A new drug usually starts with a tragedy.
Peter Ray knows that. Born in what is now Zimbabwe, the child of a mechanic and a radiology technician, Ray fled with his family to South Africa during the Zimbabwean War of Liberation. He remembers the journey there in 1980 in a convoy of armored cars. As the sun blazed down, a soldier taught 8-year-old Ray how to fire a machine gun. But his mother kept having to stop. She didn’t feel well.
Doctors in Cape Town diagnosed her with cancer. Ray remembers going to her radiation treatments with her, the hospital rooms, the colostomy bags. She loved the beach, loved to walk along the line where the water met the land. But it got harder for her to go. Sometimes she came home from the hospital for a while and it seemed like things would get better. Ray got his hopes up. Then things would fall apart again. Surgery, radiation, chemotherapy—the treatments that were on the table in the 1980s—were soon exhausted. As she lay dying, he promised her he was going to make a difference, somehow. He was 13 years old.
Ray studied to become a medicinal chemist, first in South Africa, taking out loans to fund his studies, then at the University of Liverpool. He worked at drug companies across the UK, on numerous projects. Now, at 53, he is one of the lead drug designers at a pharmaceutical company called Recursion. He thinks about that promise to his mom a lot. “It’s lived with me my whole life,” he says. “I need to get drugs on the market that impact cancer.”
The desire to stop your own tragedies from happening to someone else may be a strong motivator. But the process of drug discovery has always been grindingly, gruelingly slow. First, chemists like Ray zero in on their target—usually a protein, a long string of amino acids coiled and folded upon itself. They call up a model of it on their computer screen and watch it turn in a black void. They note the curves and declivities in its surface, places where a molecule, sailing through the darkness like a spaceship, could dock. Then, atom by atom, they try to build the spaceship.
When the new molecule is ready, the chemists pass it along to the biologists, who test it on living cells in warm rooms. More tragedy: Many cells die, for reasons that are not always clear. Biology is complex, and the new drug doesn’t work as expected. The chemists will have to create another, and another, tweaking, adjusting, often for years. One biologist, Keith Mikule of Insilico Medicine, told me of his experience at a different drug company. After five years of work, their best molecule had unforeseen, dangerous side effects that meant they could take it no further. “There was a large team of chemists, a large team of biologists, thousands of molecules made, and no real progress,” he said.
If a team is very lucky, they get a molecule that, in mice, does what it’s supposed to. They get a chance to give it to a small group of healthy human volunteers, a phase I trial. If the volunteers stay healthy, then they give it to more people, including those with the disease in question, in a phase II. If the sick people don’t get sicker, they get a chance—phase III—to give it to more sick people, as many as they can find, as diverse a group as possible.
At each stage, for reasons few people understand and fewer can predict, great rafts of drugs drop out. More than 90 percent of hopefuls fail along the way. When you meet drug hunters, you might ask them, cautiously, tenderly, if they’ve ever had a drug make it. “It’s very rare,” says Mikule, who has one drug (niraparib, for ovarian cancer) to his name. “We’re unicorns.”
But Mikule, Ray, and other chemists and biologists are trying a new approach. When I talk to Ray, he’s excited to show me a molecule he and his colleagues at Recursion have been working on. It’s a so-called MALT1 inhibitor, designed to interfere with the growth of blood cancer cells. On his screen, REC-3565 is a series of rings and lines, another skeletal spaceship floating in the void. But it exists in the real world too: Just a few weeks before my chat with Ray, the first phase I volunteers swallowed it in a little pill. What’s special about this molecule, Ray says, isn’t just that it has survived the gauntlet thus far. It’s that REC-3565 “wouldn’t have come by human design.” Ray’s team, he believes, would not have made the logical leaps required to reach this point without using artificial intelligence.
As the world’s pharma giants get caught up on AI, Recursion is among a group of startups betting everything on the technology. Founded 12 years ago by academics in Utah, the company made its name by taking snapshots of cells under various conditions, creating a vast database of pictures, and turning AI on them to identify potential new targets. Last year, Recursion acquired another decade-old startup—Ray’s former employer, Exscientia—which pioneered the use of AI to design small molecules. There are others, including Mikule’s employer Insilico, which was founded in 2014. Just last year, Xaira Therapeutics launched with $1 billion in venture capital—the biggest biotech funding round in years. (The only other new startup that pulled in as much in 2024 was Safe Superintelligence, cofounded by a former top OpenAI researcher.)
