Claude generated hundreds of thousands of designs for proteins intended to attach to scorpion neurotoxins. Lab testing found a smaller set that met the researchers’ criteria for binders, including two that bound strongly to a toxin from the Mexican scorpion. The results are an early step in exploring whether AI-designed proteins could contribute to future antivenom research; they do not show that an antivenom works.
From toxin targets to candidate proteins
Protein binders are small protein molecules shaped to latch onto a particular target. In this study, the targets were five toxins from four scorpion species. The article says scorpion neurotoxins kill more than 3,250 people every year, giving the research a high-stakes medical context.
After receiving an initial prompt, Claude worked autonomously for 24 hours. The research team then guided the work with more instructions. Across the effort, Claude generated 318,484 possible binder designs.
That large design set was narrowed before the lab stage. Claude selected 180 candidates, and the research team added 31, bringing the number tested to 211. Of those, 17 met the criteria for a binder.
What the strongest candidates did
The two strongest binders attached very tightly to Cn2, a toxin from the Mexican scorpion. Lab tests also suggest that they dock at or near the same site as an antibody known to neutralize Cn2 in mice.
This resemblance is a useful clue about how the binders may interact with the toxin. It is not evidence that the designed proteins themselves neutralize the toxin in an animal or can serve as an antivenom. The source reports that this has not been tested.
That distinction matters when interpreting the findings. A protein that binds to a toxin has shown a connection at the molecular level; whether that connection can prevent harm is a separate question. The reported experiments establish candidate binders, while their potential role in treatment remains open.
A promising result with clear limits
External experts Susana Vázquez Torres and Timothy P. Jenkins described the work as a promising first step. Two independent labs confirmed the results, adding support for the reported binder findings.
The research took place in Claude Science, Anthropic’s AI environment for researchers. The source also points to an earlier study in which Claude designed mini-binders against 14 of 15 target proteins. That study reported hit rates of 22 to 35 percent, compared with a typical 10 to 15 percent.
Those earlier results provide context for the scorpion project, but they do not answer whether these particular candidates could work as a treatment. The key next question raised by the findings is whether the strongest toxin binders can demonstrate antivenom activity in further testing.
What this could mean for research
The project shows a path from AI-generated protein designs to candidates that can be examined in a lab. The process generated a large pool, selected a fraction for testing, and identified some that met the study’s binder criteria. That is a concrete research outcome, even while the medical application remains unproven.
For now, the findings support cautious interest in AI-assisted protein design for toxin research. They do not establish an available treatment or demonstrate that the binders protect against scorpion stings. The work’s promise lies in producing candidates for the next stages of investigation.