Medicine by machine: Is A.I. the cure for the world’s ailing drug industry?

Computers are sifting through an endless archive of biological data and quickly finding patterns that it would take a human a lifetime to discover.
January 20, 2020 9:30 AM UTC
Jamie Chung—Trunk Archive

This article is part of a Fortune Special Report on Artificial Intelligence.

When a Canadian company called Deep Genomics announced in September that it had used artificial intelligence to solve a long-standing mystery about a genetic disorder called Wilson’s disease—and, what’s more, had used another deep-learning platform to identify a potential treatment—there was a flurry of excitement in the drug development world. The apparent milestone, which the company hailed as the “first-ever A.I.-­discovered therapeutic candidate,” got echoing headlines from dozens of news outlets, and in January, the five-year-old startup received a $40 million endorsement in the form of a fresh round of venture capital funding.

But the discovery itself is far more nuanced than most of the press reports have made it seem—and illustrates both the remarkable potential of A.I. in drug development and its inevitable limitations, at least in the near term.

The team at Deep Genomics began by feeding massive amounts of data related to more than 200,000 gene mutations into so-called training algorithms—teaching their computers to find connections between those misspelled snippets of human DNA and the faulty proteins they encode, which in turn seem to drive certain human diseases. “The A.I. learned to understand the molecular biology,” says Deep Genomics CEO ­Brendan Frey, who did his doctoral studies at the University of Toronto under A.I. pioneer Geoffrey Hinton and later became a leading researcher in neural networks himself. 

In the case of Wilson’s disease, a rare disorder that prevents sufferers from metabolizing trace amounts of copper found in food (and which can lead to liver disease and a host of neurological and other problems), the company’s A.I. found just such a connection with blistering speed. It has deciphered precisely how a mutation known as Met645Arg leads to a crucial defect in an essential copper-metabolizing protein.

Next, the company used another set of A.I. tools to sort through billions of molecules and rapidly identify nontoxic compounds that could correct the error made by the genetic glitch and enable a functional protein to be produced. The algorithms came up with no fewer than 12 drug candidates, each of which appeared to work in both cell models and mice. Deep Genomics hopes to put one of them, known as DG12P1, into human clinical trials as soon as next year.

From a scientific perspective, “it’s a big leap forward,” says Dr. Fred Askari, who directs the Wilson’s Disease Program at the University of Michigan School of Medicine. And when seen through the traditional lens of the drug industry, it may seem to be an even grander achievement: The entire process took just 18 months, Frey says, compared with a more typical preclinical development timeline of three to six years. If this can be repeated again and again, as Deep Genomics claims it can, then the effect could be revolutionary. Shrinking the time to discovery, after all, means quicker help for patients—and, in theory, less cost to drugmakers.

No wonder much of Big Pharma is racing into the field. Novartis is now collaborating with Microsoft on A.I.-driven drug discovery; Pfizer is using IBM’s Watson; and others (including Johnson & Johnson, Merck, ­AstraZeneca, and GlaxoSmithKline) have inked drug development partnerships with smaller A.I. companies. All told, investors have put $2.4 billion into hundreds of such startups since 2013, according to data from PitchBook.

Yet as promising as all this sounds, there are caveats to consider before you grab those A.I. pitchforks and storm the barricades. First is that Deep Genomics’ drug candidate has yet to be tested in a single patient, so—as the company itself acknowledges—we don’t yet know that it works.

Dr. Eric Topol, a prominent cardiologist and geneticist at Scripps Research and author of the 2019 book Deep Medicine: How Artificial Intelligence Can Make Healthcare Human Again, has been carefully reviewing the A.I. hope and hype for years. “We’re long on promise and short on proof,” he says. 

 “A.I. is a term that conflates ideas that we’re pretty sure will work in the future with ideas that are really just proof of concept,” agrees the University of Toronto’s David Duvenaud, whose celebrated research was crowned “Best Paper” at last year’s largest academic A.I. conference, the NeurIPS. At the gathering, Duvenaud surprised many by giving a blunt critique of the shortcomings of his own theoretical paper—a commentary that data scientist Mostapha Benhenda says reflects the more sober view of the field that many of his colleagues now share. “Such open self-criticism was unthinkable a couple years ago,” marvels Benhenda, whose Paris-based crowdsourcing startup, Melwy, fact-checks and fixes drug development algorithms.

One critical limitation, say experts, is simply the lack of good data on which to train the machine-learning systems. There are academic sources on genetic and molecular interactions to plumb, but comprehensive information about many chemical compounds—and more important, patient data from many clinical trials—is often proprietary to drugmakers and is rarely shared. Much of the data that is available, moreover, is poorly characterized or poorly structured, says Dr. David Agus, professor of medicine and engineering at the University of Southern California and the founding CEO of USC’s Lawrence J. Ellison Institute for Transformative Medicine. “It’s garbage in, garbage out,” he says.

