Anthropic reports that an internal Claude research model generated 318,484 protein designs aimed at scorpion neurotoxins during a 64-hour design campaign. According to Anthropic's account, partner labs tested 211 selected designs, and 17 met its criteria for binding to a toxin. Anthropic does not claim to have produced an antivenom: the tests measured binding, not whether the designs stop toxins from acting.
Working with investment research firm Citrini Research, Anthropic gave the model an initial prompt specifying five toxin targets, design tools and filtering rules, then directed it with four follow-ups. Claude chose toxin structures and sites for the designs to target. It used 12 design tools and produced a shortlist of 180 designs 64 hours after the initial prompt. Anthropic's team later added 31 designs with Claude's help, bringing the lab-testing set to 211. Citrini's account also describes the initial 180 designs and the later testing set.
According to Anthropic, Adaptyv Bio and Twist Bioscience tested the designs for binding. Seventeen of the 211 met its criteria, across four of the five targeted toxins; none qualified against Brazil's Ts1. No design intended to target multiple toxins qualified as a binder for more than one. Anthropic says its criteria generally required binding in every replicate at both labs, though LqhIII was assessed only by Adaptyv Bio.
The tested toxins were produced in a lab rather than purified from venom. Anthropic says it did not test any design for neutralization in cells or for protection in animals. Its proposed next steps include checking binding to toxin purified from venom, testing whether a binder stops toxin action in cells, and testing protection in mice. Anthropic says it is releasing the 211 designs and their binding data for further research.
