news

AI drug discovery booms in Chinese upstart companies and IT goliaths hope to turn country’s artifcial intelligence prowess into world-leading drug innovation. Do they have an edge over Western players?

hinese researchers are abuzz about the potential of artificial intelligence C(AI) in drug design and development. Just last month, three Chinese players raised hefty sums to fuel their drug discovery plans. Insilico Medicine, a Hong Kong–based company with half its global operations in , raised $255 million to push drug candidates discovered by its AI into clinical trials and tune its algorithms to find new targets. That followed $100 million pocketed by StoneWise Technology in April and $319 million that Shenzhen-based XtalPi added to its coffers last September. IT goliaths like , and ByteDance have also been turning their formidable AI capacity to drug design. With an ambitious national strategy aimed at AI dominance, a massive talent pool of AI researchers, a plethora of contract research organizations (CROs) with which to partner, and evolving health data privacy protections, Chinese AI companies are T he Chinese government’s regulatory reform is designed to boost AI-driven drug discovery, but attitudes poised to compete in the global race to toward data privacy continue to cause international concern. Credit: THANANIT SUNTIVIRIYANON / make more drugs faster and cheaper. But Alamy Stock Photo even if a population of 1.4 billion can supply the wealth of data that will give Chinese AI companies a head start over Western counterparts, questions remain as to the Since then, though, the situation has for new pipelines and with it a curiosity to quality of the preclinical and clinical data changed rapidly. AlphaGo’s victories over see what AI had to offer. In 2019 Hansoh being used to train company algorithms. top human Go players in 2016 and 2017 Pharmaceutical Group, for example, agreed “If you come here and see what is focused industry attention on the potential to pay fees that could amount to over a happening, it impresses you,” says Alex of AI as a path to success in the minds billion dollars to Atomwise to design Zhavoronkov, founder and CEO of Insilico of many young Chinese. The following drugs targeting up to 11 proteins involved Medicine. Kai-Fu Lee, CEO of Sinovation year, the central government launched an in cancer and other disorders. And while Ventures, a venture capital firm based in ambitious framework to make China a XtalPi has received most of it partnering Beijing that invests in AI companies, agrees: global leader in AI by 2030 that led money from US companies, China is “China does have a state-of-the-art ability to to, among other things, new university now its fastest growing market. “For build drug discovery systems.” training programs. [Chinese] companies with little experience But Chinese AI drug hunters are of innovation, AI represented a way to latecomers. Over the past decade, companies “Top AI experts in China cost leapfrog,” says Ma. like Atomwise, BenevolentAI (London), more than in the Researchers in China responded to Exscientia (Oxford), Relay Therapeutics and the call. In a “storm of activity,” computer Numerate were working out big deals with and tend to move around.” scientists, physicists, mathematicians and multinational pharmaceutical companies. traditional drug discovery researchers Even the two most prominent AI drug Around the same time, Chinese launched over 50 AI drug discovery design companies now associated with regulatory reform gave innovative drug companies, says Zhang Jian, who runs a China—Insilico Medicine and XtalPi—were discovery, including that based on AI, a molecular design laboratory at Shanghai both established in 2014 in the United States. boost. Updated bioequivalence rules in 2015 Jiaotong University and has worked with a At the time, Chinese AI drug research was and new volume-purchasing stipulations handful of AI drug design startups. In 2019, getting little traction. “When I first came have meant that innovative drugs are Zhang helped to reposition Shanghai-based back to China in 2015 and started talking increasingly favored over generics, which Nutshell Therapeutics, which just raised its with Chinese pharmaceutical companies, had previously dominated China’s drug second installment of $20 million in June, it was apparent that most of them were not market. The new rules tip the scales in favor as an AI drug design company. In 2018, interested in or were not ready for AI in of the quickest and most forward-looking Zhou Jielong, a key force behind the drug R&D,” says XtalPi co-founder Ma Jian. manufacturers. That created a huge demand evolution of Baidu’s search engine,

