So How Is AI Drug Discovery Doing, Really?

gist

Derek Lowe argues that AI drug discovery still has little evidence of clinically meaningful impact. The decisive test is not better performance on convenient molecular benchmarks but improved Phase II success, where biological uncertainty and development cost become severe. Progress requires expensive data generation, realistic systems, and careful causal evaluation. Noisy assay labels, confounding variables, selective storytelling, and investor incentives currently make technical activity easy to mistake for better medicines.

ideas

• Clinical translation is the governing metric. Faster early discovery matters little unless it improves candidate selection and raises success in costly human trials.
• Absence of evidence is not evidence of failure. Current results do not rule out future gains, but they cannot support the stronger claims common in press releases.
• Drug-discovery data is conditionally true. Assay conditions, biological context, confounding, and opaque labels can make benchmark improvements fail to transfer to real systems.
• A ligand is not a drug. Each move from isolated protein to cells, animals, toxicity studies, and patients introduces new constraints that a useful system must survive.
• Incentives favor legible progress. Generating the data needed for clinical impact is slower and riskier than optimizing existing benchmarks, making hype the cheaper strategy.