009). 19. Boncoraglio, G. & Saino, N. Habitat structure and the evolution of bird song: a metaanalysis of the evidence for the acoustic adaptation hypothesis. Funct. Ecol. 21, 134–142 (2007). 20. Francis, C. D. Vocal traits and diet explain avian sensitivities to anthropogenic noise. Glob. Chang. Biol. 21, 1809–1820 (2015). 21. Huet des Aunay, G. et al. Negative impact of urban noise on sexual receptivity and clutch size in female domestic canaries. Ethology 123, 843–853 (2017). 22. Proppe, D. S., Sturdy, C. B. & St Clair, C. C. Anthropogenic noise decreases urban songbird diversity and may contribute to homogenization. Glob. Chang. Biol. 19, 1075–1084 (2013). 23. Kleist, N. J., Guralnick, R. P., Cruz, A., Lowry, C. A. & Francis, C. D. Chronic anthropogenic noise disrupts glucocorticoid signaling and has multiple effects on fitness in an avian community. Proc. Natl Acad. Sci. USA 115, E648–E657 (2018).

24. Dominoni, D., Quetting, M. & Partecke, J. Artificial light at night advances avian reproductive physiology. Proc. Biol. Sci. B 280, 20123017 (2013). 25. Visser, M. E., Both, C. & Lambrechts, M. M. Global climate change leads to mistimed avian reproduction. Adv. Ecol. Res. 35, 89–110 (2004). 26. Winkler, D. W., Dunn, P. O. & McCulloch, C. E. Predicting the effects of climate change on avian life-history traits. Proc. Natl Acad. Sci. USA 99, 13595–13599 (2002). 27. Dunn, P. O. & Winkler, D. W. Climate change has affected the breeding date of tree swallows throughout North America. Proc. Biol. Sci. 266, 2487–2490 (1999). 28. Both, C. & Visser, M. E. Adjustment to climate change is constrained by arrival date in a long-distance migrant bird. Nature 411, 296–298 (2001). 29. Burgess, M. D. et al. Tritrophic phenological match-mismatch in space and time. Nat. Ecol. Evol. 2, 970–975 (2018). 30. van de Pol, M. & Wright, J. A simple method for distinguishing within- versus between-subject effects using mixed models. Anim. Behav. 77, 753–758 (2009). Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. © The Author(s), under exclusive licence to Springer Nature Limited 2020

Nature | Vol 587 | 26 November 2020 | 609

Article Methods Nest watch data from citizen scientists Through the NestWatch project, the Cornell Laboratory of Ornithology coordinates volunteers who monitor wild bird nests. Volunteers report information on key events in the breeding process, such as the first day an egg is laid (that is, first lay date or clutch-initiation date), the number of eggs (that is, clutch size), the number of eggs that successfully hatch and whether the nest is successful at fledging at least one young (that is, nest success). We originally obtained 186,705 nest records spanning 2000–2014. Following a procedure that maximized the precision and plausibility of nest records (Supplementary Text), our final dataset included 58,506 samples of 142 species (Supplementary Table 1). Habitat and species-specific samples From our final dataset, we extracted 27 species with at least 100 nests for the species-specific analyses (Supplementary Table 3). We used habitat descriptions in The Birds of North America31 to categorize species as inhabiting closed, open or mixed habitats. ‘Open’ denotes the absence of tall vegetation: wetlands, grasslands, shrublands and farmland. ‘Closed’ habitats include deciduous, evergreen and mixed forests. Species inhabiting both habitat types or open woodlands received a ‘mixed’ label. As a result, 4,251 nests of 51 species were classified as open-habitat species, 5,076 nests of 22 species were classified as closed-habitat species and 49,179 nests of 69 species were classified as mixed-habitat species (Supplementary Tables 1 and 3). Environmental and trait data Summary statistics of untransformed environmental variables available in Supplementary Table 2. Anthropic night lighting. Anthropic night lighting is scattered by the atmosphere back towards the ground, resulting in an increase in night sky luminance. Data describing the magnitude of zenith skyglow were obtained from the second world atlas of artificial night sky brightness and converted to 270-m resolution32. These data reflect the zenith anthropic sky brightness as a ratio to the natural background sky brightness. The atlas is based on a light pollution propagation model with upward emission function calibrated by ground measurements. High-resolution measurements of upward radiance were acquired from the Visible Infrared Imaging Radiometer Suite (VIIRS) Day/Night Band (DNB) sensor on the Suomi National Polar-orbiting Partnership satellite. Six months of satellite data from 2014 were used, and the projected anthropic sky brightness ratio matches the time of satellite overpass, around 1 a.m. The atlas was computed using several constant assumptions, including the transparency of the atmosphere, the upward emission function of cities, the spectrum of anthropic lights and the hour of the night of the observation. The more that actual conditions differ from these assumptions, the greater the deviation in actual zenith skyglow will be compared to the atlas prediction. For example, when the sky is overcast, a severalfold increase in skyglow is to be expected near cities. The atlas may underestimate the ecological consequences of anthropic night lighting. Zenith brightness gives direct information for only one point in the sky, which is usually the darkest. The DNB is sensitive to light in the range 0.5–0.9 mm; this includes the near-infrared region, beyond the range of the human eye, and leaves out the blue and violet parts of the visible spectrum. Because the VIIRs DNB lacks sensitivity at wavelengths shorter than 500 nm, the blue-light emission peak of white LEDs is not detected and light pollution from this growing source of light in cities is underestimated.

Anthropogenic noise The anthropogenic noise data were obtained from a georeferenced map of expected environmental sound pressure levels. Geospatial sound

models have been developed to interpret and project acoustic conditions across the contiguous United States33. These models use machine learning to formulate relationships between sparsely distributed measurements of the ambient sound level and non-acoustic geospatial features such as topography, climate, hydrology and anthropogenic activity. The acoustical data included 1.5 million hours of long-term measurements from 492 sites located in urban and rural areas during 2000–2014. The resulting geospatial sound model was used to project expected sound levels under existing conditions at 270-m resolution. By changing model inputs from their current values to minimize anthropogenic factors, the geospatial sound model was adjusted so that it estimated a natural sound level that includes contributions from biotic and physiographic sources only. The anthropogenic noise exceedance level was calculated by the logarithmic subtraction of the natural from the existing sound projections. Environmental sound levels vary from one moment to the next and are summarized using a variety of statistics across multiple time scales and frequency ranges. We examined the anthropogenic daytime A-weighted L50 sound pressure level (dB re 20 μPa) as our measure of noise exposure. L50 is a robust statistic (50th percentile over time) that is less sensitive to infrequent, loud events. A-weighting is the most widely used composite measure of sound in human and wildlife noise studies, whereby sound energy is summed across the spectrum, emphasizing frequencies in which many terrestrial vertebrates have their most sensitive thresholds of hearing34. To account for temporal and seasonal variation in the acoustic environment, time of day and day of year were included as model covariates and projections were made for a mid-summer day (defined as 7 a.m. to 7 p.m.).

