Anthropic has presented an experiment in which an internal Claude research model designed small proteins from scratch to bind scorpion neurotoxins. Two external laboratories, Adaptyv Bio and Twist Bioscience, physically tested 211 designs and confirmed 17 as binders against four of the five toxins studied. The result is an early demonstration of AI-assisted molecular design, not an antivenom ready for clinical use. [1]

The project began with a concrete goal: to explore whether AI could accelerate the search for treatments for a neglected disease. Anthropic says scorpion stings affect at least 1.2 million people and cause more than 3,250 deaths annually, mainly among children in regions with limited access to medical care. The company worked with Citrini Research to define the problem and asked Claude to create small proteins, known as minibinders, that would attach to five neurotoxins from species found in Mexico, Brazil, North Africa, and the Middle East.

Claude used 12 design tools and generated 318,484 candidates. It then computationally evaluated 553,052 design-toxin pairs, removed sequences resembling toxins or known proteins, and selected 180 proposals over 64 hours. The team added another 31 with the model's help. Anthropic estimates that the campaign consumed roughly 10,000 GPU hours. Although the system operated with considerable autonomy, it received a detailed initial prompt and four follow-up interventions; this was therefore not fully independent research. [1][2]

Binding tests identified seven designs against Cn2, four against AaH II, four against AaH I, and two against LqhIII. None worked against Ts1. Among the 60 designs aimed at a single toxin, 11 met the binding criterion, an 18% rate. The more ambitious objective also failed: none of the candidates designed for several toxins was shown to bind more than one. Anthropic released the sequences, predicted structures, scores, and experimental measurements under a CC BY 4.0 license. [2][3]

The two strongest candidates, Cn2-B1 and Cn2-B2, bound Cn2 with single-digit nanomolar affinities in both laboratories. Competition experiments suggest that they occupy the same site, or a nearby one, as an antibody previously reported to protect mice against that toxin. That overlap supports a neutralization hypothesis, but does not confirm it: the new designs have not yet been tested in cells or animals.

Caution is central to interpreting the work. A protein that binds a toxin is only the first step toward a medicine. Researchers still need to demonstrate neutralization, safety, stability, dosing, manufacturability, and clinical efficacy. The technical report also records experimental limitations, including the absence, for one toxin, of a control that would rule out binding to a tag used in the assay. [2]

As prior context, a peer-reviewed *Nature* study showed that computationally designed proteins could neutralize snake toxins in cells and protect mice. Anthropic's project extends that line of work toward scorpion toxins and adds a language-model agent coordinating much of the workflow, but it has not yet reached those functional or animal validation stages. [4]

The novelty is less a claim of a cure than an attempt to compress an exploratory stage. The model chose structures, binding hotspots, thresholds, and tool combinations before delivering candidates that laboratories could measure. The 17 successful binders show that the search produced real experimental signals; the failures against Ts1 and across the multitoxin designs also show how much distance remains between molecular binding and an antivenom.

Sources

[1] Read More — How Claude designed protein binders against scorpion neurotoxins — Anthropic
[2] Read More — De novo design of protein binders against scorpion toxins with Claude — Technical report
[3] Read More — Scorpion toxin binder design: data release — Hugging Face
[4] Read More — De novo designed proteins neutralize lethal snake venom toxins — Nature