AI Drug Discovery: Plenty of Benchmarks, Little Proof It Reaches the Clinic

Derek Lowe argues that the sheer volume of press releases and papers on AI in drug discovery has outrun anyone’s ability to judge it. The marketing flood makes it hard to tell what these tools actually do, and while researchers have built and benchmarked no shortage of models against one another, there is a conspicuous absence of evidence that any of them has changed the course of a real drug program. Doing well on a leaderboard is not the same as delivering a molecule that survives human trials.

The useful distinction Lowe draws is between domains rich in clean, structured data and everything else. Protein-structure prediction and antibody design have genuinely benefited, because resources like the Protein Data Bank give models something consistent to learn from. Small-molecule discovery and, especially, clinical translation lag far behind — the underlying data is messy and inconsistent, and lab-assay results are poor predictors of what happens in people. That gap is why roughly 85–90% of candidates still fail in the clinic; it is fundamentally a prediction problem, and AI has not cracked it.

His prescription is to stop optimizing for what is easy to model and start tackling what matters: target validation and the translation of preclinical signals into clinical outcomes. Lowe casts himself as a short-term skeptic and long-term optimist — the real bottleneck is a genuinely hard scientific problem, and progress will take longer than the current hype cycle admits.