A single bad trial design may have cost ~$30B — where translational AI could help
hardimanjames · x · 2026-09-25
An argument for translational AI in drug development, using the PD1 race as a counterfactual: BMS's Opdivo targeted a broader population (>5% PDL1 expression) while Merck's Keytruda was more selective (>50%). Merck's head-to-head readout dominated (OS hazard ratio 0.6 vs 1.02), making Keytruda a $30B drug while Opdivo trailed at $10B.
The author estimates the NPV lost from that single poor trial-design decision could be on the order of $30B — exactly the kind of upfront decision strong translational AI could help with.
Related event: A Flawed Trial Design Cost $30 Billion — Could AI Have Prevented It?(2 posts)→
More from Research
- Lab lands 7 NeurIPS main-track acceptances spanning foundation models, RL traffic control and optimizers — flosalim · 2026-09-25
- YorkU-led team scores 3 NeurIPS 2026 acceptances on circuit discovery, hallucination mitigation and test-time scaling — CSProfKGD · 2026-09-25
- Primacy bias exists in RLVR training: new paper accepted to NeurIPS 2026 — ParshinShojaee · 2026-09-25
- New NeurIPS 2026 paper reveals primacy bias in RLVR training of LLMs — ParshinShojaee · 2026-09-25
- Historian uses GPT-6 and Opus 5.5 to crack John Dee's ciphers, urges lab funding — emollick · 2026-09-25
- Judea Pearl turns 90: UCLA hosts tribute with Koller, Horvitz, Domingos, videos online — AlexTensor · 2026-09-25