Anthropic's Claude Designs 17 Antivenom Proteins

An internal research build of Anthropic's Claude recently directed an automated biopharmaceutical campaign against five scorpion neurotoxins: Mexico’s Cn2, Brazil’s Ts1, North Africa’s AaH I and AaH II, and the Middle East’s LqhIII. Working from an initial prompt and four follow-up instructions, the AI operated 12 distinct software tools over a 64-hour run requiring roughly 10,000 GPU hours. The pipeline screened 318,484 candidate minibinders, co-folding and scoring 553,052 protein-toxin pairs to identify top candidates.

Claude recommended 180 synthetic candidates, to which human researchers added 31, creating a final testing batch of 211 proteins. Synthesis and surface plasmon resonance evaluations conducted by Adaptyv Bio and Twist Bioscience confirmed 17 successful binders across four targets, marking an overall success rate of 8.1%. For proteins designed specifically against a single target, 11 of 60 succeeded—an 18% hit rate. Top performers Cn2-B1 and Cn2-B2 bound the Mexican Cn2 neurotoxin with single-digit nanomolar affinity, competing for the same epitope site as a known neutralizing antibody. However, no multi-toxin binders succeeded, and zero designs bound the Brazilian Ts1 toxin.

For biomedical researchers, this experiment proves LLM agents can orchestrate complex multi-step bio-design workflows rather than merely generating sequences. Traditional antivenom production requires harvesting venom from some of the 1.2 million annual scorpion stings—which cause over 3,250 yearly deaths—and purifying antibodies from immunized horses over several months. While the candidates still require functional patch-clamp testing, cell assays, animal protection trials, and drug development, Anthropic released all 211 structures, scores, and prompts publicly. Additionally, Anthropic noted a scaled-down single-agent pipeline running on one GPU for 24 hours yields comparable computational scores, drastically lowering early-stage discovery hurdles.