AI Drug Discovery

AI in drug discovery attracts more capital and more overstatement than any other area covered on this site, largely because the feedback loop is so long that claims cannot be checked for years.

The genuine advances are narrower than the headlines. Protein structure prediction moved from a decade-long experimental problem to a computational one, and that is a real change in what is possible. Generative chemistry can propose candidate molecules with desired properties faster than medicinal chemists can enumerate them. Target identification from multi-omic data surfaces hypotheses humans would not have prioritized.

Where the bottleneck actually is

None of those steps is where drug development fails. Attrition is concentrated in clinical trials, and the dominant causes are lack of efficacy in humans and unanticipated toxicity ... both of which are failures of biological understanding rather than of chemistry throughput. Generating candidate molecules faster does not address either.

The honest position is that AI has compressed the preclinical stage of a process whose expensive failures happen later. Whether that translates into more approved drugs is an empirical question that will be answered by readouts over the next several years, not by preprints now.

Reading a claim carefully

"AI-discovered drug" is a marketing phrase with no fixed meaning. It has been applied to compounds where a model proposed the structure, where a model identified the target, and where a model was used somewhere in an otherwise conventional program. Ask which step the model performed, and what the comparator was.