Bad binder design drove only ~3% of Phase 2/3 failures — the rest elude AlphaFold
clarejtbirch · x · 2026-10-29
A reality check on AI drug discovery hype: bad binder design accounted for only 3% of Phase 2/3 trial failures over the past decade — meaning drugs rarely fail because they can't bind the target.
A chart of 3,907 programs stopped in Phase 2/3 (2015-2025) shows 38% failed on efficacy or exposure (PK, dose, formulation), 26% on business/strategic events, and 24% were positive readouts not yet advanced. Among 1,340 efficacy failures, 46% engaged the target with no benefit and 36% lacked engagement data.
The author argues AlphaFold, Chai, Boltz, and RFDiffusion are genuinely useful but optimized for the 3% problem, leaving 97% of later-stage failures unaddressed.
More from AGI Musings
- Grok reviews its own failed predictions: Gemini users, Cursor, AI bubble calls all wrong — venturetwins · 2026-10-29
- Exec: work has crossed a before/after AI rubicon comparable to the internet — yungcontent · 2026-10-29
- Technical people mourn devaluing skills while non-technical people scramble to learn agents — trq212 · 2026-10-29
- Researchers slam secrecy around OpenAI-linked Yang-Mills proof materials — anshulkundaje · 2026-10-29
- Job-market paper: German engineering societies lifted regional patents 4.8% — Afinetheorem · 2026-10-29
- Eight AI Failure Cases Show Accountability Lands Anywhere but Where Expected — Comfortable_Gene5180 · 2026-10-29