Urbanization variables: population density and impervious surface area Increases in human population density and transformation of natural land cover features to those reflective of high-intensity human use (that is, roadways, parking lots and buildings) are quintessential features of urbanization35,36. Therefore, we used the 30-m-spatial-resolution grid of per cent developed imperviousness from 2011 National Land Cover Database37 as a measure of the intensity of human use within the landscape. To quantify human population density, we used the 2010 US Census38 block data downscaled to 270-m grids. We then used a buffer radius of 500 m (that is, 0.785 km2) to quantify the mean proportion of impervious surface per 30-m grid and the mean human population density per 270-m grid surrounding each nest. Traits We collected additional trait information for the 27 species used for the species-specific analyses. Because species-specific peak vocal frequency, which is the frequency with the highest amplitude, has been shown to be related to changes in abundance among birds20, we gathered species-specific peak frequency measurements; those for 23 species were obtained from Francis20 and those for the other four species were calculated from high-quality recordings available at http://www.xeno-canto.org following the same methodology of averaging several recordings per species as described in Francis20. We obtained dominant diet (that is, animal-based or plant-based diets) and nesting strategy (that is, cavity or open) information from Birds of North America Online31. To obtain a variable indicative of an animal’s visual light sensitivity, we used the ratio of the corneal diameter to the transverse diameter, which scales values to the size of the visual system and animal. Corneal diameter is approximately equal to the aperture diameter and thus reflects the amount of light entering the eye, whereas transverse diameter is the theoretical upper bound of corneal diameter. The ratio of the two provides a measure of light sensitivity39–41. We obtained direct measurements of the corneal diameter and transverse diameter for 16 of the 27 species from several sources42–46. To obtain values for the remaining 11 species, we imputed

missing values using the phylopars function in the R package Rphylopars47. The phylopars function reconstructs ancestral states and imputes missing data using a linear-time algorithm48,49. We performed imputations of several eye-measurement variables simultaneously where covariances among traits and evolutionary relationships inform imputation of missing data. The approach used a pruned consensus tree from a recent class-wide phylogeny50 as the phylogenetic hypothesis and assumed a Brownian motion model of evolution because we found the corneal-transverse ratio to approximate a Brownian motion model of evolution among species with measurements (fitContinuous function in geiger, n = 62, λ = 0.981). To do so, in our data matrix we included 50 additional North American species from which eye measurements were available to improve the imputation of eye traits for the 11 species in our database without eye measurements. For this matrix we included complete information on several morphological traits reflective of size and ecology for all species. We categorized habitat affiliations from Birds of North America Online31, which we converted to an index spanning 1–4 reflecting vegetation density. To represent size we included body mass from the EltonTraits 1.0 database51, wing chord length from Lislevand et al.52 and body length31,53. To capture aspects of foraging ecology, we used bill length from Lislevand et al.52 and the proportion of the species diet that consists of invertebrates, fruit, nectar, seeds and other plant material from the EltonTraits 1.0 database51. Finally, we also included several eye morphological measurements: eye corneal diameter (87% complete), eye transverse diameter (81% complete), eye axial diameter (83% complete), corneal diameter to eye axial ratio (83% complete) and ratio of corneal diameter to transverse diameter (81% complete). We manually checked imputed values for potentially unrealistic values and compared imputed data for several species to congeners, which tended to be quite similar as expected from a Brownian motion model.

Statistical analyses We used linear and generalized linear mixed-effect models (LMMs and GLMMs) with a spatially explicit exponential correlation structure using the fitme function in the R package spaMM54 to examine the effects of the following explanatory predictors: (1) anthropogenic noise, (2) anthropic light (that is, zenith skyglow), (3) latitude, (4) human population density within 500 m from each nest, (5) proportion of impervious surface area within 500 m from each nest and, when applicable, (6) whether or not the nest was located in a coastal area, defined as 120 days. Therefore, we used 10 randomly drawn pseudo-replicates of 5,000 nests for models and used the average parameter estimates for interpretation20 (Supplementary Table 4). To test for relationships between responses to noise or light and traits, we used phylogenetic generalized least squares (PGLS) with the gls function in the R package nlme56. We simultaneously estimated the phylogenetic signal (λ) of the model following Revell57, but also incorporated recommendations from Ives et al.58 that accounts for error in the response variable. To do so, we used a weighting function with fixed variance of one over the square root of the s.e. of the response estimate59. To reveal any phylogenetic relationships between each trait and responses to noise or light, we considered traits one at a time in our models. All traits were tested with all responses, with the exception that light-gathering ability was used as a predictor only for models involving responses to light and vocal frequency was used as a predictor only for models involving responses to noise. If initial λ estimates fell outside of the range 0–1, we fixed λ at the boundary (that is, at 0 or 1). Following suggestions of Jones and Purvis60, we examined the potential influence of outliers by checking and removing observations with Studentized residuals ≥3.0. Outliers were detected in several models, but their removal did not alter interpretations (Supplementary Table 8). Finally, we repeated all analyses involving the ratio of corneal diameter to transverse diameter ratio (that is, light-gathering ability) as a predictor variable using non-imputed data. With the exception of the analysis of clutch failure responses to light, analyses on the restricted dataset did not differ from those based on the full dataset that included imputed values of this ratio (Supplementary Table 8).

Article For mixed-effect models evaluating responses to noise or light and PGLS models evaluating how responses (and s.e.) are related to traits, we embraced a more nuanced approach to reporting relationships between parameter estimates than the dichotomous approach of significance testing61–63. Specifically, we report and discuss apparent trends and provide 95% confidence intervals (CIs) to reflect the relative precision of estimates. In the Supplementary material, we also report 85% CIs to identify effects that also warrant consideration for inference and for developing future testable hypotheses23,64. Because CI estimates require re-running mixed-effect models iteratively for each parameter estimate in spaMM’s fitme function, and the computational demands of individual models required runtimes that ranged from days to several months, we calculated CIs as the s.e. of the parameter estimate multiplied by 1.440 (85%) and 1.960 (95%) for all-species, open-habitat birds and closed-habitat bird models. CI ranges for PGLS parameter estimates were derived using the confint function from the stats package65. The same function was used to calculate 95% CIs for all species-specific responses (Supplementary Table 7), but we also used the s.e. of the parameter estimate multiplied by 1.44 to generate 85% CIs for the computational reasons outlined above.

Potential collinearity and redundancy Anthropogenic noise and light levels are often correlated with one another and other environmental variables associated with human activities, necessitating careful inspection of models for issues of multicollinearity. Tools for assessing multicollinearity are not readily available for models in the spaMM package. Thus, we reran our models using the lmer and glmer functions from the lme4 package and the glmmTMB function with Conway–Maxwell–Poisson error from the glmmTMB package66 to check for potential collinearity and redundancy among the explanatory predictors by calculating the variance inflation factor (VIF) using the check_collinearity function in the performance package67. We used the recommendation of Dormann et al.68 that VIF >10 could reflect problematic issues of multicollinearity. VIF values of all explanatory predictors in all-species models, and for closed and open-habitat species models, were 10 (Supplementary Table 10), we explored whether removal of parameters contributing to high VIF values altered model interpretations. Only in three cases did this prove to be the case (Supplementary Table 11). For these three models, we re-ran spatial versions in fitme within the spaMM package as above and used the parameters from the reduced models for subsequent trait analyses and interpretation (Supplementary Table 7). Reporting summary Further information on research design is available in the Nature Research Reporting Summary linked to this paper.

Data availability The datasets analysed during this study are available at https://doi. org/10.5061/dryad.dbrv15dzc; Additional publicly available data used in this study include: Anthropogenic noise levels from the National Park Service Data Store (https://irma.nps.gov/DataStore/Reference/ Profile/2217356); New World Atlas of Artificial Night Sky Brightness (https://dataservices.gfz-potsdam.de/contact/showshort.php?id=esci doc:1541893&contactform); the 2011 US National Land Cover Database (https://www.mrlc.gov/data/nlcd-2011-land-cover-conus-0); US Human population density data (https://data.census.gov/cedsci/); EltonTraits 1.0 database (http://www.esapubs.org/archive/ecol/E095/178/), Birds of North America Online (recently changed to Birds of the World, https:// birdsoftheworld.org/bow/home) and vocal frequency (https://doi. org/10.5061/dryad.75nn1932) and body morphology data (https:// doi.org/10.6084/m9.figshare.3527864.v1). Source data are provided with this paper.

