Drug Discovery Has No Magic Wands

Daphne Koller has spent nearly three decades straddling AI and biology, the last ten years running a drug discovery company, and she's used that vantage point to puncture one of Silicon Valley's favorite fantasies: that a sufficiently smart AI will simply read the human body like a text file and hand back cures. Her argument, laid out in a lengthy essay published on insitro's Substack, is that this framing gets the actual bottleneck in medicine backwards.

Drug discovery breaks into three stages, she explains: finding the right biological mechanism behind a disease, engineering a molecule that hits that mechanism, and then proving it works in patients through clinical trials. Almost all the AI hype, from AlphaFold's protein-folding breakthrough onward, has piled into stage two, the molecule-design part. And it's genuinely gotten better and faster. But Koller points out that over 90% of drugs entering clinical trials still fail, a number that hasn't budged in decades, and in most of those cases the molecule was engineered fine. The target was wrong. Better key-cutting doesn't help if you're making keys for the wrong locks, as she puts it, and the number of novel drug targets pursued industry-wide dropped from around 100 in 2015 to roughly 30 in 2024.

She's equally skeptical of the idea that large language models can just reason their way through the published literature to new mechanistic insights. That assumes we've already collected enough biological data to contain the answers. We haven't. Human biology spans DNA, proteins, cells, tissues, whole organisms, all shaped by billions of years of messy evolution rather than clean engineering. The largest cell atlases built so far, spanning hundreds of millions of cells, are still orders of magnitude too small to map how biological systems respond to intervention, and diseases like Alzheimer's or ALS are frustratingly human-specific, meaning animal models and existing datasets barely apply.

Koller also takes aim at the vision of AI agents running closed-loop experiments in automated labs, iterating like a coding assistant against a compiler. That model works beautifully when feedback is instant and cheap, which is exactly why AI has crushed coding and molecular design tasks. Drug development offers no such shortcut. The only real scorecard is a human clinical trial, and that takes years, costs millions, and is bound by ethics and biology in ways compute can't route around. Speeding up trial logistics or drafting regulatory paperwork faster, she notes, only accelerates failure if the underlying mechanism was wrong to begin with.

Her prescription isn't anti-AI, though. It's a call to redirect it toward building targeted causal models of disease and identifying clinical readouts that pick the right patients and catch failures earlier, rather than chasing the fantasy of AI as a shortcut around biology itself.