What we learned hosting the top minds in AI for biology in San Francisco

Last week in San Francisco, Owkin put a single question to five people unusually well placed to answer it: will AI scientists solve drug R&D, or merely make the existing process faster? The occasion was a thought-leadership panel we hosted in partnership with Amazon Web Services (AWS). Pascal Weinberger, Owkin's co-chief executive, moderated a conversation with Michelle Chen, CEO of Form Bio; Vivek Natarajan, AI Researcher at Google; Sam Rodrigues, Co-Founder and CEO of FutureHouse and Edison Scientific; David Yang, investor at Lux Capital; and James Zou, Associate Professor of Biomedical Data Science of Stanford University.

I came away fixed on one point the discussion kept circling back to. Ask where AI is already saving time in biotech and the answer is almost always the same: early hypothesis generation and target identification. Ask where the industry is still stuck, and the answer is clinical trials. The real bottleneck sits between those two points: testing, not thinking. This raises an uncomfortable question the panel kept returning to: is the industry lavishing too much attention on the early phase of the pipeline?

‍Why is generating ideas easy and proving them hard?

Generating hypotheses about which molecules might work has become cheap. Testing whether they actually do has not. Identifying possible candidate drug targets now moves at the pace of software; clinical validation still moves at the pace of biology, which is to say slowly, expensively and, critically, with patients involved. AI can widen the search for candidate biology far beyond what a human team could manage, and it can even design trials to run more efficiently, with a better chance of success. What it cannot do is make a patient respond to a new drug faster. That mismatch, more than any model's cleverness, sets the ceiling on progress.

The numbers on the panel put a frame around it. The thinking is that AI is currently delivering gains of roughly 10-15% across drug discovery and development timelines, and up to 50% on narrower tasks such as hypothesis generation and literature triage. That is a genuine improvement, not a revolution. The 14 years it typically takes to go from target to approval could, with AI deployed well across the pipeline, fall to six or seven, the panelists reckoned. That saving does not come from hypothesis generation alone, it comes from AI compressing the many steps in between, including the rate at which experiments can be run and results validated. What it does not touch are the harder, irreducible steps of biology, manufacturing and patient response. Those are precisely what set the floor: they still resist software and no amount of it makes a cell divide or a patient respond any faster.

Is judgement a machine skill yet?

Not in the panelists' view. Models are good at finding patterns and proposing biology that no human team would have considered. They are poor at knowing which of those proposals is worth pursuing. That discernment, sometimes called "taste", remains a human responsibility, at least for now. The likelier near-term arrangement is a division of labour: AI widening the set of options, humans still deciding which to back.

Where the technology is deployed, I’d argue, matters as much as whether it is used at all. A model bolted onto an existing workflow yields incremental gains. Embedding AI as an agent within that workflow, or building whole systems of AI scientists working in concert, is where the panel expected the larger gains to appear - a view shared most strongly, unsurprisingly, by those panelists who build such systems for a living. None of it works, though, without data of sufficient quality and specificity; more data is not the same as better data, and the latter remains scarce.

What still hangs over the field?

What struck me the most are the two structural questions that loomed over the discussion without being resolved. The first was control: biotech firms want to separate the model layer from the application layer and to avoid dependence on a single vendor, even as open-source models close the gap with proprietary ones. The second was consolidation. Lower barriers to entry could spawn a wave of new biotech startups, or, just as easily, accelerate the rise of a handful of AI-native pharma companies that absorb the rest.

Asked when an AI system might design a drug end-to-end, panelists converged on the early 2030s, with 2033 the rough average - and even then, with a human validating the result. That caveat, for me, is the whole argument. The constraint was never imagination: it is proof. I left the room convinced that until the pace of testing catches up with the pace of discovery, the value still to be unlocked in drug development lies less in generating better ideas than in validating them in patients faster. That is a problem no model solves alone. It will take AI builders, biologists, and clinicians working the same problem together. Watch this space.

Discover how K Pro, Owkin's autonomous AI scientist, is designed to address these challenges.