On October 20, 2026, Anthropic’s Science team published “How Claude designed protein binders against scorpion neurotoxins”. An internal research model running in Claude Science worked mostly autonomously for 64 hours, guided by an initial prompt developed with Citrini Research and four follow-up prompts, to design de novo minibinders against five scorpion toxins from four regions: Cn2 (Mexico), Ts1 (Brazil), AaH I and AaH II (North Africa) and LqhIII (Middle East). Claude chose the toxin structures and five hotspot residues per toxin itself.
The model used 12 design tools to generate 318,484 designs, filtered out those resembling a toxin or known protein, co-folded and screened 553,052 design-toxin pairs, and picked 180 designs; the team added 31 for 211 total. The campaign used on the order of 10,000 GPU hours. Adaptyv Bio and Twist Bioscience tested binding independently by surface plasmon resonance. A design counted only if it was a hit at both labs, had a KD of 1 micromolar or tighter in every replicate, and showed no off-target signal.
Seventeen of 211 designs passed: 7 against Cn2, 4 against AaH II, 4 against AaH I and 2 against LqhIII. Among single-toxin designs the hit rate was 11 of 60 (18%). The strongest, Cn2-B1 and Cn2-B2, showed single-digit nanomolar affinity, and competition experiments suggest they bind on or near the epitope of an antibody reported to neutralize Cn2 in mice. All 211 designs and their data are released on Hugging Face.
Why it matters: it extends Claude’s earlier binder-design results to a neglected-disease target (the post cites more than 3,250 scorpion-sting deaths a year, mostly children) with an unusually strict two-lab validation bar, showing an agent running a full computational campaign end to end.
What it does not show: binding was tested only against recombinant toxins, not venom-purified ones, and no design was tested for neutralization in cells or animals; the authors state “we do not claim to have produced an antivenom.” The goal of multi-toxin binders for a universal cocktail failed: none of 91 multi-toxin designs bound more than one toxin.