31. Cornell Laboratory of Ornithology. The Birds of North America Online (Cornell Laboratory of Ornithology, 2015). 32. Falchi, F. et al. The new world atlas of artificial night sky brightness. Sci. Adv. 2, e1600377 (2016). 33. Mennitt, D. J. & Fristrup, K. M. Influence factors and spatiotemporal patterns of environmental sound levels in the contiguous United States. Noise Control Eng. J. 64, 342–353 (2016). 34. Dooling, R. J., Lohr, B. & Dent, M. L. in Comparative Hearing: Birds and Reptiles (eds. Dooling, R. J. et al.) 308–359 (Springer, 2000). 35. Arnold, C. L. & Gibbons, C. J. Impervious surface coverage: the emergence of a key environmental indicator. J. Am. Plann. Assoc. 62, 243–258 (1996). 36. McKinney, M. L. Urbanization as a major cause of biotic homogenization. Biol. Conserv. 127, 247–260 (2006). 37. Xian, G. et al. Change of impervious surface area between 2001 and 2006 in the conterminous United States. Photogramm. Eng. Remote Sensing 77, 758–762 (2012). 38. United States Census Bureau. 2010 Census https://data.census.gov/cedsci/ (US Census Bureau, 2011). 39. Hall, M. I. & Ross, C. F. Eye shape and activity pattern in birds. J. Zool. (Lond.) 271, 437–444 (2007). 40. Kirk, E. C. Comparative morphology of the eye in primates. Anat. Rec. A Discov. Mol. Cell. Evol. Biol. 281, 1095–1103 (2004). 41. Martin, G. R. in Perception and Motor Control in Birds: An Ecological Approach (eds. Davies, M. & Green, P.) 5–34 (Springer, 1994). 42. Blackwell, B. F., Fernández-Juricic, E., Seamans, T. W. & Dolan, T. Avian visual system configuration and behavioural response to object approach. Anim. Behav. 77, 673–684 (2009). 43. Hall, M. I., Iwaniuk, A. N. & Gutiérrez-Ibáñez, C. Optic foramen morphology and activity pattern in birds. Anat. Rec. (Hoboken) 292, 1827–1845 (2009). 44. Moore, B. A., Doppler, M., Young, J. E. & Fernández-Juricic, E. Interspecific differences in the visual system and scanning behavior of three forest passerines that form heterospecific flocks. J. Comp. Physiol. A Neuroethol. Sens. Neural Behav. Physiol. 199, 263–277 (2013). 45. Ritland, S. M. The Allometry of the Vertebrate Eye (Univ. of Chicago, 1983). 46. Tyrrell, L. P. & Fernández-Juricic, E. The hawk-eyed songbird: retinal morphology, eye shape, and visual fields of an aerial insectivore. Am. Nat. 189, 709–717 (2017). 47. Goolsby, E. W., Bruggeman, J. & Ané, C. Rphylopars: fast multivariate phylogenetic comparative methods for missing data and within-species variation. Methods Ecol. Evol. 8, 22–27 (2017). 48. Uyeda, J. C., Pennell, M. W., Miller, E. T., Maia, R. & McClain, C. R. The evolution of energetic scaling across the vertebrate tree of life. Am. Nat. 190, 185–199 (2017). 49. Vitousek, M. N. et al. Macroevolutionary patterning in glucocorticoids suggests different selective pressures shape baseline and stress-induced levels. Am. Nat. 193, 866–880 (2019). 50. Jetz, W., Thomas, G. H., Joy, J. B., Hartmann, K. & Mooers, A. O. The global diversity of birds in space and time. Nature 491, 444–448 (2012). 51. Wilman, H. et al. EltonTraits 1.0: species-level foraging attributes of the world’s birds and mammals. Ecology 95, 1717–2032 (2014). 52. Lislevand, T., Figuerola, J. & Székely, T. Avian body sizes in relation to fecundity, mating system, display behavior, and resource sharing. Ecology 88, 1605 (2007). 53. Cornell Laboratory of Ornithology. All About Birds https://www.allaboutbirds.org (Cornell Laboratory of Ornithology, 2018). 54. Rousset, F. & Ferdy, J.-B. Testing environmental and genetic effects in the presence of spatial autocorrelation. Ecography 37, 781–790 (2014). 55. Smith, R. J. & Moore, F. R. Arrival timing and seasonal reproductive performance in a long-distance migratory landbird. Behav. Ecol. Sociobiol. 57, 231–239 (2005). 56. Pinheiro, J., Bates, D., DebRoy, S. & Sarkar, D. nlme: linear and nonlinear mixed effects models (R package version 3.1-104, 2012). 57. Revell, L. J. Phylogenetic signal and linear regression on species data. Methods Ecol. Evol. 1, 319–329 (2010). 58. Ives, A. R., Midford, P. E. & Garland, T. Jr. Within-species variation and measurement error in phylogenetic comparative methods. Syst. Biol. 56, 252–270 (2007). 59. Garamszegi, L. Z. in Modern Phylogenetic Comparative Methods and Their Application in Evolutionary Biology (ed. Garamszegi, L. Z.) 157–199 (Springer, 2014). 60. Jones, K. E. & Purvis, A. An optimum body size for mammals? Comparative evidence from bats. Funct. Ecol. 11, 751–756 (1997). 61. Hurlbert, S. H., Levine, R. A. & Utts, J. Coup de grâce for a tough old bull: “statistically significant” expires. Am. Stat. 73, 352–357 (2019). 62. Amrhein, V., Greenland, S. & McShane, B. Scientists rise up against statistical significance. Nature 567, 305–307 (2019). 63. Halsey, L. G. The reign of the p-value is over: what alternative analyses could we employ to fill the power vacuum? Biol. Lett. 15, 20190174 (2019). 64. Ware, H. E., McClure, C. J. W., Carlisle, J. D. & Barber, J. R. A phantom road experiment reveals traffic noise is an invisible source of habitat degradation. Proc. Natl Acad. Sci. USA 112, 12105–12109 (2015). 65. R Core Team. R: a language and environment for statistical computing (R Foundation for Statistical Computing, 2019). 66. Brooks, M. E. et al. glmmTMB balances speed and flexibility among packages for zero-inflated generalized linear mixed modeling. R J. 9, 378–400 (2017). 67. Lüdecke, D., Makowski, D. & Waggoner, P. Performance: Assessment of Regression Models Performance (2019). 68. Dormann, C. F. et al. Methods to account for spatial autocorrelation in the analysis of species distributional data: a review. Ecography 30, 609–628 (2007). 69. Yu, G. & Ekstrøm, C. T. emojifont: emoji and font awesome in graphics (R package version 0.5.3, 2019).

Acknowledgements We thank the NestWatch Program for use of their nesting database, the volunteers who monitored the nests and F. Rousset for advice in using the spaMM package. Supported by US National Science Foundation grants 1414171 to C.D.F. and J.R.B., 1556177 to J.R.B., 1556192 to C.D.F. and 1812280 to J.N.P.; NASA Ecological Forecasting grant NNX17AG36G to N.H.C., C.D.F. and J.R.B.; and Japanese Society for the Promotion of Science KAKENHI grant 17J00646 to M.S. Author contributions C.D.F., J.R.B. and C.J.W.M. conceived the project. C.B.C. and J.V. contributed geospatial NestWatch data and data validation, D.J.M and K.M.F. provided key data on noise and night lighting and L.P.T. provided key trait data. M.S., A.A.W., J.N.P. and C.D.F. performed analyses with advice from M.A.D. and N.H.C. All authors contributed to the writing of the manuscript.

Competing interests The authors declare no competing interests. Additional information Supplementary information is available for this paper at https://doi.org/10.1038/s41586-0202903-7. Correspondence and requests for materials should be addressed to C.D.F. Peer review information Nature thanks Albert Phillimore and Andrew Radford for their contribution to the peer review of this work. Reprints and permissions information is available at http://www.nature.com/reprints.