There are no drugs on the market designed using AI. But both Recursion and Insilico have gotten candidates through phase II clinical trials, which means they’re safe in patients. REC-994 is for cerebral cavernous malformation, a disease that causes brain lesions, and ISM001-055 is for idiopathic pulmonary fibrosis, a progressive, fatal lung condition. More AI-linked drug candidates are in development, from Insilico, Recursion, and other companies, including the one Ray showed me.
All of these molecules, right now, are like cards lying face down on the table. Can AI help make drugs that actually work, faster and cheaper than usual, or are the drug hunters about to be dealt another losing hand?
In the summer of 1981, a headline on the cover of Fortune magazine proclaimed that the age of digital drug discovery was at hand. The story explored how scientists were using computer visualization to select the best molecules to try in cells, hoping to break through the gridlock. Derek Lowe, a medicinal chemist who writes the long-running blog In the Pipeline, recalls that the Fortune article made some drug hunters at the time nervous. At the pharmaceutical company Schering-Plough, where he worked, there was a room labeled “Computer-Aided Drug Discovery (CADD),” packed with expensive equipment. “The medicinal chemists across the hall didn’t think too much of that,” Lowe told me, “so they put a sign over their door that said ‘BADD: Brain-Assisted Drug Discovery.’”
Computers did revolutionize everything. But the hard problems of drug discovery didn’t evaporate with the touch of a cursor. Seasoned drug hunters refer in a jaundiced tone to combinatorial chemistry, an attempt to stumble across new kinds of drugs by assembling molecular pieces in random order. (It didn’t work, in part because the costs of such a wildly democratic approach were crippling.) Computational chemistry, which allows scientists to simulate how a target and a molecule will interact, gained grudging acceptance—but its success depends on accurate models of the target and the candidates, and for that you need old-fashioned elbow grease in the lab.
If anything, the hard problems have grown harder as the full complexity of biology has come into focus. “We have more things to worry about than we used to,” says Lowe. Cancers with different mutations driving them respond to different therapies. Drugs that attached to a certain receptor were linked to heart problems, and thus any new drug candidate, no matter how promising, must be removed from the running if it shows affinity for that receptor.
Karen Billeci, a principal biologist at Recursion, still remembers one of the first times she heard a drug hunter mention artificial intelligence. One dawn in 1993, Billeci was walking across her company’s parking lot on the edge of the San Francisco Bay with a couple of other employees. They worked at a scrappy startup called Genentech (later acquired by Roche for $47 billion). Billeci’s programmer friends were exploring whether neural networks—a form of machine learning—could be used to find patterns in patient information and help reveal why some responded to a drug and others didn’t. “These great drugs would go into humans, and they would fail,” Billeci says. They talked in the parking lot about whether, someday, there would be software that could learn to see patterns they couldn’t. “We didn’t say ‘train,’” Billeci recalls. “We didn’t have the words for that yet.”
It gradually became clear, over the next several decades, that AI might do more than pick out patterns in patient data. In 2020, something happened that crystallized what might be possible. In a global competition that fall, an AI built by Alphabet’s DeepMind showed it could correctly predict how a protein would fold up into its final form—a canonical hard problem in biology and a key task for drug hunters. DeepMind’s AI easily beat out all the other contestants. David Baker, a biochemist at the University of Washington, was inspired to dig deeper into using AI to design new drug proteins, work that later won him the 2024 Nobel Prize for Chemistry. “It didn’t take us long to develop methods that surpassed the ones we had been developing before,” he says. (Baker is one of the founders of Xaira.)
After that, what else might be possible with AI? What if it were shown all the drugs that have ever existed, with all the data about how they work, and then set loose on a database of untried molecules to identify others to explore? What if—and this is where the discussion around machine learning has gotten to now, in 2025—the software could take in a decent chunk of all the information about biology generated by humankind and, in an act both spooky and profound, suggest entirely new things?
Sometimes, humanity is learning, AI produces things that look good at first glance but turn out to be whimsical potpourris of words or thoughts, mere nothingburgers. The fact that drug discovery involves extensive real-life testing makes it unlikely that such suggestions would survive the process. The biggest risk of AI hallucinations might be wasted time and resources. But the failure rate of new drugs is already so high that scientists at these startups think the risk is worth taking.
Peter Ray looks at the MALT1 inhibitor floating in the void. “If I get a drug to market, I would feel I had fulfilled my promise,” he says. He points out where the AI revealed a way to remove a section of the molecule that could cause toxicity. It was a reaction that had not occurred to any of the humans involved.
