Anthropic said Claude designed 17 lab-confirmed protein binders against scorpion neurotoxins after working mostly autonomously for 64 hours. An internal research model, run inside Claude Science, received one initial prompt and four follow-ups, then generated the designs, analyzed the lab data, produced the figures and drafted the report. Anthropic published the results on October 20 as an early step toward a possible scorpion antivenom.
The targets were five neurotoxins from four regions: Cn2 from Mexico, Ts1 from Brazil, AaH I and AaH II from North Africa, and LqhIII from the Middle East. Claude pulled the structures from the Protein Data Bank and the AlphaFold database and picked five hotspot residues per toxin on the site where each toxin binds sodium channels in nerve cells. The toxins share one overall fold but only 26% to 62% pairwise sequence identity, which made binders aimed at several toxins at once a demanding test.
The medical case is specific. Scorpions sting 1.2 million or more people every year, causing more than 3,250 deaths, mostly among children in areas with limited medical care. Current antivenoms are made from horse-derived antibodies, take months to produce, usually work against only a few neurotoxins, and are costly or scarce in many regions.
Key facts
- 0117 of 211 tested designs, about 8%, met Anthropic's binder criteria across four scorpion neurotoxins.
- 02Claude worked mostly autonomously for 64 hours from one initial prompt and four follow-ups, using 12 design tools.
- 03The run generated 318,484 designs and screened 553,052 design-toxin pairs on roughly 10,000 GPU hours.
- 04Lead binders Cn2-B1 and Cn2-B2 bound the Mexican toxin Cn2 with single-digit nanomolar affinity at both test labs.
- 05No multitoxin design passed, and no design has yet been tested for neutralization in cells or animals.
On the computational side, Claude used 12 design tools to generate 318,484 candidate designs, filtering out any that resembled the toxins or known proteins. It co-folded and screened 553,052 design-toxin pairs and selected 180 designs at the 64-hour mark, and the team later added 31 for a total of 211. The campaign used roughly 10,000 GPU hours, though Anthropic noted that a simplified single-agent run on one GPU over 24 hours reached comparable in silico scores.
Claude set its own quality thresholds, chose how to use the tools and developed the strategies for multitoxin binders. Most of the 180 designs, including the strongest binder, were selected after the third follow-up prompt. The prompt was adapted with Citrini Research, and the project tested whether motivated nonexperts could use Claude to take on a real health problem.
Two labs ran the experiments. Adaptyv Bio tested designs in a cell-free expression system, Twist Bioscience expressed them in human kidney cells as Fc fusions, and binding was measured with surface plasmon resonance. A design counted as a binder only if it hit at both labs, bound in every replicate, reached a dissociation constant of 1 µM or tighter, and showed no signal against the off-target protein CLEC12A.
Seventeen of the 211 designs, about 8%, cleared that bar: seven against Cn2, four against AaH II, four against AaH I and two against LqhIII. Among the 60 single-toxin designs, 11 were binders, an 18% hit rate Anthropic described as consistent with typical protein design campaigns. The two strongest, Cn2-B1 and Cn2-B2, bound Cn2 with single-digit nanomolar affinity at both labs.
“Designing a binder against almost any target no longer seems unrealistic.”— Susana Vázquez Torres, Postdoctoral Researcher, Spanish National Cancer Research Centre (CNIO)
Competition experiments suggest those two binders attach at or near the same site as antibody LR, which has been reported to neutralize Cn2 in mice. That site covers the part of Cn2 that interacts with the sodium channel.
The multitoxin goal largely failed. No design met the criteria for more than one toxin, and Anthropic listed prompting, tool limitations and biological feasibility as possible reasons. Binding was also tested only against recombinant toxins rather than toxins purified from venom, and no design has been tested for neutralization in cells or protection in animals, so Anthropic makes no antivenom claim.
Timothy P. Jenkins, an associate professor at the Technical University of Denmark, said designing credible binders against five toxins in parallel within days showcases an AI-driven shift in therapeutic discovery. Vázquez Torres said binding is no longer the main bottleneck and that the challenge now is turning such molecules into real drugs.
Next steps for the lead candidate, Cn2-B1, would include testing against venom-derived toxin, patch-clamp electrophysiology and mouse studies. Anthropic has released all 211 designs, their predicted structures and scores, and the experimental data on Hugging Face alongside a technical report.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
