← Back to feed

Anthropic published results showing its internal research model designed protein binders against scorpion neurotoxins in a 64-hour autonomous run. The work was done within Claude Science, the company’s multi-database research platform, in collaboration with Citrini Research, a financial investment research firm whose science team approached Anthropic about applying AI drug discovery to neglected tropical diseases.

The Numbers

• 318,484 protein designs generated in total
• 180 selected by Claude for lab testing
• 31 added by the research team with Claude’s assistance
• 211 total designs sent to Adaptyv Bio and Twist Bioscience for experimental testing
• 17 confirmed binders — one or more against each of four scorpion species

The 17 confirmed binders each attach to one of four scorpion neurotoxins known to cause serious harm to humans. The toxins were drawn from four regions: Mexico (Cn2), Brazil (Ts1), North Africa (AaH I and AaH II), and the Middle East (LqhIII). Claude targeted each toxin’s sodium channel-binding site — the structural feature that causes nerves to fire incorrectly and, in severe cases, triggers heart failure.

Why Scorpion Venom

Scorpions sting more than 1.2 million people per year. Around 3,250 die, most of them children in regions with limited access to medical care. Existing antivenoms are produced by immunizing horses with extracted venom — a process that takes months, produces antibodies effective against only a narrow set of toxins, and results in treatments that are expensive and often unavailable where they are most needed.

De novo protein design offers an alternative path. Rather than antibodies drawn from animal blood, binders are designed computationally against a specific target, optimized for stability and potency, and can in principle be produced without biological feedstocks. An earlier project from Nature demonstrated this approach for snake neurotoxins and protected mice; no de novo binders against scorpion neurotoxins had previously been described.

What Claude Did

Anthropic started with the same base prompt used in its August protein-binder study — a general framework for running autonomous protein engineering campaigns — modified in collaboration with Citrini. An internal research model then worked mainly independently, with four follow-up prompts from the team over 64 hours.

Claude searched the Protein Data Bank and the AlphaFold database to identify toxin structures, selected five amino acid hotspots per toxin as binding targets, generated hundreds of thousands of candidate designs, and then selected the 180 it judged most promising for experimental testing. The research team reviewed Claude’s work and added 31 more designs before the batch was sent to the two contract research organizations.

The team sent Claude’s final report to two independent structural biology experts, who described the results as “a promising first step.”

What This Is Not

Anthropic explicitly does not claim to have produced an antivenom. The 17 binders have not been tested in a cellular environment to confirm whether they neutralize venom, which is the next step before any therapeutic claim is possible. The gap between “confirmed binder” and “drug” involves toxicology, formulation, delivery mechanism, manufacturing at scale, and clinical trials — none of which were part of this study.

The comparison made in the paper is structural: binding at the sodium channel site is the mechanism by which the toxins cause harm, and blocking that site is the principle underlying antivenom design generally. Whether these specific binders perform that function in living tissue remains to be tested.

The Larger Context

This is the second protein engineering result Anthropic has published using Claude Science. The August paper described a broader campaign across 10 targets with 1,440 designs and reported successful binders across all ten. The scorpion toxin work differs in that it was initiated by an external partner, targeted a specific neglected disease category, and produced results on a new class of toxins not previously addressed by de novo design.

All 211 designs and experimental results are available on HuggingFace. Anthropic has framed this as a resource for the field rather than a proprietary finding.

The cost of the compute was not reported. The 64-hour runtime gives some indication of the scale: this was not a short agent run.