Article 309: Claude Ran a 64-Hour Protein Campaign Alone. 17 of 211 Designs Bound Scorpion Toxins in the Lab

Anthropic set an internal research model the task of designing miniprotein binders against scorpion neurotoxins, aiming each design at the toxin's channel-binding site. Claude worked mainly autonomously for 64 hours from a single prompt plus four follow-ups, chose the toxin structures from the Protein Data Bank and AlphaFold database, selected five hotspot residues per toxin, and generated 318,484 designs. It picked 180 for lab testing; Anthropic's team added 31. Contract research organisations Adaptyv Bio and Twist Bioscience tested all 211. Of those, 17 met Anthropic's criteria for a binder, covering four of the five targeted neurotoxins — Cn2 (Mexico), Ts1 (Brazil), AaH I and AaH II (North Africa), LqhIII (Middle East). Best affinities reached nanomolar range, and seven hit Cn2.

The framing matters. Scorpions sting more than 1.2 million people a year and kill over 3,250, mostly children in places with limited medical access. Existing antivenoms are horse-derived, take months to make, are effective against only a narrow set of neurotoxins, and are expensive and often short in supply. That economics is why envenomation is classed as a neglected tropical disease — nobody expects to make their money back. Anthropic ran the campaign with Citrini Research, which asked whether motivated non-experts with AI could attack a real health problem.

Anthropic is explicit that it has not shown these binders neutralise venom in cells or animals, does not claim an antivenom, and released all 211 designs and their data on Hugging Face for others to check. Outside experts Susana Vázquez Torres (CNIO) and Timothy P. Jenkins described it as a promising first step. Claude also analysed data, generated figures and drafted the report.

Key Takeaway: The 8% hit rate is not the news. The news is that binding has stopped being the bottleneck and the claim has stopped being inflated — and those two things arrived together, which is unusual enough to note. A year ago a result like this would have been announced as an antivenom. This time the lab led with what it had not proven, published the negative 92%, and named the remaining pipeline — neutralisation, cell assays, animal models, safety, formulation, cold chain, distribution — as the actual problem. That candour is what makes the number credible. The strategic point is that the neglected-disease case is the cleanest demonstration of AI research leverage available: no commercial incentive to protect, a hard technical target to test against, a public benchmark in the form of a wet lab, and a real mortality figure at the end of it. If agentic design keeps working here, the argument for applying it to the next twenty neglected diseases is straightforward. If it stops at binding, the field will have learned that in public, on data everyone can download — which is also the right way to learn it.

Sources: Anthropic research post and technical report, 20 October 2026; Adaptyv Bio; Twist Bioscience; ProteinBase collection