AI is getting drugs into the clinic faster. But is it making them work better? The clinical numbers tell a surprisingly uncomfortable story: Phase I: 80–90% success vs. 40–65% historically. Phase II: ~40% vs. ~37% historically.
The AI advantage appears to disappear exactly where human biology gets tested: Phase I asks whether a molecule is safe and behaves as expected—areas where AI is genuinely good at optimizing molecular properties. Phase II asks the harder question: “Did we choose the right biology?”
So far, the answer is: AI has not demonstrated an advantage here. Rentosertib is an interesting counterpoint: an AI-originated drug that reached Phase 2a in IPF and showed a real FVC signal. But even there, biology remained unforgiving: 7 of 12 discontinuations involving adverse events were associated with liver injury. We can search chemical space faster and get candidates to the clinic faster. But the deeper issue is this: AI may be much better at optimizing a molecule than validating the biology behind it. That distinction matters when deciding where AI investment should go next.
Detailed report: https://lnkd.in/evxP52d4
Where in the drug-development pipeline does the clinical evidence actually show an AI advantage?
AI has proven its strength in Phase I by optimizing molecular properties and pharmacology (80–90% success). However, Phase II success rates (~40%) remain identical to historical baselines because the fundamental bottleneck is choosing the right disease biology, not designing molecules.