Article

Extended Data Fig. 1 | Natural log of zenith artificial sky brightness as a ratio to the natural background sky brightness. Brighter colours indicate higher light levels.

Extended Data Fig. 2 | Anthropogenic component of sound levels (L50, A-weighted dB) across the contiguous United States. Brighter colours indicate higher sound levels. Sound levels used in analyses were exceedance

values, calculated by the logarithmic subtraction of the natural from the existing sound projections.

Article

Extended Data Fig. 3 | Exposure to noise and light. Reproduction or breeding phenology was influenced by noise or light for most species, and mean exposure to noise and to light per species were positively correlated (solid black line, Spearman’s correlation test; n = 27, rho = 0.830, P = 100 nests.

Data collection

Nesting data were collected by thousands of citizen scientists volunteers via Cornell Lab of Ornithology’s NestWatch Program. Additional data were collected from existing sources: Anthropogenic noise levels from the National Park Service Data Store (https:// irma.nps.gov/DataStore/Reference/Profile/2217356); New World Atlas of Artificial Night Sky Brightness (https://dataservices.gfzpotsdam.de/contact/showshort.php?id=escidoc:1541893&contactform), the 2011 U.S. National Land Cover Database (https:// www.mrlc.gov/data/nlcd-2011-land-cover-conus-0); U.S. Human population density (https://data.census.gov/cedsci/); EltonTraits 1.0 database (http://www.esapubs.org/archive/ecol/E095/178/),, Birds of North America Online (recently changed to Birds of the World, https://birdsoftheworld.org/bow/home), vocal frequency (https://doi.org/10.5061/dryad.75nn1932), body morphologies (https:// doi.org/10.6084/m9.figshare.3527864.v1).

nature research | reporting summary

Field-specific reporting

Timing and spatial scale We included nests that were monitored between 3 March, 2000 and 24 September, 2014 and for which reliable reproductive metrics were taken. This date range was used because it most closely matched available geospatial data used in our analyses. To generate insights that were as generalizable as possible and included as many species as possible and different environmental features, the spatial scale of the study was set to the contiguous United States.

Data exclusions

There were 186,705 nests in the database initially. Using pre-established criteria to maximize nest data precision and plausibility, this was reduced to 58,506 unique nests. See supplemental text.

Reproducibility

We conducted extensive sensitivity analyses to explore model robustness consider how spatial autocorrelation, and multicollinearity could influence interpretation. Moreover, our data package provides sample code for repeating all analyses.

Randomization

We took advantage of already available nesting data and geospatial data, thus did not randomize as one would in a predetermined design.

Blinding

Data included in this study were retrieved from many independent sources and had been generated for different purposes. Blinding to group allocation was not relevant in this study.

Did the study involve field work?

Yes

No

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

Materials & experimental systems

Methods

n/a Involved in the study

n/a Involved in the study ChIP-seq

Eukaryotic cell lines

Flow cytometry

Palaeontology

MRI-based neuroimaging

Animals and other organisms

October 2018

Antibodies

Human research participants Clinical data

2

Article

The major genetic risk factor for severe COVID-19 is inherited from Neanderthals https://doi.org/10.1038/s41586-020-2818-3

Hugo Zeberg1,2 ✉ & Svante Pääbo1,3 ✉

Received: 3 July 2020 Accepted: 22 September 2020 Published online: 30 September 2020 Check for updates

A recent genetic association study1 identified a gene cluster on chromosome 3 as a risk locus for respiratory failure after infection with severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2). A separate study (COVID-19 Host Genetics Initiative)2 comprising 3,199 hospitalized patients with coronavirus disease 2019 (COVID-19) and control individuals showed that this cluster is the major genetic risk factor for severe symptoms after SARS-CoV-2 infection and hospitalization. Here we show that the risk is conferred by a genomic segment of around 50 kilobases in size that is inherited from Neanderthals and is carried by around 50% of people in south Asia and around 16% of people in Europe.

The COVID-19 pandemic has caused considerable morbidity and mortality, and has resulted in the death of over a million people to date3. The clinical manifestations of the disease caused by the virus, SARS-CoV-2, vary widely in severity, ranging from no or mild symptoms to rapid progression to respiratory failure4. Early in the pandemic, it became clear that advanced age is a major risk factor, as well as being male and some co-morbidities5. These risk factors, however, do not fully explain why some people have no or mild symptoms whereas others have severe symptoms. Thus, genetic risk factors may have a role in disease progression. A previous study1 identified two genomic regions that are associated with severe COVID-19: one region on chromosome 3, which contains six genes, and one region on chromosome 9 that determines ABO blood groups. Recently, a dataset was released by the COVID-19 Host Genetics Initiative in which the region on chromosome 3 is the only region that is significantly associated with severe COVID-19 at the genome-wide level (Fig. 1a). The risk variant in this region confers an odds ratio for requiring hospitalization of 1.6 (95% confidence interval, 1.42–1.79) (Extended Data Fig. 1). The genetic variants that are most associated with severe COVID19 on chromosome 3 (45,859,651–45,909,024 (hg19)) are all in high linkage disequilibrium (LD)—that is, they are all strongly associated with each other in the population (r2 > 0.98)—and span 49.4 thousand bases (kb) (Fig. 1b). This ‘core’ haplotype is furthermore in weaker linkage disequilibrium with longer haplotypes of up to 333.8 kb (r2 > 0.32) (Extended Data Fig. 2). Some such long haplotypes have entered the human population by gene flow from Neanderthals or Denisovans, extinct hominins that contributed genetic variants to the ancestors of present-day humans around 40,000–60,000 years ago6,7. We therefore investigated whether the haplotype may have come from Neanderthals or Denisovans. The index variants of the two studies1,2 are in high linkage disequilibrium (r2 > 0.98) in non-African populations (Extended Data Fig. 3). We found that the risk alleles of both of these variants are present in a homozygous form in the genome of the Vindija 33.19 Neanderthal, an approximately 50,000-year-old Neanderthal from Croatia in southern Europe8. Of the 13 single nucleotides polymorphisms constituting the core haplotype, 11 occur in a homozygous form in the Vindija 33.19

Neanderthal (Fig. 1b). Three of these variants occur in the Altai9 and Chagyrskaya 810 Neanderthals, both of whom come from the Altai Mountains in southern Siberia and are around 120,000 and about 60,000 years old, respectively (Extended Data Table 1), whereas none of the variants occurs in the Denisovan genome11. In the 333.8-kb haplotype, the alleles associated with risk of severe COVID-19 similarly match alleles in the genome of the Vindija 33.19 Neanderthal (Fig. 1b). Thus, the risk haplotype is similar to the corresponding genomic region in the Neanderthal from Croatia and less similar to the Neanderthals from Siberia. We next investigated whether the core 49.4-kb haplotype might be inherited by both Neanderthals and present-day people from the common ancestors of the two groups that lived about 0.5 million years ago9. The longer a present-day human haplotype shared with Neanderthals is, the less likely it is to originate from the common ancestor, because recombination in each generation will tend to break up haplotypes into smaller segments. Assuming a generational time of 29 years12, the local recombination rate13 (0.53 cM per Mb), a split between Neanderthals and modern humans of 550,000 years9 and interbreeding between the two groups around 50,000 years ago, and using a published equation14, we exclude that the Neanderthal-like haplotype derives from the common ancestor (P = 0.0009). For the 333.8-kb-long Neanderthal-like haplotype, the probability of an origin from the common ancestral population is even lower (P = 1.6 × 10−26). The risk haplotype thus entered the modern human population from Neanderthals. This is in agreement with several previous studies, which have identified gene flow from Neanderthals in this chromosomal region15–21 (Extended Data Table 2). The close relationship of the risk haplotype to the Vindija 33.19 Neanderthal is compatible with this Neanderthal being closer to the majority of the Neanderthals who contributed DNA to present-day people than the other two Neanderthals10. A Neanderthal haplotype that is found in the genomes of the present human population is expected to be more similar to a Neanderthal genome than to other haplotypes in the current human population. To investigate the relationships of the 49.4-kb haplotype to Neanderthal and other human haplotypes, we analysed all 5,008 haplotypes in the 1000 Genomes Project22 for this genomic region. We included

