AI in Drug Discovery: Evidence of Real Impact Still 'Disappointingly Limited'
A new review in Nature Reviews Drug Discovery concludes the clinical impact of AI in drug discovery is so far 'disappointingly limited'. Where would real evidence show up — and why Phase II success rates are the number that matters.
What happened
On 10 August 2026, chemist Derek Lowe published an analysis on Science magazine's "In the Pipeline" blog of a new article in Nature Reviews Drug Discovery, written by field experts who are not trying to sell anything. Their conclusion: although a wide variety of AI methods have been developed, applied and benchmarked, evidence of clinically relevant impact is, so far, "disappointingly limited."
Where the real effect would show up
The improvement everyone wants is better decision-making: choosing disease areas, targets, lead molecules, clinical candidates and trial designs. Lowe notes current AI techniques are least equipped for exactly these calls — but stresses the word "current": gains are possible, just slower and more expensive than press releases claim.
Phase II success rates are the metric to watch, since trials consume far more time and money than early discovery work. So far, AI has not visibly moved them.
Why the evidence is hard to read
Successful projects get credited to AI while failures are rarely discussed, so the field is still at an "absence of evidence" stage — not proof AI cannot work, but no proof yet that it does.