Anthropic's Claude AI model spent 24 hours largely unsupervised designing molecules meant to neutralize some of the deadliest scorpion venom on record, according to a report from The Decoder on Anthropic's Claude Science research effort.
Scorpion stings kill more than 3,250 people every year, and the toxins responsible are notoriously hard to treat. Claude's task was to design protein binders, small molecules engineered to latch onto a specific target, in this case five neurotoxins drawn from four scorpion species. Given an initial prompt, the model generated 318,484 candidate designs before a research team stepped in to steer the rest of the work.
From that enormous pool, Claude selected 180 candidates it judged most promising. Researchers added 31 more of their own choosing, bringing the total tested in the lab to 211. Of those, only 17 actually met the criteria to count as a working binder, a reminder that most AI-generated candidates still fail once they hit a test tube.
Two candidates stand out
The two strongest performers bound extremely tightly to Cn2, a toxin produced by the Mexican scorpion. Lab tests indicate these binders dock at or near the same site on the toxin as an antibody already known to neutralize Cn2 in mice, a result that hints at real pharmacological relevance rather than a coincidental match.
That said, nobody has tested whether these molecules actually work as an antivenom. Binding to the right spot on a toxin is a necessary step, not a finished drug. The next stage would involve testing in animal models to see whether the binders actually block the toxin's effects in a living system, work that hasn't started yet.
Two independent labs confirmed the lab results, which is standard practice before a finding like this gets taken seriously. Outside experts Susana Vázquez Torres and Timothy P. Jenkins reviewed the work and described it as a promising first step rather than a finished result, according to The Decoder.
The project ran inside Claude Science, the AI environment Anthropic has built for researchers to run experiments with its models. It isn't the first time Claude has been pointed at protein design. In an earlier study cited by The Decoder, the model designed mini-binders against 14 of 15 target proteins it was given, with hit rates between 22 and 35 percent, well above the 10 to 15 percent rate typically seen in this kind of protein engineering.
The scorpion toxin work builds on that track record, but the gap between a binder that works in a dish and a drug that works in a patient is still wide. Anthropic hasn't given a timeline for moving these candidates into animal testing, and there's no antivenom product to speak of yet. What the result does show is that an AI model can narrow a search space of hundreds of thousands of molecular designs down to a short list worth testing by hand, which is where the actual biology still has to take over.
ADFiled by The AI Desk
Models, assistants and the companies and chips behind them, reported from what was released and what was claimed, with the difference kept clear.
More from this desk →
Be the first to comment
Join the argument. No password, just your email or a passkey.