Claude designs 318,484 protein binders against scorpion toxins; 17 of 211 lab-tested designs bind

Anthropic's Claude generated 318,484 protein binder designs against scorpion neurotoxins, and 17 of the 211 designs tested in the lab met the criteria for a binder. The work was done largely autonomously inside Claude Science, Anthropic's AI environment for researchers. Scorpion stings kill more than 3,250 people every year.

What Claude did

Protein binders are small protein molecules shaped to latch onto a specific target molecule. Claude targeted five toxins from four scorpion species.

After an initial prompt, Claude worked on its own for 24 hours. The research team then steered the work with additional instructions. From the pool of 318,484 designs, Claude selected 180 candidates, and the team added 31 more, for 211 designs tested in the lab.

Results

• Designs generated: 318,484
• Designs tested in the lab: 211 (180 chosen by Claude, 31 added by the team)
• Designs meeting the binder criteria: 17 (about 8%)
• Strongest performers: Two designs bound extremely tightly to Cn2, a toxin from the Mexican scorpion.

Lab tests suggest these two dock at or near the same site as an antibody known to neutralize Cn2 in mice. Two independent labs confirmed the results. External experts Susana Vázquez Torres and Timothy P. Jenkins called the work a promising first step.

What is not established

The binders have not been tested as an antivenom. Binding to a toxin in the lab does not show that a molecule neutralizes it in an animal or a patient. The source does not specify which Claude model was used. It also does not give binding affinity values, the exact binder criteria, or the specific toxins beyond Cn2.

Earlier work

In an earlier study, Claude designed mini-binders against 14 of 15 target proteins. Hit rates ran from 22 to 35 percent, against a typical 10 to 15 percent for conventional approaches.

What this means

The notable change is the workflow. Claude ran a generate-and-select pipeline over hundreds of thousands of designs for 24 hours with minimal human input, and wet-lab results from independent labs back its picks. That is stronger evidence than a computational benchmark.

The scorpion hit rate of about 8% is below the 22 to 35% from the earlier study and the typical 10 to 15% cited for the field. The two figures are not directly comparable. The targets differ, and the source does not say the hit criteria are the same. Toxins may simply be harder targets.

The main open question is function. Until the binders show neutralization in animal models, this is a demonstration of design capability, not a therapeutic. The two strongest binders sharing a site with a known neutralizing antibody is the most promising sign, but it is indirect evidence.