900 Nature Biotechnology | VOL 39 | August 2021 | 897–907 | www.nature.com/naturebiotechnology news established StoneWise to aid in the discovery China-based researchers account for 22.4% of small-molecule drugs. of the total, compared with 16.4% from WHO takes up CRISPR Industry insiders and drug discovery Europe and 14.6% from the United States, babies researchers in the United States have according to a recent report from Stanford More than two years after the birth remained skeptical of the AI hype and University. And for the first time last year, of twin girls with edited genomes, an claims that AI can improve the attrition AI publications from China garnered more advisory panel for the World Health rate once drug programs reach human citations than US ones. China still lacks Organization (WHO) has issued proof-of-concept studies. Chinese scientists who will make fundamental new recommendations that uphold researchers argue that, given this breakthroughs, represented by the top 1% the agency’s initial reaction: that “it conservative thinking, Chinese companies of the papers—especially the highly cited would be irresponsible to proceed with might be able to leverage the move to AI conference papers, says Lee. “When it comes clinical applications of human germline faster than Western companies: “China to really the most creative out-of-the-box genome editing.” Such concerns over doesn’t have a traditional way to do thinkers, the US still leads,” he says. manipulation of human embryos are discovery, so there’s no box. Those people But that might not matter for drug shared by the international scientific don’t have other rules in their mind,” says Yu discovery. Chinese companies can tinker community, but few countries have Zhengtian, co-founder of Nutshell, which with AI algorithms based on generative regulations in place for the responsible uses AI to predict allosteric properties of adversarial networks or generative use of genome editing techniques, and disease targets. cooperative networks—most of which were when they do have guidelines, they China’s large population and massive developed in the United States, Canada or are often unenforceable. The WHO’s hospitals make it easy to aggregate large data Europe—because they are often open source 18-person expert panel has now come up sets—a crucial requirement for training AI. and publicly accessible. “You change a few with broad recommendations that address There is also less concern about privacy in words and change the whole concept. You the scientific, ethical, social and legal China, meaning that the data are easier to can find a way to address the question you implications raised by the prospect of access. Lee says that AI companies in the have,” says Nutshell’s Yu. “The science is editing the genome of a human embryo. United States in particular are hobbled by from there [the West], but the technology The first statement sets out five regulations, such as the Health Insurance is from here.” scenarios where gene editing might be Portability and Accountability Act of 1996, Lee says that the approach of Chinese AI applied—somatic genome editing for which sets standards for how electronic drug discovery may also be more pragmatic, sickle cell disease, somatic gene editing health records can be shared. HIPAA makes something he thinks gives China an edge. for Huntington’s disease, somatic gene it difficult, even when patients have given DeepMind, a subsidiary that has editing by “unscrupulous” actors, consent, to pull data together, says Lee. stunned researchers by solving a decades somatic gene and epigenetic editing for Although the Chinese government is about old protein folding riddle, might win the enhancement, and germline editing— to come out with severe penalties for those Nobel Prize or Turing award, but Chinese and lays out the technical, ethical and who sell leaked personal data, companies companies are “very good in monetizing societal issues that each faces. The second like Yidu Cloud, which brings together AI, creating value with AI, building new set of recommendations outlines the consented patient data into a research tool, products with AI,” he says. role that the WHO should undertake are emerging to ensure the data can be The explosion of interest and in establishing a moral and scientific accessed, says Lee. Their counterparts in opportunities in AI has, however, driven leadership role. Among them are creating the US—Flatiron Health, Tempus, Aetion up wages and made it difficult to retain a registry of clinical trials involving and Palantir—are similarly consolidating expertise. Top AI experts in China cost gene editing and working with patent real-world electronic health record data on a more than in the United States and tend to holders and regulatory agencies to insure massive scale for use in drug discovery. move around, says Zhavoronkov. Insilico that the technology, once developed, is But many of the data are of poor quality, mainly hires AI expertise outside mainland widely accessible. which can mislead the AI. “Data quality is China. Insilico, whose research stretches While these guidelines present absolutely critical. Most Chinese companies from a biological hypothesis and the a go-slow approach to gene editing, working in drug discovery don’t have good identification of new biological targets to the opposite might be said about the quality data for training,” says Ye Tao, a the production and testing of candidate International Society of Stem Cell medicinal chemist who uses computer-aided molecules, mainly hires AI expertise outside Research (ISSCR) and its recommendation drug design at Shenzhen mainland China and leans more heavily on to relax the Warnock rule, which since Graduate School. Chinese research for later stage chemical 1984 has held that research on human Zhang agrees that data quality is a design and molecule production, areas in embryos should not proceed beyond problem, especially outside of a few medical which China excels. “We do geographic 14 days. The society is now calling for research hubs like Beijing and Shanghai. “It arbitrage,” he says. public discussions among stakeholders in may be only possible to get good data in big The bounty of CROs—some 3,000 different jurisdictions to decide whether cities. It’s a little complex,” he says. altogether—in China is perhaps the most experiments should be allowed to proceed China’s formidable pool of AI appealing aspect of China-based AI drug beyond the 14 day limit. researchers—comprising returnees from discovery firms, says Zhavoronkov. Insilico the United States and Europe; veterans of is partnering with 80 of them. This allows Published online: 10 August 2021 industrial research at Baidu, Alibaba and the company to avoid the loss of time and https://doi.org/10.1038/s41587-021-01027-x Tencent; and those being trained under intellectual property that would happen if new government programs supporting they worked with an academic institution. the country’s long-term AI goals—should It also allows multiple experiments to be also give the country’s sector a competitive run in the development chain in parallel at edge. Among global AI research papers, multiple CROs. Comparing the results “is a

