Save this storySave this storySave this storySave this story
When David Fajgenbaum was a twenty-five-year-old medical student, at the University of Pennsylvania, he started to feel so tired that he could barely stand. Fajgenbaum, a former college quarterback, could still bench-press three hundred and seventy-five pounds; he was known for doing pullups on a tree near his workplace. But now he was desperately ill. The lymph nodes in his groin and neck swelled. Small red dots—blood moles—emerged on his chest, and he woke up soaked in sweat. One day, at the hospital where he was doing his rotation, he stumbled down the hall into the emergency room, and doctors told him that his liver, bone marrow, and kidneys were failing. Fluid had leaked out of his blood vessels, into his abdomen and around his heart; bleeding in his retina temporarily blinded him in his left eye. Fajgenbaum was admitted to the I.C.U. and given enormous doses of steroids to try to quell runaway inflammation. It took him a month to recover enough to leave the hospital.
Fajgenbaum remembers asking a doctor what had caused his strange illness. “We don’t know,” the doctor said. “But let’s just hope it doesn’t come back.” When it did, a month later, he was diagnosed with a form of Castleman disease—a rare, often fatal condition that straddles the line between cancer and autoimmune disorder. There were few studies of the disease and no approved treatments. Doctors started Fajgenbaum on an aggressive chemotherapy regimen, but his condition deteriorated so much that a priest delivered last rites. Miraculously, and mysteriously, the illness receded. And yet it kept coming back.
Fajgenbaum scoured his medical records and the available scientific literature. He requested a series of tests. Eventually, he noticed something curious. A few months before one of his hospitalizations, the number of activated T cells in his blood had started to rise, and so did the amount of a protein called VEGF, which promotes the growth of blood vessels. Fajgenbaum performed experiments on his own blood samples and lymph nodes that identified a link between the two processes—a signalling cascade known as mTOR—and convinced his doctors to give him a potent suppressor of the pathway called sirolimus. The drug had been on the market for more than a decade; it was often given to transplant patients to stop their immune systems from attacking a new organ. After Fajgenbaum started the medication, months passed without a relapse, then years. When I spoke with him, a few weeks ago, his disease had been in remission for more than a decade.
Fajgenbaum told his story in a 2019 memoir, “Chasing My Cure: A Doctor’s Race to Turn Hope Into Action.” He focussed on a central irony: the medication that saved his life already existed, and nobody had thought to give it to him. “There is a systemic problem,” he told me. “There are all these drugs available, sitting in your local pharmacy, but they aren’t being used to treat all the conditions they could.” But in the years since the book came out—and, more specifically, since the rise of powerful A.I. models—Fajgenbaum has started to find solutions for that problem. He co-founded a nonprofit called Every Cure, which trained an A.I. model on what he described as “the world’s knowledge of every disease, gene, protein, and molecule, as well as the interactions between them.” The algorithm began to propose previously unknown applications for known treatments.
Talking to Fajgenbaum, I thought about how I’d often been in the uniquely discouraging position that his doctors were once in, with little to offer someone for a perplexing or intractable condition. Doctors have long prescribed off-label medications, usually through trial and error or in clinical trials, but now A.I. appears poised to supercharge the practice. Earlier models needed examples of effective drugs to learn, so they were unlikely to identify promising candidates unless treatments already existed. But more advanced models can conduct what’s called “zero-shot inference,” nominating drug candidates for conditions without any known treatments. “We’re at an inflection point,” Marinka Zitnik, a computer scientist at Harvard and a member of Every Cure’s advisory board, told me. “I think we’ve entered a new era.”
More than eighteen thousand diseases afflict humans, by Every Cure’s count, including upward of nine thousand that don’t have a single approved treatment. Meanwhile, there are thousands of medications on the market. The tally of conceivable drug-disease combinations numbers in the tens of millions. Some combinations obviously won’t work—penicillin won’t treat cancer—but many others could save lives if only we had a way to identify them. Of course, few patients can conduct the sort of rigorous self-study that Fajgenbaum did, and pharmaceutical companies have little incentive to find novel uses for old drugs, which bring in little money. “Who’s going to invest in uncovering all the potential matches?” Fajgenbaum said. “And how do you rank and prioritize all the possibilities?”
Every Cure’s A.I. platform, dubbed the MATRIX, is trained on what are known as “knowledge graphs”—networks of data representing the relationships between genes, proteins, drugs, and diseases. Knowledge graphs pull their information from the scientific literature and from curated medical sources: biobanks with health data from millions of people, for example, or repositories of chemicals and their safety profiles. The power of these networks has grown considerably over the past decade, allowing models such as Every Cure’s to generate a ranked list of the likelihood that a medication will treat a disease, on a scale from zero to one. Fajgenbaum’s team then undertakes a lengthy review process to determine which drug-disease pairs warrant further study or promotion. “At this stage, it’s less that the A.I. is outsmarting us humans and more that it’s really good at highlighting things that we’ve already discovered,” Fajgenbaum said. “Often, we just aren’t making the connections. Scientists might write something up in a journal somewhere, but then there’s no one to take it forward.”
