AI designs 17 proteins that bind scorpion toxins, in lab tests

A chip, a protein, a toxin: the test is simple to describe and slow to run. In the Anthropic study, 318,484 protein designs were generated by an internal Claude research model in 64 hours, and only a small slice of them ever reached a bench. Of the 211 candidates sent to the lab, 17 passed the binding tests against dangerous scorpion neurotoxins.

The antivenom problem explains why the test matters. Scorpion stings affect at least 1.2 million people worldwide each year, causing more than 3,250 deaths, mainly among children in regions with limited access to medical care. Existing treatments are made by immunizing animals, including horses, with venom and extracting antibodies from their blood. Manufacturing can take months, and a treatment may work against only a limited range of toxins. The design task Anthropic set was to create minibinders, small engineered proteins that attach to specific targets. The project focused on five neurotoxins from scorpions found in Mexico, Brazil, North Africa and the Middle East.

The model worked largely on its own, using 12 protein design tools to produce 318,484 candidates and to screen 553,052 protein-toxin pairs, a campaign that consumed about 10,000 GPU hours. After an initial prompt and four follow-ups, Claude selected 180 designs for the lab, and researchers added 31 more with its assistance. Adaptyv Bio and Twist Bioscience produced the proteins using different laboratory methods and measured binding by surface plasmon resonance, a technique that tracks how strongly molecules attach and how fast they separate. To qualify, a design had to bind consistently at both labs, reach a set affinity threshold, and show no detectable binding to an unrelated human protein used as a control.

Seven designs bound the Mexican scorpion toxin Cn2, four each bound AaH I and AaH II, and two bound LqhIII. Among the 60 designs built for a single toxin, 11 succeeded, a hit rate of about 18 percent. The multitoxin designs did not qualify against more than one toxin. The two leading candidates, Cn2-B1 and Cn2-B2, bound particularly tightly, and competition experiments suggest they attach near the same site as an antibody previously shown to neutralize Cn2 in mice. The researchers stress the limits: they tested binding only against toxins produced in the laboratory, not venom taken from scorpions, and they have not established whether the candidates block toxin activity in cells or protect animals.

For the people bitten each year, nothing changes yet. What changes is the starting point. Venom neutralization, safety, stability and therapeutic effectiveness all still have to be tested before any clinical development could be considered, and no timeline has been announced. Anthropic is releasing all 211 designs, their predicted structures and the experimental results so that other groups can build on them. The findings appear in a technical report rather than a peer-reviewed journal, which leaves the claims open to independent checking.