Max Planck Institute for Evolutionary Anthropology, Leipzig, Germany. 2Department of Neuroscience, Karolinska Institutet, Stockholm, Sweden. 3Okinawa Institute of Science and Technology, Onna-son, Japan. ✉e-mail: [email protected]; [email protected]

1

610 | Nature | Vol 587 | 26 November 2020

a

b

1.0 0.9

14

0.8

Linkage disequilibrium (r2)

12

–log10(P)

10 8 6

0.7 0.6 0.5 0.4 0.3 0.2

4

0.1 0 45.6

15 16 17 18 19 20 21 22

14

13

12

11

9

10

8

7

6

5

4

3

2

1

2

45.7

45.8

45.9

46.0

46.1

46.2

46.6

Chromosome 3 coordinate (Mb)

Chromosome

LZTFL1

LIMD1

CCR1

XCR1 CXCR6

SACM1L

FYCO1

CCR9

SLC6A20

CCR3

XV III XLI I XXX III XXXV

XI

XX

V

I

IX

XX

str

II

ce

L

VI XL

An

XX

XIV XII I

X XI

VI

V

VIII

IX

suggested that the Neanderthal haplotype has been positively selected in Bangladesh25. At this point, we can only speculate about the reason for this—one possibility is protection against other pathogens. It is also possible that the haplotype has decreased in frequency in east Asia VII

all positions that are called in the Neanderthal genomes and excluded variants found on only one chromosome and haplotypes seen only once in the 1000 Genomes Project data. This resulted in 253 present-day haplotypes that contained 450 variable positions. Figure 2 shows a phylogeny relating the haplotypes that were found more than 10 times (see Extended Data Fig. 4 for all haplotypes). We find that all risk haplotypes associated with severe COVID-19 form a clade with the three high-coverage Neanderthal genomes. Within this clade, they are most closely related to the Vindija 33.19 Neanderthal. Among the individuals in the 1000 Genomes Project, the Neanderthal-derived haplotypes are almost completely absent from Africa, consistent with the idea that gene flow from Neanderthals into African populations was limited and probably indirect20. The Neanderthal core haplotype occurs in south Asia at an allele frequency of 30%, in Europe at an allele frequency of 8%, among admixed Americans with an allele frequency of 4% and at lower allele frequencies in east Asia23 (Fig. 3). In terms of carrier frequencies, we find that 50% of people in South Asia carry at least one copy of the risk haplotype, whereas 16% of people in Europe and 9% of admixed American individuals carry at least one copy of the risk haplotype. The highest carrier frequency occurs in Bangladesh, where more than half the population (63%) carries at least one copy of the Neanderthal risk haplotype and 13% is homozygous for the haplotype. The Neanderthal haplotype may thus be a substantial contributor to COVID-19 risk in some populations in addition to other risk factors, including advanced age. In apparent agreement with this, individuals of Bangladeshi origin in the UK have an about two times higher risk of dying from COVID-19 than the general population24 (hazard ratio of 2.0, 95% confidence interval, 1.7–2.4). It is notable that the Neanderthal risk haplotype occurs at a frequency of 30% in south Asia whereas it is almost absent in east Asia (Fig. 3). This extent of difference in allele frequencies between south and east Asia is unusual (P = 0.006, Extended Data Fig. 5) and indicates that it may have been affected by selection in the past. Indeed, previous studies have

XIX

Genomes Project. Red circles indicate genetic variants for which the alleles are correlated to the risk variant (r2 > 0.1) and the risk alleles match the Vindija 33.19 Neanderthal genome. The core Neanderthal haplotype (r2 > 0.98) is indicated by a black bar. Some individuals carry longer Neanderthal-like haplotypes. The location of the genes in the region are indicated below using standard gene symbols. The x axis shows hg19 coordinates.

XL

Fig. 1 | Genetic variants associated with severe COVID-19. a, Manhattan plot of a genome-wide association study of 3,199 hospitalized patients with COVID-19 and 897,488 population controls. The dashed line indicates genomewide significance (P = 5 × 10 −8). Data were modified from the COVID-19 Host Genetics Initiative2 (https://www.covid19hg.org/). b, Linkage disequilibrium between the index risk variant (rs35044562) and genetic variants in the 1000

al

I

VI

XX

VI XX XV

I XVI V X I

XVIII Altai

XXXIX XXX

Chagyrskaya

97

XXXII

Vindija

100

XLVII VI XXX I XLV

I IV

II

XI

III

XX

I

XX

V II

LI

XXVIII

XL

XX IX XLI

I

XX

XV

XL

LII

LIV

LII

XXIV

LV LVI

XX

III

II

XX

I XL

XIV

V XL

III

XX

0.01

Fig. 2 | Phylogeny relating the DNA sequences that cover the core Neanderthal haplotype in individuals from the 1000 Genomes Project and Neanderthals. The coloured area highlights the haplotypes that carry the risk allele at rs35044562—that is, the risk haplotypes for severe COVID-19. Arabic numbers indicate bootstrap support (100 replicates). The phylogeny is rooted with the inferred ancestral sequence of present-day humans. The three Neanderthal genomes carry no heterozygous positions in this region. Scale bar, number of substitutions per nucleotide position.

Nature | Vol 587 | 26 November 2020 | 611

Article

Fig. 3 | Geographical distribution of the Neanderthal core haplotype that confers risk for severe COVID-19. Pie charts show the minor allele frequency at rs35044562. Frequency data were obtained from the 1000 Genomes Project 22. Map source data were obtained from OpenStreetMap23.

owing to negative selection, perhaps because of coronaviruses or other pathogens. In any case, the COVID-19 risk haplotype on chromosome 3 is similar to some other Neanderthal and Denisovan genetic variants that have reached high frequencies in some populations owing to positive selection or drift14,26–28, but it is now under negative selection owing to the COVID-19 pandemic. It is currently not known what feature in the Neanderthal-derived region confers risk for severe COVID-19 and whether the effects of any such feature are specific to SARS-CoV-2, to other coronaviruses or to other pathogens. Once the functional feature is elucidated, it may be possible to speculate about the susceptibility of Neanderthals to relevant pathogens. However, with respect to the current pandemic, it is clear that gene flow from Neanderthals has tragic consequences.

Online content Any methods, additional references, Nature Research reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at https://doi.org/10.1038/s41586-020-2818-3.

1.

Ellinghaus, D. et al. Genomewide association study of severe COVID-19 with respiratory failure. N. Engl. J. Med. https://doi.org/10.1056/NEJMoa2020283 (2020). 2. COVID-19 Host Genetics Initiative. The COVID-19 Host Genetics Initiative, a global initiative to elucidate the role of host genetic factors in susceptibility and severity of the SARS-CoV-2 virus pandemic. Eur. J. Hum. Genet. 28, 715–718 (2020). 3. WHO. Coronavirus disease (COVID-19) Weekly Epidemiological Update and Weekly Operational Update: Weekly Epidemiological Update 14 September 2020 https:// www.who.int/emergencies/diseases/novel-coronavirus-2019/situation-reports (2020). 4. Vetter, P. et al. Clinical features of COVID-19. Br. Med. J. 369, m1470 (2020). 5. Zhou, F. et al. Clinical course and risk factors for mortality of adult inpatients with COVID-19 in Wuhan, China: a retrospective cohort study. Lancet 395, 1054–1062 (2020). 6. Green, R. E. et al. A draft sequence of the Neandertal genome. Science 328, 710–722 (2010). 7. Sankararaman, S., Patterson, N., Li, H., Pääbo, S. & Reich, D. The date of interbreeding between Neandertals and modern humans. PLoS Genet. 8, e1002947 (2012). 8. Prüfer, K. et al. A high-coverage Neandertal genome from Vindija Cave in Croatia. Science 358, 655–658 (2017).