A common heart drug called propranolol might be useful for treating angiosarcoma, an aggressive form of cancer. (The medication inhibits a receptor that’s expressed in such tumors.) A Botox injection between the eyebrows could theoretically reduce the symptoms of depression. (Perhaps making it harder to frown could disrupt physical feedback loops that reinforce negative emotions.) Several studies suggest that giving folinic acid, a kind of vitamin, to certain autistic children could enhance their verbal abilities. A recent randomized trial in India found that injecting high doses of lidocaine around breast tumors at the time of surgery significantly improved survival rates for women with breast cancer. Some of these pairings may not work, but it wouldn’t take many successes to justify the effort. “Lidocaine is one of the most inexpensive substances in all of medicine,” Fajgenbaum said. “With breast-cancer surgery, it’s already going to be injected at the site of the incision. We’re not even talking about bringing a new substance into the surgery. We’re just talking about injecting the same substance at a higher volume!”
Last year, Fajgenbaum received an e-mail from a woman named Tara Theobald, whom he’d met at a conference for Castleman disease. Her boyfriend, Joseph Coates, had initially been diagnosed with the condition, but was later found to have a rare blood disorder called POEMS syndrome. Coates was now entering a terminal stage: his heart and kidneys were failing, and fluid had to be drained from his belly several times a week. He was too sick to undergo a stem-cell transplant, which could save his life. Every Cure doesn’t usually offer treatment recommendations for individual cases, but this time Fajgenbaum used the organization’s A.I. platform to recommend a combination of drugs normally used to treat a blood cancer: a steroid called dexamethasone; a chemotherapy called cyclophosphamide; an immunotherapy called carfilzomib. “It was a Hail Mary, last-ditch effort,” Fajgenbaum told me. The doctor treating Coates thought that the regimen sounded “a little bit crazy.” But, with little to lose, they gave it a try—and within days Coates started to improve. A few months later, he’d recovered sufficiently to receive a stem-cell transplant. He is now in remission.
Stories like this have earned Fajgenbaum a reputation as a miracle worker. Each week, he receives hundreds of inquiries from patients and families asking whether he can conjure a treatment to help them. He is in touch with as many as he can be, but “we’re not naïve enough to think that one hundred per cent of these matches will be successful,” he said. “Plus, a one-off approach is not how we can help the most people.” Fajgenbaum and his team manually sift through around a thousand drug-disease combinations per month. Is there a plausible biological mechanism by which a medication can treat a disease? How many patients would benefit, and how much would they benefit? What are the chances of being able to prove that a medicine created for one condition works for another? Drug development is a bit like foraging in an unfamiliar forest; some fruits will nourish you and others might poison you. Drug repurposing is more like wandering around an overgrown orchard. Every candidate has already been vetted for safety; A.I. can identify low-hanging fruit, and then humans can decide whether the pairing is ripe for picking.
Every Cure classifies high-scoring drugs into one of four buckets: frontier explorers, clinical gems, unsung heroes, and known entities. A frontier explorer does well in the model and has a strong biological rationale, but there’s little research supporting its use for the target condition. A clinical gem has some evidence, for example in cells or animals, but requires more study in people. Unsung heroes have proved themselves—they’re just not being used. And known entities are already widely repurposed in clinical practice. Fajgenbaum’s goal is to move as many drugs up the ladder as possible, whether by prompting investigations of little-studied treatments or by helping proven pairings become more broadly adopted. To do so, the nonprofit has secured more than a hundred million dollars in awards from the TED Audacious Project and the federal government’s Advanced Research Projects Agency for Health, or ARPA-H.
A few weeks ago, I sat in on a meeting at which Every Cure employees discussed the drug-disease pairs that the MATRIX had deemed worthwhile. Much of the medical team sat around a conference table in Boston with Fajgenbaum; members of the tech team Zoomed in from London. A physician named Elliott Sharp, with neat brown hair and red-rimmed glasses, presented a summary of Rosai-Dorfman disease, or R.D.D., a rare and little-understood condition that can lead to enlarged lymph nodes and serious skin problems, such as nodules and ulcerations. A doctor named Luke Chen, in Canada, had used a cancer drug called lenalidomide to treat some of his patients with the disease, Sharp said, and had suggested it to Every Cure. The medication scored highly in the A.I. model—in the top 0.3 per cent of all drug-disease matches—and ranked twenty-third on the A.I.’s list of possible treatments for R.D.D. “Pretty much everything above it is a chemotherapy, steroid, or a known treatment,” Sharp told the group. There seemed to be a clear biological mechanism by which it could work—it suppressed an inflammatory molecule that was thought to play a critical role in the disease—and by 2026 the medication would fall in price and be more widely available. Sharp pulled up an A.I.-generated graphic that displayed linkages between the drug and the disease; it looked like a jumble of yarn and multicolored Christmas ornaments, each labelled with a word such as “papule,” “prednisone,” or “erythema.”