The real question is whether molecules designed using AI are any better at getting to market. The last few stages of the process are the most expensive, the most unpredictable. In any clinical trial, it’s hard to find the right people, says Carol Satler, the vice president of clinical development at Insilico. It’s slow. She worries about it—hopes she has made the right choices, contacted the right doctors, excluded the people who would not benefit, included those who might, to see what the drug can do. By the time a drug reaches trials, it represents a billion dollars and a decade in the lives of hundreds, if not thousands, of scientists. One patient signs up. Then two. Months pass. Time crawls. “The meter is always running,” Satler says. “It’s so expensive.”
Late last year, soon after Recursion finalized its acquisition of Exscientia, 300-odd drug hunters from both companies converged on an event space in London.
The pink-lit conference hall buzzed with news of an announcement made just days before by Recursion’s chief scientific officer. Molecule REC-617, developed by Exscientia, had been given to 18 patients whose terminal cancers had stopped responding to other treatments. The phase I clinical trial was designed to see both whether patients could tolerate the drug candidate and whether it had any effect. One patient—a woman with ovarian cancer that had come back three times—surprised everyone: She lived. She was still alive after six months of the treatment. Because the trial is blinded, no one at Recursion or Exscientia has any idea who this woman is and whether she is still alive today. But in that room, she seemed to radiate with life.
The announcement contained another noteworthy detail. Because Exscientia used AI to narrow down the number of candidate molecules before any of them were made, it was not thousands but a mere 136 that were finally manufactured and tested in cells. (Ray’s MALT1 inhibitor involved making only 344, also a tiny fraction of what would have happened in a traditional setting.) Chris Gibson, Recursion’s cofounder and CEO, underscored that number in his talk to the assembled crowd, emphasizing the savings in time and resources. By failing faster, goes the logic—by using AI not only to invent new molecules but also to rule most of them out in advance—it might be possible to bring down the cost of the first stages of this extremely costly process.
In the center’s lobby, a breakout group with David Mauro, Recursion’s chief medical officer, Jakub Flug, an Exscientia medicinal chemist, and a handful of others stood in a circle. The employees were having what amounted to an enormous blind date. They were meeting people they’d never seen in person, telling their stories, trying to see how they would all fit together. They took turns introducing themselves and saying why they had chosen to join these companies. One person said: I’m here to have fun. Another said: I’m here because I was tired of doing something that I didn’t believe in anymore. Another: I am here because I want to actually release a drug onto the market. Everyone nodded at this one.
Downstairs, in a basement room, Gibson was thinking about the future too. His hope is that Recursion is laying the groundwork for what drug discovery will someday be like across the industry, starting with the eight drugs that have advanced to clinical trials and the handful behind them, in the preclinical stage. “If we’re doing this right, if we’re building a learning system, the next 10 drugs after that have a higher probability of success. Next 10 drugs after that, higher probability of success. We keep refining this thing,” he said.
I asked him about his claim, last summer, that there would soon be information about 10 or so different candidates. This critical mass, with information going public in a large bolus, is a calculated goal, he said: If around 90 percent of drugs fail, then Recursion needs to show results of about 10 different programs just to see if they are doing what they hope. “At the end of the day, it’ll be fair to judge us by the first 10,” Gibson says. “That’s enough of an n.” Enough of a sample size, in other words, to see what this approach can do.
One cold morning late last year, I went to see one of Recursion’s discovery engines. Patrick Collins, the director of automation, and Su Jerwood, a principal scientist in pharmacology, showed me into a room the size of a small supermarket with aisles of machines in plate-glass cases. White lamps like halos hung above them. “We’ve got biology on one side, chemistry on the other,” Collins said. A magnetic railway threaded through the machines, connecting pipetting robots to incubator chambers. “It’s about design, make, test, learning, loop,” Collins said. He indicated cases of bottles and powders, “all the building blocks, reagents and things.” Humans keep the machines topped up.
These machines, Jerwood explained, dispense molecules created from raw atoms, molecules that AI systems have already tested and explored in virtual spaces. The candidate drugs drip onto trays of cells, and the system evaluates their effects. It’s new, and there are kinks to be worked out. Some parts of the automated process still need humans to move them along, Collins said, and Recursion is figuring out how to streamline the flow of information to and from the AI. But when it works, scientists will have the results of thousands of tests glowing on their screens. The automated system has been up and running for about a year, so it wasn’t involved in making the candidates currently in clinical trials. But it is helping to make future drugs.
