News

For three weeks last May, employees of the AI giant Nvidia and Recursion Pharmaceuticals slept on the floor of a data center in Salt Lake City. They were there to build a machine that Recursion, a decade-old biotech company, believes will give it an edge in the contest to develop the next great new medicines: BioHive-2, the largest and fastest supercomputer ever to be used in the biopharmaceutical industry. It's an audacious bet that the future of America’s pharmaceutical industry will be as much about computing power as it is about scientific talent.

There’s a reason for the rush: AI-powered drug discovery has been in development for years, but ever since ChatGPT rocketed into the public consciousness in late 2022, the hope and hype around its potential has reached a fever pitch. The question the tech and medical worlds want answered is, when will AI bring its magic to the long, hard, terribly expensive business of pharmaceutical research and development? Is it possible—as OpenAI’s Sam Altman has mused—that one day we’ll simply ask ChatGPT to cure cancer, or Alzheimer’s, or any number of other intractable human diseases?

The race to achieve that sci-fi scenario is well underway. As of June 2023, more than $18 billion had poured into some 200 “AI-first” biotechs, and by January 2024, at least 75 drugs or vaccines from those companies had entered clinical trials, according to Boston Consulting Group . Citeline, a pharmaceutical market research firm, meanwhile, has counted 446 financing rounds totaling $30.6 billion in the AI-driven life sciences space since 2020.

Recursion is hoping to pull ahead of the crowd in a field that has so far been more promise than performance. While there has been a boom in AI-discovered compounds, none so far have made it to market as approved drugs.  Most are still in early stages of development, but some AI-discovered drugs have suffered the same dreaded fate as many traditionally developed ones: They’ve failed in human clinical trials.

It’s too early to judge the whole sector based on those setbacks, but many have been tempted to, given the sky-high expectations driven by AI enthusiasts and the success of large language models. It leaves the industry in an awkward place: Generative AI in its current form is mostly built around language processing; it hasn’t proved to be that helpful in the world of molecules—at least not yet.

But even if it isn’t yet originating new drugs, there’s no question that AI is significantly changing the drug development process. Modern drug development is a crazily inefficient pursuit: It takes, on average, well over a decade and an estimated $2.6 billion to create a single medicine. And making it to the finish line with an FDA-approved drug is no sure thing—only 5% of experimental drugs that scientists design in the lab ever get there.

AI can—and increasingly does—help do this work faster, cheaper, and with greater odds of success. And many pharmaceutical companies tell Fortune that AI is already saving time and money in several ways. Moderna , which has used machine learning tools for nearly a decade, points to a number of use cases, from optimizing mRNA sequence design to writing a several-hundred-page regulatory filing. What once engaged a whole team now just requires one human to review the computer’s work.

But the vision of companies like Recursion is grander than that: Their bet is that by pairing massive amounts of scientific data with powerful new computing tools, AI can unlock the mysteries of biology and design drugs to cure the diseases we’re plagued by.

Much like OpenAI shoveling the world’s text into a large language model to create ChatGPT, Recursion believes that by feeding its fast-accumulating cellular and medical data to BioHive-2, it can break open biology—providing the insights, long elusive to humans, needed to understand hard-to-treat conditions from cancer to neurodegenerative and autoimmune diseases.

Bullish investors convinced of this thesis include Nvidia founder and CEO Jensen Huang, whose company—the third-most valuable in the world—made a $50 million investment in Recursion in 2023. In a conversation with the biotech’s employees and investors last June, Huang compared Recursion’s opportunity to Nvidia’s at the beginning of the chip revolution a few decades ago.

“This is such a fun time for you guys… I’m jealous,” he told them. “You might be within a click or two away from really being able to understand the meaning of life.”

A little company “making bets where others are scared”

In some ways, Recursion, an 11-year-old biotech with 800 employees and zero approved medicines, is an unlikely steward for the drug industry’s mightiest computer. In 2024, the Utah-based biotech—or “TechBio,” as Recursion prefers to call itself—took in just under $59 million in revenue and recorded a net loss of $464 million. In the company’s early days, Chris Gibson, Recursion’s CEO and cofounder, frequently predicted that his company would cure 100 diseases in 10 years. Like all of its AI-native peers, it’s still working on its first.

