In a campaign Claude ran largely on its own in 64 hours, 17 of 211 protein designs bound scorpion neurotoxins

Anthropic said on October 20 that an internal research model, working in Claude Science, designed small binding proteins called minibinders against scorpion neurotoxins. The idea came from the investment research firm Citrini Research: could motivated nonexperts use AI to tackle a real health problem? Working from an initial prompt and four follow-ups, the model ran mostly autonomously for 64 hours against five toxins from Mexico, Brazil, North Africa and the Middle East. Per Anthropic, scorpion stings kill more than 3,250 people a year.

Claude used 12 design tools to generate 318,484 designs, screened 553,052 design-toxin pairs and picked 180. The team added 31 more with Claude’s help; the campaign used about 10,000 GPU hours. In surface plasmon resonance (SPR) tests at Adaptyv Bio and Twist Bioscience, 17 of the 211 designs met strict binding criteria. Each bound a single toxin, and none of the multi-toxin designs succeeded. The two strongest bound Cn2, from a Mexican scorpion, with single-digit nanomolar affinity, apparently at the same site as LR, an antibody reported to neutralize Cn2 in mice.

No design has been tested for neutralization in cells or animals, so Anthropic does not claim to have produced an antivenom. Tests used lab-made toxins rather than toxins purified from venom. All 211 designs and their data are on Hugging Face. Two outside experts, Susana Vázquez Torres of CNIO and Timothy P. Jenkins of the Technical University of Denmark, called the work a promising first step.