Claude Designed 17 Proteins Against Scorpion Toxins in 64 Hours

Anthropic's internal research model designed proteins that bind scorpion neurotoxins in 64 hours. It generated 318 484 designs and selected 180, and the team added another 31. Adaptyv Bio and Twist Bioscience tested all 211 in the lab, and 17 passed strict binding criteria. No one measured venom neutralization, so this is not an antivenom. It is a pitch for a method.

The task was devised together with the analytics firm Citrini Research. Its representatives asked whether AI and a group of motivated non-specialists could take on a real public health problem. Scorpions sting at least 1.2 million people a year, and more than 3250 die, mostly children in regions with poor medical care. Conventional antivenoms are made from the antibodies of immunized horses, a process that takes months. They work against a narrow set of toxins and are expensive. New treatments are hard to make financially viable, so the field is considered neglected.

How the campaign worked

The model ran inside Claude Science from an initial prompt and four follow-up refinements. It first chose toxin structures from the PDB and AlphaFold DB on its own and assigned five "hotspots" to each, the points where a binder should attach. The target was the site the toxin uses to latch onto sodium channels in nerve cells. The model then set quality thresholds and distributed the workload across tools. In total it used 12 design tools, filtered out anything that resembled a toxin or a known protein, and ran 553 052 design–toxin pairs through co-folding prediction.

For the first 24 hours the model worked fully autonomously. The remaining time went to refinements, mostly about designing against several toxins at once. The model picked the strongest binder, Cn2-B1, only after the third refinement. At the researchers' request, it submitted the sequences to Twist Bioscience itself, and the emails were labeled as written by Claude. The campaign used roughly 10 000 GPU-hours.

Numbers from the lab

A design counted as a binder if it produced a hit in both labs in every replicate, had a KD no worse than 1 µM, and showed no signal against the control protein CLEC12A, which tests for nonspecific "stickiness". Binding was measured by surface plasmon resonance. Adaptyv Bio used cell-free expression, while Twist Bioscience produced the proteins in human kidney cells as Fc fusions.

• 17 of 211 designs passed, each against a single toxin.
• By target: 7 against Cn2, 4 each against AaH I and AaH II, and 2 against LqhIII. None of the passing designs targets Ts1.
• Among 60 single-toxin designs (12 per toxin), 11 became binders, or 18%. That is a typical rate for protein design campaigns.
• The best two, Cn2-B1 and Cn2-B2, showed a KD in the single-digit nanomolar range in both labs.
• Among 91 designs against all five toxins and 60 against toxin families, none met the criteria for more than one toxin.

Competition experiments showed that the Cn2 binders attach at or near the same epitope where the neutralizing antibody LR, which protected mice, acts. For comparison, three known antibodies and nanobodies reach a KD below 1 nM against Cn2 or AaH II. The bar for de novo design is still higher than this result.

What the study leaves out

The authors name the main limitations themselves. Binding was tested only against recombinant toxins, not toxins purified from venom. Neutralization in patch-clamp cell assays and efficacy in animals were not tested. Multi-toxin design failed, and the cause could be the prompt wording, the tools' capabilities, or the biology itself.

Outside experts Susana Vázquez Torres of CNIO and Timothy Jenkins of DTU called the work a promising first step. They see the bottleneck as no longer binding, but turning a binder into a safe and accessible drug. For developers and research groups, the practical takeaway is simple. All 211 designs and the data are open, so they can be used as a benchmark and tested for neutralization. For Cn2-B1, the next steps are clear: confirm binding to the toxin from venom, run patch-clamp experiments, and then test in mice.