Claude autonomously generates 318,000 protein designs in 64 hours, hits 14 of 15 targets
kimmonismus · x · 2026-10-20
Anthropic's blog details two results showing Claude accelerating life science:
- Autonomous protein design: Given one starting prompt and four follow-ups, Claude (Mythos Preview and Opus 4.8) used specialized tools to generate 318,000+ protein binder designs against scorpion toxins in 64 hours, succeeding on 14 of 15 targets; 22–35% of individual designs bound successfully vs the typical 10–15%, with some binding several times tighter than the best published results. Lab-confirmed.
- Analytical chemistry: Claude Opus 5, given only raw NMR and LC-MS files and a two-sentence prompt, returned finished analyses in 23 and 19 minutes, matching the lab's own hydrogen counts and purity (96.4% vs 96.33%).
The author frames this as the start of autonomous AI scientists; the post also notes agents solving decades-old Erdős problems at several per month.
Related event: Claude Autonomously Designs 318K Proteins in 64 Hours, Lab-Validated(2 posts)→
More from AGI Musings
- John Smart's new talk: Biostasis drops to $36K as AI reshapes the future of self — Chris_Armstrong · 2026-10-20
- Mathematicians aren't immune to sycophantic AI; weak verification is the culprit — gerardsans · 2026-10-20
- noOS: the OS's only job will be handing flops to the agent, says Papailopoulos — DimitrisPapail · 2026-10-20
- The problem with the "AI as a normal technology" slogan — intellectronica · 2026-10-20
- AI existential dread has a cognitive core that soothing emotions won't fix — zetalyrae · 2026-10-20
- Scoble: As AI Makes Everyone Look Like an Expert, Advantage Shifts to Judgment and Taste — Scobleizer · 2026-10-20