Anthropic says an internal research version of Claude has designed small proteins, called minibinders, that latch on to scorpion neurotoxins. Working largely on its own for 64 hours, with four follow-up prompts from Anthropic staff, the model chose the toxin targets, ran 12 design tools, generated 318,484 candidate designs and picked 180 for lab testing. Anthropic added 31 more.
Two outside labs, Adaptyv Bio and Twist Bioscience, tested the 211 designs. Seventeen met Anthropic’s bar for a binder, each against one of five toxins from scorpions found in Mexico, Brazil, North Africa and the Middle East. The two strongest, aimed at the Mexican toxin Cn2, bound at single-digit nanomolar strength in both labs. No design managed to bind several toxins at once.
Why it matters: the report frames this as a possible route towards a scorpion antivenom, and it shows an AI system running much of a protein design campaign end to end. But it is an early step. The proteins were tested only against lab-made toxins, not real venom, and not in cells or animals. Anthropic is explicit: “we do not claim to have produced an antivenom.” Timothy P. Jenkins, one of the outside experts who reviewed the work, said “binding itself is no longer the main bottleneck”.
Read Anthropic’s report: Designing scorpion toxin binders with Claude (Anthropic)