Nature Biotechnology | VOL 39 | August 2021 | 897–907 | www.nature.com/naturebiotechnology 901 news

huge learning experience for your AI,” says The entrance of Tencent, Baidu and CRISPR super-sizes corn Zhavoronkov. ByteDance into AI drug discovery will “You’ve got enormous number of labs add weight to China’s efforts. In July 2020, that can perform all experiments for you, Tencent launched its own AI-driven drug at a reasonable cost, but at the same time in design platform, called iDrug. It already parallel,” says Zhavoronkov. “The only place has already launched over ten projects, where you can do it is China.” including a quest for drugs that work So far Insilico and XtalPi, like their against coronaviruses. The company plans Western counterparts, have reached some to cover the gamut of preclinical research, preclinical milestones. Insilico uncovered predicting protein structures, screening a biological target linked to idiopathic candidates, designing and optimizing pulmonary fibrosis and then designed a molecules, and characterizing the functions ‘novel’ molecule — at least, novel enough for of the proteins. Gene-edited corn has more rows per ear. an intellectual property claim—from Credit: Kerem Arikan / Alamy Stock Photo scratch to match that target and potentially “There’s a lot of hype, a lot of treat the disease. According to the company’s people try to sell you things press releases, the feat took less than 18 Biotech startup Pairwise Plants has months and cost only about $2 million. that just aren’t there.” launched field trials in the US Midwest, That’s a fraction of what these stages with backing from Bayer Crop Science, typically cost. And according to a press Lee says these IT giants will act as of corn with more kernel rows. The release from XtalPi, its collaboration with China’s version of Google, producing huge genetically edited corn, obtained with Signet Therapeutics identified a gastric machines like DeepMind that can take on CRISPR base editing technology, has cancer treatment in six months. The problems, like protein folding, that require more than the average 16 rows on an company says it is now moving toward more computational power than smaller ear, a trait, designed to help farmers clinical trials. Shanghai Jiaotong University’s startups can muster. “So now that we have a increase efficiency and boost yields. Zhang says two compounds he has worked small, nimble companies that may be more Pairwise was founded by CEO Tom on are awaiting approval to begin clinical pragmatic and faster to make money, and Adams and Chief Business Officer Haven trials; he expects one of the trials to start then big giants,” says Lee. “So it’s a good, Baker, with co-founders David Liu of next year. Nutshell’s oncology candidate, good division of labor.” Harvard University, Feng Zhang of the meanwhile, is couple years away from While there is some healthy competition Massachusetts Institute of Technology and clinical trials, Yu says. between China and the United States, J. Keith Joung of Massachusetts General But the success of AI drug design in many researchers are more worried Hospital, to use CRISPR base editing China, as elsewhere, will depend on that current political tensions could technology to produce new varieties whether its candidates are successful in the compromise synergies that benefit both of fruit and vegetables with improved clinic, something that no AI-designed countries in this emergent arena. Most of taste, longer shelf life, better yield, easier drug has yet achieved. “There’s a lot of the investment in Chinese AI drug design harvesting, and better adaptation to hype, a lot of people try to sell you things is from the US while a third of authors different weather conditions and season that just aren’t there,” says Tudor Oprea, a on influential AI conference papers were availability. In 2018, Pairwise raised computational biologist at the University of educated in China but are mostly working in $25 million from investors and entered New Mexico. “The proof of the pudding the United States. “So I’m really hoping that a $100 million R&D collaboration with is in the eating. So you don’t really know healthcare or drug discovery are areas that Monsanto (now Bayer Crop Science) until it’s on the market.” don’t become a part of this decoupling talk, focused on corn, soybeans, wheat, cotton Oprea says Insilico, for which he because this research is all for the good of and canola. In February 2021, Pairwise became scientific advisor, is a promising humanity,” says Lee. ❐ raised $90 million to develop its own first model because it takes on an “end-to-end” product: a salad leaf with the nutritional validation process, whereas many of the David Cyranoski elements of kale but the eating qualities others look at a very small piece of the Kyoto, Japan of green leaves that people find delicious, puzzle, making it difficult to know whether like romaine lettuce. Pairwise is also part the AI is giving a meaningful result. “There’s Published online: 3 August 2021 of a public–private partnership to identify a lot of narrow-band vision,” he says. https://doi.org/10.1038/s41587-021-01016-0 new traits in berries in the genus Rubus that allow low-labor picking and year round growth with maximum yield. Because CRISPR-edited plants can reach the market without the US Department of Agriculture’s oversight, Pairwise and its competitors, such as Caribou Biosciences and collaborator DuPont Pioneer, expect to avoid the hurdles faced by developers of traditional genetically modified crops.

Published online: 10 August 2021 https://doi.org/10.1038/s41587-021-01028-w

902 Nature Biotechnology | VOL 39 | August 2021 | 897–907 | www.nature.com/naturebiotechnology