That said, if A.I. can truly slash the discovery time for new medicines by half, it might finally offer an antidote to ever-soaring drug prices. That’s because a big area of promise for the technology is in tackling the mysteries of genetic disorders, like Wilson’s disease, in which the patient population is relatively small and where Big Pharma isn’t likely to bother with investing in new medicines. (Wilson’s disease affects an estimated one out of every 30,000 people worldwide, and Met645Arg, the gene variant that Deep Genomics’ A.I. programs helped decipher, is just one of more than 500 mutations that may play a role in causing it—making the potential market for the company’s experimental drug likely even tinier.)

In the past two decades or so, drugs for such “orphan diseases” (those for which the U.S. patient population is fewer than 200,000) have represented an increasing share of newly approved medicines in the U.S., rising from 10% of the total in 1998 to 44% in 2017, according to a study this past fall by America’s Health Insurance Plans. At the same time, their prices have skyrocketed, leaping from an average per person annual cost of $7,136 in 1998 to $186,758 in 2017—a 26-fold increase over 20 years. Today, nine of every 10 orphan drugs cost more than $10,000 annually per patient.

If A.I. can reverse that frightening trend—and shrink both the time and cost of discovering new targeted treatments for these uncommon maladies—it will be worth all of the hype, and then some. 

Panna Sharma, Lantern Pharma: “If we can do this faster, cheaper, and make drugs more personalized, then that changes everything.” | Brendan Frey, Deep Genomics: The Toronto company’s “A.I. learned to understand the molecular biology” of Wilson’s disease, he says. Abraham Heifets, Atomwise: “It’s just like testing a prototype of an airplane before making it,” he says. “You’ll test a thousand wings in the computer for every one you build.”
Courtesy of Lantern Pharma, Deep Genomics, and Atomwise

A.I.’s pharm team

Entrepreneurs are employing machine learning for every phase of drug development—from identifying better biological targets and experimental compounds to designing smarter clinical trials. Here, a few pioneers.

Deep genomics: discovering medicines

In September, the Toronto startup claimed to have developed the first-ever A.I.-discovered therapeutic candidate, when it announced a new investigational compound for Wilson’s disease, a genetic disorder that limits a person’s ability to metabolize copper. CEO Brendan Frey says human trials on the compound will start next year. By then, he predicts, the company will have up to five new drug candidates in the works as well.

Atomwise: virtual drugmaking

The San Francisco company is employing “convolutional neural networks,” the same technology used in facial recognition and self-driving cars, to find new treatment indications for existing medicines. Its computers create a 3D grid of a disease protein’s biochemistry and then run it through various combinations of 11 billion different molecules—many of which are drawn from public sources, such as the National Institutes of Health PubChem library. The algorithms then identify combinations of molecules that would match that protein’s biochemistry.

The MIRO-1 protein structure which has been implicated in Parkinson’s disease. The area in white represents the screening site where AtomNet evaluated 6.8M molecules, identifying a drug-like molecule now known as Miro1 Reducer.
Courtesy of Atomwise

Atomwise buys the promising compounds and sends them to academic and pharma researchers to test in cell lines and animal models; the company now has 550 projects with 250 research hospitals underway. “It’s just like testing a prototype of an airplane before making it, says Atomwise founder and CEO Abraham Heifets. “You run simulations to ensure the planes will fly, be quiet, and fuel-efficient. You’ll test a thousand wings in the computer for every one you build.”

Lantern Pharma: revitalizing discarded compounds

The Dallas company combs through data from hundreds of drugs in “failed” clinical trials to identify ones that might work for certain genetic subsets of patients. Such drug rescue isn’t new, but Lantern hopes its algorithms will dramatically speed things up. It recently acquired four compounds that were previously shelved by Big Pharma companies—including an abandoned chemotherapy agent that Lantern’s A.I. platforms suggest might be effective in a particular subset of patients with lung cancer. After digesting reams of clinical and other data, including some from the drug’s initial owner that was a decade old, Lantern’s A.I. engines identified a six-gene “signature” in a specific group of patients—females with adenocarcinoma who had never smoked in their lives—who were most likely to respond to the drug. (The company hopes to work with the FDA to design a new clinical trial to test the therapy specifically in this cohort of patients.)

Lantern’s investment so far: less than $2 million, a fraction of the typical outlay by drugmakers for a compound in late-stage testing. “Drugs shouldn’t cost $300,000 per dose,” says CEO Panna Sharma. “If we can do this faster, cheaper, and make drugs more personalized, then that changes everything.”

A version of this article appears in the February 2020 issue of Fortune with the headline “Medicine by Machine.”

More from Fortune’s special report on A.I.:

—Inside big tech’s quest for human-level A.I.
—A.I. breakthroughs in natural-language processing are big for business
—Facebook wants better A.I. tools. But superintelligent systems? Not so much.
A.I. in China: TikTok is just the beginning
—A.I. is transforming HR departments. Is that a good thing?
Subscribe to Eye on A.I., Fortune’s newsletter covering artificial intelligence and business.