“It seems promising,” Matt Goddeeris, a cell biologist and former biotech researcher, who serves as the vice-president of discovery at Every Cure, said. “Looks like the N.C.C.N.”—National Comprehensive Cancer Network—“guidelines have it as a late-line option. The question is: Should it actually be higher? And do we think this is a clinical gem or an unsung hero?” For the former, Every Cure might conduct or facilitate a trial; for the latter, the organization would focus on outreach and advocacy. The group decided to advance lenalidomide for a “deep dive”: the medical team would spend weeks or months researching and, if needed, consult a scientific advisory committee about whether the organization should invest in further research or advocacy.
Every Cure’s A.I. also provides an assessment of a disease’s level of unmet need—though this assessment depends in part on the subjective values that the programmers have encoded. A condition such as R.D.D. can be debilitating and has few effective treatments; on the other hand, it generally isn’t lethal, and only about a hundred cases are diagnosed in the U.S. each year. Fajgenbaum said that the algorithm prioritizes how much an individual suffers from a disease, and whether any treatments exist, in addition to population-wide metrics such as the total number of people who suffer and die from it. R.D.D. had a score of 21.5 out of twenty-eight, placing it in the top one-third of conditions. (“David should get a tattoo of ‘No disease is too rare,’ ” one of his colleagues wrote in the Zoom chat.)
The group turned to the next potential pairing: a common immunosuppressant that had been flagged as a possible therapy for Alzheimer’s disease. “This was very, very high in MATRIX,” Nick Fragola, a research fellow at Every Cure, said. “And, in a vacuum, the mechanism makes sense.” But, Fragola went on, the few existing studies haven’t demonstrated results, and a small trial found that the drug didn’t easily enter the brain. “That’s probably a deal-breaker,” he said. “When you look at the molecule, it definitely doesn’t scream ‘brain penetrant,’ ” someone wrote in the Zoom chat. This time, the group decided to table the drug.
In early 2020, a biotech company called BenevolentAI used its knowledge graph to identify baricitinib, a rheumatoid-arthritis drug, as a potential treatment for COVID-19. The medication worked, and was formally approved for use in seriously ill COVID patients. Five years later, baricitinib remains one of the only medications that’s successfully been repurposed on a large scale as a result of A.I. “Finding medicines that work is tremendously difficult,” Derek Lowe, a longtime drug researcher and a blogger for Science, said. “Whenever someone says, ‘Ta-da, we’ve solved it!,’ I just make sure to keep my hand on my wallet.” Companies that make such claims sound like they’re selling snake oil. Lowe told me that he’s a “short-term pessimist but a long-term optimist” when it comes to using machine-learning tools to repurpose medicines. In his view, the approach may be most useful in treating certain rare diseases, for which “the standard of care is just to sit back and shake your head. Anything is going to be an improvement.” A.I.’s utility for tackling common and widely researched conditions is less certain. “A lot of what the models are doing has already been done by humans,” he said.
Next year, Fajgenbaum plans to publicly release scores for the tens of millions of drug-disease pairs that Every Cure’s A.I. platform has examined. For virtually any disease, researchers, doctors, and patients will be able to visit a website and see much of what his team saw—a prospect that, to me, sounded both empowering and overwhelming. The database is sure to offer useful ideas, but it could also generate floods of inquiries from patients and families about untested remedies. During the pandemic, the antiparasitic drug ivermectin gained traction as a possible COVID treatment because there were mechanistic reasons to think the drug might work. Studies repeatedly showed that it didn’t, but ivermectin remains a cause célèbre in certain circles—and, for some people, doctors’ unwillingness to prescribe it suggests that the medical establishment can’t be trusted. “I applaud the transparency,” Eric Topol, the director of the Scripps Translational Research Institute, told me when I described Every Cure’s plans. “But I’m also concerned that this could lead to wild-goose chases.”
Around the turn of the twenty-first century, a theoretical biologist named Stuart Kauffman introduced the concept of the “adjacent possible.” Single-celled organisms don’t suddenly evolve into fish and then primates; instead, a cell might evolve into a slightly more complex cell, and then perhaps into a multicellular organism, and that organism might gradually evolve new ways of using its biological machinery. Kauffman argued that natural selection is fundamentally a process of repurposing. Useful adaptations increase the complexity of a system; they expand the set of novel and nearby states available to an organism and its descendants. Dinosaur scales might evolve, step by step, into jury-rigged flight feathers; fish fins might eventually become mammalian limbs. “Anything can be used for more than one thing,” he has said.
Kauffman’s theory has since been adapted to explain innovation of all kinds, from the printing press to the iPhone. “The adjacent possible is a kind of shadow future, hovering on the edges of the present state of things,” the author Steven Johnson writes in his book, “Where Good Ideas Come From.” “Each new combination ushers new combinations into the adjacent possible.” Johnson likens the adjacent possible to a house that continually expands as you stroll through it. Walk through a door and another room materializes. This is one theory of scientific advancement. “Keep opening new doors,” Johnson writes, “and eventually you’ll have built a palace.” In this metaphor, drug repurposing may be less about building and more about discovering secret hallways and hidden doors. The room we are looking for might already exist out there, but the palace of biomedicine is so sprawling that simply assembling a blueprint—let alone exploring the whole structure—would be a superhuman task. Human intelligence built this place; artificial intelligence will help us navigate it. ♦
Sourse: newyorker.com