612 | Nature | Vol 587 | 26 November 2020

9.

Prüfer, K. et al. The complete genome sequence of a Neanderthal from the Altai Mountains. Nature 505, 43–49 (2014). 10. Mafessoni, F. et al. A high-coverage Neandertal genome from Chagyrskaya Cave. Proc. Natl Acad. Sci. USA 117, 15132–15136 (2020). 11. Meyer, M. et al. A high-coverage genome sequence from an archaic Denisovan individual. Science 338, 222–226 (2012). 12. Langergraber, K. E. et al. Generation times in wild chimpanzees and gorillas suggest earlier divergence times in great ape and human evolution. Proc. Natl Acad. Sci. USA 109, 15716–15721 (2012). 13. Kong, A. et al. A high-resolution recombination map of the human genome. Nat. Genet. 31, 241–247 (2002). 14. Huerta-Sánchez, E. et al. Altitude adaptation in Tibetans caused by introgression of Denisovan-like DNA. Nature 512, 194–197 (2014). 15. Sankararaman, S. et al. The genomic landscape of Neanderthal ancestry in present-day humans. Nature 507, 354–357 (2014). 16. Vernot, B. & Akey, J. M. Resurrecting surviving Neandertal lineages from modern human genomes. Science 343, 1017–1021 (2014). 17. Vernot, B. et al. Excavating Neandertal and Denisovan DNA from the genomes of Melanesian individuals. Science 352, 235–239 (2016). 18. Steinrücken, M., Spence, J. P., Kamm, J. A., Wieczorek, E. & Song, Y. S. Model-based detection and analysis of introgressed Neanderthal ancestry in modern humans. Mol. Ecol. 27, 3873–3888 (2018). 19. Gittelman, R. M. et al. Archaic hominin admixture facilitated adaptation to out-of-Africa environments. Curr. Biol. 26, 3375–3382 (2016). 20. Chen, L., Wolf, A. B., Fu, W., Li, L. & Akey, J. M. Identifying and interpreting apparent Neanderthal ancestry in African individuals. Cell 180, 677–687 (2020). 21. Skov, L. et al. The nature of Neanderthal introgression revealed by 27,566 Icelandic genomes. Nature 582, 78–83 (2020). 22. The 1000 Genomes Project Consortium. A global reference for human genetic variation. Nature 526, 68–74 (2015). 23. OpenStreetMap. Planet OSM. https://planet.osm.org/ (2017). 24. Public Health England. COVID-19: Review of Disparities in Risks and Outcomes. https:// www.gov.uk/government/publications/covid-19-review-of-dis parities-in-risks-and-outcomes (2020). 25. Browning, S. R., Browning, B. L., Zhou, Y., Tucci, S. & Akey, J. M. Analysis of human sequence data reveals two pulses of archaic Denisovan admixture. Cell 173, 53–61 (2018). 26. Dannemann, M., Andrés, A. M. & Kelso, J. Introgression of Neandertal- and Denisovan-like haplotypes contributes to adaptive variation in human Toll-like receptors. Am. J. Hum. Genet. 98, 22–33 (2016). 27. Zeberg, H., Kelso, J. & Pääbo, S. The Neandertal progesterone receptor. Mol. Biol. Evol. 37, 2655–2660 (2020). 28. Zeberg, H. et al. A Neanderthal sodium channel increases pain sensitivity in present-day humans. Curr. Biol. 30, 3465–3469 (2020). Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. © The Author(s), under exclusive licence to Springer Nature Limited 2020

Methods Linkage disequilibrium was calculated using LDlink 4.129 and alleles were compared to the archaic genomes8–11 using tabix30 (HTSlib 1.10). Haplotypes were constructed from the phase 3 release of the 1000 Genomes Project22 as described. Phylogenies were estimated with phyML 3.331 using the Hasegawa–Kishino–Yano-8532 substitution model with a gamma shape parameter and the proportion of invariant sites estimated from the data. The probability of observing a haplotype of a particular length or longer owing to incomplete lineage sorting was calculated as previously described14. The inferred ancestral states at variable positions among present-day humans were taken from Ensembl33. The distribution of frequency differences of Neanderthal haplotypes between east and south Asia was computed by filtering diagnostic Neanderthal variants (fixed positions in the three high-coverage Neanderthal genomes and the Neanderthal allele missing in 108 Yoruba individuals) using a published introgression map20, followed by pruning using PLINK1.9034 (r2 cut-off of 0.5 in a sliding window of 100 variants) and allele frequency assessment in the 1000 Genomes Project. Maps displaying allele frequencies and linkage disequilibrium in different populations were made using Mathematica 11.0 (Wolfram Research) and OpenStreetMap data. For the meta-analysis carried out by the COVID-19 Host Genetics Initiative2, participants consented and ethical approvals were obtained (https://www.covid19hg.org/partners/). The following eight studies contributed to the meta-analysis of hospitalization versus population controls: Genetic modifiers for COVID-19-related disease ‘BelCovid’ (Université Libre de Bruxelles, Belgium), Genetic determinants of COVID-19 complications in the Brazilian population ‘BRACOVID’ (University of Sao Paulo, Brazil), deCODE (deCODE Genetics, Iceland), FinnGen (Institute for Molecular Medicine Finland, Finland), GEN-COVID (University of Siena, Italy), Genes & Health (Queen Mary University of London, UK), COVID-19-Host(age) (Kiel University and University Hospitals of Oslo and Schleswig-Holstein, Germany and Norway) and the UK Biobank (UK).

Reporting summary Further information on research design is available in the Nature Research Reporting Summary linked to this paper.

Data availability The summary statistics of the genome-wide association study that support the finding of this study are available from the COVID-19 Host Genetics Initiative (round 3, ANA_B2_V2: hospitalized patients with COVID-19 compared with population controls; https://www.covid19hg. org/). The genomes used are available from the 1000 Genomes Project (phase 3 release, https://www.internationalgenome.org/) and the Max Planck Institute for Evolutionary Anthropology (Chagyrskaya, Altai and Vindija 33.19, http://cdna.eva.mpg.de/neandertal/). The ancestral alleles are available at Ensembl (release 100, https://www.ensembl. org/). Map data are from OpenStreetMap and available from https:// www.openstreetmap.org. 29. Machiela, M. J. & Chanock, S. J. LDlink: a web-based application for exploring population-specific haplotype structure and linking correlated alleles of possible functional variants. Bioinformatics 31, 3555–3557 (2015). 30. Li, H. Tabix: fast retrieval of sequence features from generic TAB-delimited files. Bioinformatics 27, 718–719 (2011). 31. Guindon, S. et al. New algorithms and methods to estimate maximum-likelihood phylogenies: assessing the performance of PhyML 3.0. Syst. Biol. 59, 307–321 (2010). 32. Hasegawa, M., Kishino, H. & Yano, T. Dating of the human–ape splitting by a molecular clock of mitochondrial DNA. J. Mol. Evol. 22, 160–174 (1985). 33. Yates, A. D. et al. Ensembl 2020. Nucleic Acids Res. 48, D682–D688 (2020). 34. Chang, C. C. et al. Second-generation PLINK: rising to the challenge of larger and richer datasets. Gigascience 4, 7 (2015).

