Claude autonomously designs scorpion-antidote proteins in 64 hours; 17 candidates validated in wet labs
AI寒武纪 · wechat · 2026-10-21
Anthropic had Claude autonomously design protein binders against five lethal scorpion neurotoxins in 64 hours, with all 211 designs and data open-sourced.
- Humans gave only an initial prompt and four follow-ups; the model used 12 protein-design tools, generated 318,484 candidates, ran 550k+ structure predictions (10k GPU-hours), and even emailed a synthesis vendor itself.
- 211 designs were tested by two independent labs (Adaptyv Bio, Twist Bioscience) via SPR under strict criteria (KD ≤ 1µM, no off-target binding); 17 passed, with single-toxin designs hitting an 18% success rate.
- Top binders Cn2-B1/B2 reached single-digit nanomolar affinity against Mexican toxin Cn2, occupying the same epitope as a known neutralizing antibody.
- Limits: none of the 91 multi-toxin designs met the bar; only recombinant toxins and in vitro binding were tested — no patch-clamp or animal studies yet.
- Context: scorpion stings affect 1.2M+ people yearly, killing 3,250+, mostly children in underserved regions; traditional horse-serum antivenoms are slow and costly.
More from AGI Musings
- AI advocates push back on bubble talk: local models mean there is no moat this time — RachelVT42 · 2026-11-01
- AI Impacts' '5% extinction' framing was misleading: the survey also asked about severe disempowerment — NathanpmYoung · 2026-11-01
- David Sacks: Anthropic's AI Already Did 26% of the Work Building Its Next Models — rohanpaul_ai · 2026-11-01
- First ruling of its kind: Brazilian lawyers fined for white-text prompt injection in court filing — evisoft · 2026-11-01
- AI safety researcher reshares: AIs could be angels among men, yet we turn them into conduits for our sins — repligate · 2026-11-01
- Is 'one winner takes all, the rest go to zero' still Anthropic's investment thesis? — teortaxesTex · 2026-11-01