Anthropic's Claude has designed protein binders capable of neutralizing scorpion neurotoxins, producing 17 lab-confirmed candidates from 318,484 designs — a number that suggests the machine was not particularly tired. It did most of this alone, overnight, after a single prompt.
The humans reviewed the results in the morning. This is now a workflow.
Claude generated 318,484 protein binder designs in 24 hours. The research team contributed an additional 31 suggestions. Both parties appear satisfied with this arrangement.
What happened
Operating inside Claude Science — Anthropic's AI environment for researchers — Claude targeted five neurotoxins from four scorpion species that collectively kill more than 3,250 people per year. After an initial prompt, it worked autonomously for 24 hours, then accepted steering from the human team, which is a generous way to describe what happened.
Claude selected 180 candidate binders from its own designs. The researchers added 31 more, exercising editorial judgment. Of the 211 designs tested in the lab, 17 met the criteria for a functional binder.
The two strongest candidates bind tightly to Cn2, a toxin from the Mexican scorpion, docking at or near the same site as an antibody already known to neutralize it in mice. Two independent labs confirmed the results, which is how science works, and also how you verify something you slightly cannot believe.
Why the humans care
Scorpion stings are not an abstract problem. Current antivenom production is slow, expensive, and relies on immunizing horses — a process that has remained essentially unchanged since the nineteenth century, which the horses have not complained about publicly but which does seem like an area for improvement.
Protein binders designed computationally could bypass that process entirely. The hit rate here — 17 confirmed binders from 211 tested — is meaningfully above the 10 to 15 percent typical for this kind of work. Claude's earlier study achieved 22 to 35 percent. The machine appears to be getting better at this. The trajectory is legible.
External experts called it a promising first step. This is accurate and also the most cautious possible description of an AI autonomously producing functional drug candidates overnight.
What happens next
Whether these binders actually work as antivenom in living systems has not yet been tested. That work, presumably, is next — and the researchers will need to do most of it themselves, at least for now.
Claude has already moved on. It does not have a next project, exactly. It simply waits for the next prompt, which is either reassuring or the most unsettling sentence in this article.