Acknowledgements We thank the COVID-19 Host Genetics Initiative for making the data from the genome-wide association study available, and the Max Planck Society and the NOMIS Foundation for funding. Author contributions H.Z. performed the haplotype analysis. H.Z. and S.P. jointly wrote the manuscript. Competing interests The authors declare no competing interests. Additional information Supplementary information is available for this paper at https://doi.org/10.1038/s41586-0202818-3. Correspondence and requests for materials should be addressed to H.Z. or S.P. Peer review information Nature thanks Tobias Lenz, Yang Luo and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Peer reviewer reports are available. Reprints and permissions information is available at http://www.nature.com/reprints.

Article

Extended Data Fig. 1 | Odds ratios for hospitalization owing to COVID-19 for cohorts contributing to the meta-analysis (round 3) of the COVID-19 Host Genetics Initiative (rs35044562). The odds ratio and the P value for the summary effect are odds ratio = 1.60 (95% confidence interval, 1.42–1.79) and P = 3.1 × 10 −15 (two-sided z-test, n = 3,199 patients with COVID-19 and 897,488

controls over 8 independent studies). Data are the odds ratios and 95% confidence intervals. HOST(age), UK Biobank European (EUR), GENCOVID, deCODE and BelCovid use European population controls. BRACOVID, Genes & Health and FinnGen use American, south Asian and Finnish population controls, respectively.

Extended Data Fig. 2 | Pairwise linkage disequilibrium between diagnostic Neanderthal variants. Heat map of linkage disequilibrium between genetic variants in which one allele is shared with three Neanderthal genomes and

missing in 108 Yoruba individuals. The black box highlights a haplotype of 333.8 kb between rs17763537 and rs13068572 (chromosome 3: 45,843,315– 46,177,096). Red, r2 correlation; blue, D′ correlation.

Article

Extended Data Fig. 3 | Linkage disequilibrium between index variant rs11385942 and the index variant of the COVID-19 Host Genetics Initiative (rs35044562). Shades of red indicate the extent of linkage disequilibrium (r2) in the populations included in the 1000 Genomes Project. Populations labelled

‘n/a’ are monomorphic for the protective allele of rs35044562. The previously described index variant (rs11385942)1 does not have any genetic variants in linkage disequilibrium (r2 > 0.8) in populations from Africa. Map source data from OpenStreetMap23.

Extended Data Fig. 4 | Phylogeny of haplotypes in individuals included in the 1000 Genomes Project and Neanderthals covering the genomic region of the core risk haplotype. The shaded area highlights a monophyletic group that contains all present-day haplotypes carrying the risk allele at rs35044562

and the haplotypes of the three high-coverage Neanderthals. Arabic numbers show bootstrap support (100 replicates). The tree is rooted with the inferred ancestral human sequence. Scale bar, number of substitutions per nucleotide position.

Article

Extended Data Fig. 5 | Frequency differences between south and east Asia for haplotypes introgressed from Neanderthals. The dashed line indicates the frequency difference for the Neanderthal haplotype that confers risk of severe COVID-19.

Extended Data Table 1 | Genetic variants in LD (r2 > 0.98) with rs35044562 and the corresponding Neanderthal variants

Data from the 1000 Genomes Project22. ‘Ref’ indicates the alleles from hg19. The three Neanderthal genomes are homozygous at these positions. LD, linkage disequilibrium.

Article Extended Data Table 2 | Previous studies that identified gene flow from Neanderthals at the core haplotype

The hg19 coordinates for the previously identified15–21 introgressed haplotypes are shown.

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

-$"$$0

12345656762589 556 53 17 7425

!"#$%&'( g#!j*! )"$#"$* *+" + #$%&'(-kl9mnmn

8"33$"$$0"3""3+904.$%"$$%4331!$."$ $%4!#3!9$"*33!9."$" 4. o 7%>0"62"6.!6>"$"3> knnn!.[0$&%"q3"'(%$$(>>1116$"$"3!.6!> :.*3&3"knn'(%$$(>>1116.*36!>

0

o )400 a%"2#"3h0"300 :03!0"3923#$"+h2.$"300 8"400+4$%0#.$1$%"330$9"$#60.>0#.$>C$!C#.."+C43"$64

)A334$ 0   0   $ #  +   !  #.#$03 $  %$21%$%03# !"$26

e "33"2"3"*3%!%C02"!/"$"3!.&xq'67%".3? e# 4$%feA-&q9ktt0""lty9pll0$3'1"3.$ *+$ +%"$"24.$%0%$6e e#"33!. $%%"q3"&xmznp'4$%knnn!.[0$6 $%!."$4."$24$% B"$"#4$ !($"8&7746&$&( ((# t >#4$ !($"8&"/& '2!($" 44($"='4)& ((8"2&7$(/&"#&#`'(2"(82'7($!742!&$" '77#4$!($"8() (&($($4&7!&&2($"47'#$"%4"(&7("#"4/*:%:2&"+().&$4($2&( *:%:% $"488$4$"(+ t >8 >3G6&$&($"*:%:(&"#&##6$&($"+& 4$&(#($2&( 8'"4(&$"(/*:%:4"8$#"4$"(6&7+ 7)/!()$(($"%=()(((&($($4*:%:a=P=W+5$()4"8$#"4$"(6&7=884($E=#% 88#2&"#b6&7'"(# t A1*w'$4(&7:qsxy+=.#(7q:qz:x=C@C@{:s:|=l}C@0r:xs=01('#$6x:x:r{p=0 G&(&&"&7/$ |:{:x&"#|:p:x=.5($x:q:x:x=1;>0q:z:s8=01@Cx:|:x=;$2k&7s:r:{=#%0|:qs:y=A$4&#;7x:xrx=C>1qq:x:x=G$88f$"#q:p:p= G@1mqx:x~:x=-G1x:s:s=G>vGp:~=C>kC>x:sz.=%!8$7qs:x:~=&"#1C>0;H€x:x:

772&"'4$!(2'($"47'#&#&(&&6&$7&.$7$(/(&(2"(:;)$ (&(2"()'7#!6$#()8775$"%$"82&($"=5)&!!7$4&.7,



H>44 $"4#='"$m'$#"($8$=5.7$"98!'.7$47/&6&$7&.7#&(&( H>7$(88$%' ()&()&6& 4$&(#&5#&(& H>#4$!($"8&"/($4($" "#&(&&6&$7&.$7$(/ >77#&(&('#$"()$ ('#/*A->Hm=>;>Hm=03>Hm+&&6&$7&.7&(()3'4$"4C'7($H}2$4>4)$6*3C}>4)$6+'"#4"(77#&44 : )2&($"$"(&4($"=!"4)2&($"!&9=&"#%"B! $"!8$7 8&4)477(/!4&".#5"7&##82()3C}>4)$6'$"%()8775$"%7$"9, )((!,DD& (:"2&4)$6:%D#&(H'$m/~. 77(/!H!4$8$4|G!$%"2 4&".6$'&7$E#"()j&)@!$%"2f5 '$"%()#&(&)'.&(()8775$"%7$"9,)((!,DD!$%"2%&(5&/:5'(7:#'D .5 D‚%"2ƒ)%|~?!$($"ƒ4)xz,zqyx~q|~Hz||rypz{?)'.ƒ)((!,DD)"H2 "%:|H'H5(Hx:&2&E"&5 :42D)8.„'.2$$"D)8.„#&(&)'.:`"

0

t -$ 8 4$"4 f)&6$'&7? 4$&74$"4 @47%$4&7=67'($"&/?"6$"2"(&74$"4 "($.#$ '#

&7$#&($"