But what Recursion is unusually rich with is data. Every week, robots in Recursion’s automated labs run as many as 2.2 million experiments—transferring various experimental solutions into miniature samples of cells—each one resulting in a high-resolution image that captures detailed cellular morphology and features. Those experiments can run on 50 human cell types and have drawn upon millions of compounds, thousands of genetic modifications, and over a trillion lab-generated neurons. The point is not about the outcome of any one experiment but rather mining the data from so many of them. Add to those reams the around-the-clock video streaming in from the company’s animal labs, where cameras are trained on hundreds of mouse and rat cages to more precisely analyze drug-induced behavioral change. At last count, Recursion had generated 40 petrabytes of data from over 300 million experiments.  And it has no plans to slow down—which, of course, is why it needs a supercomputer.

“This little company…is making bets where others are scared,” Gibson told an audience of investors. “We've made these investments because we believe the intersection of data and compute is the future of this industry. And we intend to lead it.”

Recursion has some steep competition. New entrants to the space continue to emerge, with deeper and deeper pockets, and bigger and bigger names involved. Insitro, founded by AI pioneer and McArthur genius Daphne Koller in 2018, is backed by a who’s who of biotech investors and has a reported valuation of $2.4 billion. Xaira, which employs a fresh Nobel laureate, launched with $1 billion in funding last April. And in late January, LinkedIn founder Reid Hoffman announced he was teaming up with Pulitzer Prize–winning oncologist Siddhartha Mukherjee on Manas AI to develop cancer drugs.

Gibson marvels at the almost overnight change in industry interest compared with the dismissive skepticism he encountered just a few years ago: “We mostly got laughed at,” he recalls. But when he appeared with Nvidia CEO Jensen Huang at JPMorgan’s health care conference last year, he found a room packed with Big Pharma CEOs eager to hear his pitch.

The cusp of a boom, or an overhyped moment?

For most of history, the development of new medicines depended on a combination of astute observation and luck. The ancients made the serendipitous discovery that willow leaves and myrtle—the natural precursors to aspirin—alleviated fever and joint pain. Edward Jenner came up with the smallpox vaccine from the insight that people who worked with cattle and had been exposed to cowpox, a milder disease, weren’t affected by the virus. And the blood thinner warfarin emerged from an investigation into an epidemic of “spoiled clover disease” among a population of Wisconsin dairy cows that bled to death.

Late last century, advances in genetics and molecular biology allowed scientists to more precisely focus their efforts by identifying biological targets and designing drugs to engage them in a way that alters the course of disease. Still, the process remains largely one of trial and error that plays out over many years and across many stages—from drug discovery and design to preclinical development (testing compounds for safety and efficacy in cells and animals) to clinical trials in which the experimental medicine is tested in three successive studies in people. Staggeringly, 90% percent of all drug candidates fail in humans, meaning only one in 10 drugs makes it through that stage to approval—even after millions have been spent.

Given that status quo, if AI could help better predict which drugs are likely to work, or even which ones will not—sparing time and investment on expensive late-stage failures and increasing the percentage that succeed—that would make a meaningful difference. “What I always tell the team is, ‘If 80% of our drugs fail in the clinic, we are twice as good as the industry average, and we can be the most disruptive company in this space,” says Recursion’s Gibson.

A future in which computers predict, or even create, our next blockbuster medicine feels closer than ever, thanks to some remarkable recent tech breakthroughs, from large language models like ChatGPT, to tools of the “ resolution revolution ,” like electron cryo-microscopy, that have equipped scientists with richer data. Add to that AlphaFold, the platform developed by Alphabet’s DeepMind subsidiary, that quite accurately predicts the structure of over 200 million proteins, including the tens of thousands found in humans. Now known as Google DeepMind, the first iteration debuted in 2020; its main architects won the Nobel Prize for chemistry last year.

Some feel these innovations have put the industry on the cusp of an unprecedented productivity boom, while others regard this as an overhyped moment in a long process of incremental change. The latter camp will tell you AI is really just the latest buzzword for an evolving technology that has been expected to revolutionize their field for years. (See: machine learning, big data, and, going way back, QSAR or “quantitative structure-activity relationship”.) Indeed, this is a revolution long in coming, one that Fortune teased on a cover featuring a drug Merck designed on a computer with the text “The Next Industrial Revolution”—in October 1981.

