Anthropic asked an internal research model, working inside Claude Science, to design protein minibinders against scorpion neurotoxins, a project set up with the investment research firm Citrini Research. Over roughly 64 hours the model worked largely autonomously, directed by one prompt and four follow-ups, and generated 318,484 designs, of which it selected 180 for testing; the team added 31 more. Contract labs Adaptyv Bio and Twist Bioscience tested all 211 against five neurotoxins from four scorpion species, and 17 met Anthropic's criteria for a binder. Anthropic states that it has not confirmed whether these binders neutralise venom in a cellular environment and does not claim to have produced an antivenom; two outside researchers called the result a promising first step. All 211 designs and their data were released, along with a technical report.
Our takeThe 17-of-211 hit rate, about 8%, applies to binding in the assay used, not to neutralisation - the post is explicit about that gap, and it is the difference between a design demo and a therapeutic. There is also no stated baseline: without the hit rate an older method would get on the same five targets, the number says little about what the model added. The interesting methodological detail is the loop itself - 318,484 candidates filtered by the model, 64 hours of mostly autonomous work with four human corrections, and release of failures as well as hits.