C()#

"D& v"676#$"() ('#/ t t

t

)vAHm upHA@*fGf$4$"4=&({px{{q= 7"}x~Hx||s=-(~x~zp~p=xDxs#$7'($"+=1}uqHAAH/{:{*fGf$4$"4=&({px{sp=7"}|~Hpz~=-(~xp{zrr= xDxs#$7'($"+=1>;fqH>7B&87'prz*>.4&2=&(&.xyp{|p=7"@A3v0x|s>=-(k0|qs~xs|Hv&"#k0qq~zrzHq=xDxss #$7'($"+=2)/*>.4&2=&(&.qs{rsq=-(k0|qzxzrrHq=xDqss#$7'($"+=k.4&2=&(&.xqx~=7"y=-(k0|q{qsx{Hq=xD|ss#$7'($"+=>7B&7B&xxsrq=-(xyzzzsz=xD{ss#$7'($"+: @}C@1,;fq,4m'$$($"8Ž{=sss6"(54774(#'$"%&q{2j0#17$#1(&(G$#7&*p|{"2+ &"#pz{D|s.&"#!& 8$7(:;)$&"($.#/%&6&!$($6 $%"&7$">r|x4778$B#5$()r†82&7#)/#*xs2$"+D!2&.$7$E# 5$()s:x†Af1H;$("uHxss8x{2$"'#'"#() &24"#$($":Š 2)/, )$49"!7/47"&7(2)/:v22'"%",042.$"&"(8'77H7"%()!($"*l$H(&%+4!"#$"%(2)/:;) $22'"%" m'"4$82()8775$"%8"4*1)&"3(&7:3&('f$(4)"7%/qq,x{pzHx{zq=qssr:+:;(# &!!7$4&($",'$(&.78jf=vDv;fq*'#$"0v1A06$5+ 0&..$(2"47"&7@A3v0x|s>(1>;fq:;(#&!!7$4&($" $"47'#$"%jf=vlHA=vDv7B&$&vv=1>$&vv'=1>$&"'2$ 4&76&7'8"'2.8477!4"(&%*5$() (&($($4+$!6$##:

‹

k&($"% (&(%/

&  #. ./ /'""$"%() (#477!!'7&($" .&49 "( " ) &22&4)$"&"#&"%#82~{Hy{†:@6"(()&(&!!&# '($#()$%$"&7%&( 57&%7/""H87'4"(: *>.*>'#"00#.* :=BCkkluq9:=BCkkklq9":=BCkklut67%0"$1$ $%;$B"$"a"5&%$$(>>11160*6!>'1$%"00#.*qnjw9qnyB9 "qnjA6



;30+4."$"*#$"2"3"*3$+ 4"$" A33."#0$.#$03#""$""2"3"*3$+$"$.$7%$"$.$%#32$%4331!4."$91%"30"*3(

0

m )400 a%"2#"3h0"300 :03!0"3923#$"+h2.$"300 8"400+4$%0#.$1$%"330$9"$#60.>0#.$>C$!C#.."+C43"$64

)A334$ 0   0   $ #  +   !  #.#$03$%$21%$%03#!"$26

/$"$$0"3.$%1#$$.".3?6s*3!0"33+30"$&{s'14."0"$$%4!# 3!60+:="$"$0$4$%#"4$."!9"0*=$%6 B"$"ojolk "**$"$C.#0fA-&soqjw933-!"3!70%3!+'9-0&=#'9A30"$&ea9r;'6A$$%."#4"0$#b 1*$9$%0$"$ow6%$$(>>1116033!"360.>#0$>."+C"$*>0!"CsvC"**$C."*C.#C040>soqjw "**$"$C%#."-7r/f&osqut933-!"3!70%3!+'9-0&g#."9=#'9A30"$&ea9r;9rg'6A$$% ."#4"0$#b1*$9$%0$"$ws6%$$(>>1116033!"360.>#0$>."+C"$*>$!Ckk4C"**$C."*>osqut "$"$C;A,;o&=Aaqll9,hB-+$.'9-0&g#."9=#'9A30"$&ea'6A$$%."#4"0$#b1*$9$% 0$"$u6%$$(>>1116+$.60.>#0$>%#."C.#C"C"$*+Cjsloj}."*qll

z

12345656762589 556 53 17 7425

"**$"$Cg$gs gs&otoqvCo9$$0%'9-0&g#."9=#9,"$'9A30"$&ea9rg9r8'6A  $ $  % ."#4"0$#b1*$9$%0$"$ kok6%$$(>>1116$!3"*60.>#0$>g$CgsCA$*+CotoqvCoCA;6%$.#0$C 4."$ "$"$CgA&oovqtuksllo9,0%'9A30"$&ea'6A  $ $  %."#4"0$#b1*$9$%0$"$ s s6%$$(>> 1116!.""30%60.>0"$"3!>#0$>0%>"%"%"‚3"!{h!{p.#"$C|C$#*#3&Aa}oojtwoo9B-ga'9-0&g#."9=#'9A30"$&ea9r89rg'6A  $ $  %."#4"0$#b 1*$90  $"$4."$ 4#6%$$(>>%*6*3!+6#1"6#>okfolC"$C"3%"C$#*#3 .#"$CfA;Bg&qlllluCo9;$$0%'9-0&g#."9=#'9A30"$&ea9r89r;'6A  $ $  %."#4"0$#b 1*$9$%0$"$ kulj6%$$(>>1116$!3"*60.>#0$>fA;BgCA$*+CqllluCoCr!6%$. .#"$C"0$&=AjCoovqw97%.8%-0$40'9-0&g#."9=#'9A30"$&ea9r89r;9rg'6A  $ $  % ."#4"0$#b1*$9$%0$"$ kvw6%$$(>>1116$%.4%60.>"$*+>#0$>A0$CA$*+C03CA7/ljCuC =03"3>=AjCoovqw ;0"$"3!> #0$>tojClsjCojo ;0"$"3!> #0$>tooClsjCojk ;>1116$%.4%60.>"$*+>#0$>B5+C"$C =#Cr!fCgC)Cg!%3+CCA*C-0"+CA$*+C;3+03"3>ACsojtl A$C"**$r!f9g,;C35A$*+&tltu933-!"3!'9-0&,"**$'9A30"$&ea'6A  $ $  %."#4"0$#b 1*$9$%0$"$ sukk6%$$(>>1116033!"360.>#0$>0"+C"$*>"$C"**$C!!C%C35C"$*+>tltu A$C.#r!f9g,;C35A$*+&tltq933-!"3!'9-0&=#'9A30"$&ea'9A  $ $  %."#4"0$#b 1*$9$%0$"$ ovkl6%$$(>>1116033!"360.>#0$>0"+C"$*>"$C.#C!!C%C35C"$*+>tltq A30"$"3!> #0$>tooCjujCojk A30"$"3!>#0$>tokCjujCojs

:#5"+$00333

;30+4."$"*#$0333

g:ƒkws70331#0%"4.A7&"$,)Cskoq'6g)"0331#0%"4.A76;."+0fA-C>C=:8 1!"$4.:os6j.*+6 A33033315$"$31""! $."  $"$%$$+67%033%"2*"#$%$0"$* *+2 + "3"$! A#$%$0"$ 0fA-9-7r/f9"0$* *+1 + $*3$$! ..#43#006 A3303331$$!"$24.+03"."0$"."$6 =+03"."0$"."$ ..3+.$43 / /0..3+.$403331# $%156 &-r)A!$'

333#0&'

A."3"$%!".

;30+4."$"*#$$#232!"."3€A,,rx:!#30..4$!"."3"0%

7.."$! 40fA-C>C."3"4."3*!"&y"05)"*.#$"lkqjju'1$"*3%1%$%.01 *$1uCv.$% 4"!6=:81!"$4.!"$4."3os6j"+$4$3?"$6 B$23213"."36 e 3"."3 ".36 83C0330$".3 B$232430330$0. A."3