Computers and data have played a role in drug development since then, but their use has so far failed to translate into widespread R&D productivity gains. In fact, for the past seven decades, the pharmaceutical industry has experienced the opposite, with the development of new medicines generally taking longer and growing more expensive over time. The number of new drugs approved for every $1 billion spent on R&D has halved roughly every nine years. This phenomenon even has a name: Eroom’s Law : “Eroom” is Moore in reverse, a cheeky nod to Moore’s Law on increasing speed in computer chip development, pointing out the opposite trajectory in pharmaceutical innovation.

Several reasons are given for this trend: Government regulation is stricter now, for one. But mostly, our difficulties with R&D boil down to our limited understanding of biology: We’ve already solved the easy stuff. Solving the harder stuff requires new biological insights and models—and generally, we’ve underinvested in that science, explains Jack Scannell, the R&D productivity expert who coined the term “Eroom’s Law ” : “We're left with diseases where the models all too frequently give us the wrong answer.”

The science is just extremely challenging, given the vast, mutifaceted, heterogeneous nature of human biology. Aviv Regev, the AI-minded head of research and early development at the biotech Genentech , compares the industry’s work to develop drugs in such an enormous landscape to “looking under multiple little lampposts—a little bit here, a little bit here, a little bit here.”

This is where Regev says the convergence of new technologies with human insight is game-changing: Scientists alone cannot make sense of the exploding amounts of biological data now available to them, but an AI trained on that information—from high-resolution images of neurons to genetic sequences to patient records—can help researchers find patterns and make connections to come up with the novel insights that are needed to understand disease and develop drugs to treat them in various populations.

But can we trust AI—known for hallucinations in other realms—to give us good information on, say, the brain chemistry behind depression? Or the inner workings of a cancer cell? To provide checks and balances, Regev promotes a method of operating, widely adopted in the industry, known as “lab-in-the-loop” where an AI model’s predictions are tested in a physical lab.

The data from those real experiments is then fed back into the model, so the AI is constantly learning and refining, to make better, more accurate predictions. She adds that experiments must be done at a huge scale to reap the benefits, train effective models, and work faster and better. Genentech has partnered with Recursion on some of this work, which Regev describes as promising: “We are seeing biology that is known to disease area experts as well as potentially compelling biology that is not previously known.”

Tantalizing breakthroughs—and setbacks

Gibson got the idea for the Recursion in 2013 while doing doctoral research at the University of Utah on Cerebral Cavernous Malformation (CCM), a rare neurovascular disease that is thought to affect more than a million people globally. There’s no therapy for CCM, which can cause brain bleeds and stroke, and the typical way to develop one is difficult and laborious—requiring first the identification of a molecular target and then a drug to meaningfully interact with it.

Gibson had a different idea. Using a new machine-learning image-analysis software called CellProfiler, he could compare images of diseased and healthy cells—and see whether any compounds restored the diseased cells to health.  Using the technique, Gibson found two possible drugs that appeared to treat CCM: Vitamin D and a compound known as Tempol. These were just “hits,” drug candidates that would need to be tested and further refined, but it seemed promising. He wondered, couldn’t he apply the same method on other hard-to-cure diseases? That year he took leave from medical school to found Recursion with his professor Dean Li (now the president of Merck Research) and a friend, Blake Borgeson , with ambitions to industrialize drug discovery.

Eleven years since its founding, the company is still working to bring REC-994—the compound that Gibson first got excited about back in grad school—to market. In September, when the company first announced the results of a Phase 2 study involving 62 participants with CCM. The drug proved to be safe and well-tolerated—the point of the study—but preliminary efficacy data appeared mixed. While patient MRIs suggested that the medication was working to some extent, physicians and the patients themselves reported no improvement. Recursion’s stock fell nearly 17% that day. The company released more detailed, but not statistically significant, data in early February that signaled patients were experiencing functional improvement as well; the company’s stock rose 2.7%.

Gibson shrugs at the market’s fickleness. “I am encouraged by what we were able to show,” he told Fortune by email in February. “I believe we have a potential medicine in a space few others have even attempted to explore.” The company is discussing next steps with the FDA—there are challenges with a first-in-class drug, like figuring out how best to measure clinical improvement—while moving forward with seven other drugs that Recursion has in trials, for conditions including for cancer; C. Difficle , a stubborn bacterial infection that causes diarrhea and can be life-threatening; and neurofibromatosis type II, a disease characterized by the growth of noncancerous tumors in the nervous system.

