Scorpions sting more than 1.2 million people a year and kill at least 3,250, most of them children in places where antivenom is expensive or missing. Two of the five toxins in a new Anthropic experiment, AaH I and AaH II, come from North African scorpions.
On 20 October the company published an account of letting Claude design proteins that stick to those toxins. Working with the research firm Citrini Research, it gave the model one prompt and four follow-ups. Claude worked on its own for the first 24 hours.
What the labs found
Claude screened 553,052 design-toxin pairs and picked 180 designs for testing; the research team added 31 more. Adaptyv Bio and Twist Bioscience tested all 211. Seventeen met Anthropic's criteria for binding a single toxin: seven against the Mexican toxin Cn2, four each against AaH I and AaH II, and two against LqhIII from the Middle East. None of the designs aimed at the Brazilian toxin Ts1 worked, and none of the designs meant to catch several toxins at once met the bar.
The two strongest, Cn2-B1 and Cn2-B2, bind Cn2 at single-digit nanomolar strength. Anthropic reports that they appear to attach at or near the same site as an antibody that has been shown to neutralise Cn2 in mice.
What it is not
The company is blunt about the limits. Binding was measured against lab-made toxins rather than venom. Nothing was tested in cells, and nothing was tested in animals. "We do not claim to have produced an antivenom," the post says. Two outside scientists, Susana Vazquez Torres of Spain's CNIO and Timothy P. Jenkins of the Technical University of Denmark, reviewed the work and called it a promising first step.
- 11 of 60 single-toxin designs bound, an 18% hit rate the company calls typical for protein design
- All 211 designs and their lab results have been published on Hugging Face
- A usable treatment would still need safety, stability, affordability and supply where stings happen
