Claude ran a 64-hour autonomous campaign to design scorpion antivenom candidates

Anthropic's Science team pointed an internal model at a real problem: scorpion stings kill more than 3,250 people a year, mostly children with little access to care, and antivenom is scarce. The task was to design de novo minibinders, small proteins that latch tightly onto five scorpion neurotoxins from Mexico, Brazil, North Africa, and the Middle East. In the writeup, Claude ran for about 64 hours, mostly on its own, and worked unassisted for the first 24. Starting from a prompt adapted from an earlier binder study plus four follow-ups, it chose the target structures, used 12 design tools to generate 318,484 candidates, co-folded 553,052 design-toxin pairs, and spent roughly 10,000 GPU hours before narrowing to 211 designs for the lab.

The results are honest about what worked and what did not. Two labs, Adaptyv Bio and Twist Bioscience, measured binding by surface plasmon resonance, counting a design only if it bound in every replicate, hit a dissociation constant of 1 µM or tighter, and showed no signal against an unrelated human protein. Seventeen of the 211 designs cleared that bar, all against a single toxin, for an 18% hit rate among single-toxin designs. The two strongest, Cn2-B1 and Cn2-B2, bound with single-digit nanomolar affinity at both labs. None of the designs meant to catch two or more toxins at once succeeded. The team is careful to note the limits: testing used recombinant toxins, not venom, and no design was checked for neutralization in cells or animals, so this is not an antivenom. All 211 designs, with structures and binding data, are on Hugging Face.

Why it matters

This is a concrete look at what a general model produces when it runs a multi-day wet-lab design project end to end, including the negative results that marketing usually hides. If you work in protein design or evaluate AI for science, the open dataset gives you a real baseline to test against rather than a press-release claim.