Claude’s internal model helped scientists design a protein that could become a scorpion toxin antidote

Each year, more than a million people are stung by scorpions and thousands die, mostly children in poor regions. Existing antivenoms are made from horse blood, which is slow and expensive, and they do not work against all toxin variants.

Scorpion venom works by toxins attaching to nerve cells and causing nerves to fire incorrectly. Researchers have proposed creating small proteins that bind to the toxin at the right spot and prevent it from reaching cells.

Anthropic and Citrini Research decided to test whether Claude could help generate such decoy proteins. The agent was given a prompt and an environment and ran autonomously for almost three days. It proposed more than 300,000 candidate proteins, then filtered out weak ones and selected the 180 best. The process took about 10,000 GPU-hours.

Seventeen of the selected proteins passed strict laboratory validation, and the two best turned out to bind the toxin very tightly. However, scorpion toxins differ by region, and Claude has not yet produced a universal protein.

Even for individual toxins, these are not antidotes yet: testing continues and success is not guaranteed. Still, producing such a shortlist in three days is already a major result that previously took humans months.

All proteins were released as open .co/datasets/Anthropic/scorpion-toxin-binder-design).