An internal Claude research model spent 64 hours designing small proteins meant to grab onto scorpion toxin molecules, and 17 of the 211 designs sent to outside labs did bind their targets, Anthropic reported on October 20. That is a real result, but a narrower one than “AI designs an antivenom.” The interesting part is who did the work and how little help it needed.
Why scorpions
Scorpion stings are a neglected health problem. Anthropic’s report cites more than 1.2 million stings a year and more than 3,250 deaths, mostly among children in places with limited medical care. Today’s antivenoms are made from animal antibodies, which are expensive to produce and need cold storage.
The idea for the project did not come from a biology lab. Citrini Research, a financial research firm, asked Anthropic whether Claude could help motivated non-experts take on a real health problem, and helped write the starting prompt.
What the model did
The model worked inside Claude Science, Anthropic’s research environment, under supervision from the company’s life sciences team. It received one opening prompt and four follow-ups, and ran for about 24 hours on its own after the first.
The targets were five neurotoxins from four regions: Cn2 (Mexico), Ts1 (Brazil), AaH I and AaH II (North Africa) and LqhIII (Middle East). The goal was “minibinders,” small designed proteins that attach tightly to one spot on a target, ideally the spot the toxin uses to attack nerve cells.
Using 12 protein design tools, the model generated 318,484 candidates and computer-checked 553,052 design-toxin pairs, which Anthropic says took about 10,000 GPU hours. It picked 180 designs for testing, and the team added 31 more with its help. Two companies, Adaptyv Bio and Twist Bioscience, made and tested them.
The results, toxin by toxin
Seventeen designs met Anthropic’s bar, which required binding at 1 micromolar or tighter and a hit at both labs where tested:
- Cn2: 7 binders, including the two strongest, Cn2-B1 and Cn2-B2, in the single-digit nanomolar range
- AaH I: 4
- AaH II: 4
- LqhIII: 2
- Ts1: none
Among designs aimed at a single toxin, 11 of 60 worked, an 18% hit rate that Anthropic calls typical for protein design projects. The more ambitious part failed: 151 designs were meant to bind several toxins at once, and none met the criteria for more than one.
For comparison, existing antibodies against Cn2 and AaH II bind more tightly, below 1 nanomolar. One encouraging sign is that the Cn2 binders appear to attach at or near the same spot as an antibody that has neutralized Cn2 in mice.
What it does not show
Anthropic is direct about the limits: “We have not confirmed whether these binders neutralize scorpion venom in a cellular environment, so we do not claim to have produced an antivenom.” No design was tested for neutralization in cells or in animals. Binding was measured against lab-made versions of the toxins, not toxins taken from real venom.
The model itself is an internal research version, so outside labs cannot rerun the campaign. They can check the output, though: Anthropic released all 211 designs and the test data on Hugging Face.
How it compares with the snake venom work
The project follows a path human scientists opened last year. In January 2025, David Baker’s group at the University of Washington published in Nature proteins designed with AI tools that bind cobra and mamba neurotoxins. Those designs neutralized the toxins in lab tests and protected mice, with 80 to 100% survival. Anthropic says that study inspired its approach, and the lead author of that paper, Susana Vázquez Torres, is one of two outside experts quoted calling the scorpion work a promising first step.
Set side by side, the two projects show where the AI added something and where it did not. The snake work was led by specialists and went all the way to animal tests. The Claude campaign stopped at binding, but a model chose the binding sites, ran the design tools and filtered hundreds of thousands of candidates with four human prompts. Anthropic says no designed binders against scorpion neurotoxins had been described before.
We have covered other attempts to use AI in biology, including an AI finding that could lead to new treatments for genetic diseases and Anthropic’s own wet lab in the Bay Area.
Our take
The hit rate is ordinary; the setup is not. A finance firm with a question and a model with design tools produced lab-confirmed binders against toxins nobody had targeted this way before, in under three days. That lowers the cost of a first attempt on neglected diseases that drug companies rarely pursue. But binding is the easy step. The useful measure will be whether Cn2-B1 or Cn2-B2 protects an animal from a real sting, the step the snake venom team reached and this project has not. Until then, this is a fast, cheap shortlist, not a treatment.