Anthropic Tests Claude on Scorpion-Toxin Protein Design
Claude generated and screened hundreds of thousands of protein designs, with 17 of 211 tested candidates binding scorpion neurotoxins in the lab.
Anthropic says an internal research model operating within Claude Science designed protein binders against five scorpion neurotoxins in a campaign that ran largely autonomously for 64 hours.
From computational designs to laboratory tests
The model generated 318,484 candidate proteins and screened 553,052 design-toxin pairs using structure-prediction and protein-design tools. It selected 180 designs for testing, while Anthropic added 31 more. Adaptyv Bio and Twist Bioscience produced and tested all 211 candidates.
Seventeen met Anthropic’s criteria for binding a toxin. The successful designs targeted four toxins from Mexico, North Africa, and the Middle East. The strongest candidates against the Mexican toxin Cn2 bound in the single-digit nanomolar range and appeared to compete for a site also targeted by an antibody previously reported to neutralize the toxin in mice.
The result is notable because the campaign addressed multiple difficult targets and included an attempt at multitoxin binders. That attempt failed: none of the designs aimed at more than one toxin met the stated threshold for binding multiple toxins. The successful rate among single-toxin designs was 11 out of 60, which Anthropic says is consistent with typical protein-design campaigns.
The important limitation
Anthropic is not claiming to have produced an antivenom. The binders have not been tested in cells or live animals to establish that they neutralize venom, remain stable in the body, or are safe. They would still need optimization, formulation, manufacturing, and clinical development.
The significance is therefore methodological rather than medical. A model was able to manage a long design loop, select targets and thresholds, use a collection of specialized tools, and produce experimentally credible candidates in days. That compresses the early discovery phase, particularly for neglected diseases where commercial incentives are weak.
It also clarifies where the bottleneck has moved. Generating plausible binding molecules is becoming faster; proving that they work as affordable, deployable treatments remains the harder problem. Anthropic has released the 211 designs and their data, giving outside researchers a chance to test whether the result generalizes beyond this campaign.