Anthropic's Claude designed protein binders against scorpion neurotoxins that kill more than 3,250 people each year, working largely on its own for 24 hours after an initial prompt. The system generated 318,484 designs targeting five toxins from four scorpion species, and 17 of 211 tested designs met binder criteria in the lab. The result matters because it tests whether an AI research environment can carry a full design cycle from computation to wet-lab candidates.
How the 24-hour design run worked
After the starting instruction, Claude operated inside Claude Science, Anthropic's AI environment for researchers, with the team adding further steering instructions later. From the full pool of designs, Claude selected 180 candidates for laboratory testing, while researchers added 31 more, bringing the tested set to 211. The hit count was 17 successful binders, a conversion rate near 8 percent on this specific toxin task. Two independent laboratories confirmed the findings, which adds weight beyond a single-team result.
The binders are small protein molecules shaped to latch onto a defined target molecule, in this case scorpion neurotoxins. The two strongest candidates bound extremely tightly to Cn2, a toxin from the Mexican scorpion. Laboratory tests suggest they dock at or near the same site as an antibody known to neutralize Cn2 in mice. That positional overlap is the technical reason the candidates look interesting, even though actual antivenom function has not been tested yet.
The experiment follows an earlier Claude Science study in which Claude designed mini-binders against 14 of 15 target proteins. Hit rates there reached 22 to 35 percent, compared with typical rates of 10 to 15 percent cited in the report. External experts Susana Vazquez Torres and Timothy P. Jenkins described the scorpion work as a promising first step. The pattern is a shift from AI as an analysis assistant toward AI running extended design work with later human steering.
What this means for AI-driven R&D
For companies using AI in research, the practical change is throughput at the design stage: hundreds of thousands of binder concepts can be produced in one day before expensive laboratory work begins. A small biotech team can therefore enter testing with a ranked shortlist of around 200 candidates instead of building that list manually. A large pharmaceutical organization gains a different benefit, parallel exploration of several toxins and species within one workflow.
The limits are equally concrete and should shape procurement decisions. Only 17 of 211 tested designs qualified, and neutralization as an antivenom remains unproven, so binding alone does not equal a product. Buyers should ask which binding thresholds were used, how docking was measured, what independent confirmation covered, and what animal or safety testing comes next. The news does not by itself signal lower antivenom cost, faster approval, or transferability to other toxin families.
The marker to watch is whether the two strongest Cn2 binders advance to neutralization tests in animals and publish results. A second marker is repetition of the earlier 22 to 35 percent hit-rate range on new targets outside this toxin set. If both appear, AI-led binder design moves from demonstration to a repeatable front end for discovery pipelines.