As I examined the machines in their pristine chambers, I wondered about what it means, now, to be the kind of human who loves to think about molecules, loves to make them, whose joy comes from understanding how they work. I asked Collins this. He thought back to the moment when he first crystallized a protein by hand, first saw a drug molecule clasped against it. “I was hooked for life,” he said. Those traditional tools still have their place. But perhaps it is not here, where the focus is getting something to the clinic as fast as possible, something that works. “We’re all trying to think about patients,” Collins said.
Jerwood gave her answer: “I am so hungry for something new all the time.” Standing there, above the automated lab, she imagined the regions of chemistry where no one has yet gone, structures and reactions that lie on the far side of unknown processes. The sun was just pulling itself over the horizon. She thought of all the things the machines might do, all the things that she will do no longer. “It’s down to the untouched space, yeah? Because then I will have time to look into that space,” she said. “I will have time to take that risk.”
For some pharmaceutical researchers, though, the promise of AI goes beyond pushing scientific boundaries or even treating disease. Alex Zhavoronkov, the CEO and cofounder of Insilico, says the company favors targets that are implicated in both illness and aging. Its drug candidate for idiopathic pulmonary fibrosis, for example, is designed to prevent scarring of the lungs by dampening certain biological pathways, but it also may slow the aging of healthy cells. Zhavoronkov hopes to bring new drugs to the clinic, perhaps faster and cheaper, even as he uncovers new treatments for aging-related disease and decline.
When I speak with Zhavoronkov, he’s at a company-wide retreat in Chongqing, China. “In 20 years, I’m going to be 66,” he says. “I saw my dad when he was 66, and it’s not pretty.” He is frank about having high expectations, about his desire for speed in an industry where speed isn’t always readily available. He shows me a video of an automated lab in Suzhou, China. “We built it during Covid,” he says, explaining that some of the laboratory scientists on the project worked around the clock, sleeping in the facility, to get it up and running.
There is something vaguely science-fictional about the setup, and about Zhavoronkov’s particular form of pragmatism. Zhavoronkov has scars on his arm where he’s had skin removed to make induced pluripotent stem cells, which can be reprogrammed to grow into many types of tissue. “If you want to buy my IPSC, give us a call. We’ll ship it to you,” he says. “The more data there is about you in the public domain, the higher chances you have to get a real good treatment when you get sick, especially with cancer.”
In the lab video, the camera glides through a black hallway, then through an anteroom, past a wall of glass. The glass can be dimmed, if the work going on behind it is confidential. Behind the wall are machines loaded with trays of reagents and cells, with arms that swivel as they move components around. Humans are rarely needed.
Sooner or later, in some form, AI tools will be standard in drug discovery, suspects Derek Lowe, the medicinal chemist and blogger. He calls himself a short-term pessimist, long-term optimist about these things. It’s happened again and again in the industry: New strategies arrive, ride a wave of hype, crash and burn. Then some of them, in some form, rise again and quietly become part of what’s normal. Already, big pharmaceutical companies—the behemoths of drug discovery—are starting their own AI-related research groups. Recursion, meanwhile, is exploring the use of AI not only to dream up and test new molecules but also to find trial participants, speeding along those last, costliest steps to market.
The transformation isn’t going to be without casualties. “These techniques, both the automation part and the software, are going to make more and more things slide into that ‘humans don’t do that kind of grunt work’ category,” Lowe says. Large numbers of jobs held by human chemists will wink out of existence. Those “who know how to use the machines are going to replace the ones who don’t,” Lowe says. Even Peter Ray no longer feels it’s accurate to describe him as a medicinal chemist. “I’m something else,” he muses. “I don’t know what to call it, to be honest.”
In the months since the blind date in London, Recursion has announced two drug candidates entering clinical trials, the MALT1 inhibitor and a molecule for lung cancer. A drug for a digestive disease is already in trials. Insilico is in the process of trying to advance to a phase III trial for its idiopathic pulmonary fibrosis drug, with Carol Satler on the phone to doctors. The cards are being turned over, one by one. Ray goes running sometimes, through his neighborhood near Dundee, Scotland, and thinks of his mother.
Gibson reflected on the long game he sees Recursion playing. The way it’s tinged with urgency. Yes, they want to change the world. And personally, he thinks it’s been too long in coming. “There’s a lot of people here who have lost a loved one or multiple loved ones to a specific disease,” he said. “They’re pissed off. They’re here because they want to get revenge on the lack of opportunity that that family member, or friend, or child, had.” The meter is ticking, numbering the days, as drugs move through trials and everyone waits to see what happens.
Time is the thing we are all running out of. Some of us faster than others.
Let us know what you think about this article. Submit a letter to the editor at mail@wired.com.