Others, though, viewed Recursion’s underwhelming results as part of a pattern in the field.  BenevolentAI, a buzzy U.K.-based firm founded in 2013 and once valued at $2 billion, stopped work on its most advanced candidate, an eczema drug, when patients in a Phase 2a trial showed no clinical improvement in 2023; the stock dropped more than 80% on the results, and two rounds of layoffs later, the company’s market cap is now around $13 million. Exscientia—founded in 2012 and once valued at $3 billion—has had two of its programs dropped in late-stage development; in September, the company merged with Recursion.

Insilico Medicine, which claimed to be the first company to have a wholly AI-discovered and -designed drug in phase 2 studies, boasts that it got the molecule from concept to human trials in 18 months (compared to the industry average of 4.5 years). Founder and co-CEO Alex Zhavoronkov recalled his head of R&D waking him with a 2 a.m. phone call to excitedly share preliminary results from the study in patients with idiopathic pulmonary fibrosis. The rare but increasingly prevalent disease affects an estimated 3 million people globally, scarring and stiffening lung tissue and typically leading to death within a few years.

In November, Insilico publicly reported the results of that study in 71 patients across sites in China. Data showed the drug to be safe, and while the results were not statistically significant, patients got better on it—with improved lung function corresponding to the amount of drug they received over the 12-week study. “We didn’t expect to see that for that short period of time,” Zhavoronkov told me last fall. Another phase 2 study of the drug in the U.S. is ongoing. Preliminary as those findings are, Zhavoronkov declared the phase 2a results to be “a critical milestone in AI-powered drug discovery.”

Research by the Boston Consulting Group suggests AI is taking some of the uncertainty out of the medical trial process. The firm reviewed the pipelines of more than 100 AI-native biotech companies and found these companies have enjoyed an 80%–90% success rate in phase 1 trials (small safety studies), which is considerably better than the industry average of only 40%–65%. In phase 2 studies, success rates were comparable with the industry. There is not yet data to draw conclusions about phase 3 trials.

A high-stakes race

Who will benefit from AI’s efficiency improvements, and be the first to bring a fully AI-discovered medicine to market? Some argue that Big Pharma companies have the edge in this race because of their considerable resources and the fact that they have enormous amounts of proprietary data that they could in theory use to train an AI. But bigger isn’t necessarily better here, others point out: Much of Big Pharma’s data is messy and unstandardized, requiring considerable cleanup to use for these purposes. Plus, the organizational change required to overhaul a massive R&D operation is fraught.

AI-native startups, like Recursion, meanwhile, are building datasets from scratch expressly for the purpose of feeding them to an AI, and hiring “bilingual” teams—computer engineers as well as biologists and chemists—whom they believe are better suited to the job.

Given the disconnect between inflated expectations of AI-driven change in the sector and the reality that it will take time to deliver on that promise, the market seems unsure of how to value a company like Recursion. Four years ago, in the frothy days of the pandemic and not long after it went public, Recursion had a market value of $7 billion. Now,  in the waning days of a yearslong “biotech winter” on Wall Street, Recursion is plugging away on 20-some preclinical and clinical development programs, and is worth $2.1 billion. The company has broadened its platform considerably—acquiring multiple companies and capabilities, forming notable partnerships with pharma, drawing an enviable $50 million investment from Nvidia, amassing crazy amounts more of data, and of course building the industry’s largest supercomputer to process it. “We have been running this thing hard,” Gibson said of BioHive 2 in a February earnings call, noting that the team was building new models of biology with all that computer power. “Recursion is years ahead of almost anyone else in the space,” he said.

The real race for companies like Recursion may be one against time: Investors are restless and hungry for proof points. Insilico’s Zhavoronkov thinks skepticism of the industry will persist until the field produces an AI-generated blockbuster, but in the meantime, he encourages investors and the industry to focus on data-driven benchmarks rather than splashy funding rounds.

" Almost at the top "

Recursion’s story represents both the hopeful and humbling nature of its mission—and the necessary fortitude and inevitable hubris of those who take it on. Biology is marvelously complex, and developing drugs is incredibly hard. Resources (and investor patience) are limited—and even with the smartest, most cutting-edge technologies, certain phases of the process will take a long time. AI may revolutionize the pharmaceutical industry, and Recursion may emerge as a winner. But for now, there’s still a lot of work for humans to do.

Later in the day, Nvidia’s Huang offered a more optimistic vision: “You’re that yellow hold,” he told Gibson, pointing to a foothold high on the company’s rock-climbing wall. “You’re close.”

“For those who can’t see,” Gibson noted, “that’s almost at the top of the wall.”