What matters in AI this week

By Krystal Hu, Technology Correspondent   

U.S. dollar banknotes and medicines are seen in this illustration taken, June 27, 2024. REUTERS/Dado Ruvic/Illustration 

By now you've probably heard, more than once, that AI is going to help us cure cancer. But how far away are we, actually?

From AI labs to pharma giants to Big Tech, everyone is betting that biology is the next holy grail of AI promises. We're in the early innings, researchers and scientists told me. And one major obstacle standing between AI and drug discovery's big breakthroughs? Data.

Large language models got good by scraping the internet's text — an enormous, already-existing corpus. Biology has no equivalent. There is no "internet of cells" sitting around waiting to be ingested.

"We need to capture the language of biology, we need to capture the language of the cell. And that doesn't exist today," Alex Rives, head of science at the nonprofit Biohub, told me. Current cell datasets run to hundreds of millions of cells; an accurate predictive model will need billions, eventually trillions. That gap — several orders of magnitude — is the real bottleneck for now, Rives said.

So Biohub is trying to build that dataset from scratch, and has rallied an unusual coalition to do it: the U.S. government, Google, and Meta have joined the project, bringing total investment to $1.8 billion. Rives says the group is compressing work that would normally take decades into five years, with a first dataset expected to be ready in about a year, and usable models possible within five.

Even then, the path from dataset to actual treatment is uncertain. Rives said researchers can only begin training models once the data exists, then test whether those models can accurately predict how a cell responds to change. If that works, it could help scientists pinpoint exactly which part of a diseased cell to target.

Some already take the experiments to a physical world. Danaher, the life sciences giant, said this week it plans to launch its first AI-powered autonomous research lab by early 2027, aiming to design, build, and test new antibodies up to eight times faster than conventional methods.

Anthropic is making a similar bet. The company has quietly set up its own wet lab in the Bay Area, where human scientists and AI agents work side by side. 

Eric Kauderer-Abrams, Anthropic's head of life sciences, told my colleague that the final test for any biology claim still has to happen on a lab bench, not inside a model. Anthropic says it isn't chasing drug discovery for now; it's targeting neglected corners of basic biology instead.

While the industry waits for a full "virtual cell," it's also trying to make some corners of the predictions work today: a wave of AI companies are running virtual drug trials, simulating how an experimental medicine might perform before it ever reaches a real patient. 

The stakes are real. The biopharma industry spends roughly $140 billion a year on human clinical testing, and only about 12% of drug candidates ever win regulatory approval — a rate that's barely budged in decades. One startup, BioinvestGPT, says it correctly predicted the outcome of five of six high-profile trials ahead of time, including, this past July, forecasting that Novartis' muscular dystrophy drug would fall short — months before the actual trial missed its goal and erased $30 billion in market value.

The simulation side of this is already saving companies money and time. The "cure cancer" side is a different story—one that might require building a model that doesn't exist yet, at a scale nobody has ever attempted. Investors are pricing in the first story while trying to predict the timeline of  the second.

This newsletter was edited by Rosalba O'Brien