AI drug discovery faces a $400 billion question on Wall Street

Artificial intelligence has proven it can accelerate what scientists already know how to do, but drug development is where AI collides with the stubborn reality of human biology — and Wall Street's patience.

"AI is not smarter," Aviv Regev, a computational biologist at Roche's Genentech, said. "But what helps our scientists is that it encodes information very, very broadly."

At Genentech, Regev has built what she calls a "lab in a loop," where AI models predict promising drug targets and molecules, researchers test those predictions experimentally, and the resulting data feeds back into the models. The approach has expanded the range of research programs scientists can pursue. Yet only about one in 10 drug candidates that enter human trials ultimately reaches the market, with many failures coming after years of research and billions of dollars of investment.

Goldman Sachs estimates the present value of AI's benefits to drug development could reach as much as $400 billion over the next decade by shortening timelines, lowering costs and improving success rates. But for now, AI is making research labs more productive without changing the metric that matters most: how many successful drugs emerge from each research dollar.

The Eroom's Law problem

Drug research has moved in the opposite direction of Moore's Law: slower and more expensive over time. Scientists call this Eroom's Law — Moore's Law spelled backward. Jack Scannell, who co-wrote the 2012 paper that coined the term, argues that each new drug must outperform the ones that came before it, including cheap generic medicines that already work.

In theory, this is the kind of wall AI was built to break, by searching through accumulated scientific knowledge and uncovering possibilities no human researcher would have found. But finding possibilities is not the same as proving they work. Cell cultures, lab animals and computer models remain imperfect stand-ins for the human body.

Scannell put the challenge vividly: Training an AI on today's biological data can be "like trying to train your Waymo for San Francisco by getting a frog to ride a bike around Albuquerque." Autonomous vehicles work because real cars have logged millions of real miles. Medicine has never had anything close to a clean map of the human body.

The second inning of AI in biotech

Even AI's believers acknowledge that proving its potential will take several more years. Eric Kauderer-Abrams, who leads life sciences at Anthropic, says that AI will bend the curve by attacking multiple bottlenecks at once to boost a drug's clinical probability of success. Still, speaking as Anthropic launched its new Claude Science platform, he noted the industry is only in the "second inning."

For investors, that creates a dilemma. It is hard to give drugmakers credit when the cost of bringing a medicine to market has done little but climb. Pharma CEOs including Eli Lilly's David Ricks and Novartis's Vas Narasimhan are committing billions to computing power, partnerships and new research platforms. Yet the distinction is between today's economics and tomorrow's possibilities.

Biotech investor Rod Wong, managing partner at RTW Investments, argues that intensifying competition from China could become a catalyst for change. China's advantage is in the speed at which companies can move from research ideas to clinical evidence. That pressure could force the US to rethink a trial system that has become increasingly slow and expensive.

If those pieces come together, Eroom's Law might finally begin to bend. The winners may not be the companies that build the AI models themselves, but rather those that combine powerful tools with the deepest biological data and the ability to run global drug-development programs. Those advantages favor the largest established pharmaceutical companies, noted Citi healthcare strategist Traver Davis.

But biology runs on its own clock, not the semiconductor cycle. The revolution might yet be real, but we will not know for certain for several more years. In drug development, real and soon are rarely the same thing.

This article is for informational purposes only and does not constitute investment advice.