Gives measured hit rates (17 of 211 tested designs) for a mostly autonomous AI agentAn AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.Full definition → campaign with only four follow-up prompts, plus an open dataset for reuse.
Key takeaways · AI-distilled
The authors report an internal research model generated 318,484 candidate designs with 12 design tools and screened 553,052 design-toxin pairs, producing its final 180-design list 64 hours after the first prompt and using on the order of 10,000 GPU hours.
Of 211 designs tested, 17 met the authors' binder criteria, each against a single toxin; among the 60 single-toxin designs 11 were binders, a rate the authors call consistent with typical protein design campaigns.
Every design aimed at several toxins at once failed to qualify against more than one, and the authors list prompting, design-tool limits, and the biological feasibility of multitoxin design as possible explanations.
The authors set strict binder criteria: a hit at both contract labs with binding in every replicate, every dissociation constant at 1 micromolar or tighter, and no signal against an unrelated human protein used as a stickiness control.
Competition experiments suggest the two strongest Cn2 binders attach on or near the same site as the LR antibody, but the authors state no design was tested for neutralization in cells or in animals, so they do not claim an antivenom.
Terms in this piece · Glossary
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
