AI Can Speed Drug Discovery. But Is It Really Better Than a Human?

“It’s not like the human is not needed anymore,” says Aviv Regev, a Harvard University and Massachusetts Institute of Technology computational biologist who took a leave from her academic work to run Genentech’s research and development. “But the human all of a sudden gets the superpower.”

The project is just one example of the pharmaceutical industry’s embrace of AI to turbocharge drug development. Genentech’s Swiss parent, Roche Holding AG, has some of the most ambitious plans, including building its own generative AI tool, dubbed RocheGPT. The goal is not just to move faster but also to answer questions that couldn’t be answered before, Regev says. She describes the AI boost as like being a scientist “on steroids.”

But the true test for biotech will be whether drugs developed with the help of AI are more likely to be successful for patients than those developed by mere mortals. While that’s far from certain, money is pouring in. Over the past decade investors have pumped about $18 billion into “AI-first” biotech companies, outfits that have built their R&D workflow around AI tools, Boston Consulting Group found last year. And at January’s JPMorgan Healthcare Conference in San Francisco, the industry’s biggest meeting, nearly every chief executive officer mentioned AI in their presentations.

The AI gold rush in pharma and biotech has also attracted the interest of the world’s most valuable chipmaker, Nvidia Corp. The Silicon Valley darling saw its market capitalization more than triple in the past 12 months, to $1.56 trillion—more than twice that of the most valuable drugmaker, Eli Lilly & Co.—driven by interest in ChatGPT and other generative AI technology that relies on the kind of computing firepower Nvidia’s chips can provide.

Nvidia has been trying to persuade the pharmaceutical industry to embrace its chips, which can sell for $50,000 apiece, for more than a decade. Now the company believes it has its foot firmly in the door. It’s doing at least some business with the 20 largest pharmaceutical companies and more than 2,500 startups. And it’s signed more in-depth research deals with a few Big Pharma giants, including Roche. What’s been done so far is barely scratching the surface of what AI could do for drug development, says Nvidia CEO Jensen Huang.

At a panel during the JPMorgan conference, Huang told a standing-room-only crowd that, within the next decade, drugs could be designed almost entirely in simulation via computing platforms like the ones his company supplies. “We are determined to work with you to advance this field,” he said.

That would mark a seismic shift in the world of drug development. It typically takes 12 to 15 years to bring a drug to market, according to BCG. The consulting firm says AI-driven R&D could help cut 25% to 50% of the time and cost of bringing drug candidates to the point of human testing—but it will still require study to prove whether AI-aided drugs have a higher probability of clinical success.

Nonetheless, Kimberly Powell, head of Nvidia’s health-care segment, says the pharma business is quickly changing in response to AI, with many of the startups Nvidia works with today considering themselves “techbio” companies rather than biotechs—using data to “drive what biology they’re going after, instead of the biology influencing what technology they need to use.”

The key is AI’s ability to make sense of huge volumes of myriad types of data, she says. “Here we are in the ChatGPT moment, that makes it just obvious that the future is talking to your data,” Powell says. “The future is being able to put very, very complex data into models that allow us to reason across it in ways that human capacity just could never do. You can actually use spoken language or type in, ‘Please generate new molecules that appear like this one but have these characteristics.’ And what does it do? It goes off and it launches. And then what comes back is your top 10 that you think that you should go after.”

The chip industry hasn’t always lived up to such bullish projections. Nvidia, for instance, was one of a number of companies that predicted that, by around 2020, machine intelligence would put automated vehicles in charge of one of the most complex and dangerous human activities: driving. While the automotive industry has massively increased the amount of technology it crams into every vehicle, robotaxis and their ilk are still far from becoming part of everyday life.

“Will we in five years see full-blown drug discovery based on AI? I think that’s the million-dollar sort of question,” says Anders Romare, chief digital and information officer at Novo Nordisk A/S, the Danish maker of diabetes drug Ozempic and weight-loss shot Wegovy. Novo has already deployed AI throughout the company, using it for everything from speeding up regulatory submissions to overseeing production quality. Workers use ChatGPT inside the company’s firewall more than 50,000 times a month, Romare says. But while AI can speed up the work, “ultimately putting a drug in the hand of the patient has got to be a human decision based on human knowledge and understanding,” he says.

At Roche, the AI effort is part of a push to get back to the forefront of drug development after a series of failed late-stage clinical trials, including for cancer and Alzheimer’s disease medicines. Genentech is a biotech pioneer behind some of the best-selling cancer drugs in history, but it has struggled to develop blockbusters in recent years.

In the ulcerative colitis project, researchers already had science to suggest that an experimental drug Genentech had acquired in August 2022 from Kiniksa Pharmaceuticals might be useful against disease in the colon. They used machine learning to analyze single cells, then another digital tool to run what’s essentially a reverse image search for a specific type of cell in the colon that could be influenced by the drug.

Genentech likens it to training an algorithm to pick out a blueberry from within a fruit salad—in other words, teaching it to recognize cells that display a certain disease characteristic and then having the AI find cells with similar characteristics in other parts of the body. But only the patient trial will show how helpful the algorithm-enabled drug can actually be.

Scientists have already shown that algorithms can massively accelerate the time required to find a chemical compound that will hit a certain target on a cell—finding a key to fit a lock, so to speak, says Alpha Lee, a University of Cambridge physicist who has taken an academic leave until 2027 to co-found PostEra, a machine learning biotech company. What hasn’t been shown yet is whether AI can be more effective than humans at finding the right lock to open, Lee says. That’s why AI won’t prove itself fully in biology, he says, until human studies like the one Genentech is conducting read out.

“Sometimes when you open the door, you find out that there’s actually nothing behind it,” Lee says. AI is “not a panacea that can solve everything” in drug development, he says. “Now perhaps it is the time to take a step back and be very concrete about what are the key pain points that we can solve [with AI] right now, and what’s the roadmap toward the future.”

More stories like this are available on bloomberg.com

©2024 Bloomberg L.P.