Claude designs 318,000 proteins against scorpion toxins, with 17 confirmed binders in laboratory tests

Anthropic used Claude to design proteins targeting five scorpion neurotoxins. Seventeen candidates demonstrated binding ability.

Can an artificial intelligence assistant be entrusted with part of the research work required to design a medical treatment? Anthropic put this question to Claude by asking it to design proteins capable of binding to neurotoxins from scorpions that are dangerous to humans. In 64 hours, an internal research model generated 318,484 protein structures, then selected 180 candidates for trials. After adding 31 other proposals, the laboratories Adaptyv Bio and Twist Bioscience tested 211 proteins, of which 17 met the binding criteria defined by the researchers. Published on October 20, 2026, this experiment shows that an AI agent can lead a relatively autonomous molecular design campaign. However, it does not yet demonstrate the development of an antivenom. A neglected disease as a testing ground

Every year, scorpions sting more than 1.2 million people worldwide and cause, according to estimates cited by the researchers, more than 3,250 deaths. Children are among the most vulnerable populations, particularly in regions where access to care remains limited. Certain venoms disrupt the sodium channels present on the surface of nerve cells, destabilize their electrical activity, and trigger physiological reactions that can lead to severe cardiac or pulmonary complications.

Existing antivenom treatments often rely on antibodies obtained after immunizing animals, notably horses. Their production requires time, specialized infrastructure, and supplies of venom. Their effectiveness also depends on the species involved, as toxins differ from one region to another. In this context, designing small proteins capable of binding precisely to the active sites of toxins could open up alternative treatment pathways, provided their ability to neutralize venom is subsequently demonstrated.

The study grew out of a collaboration between Anthropic and Citrini Research, a firm specializing in financial research, which wished to explore the possibility for non-specialists to contribute to a concrete public health problem with the help of AI. After several exchanges with Claude, the team selected five neurotoxins from four scorpion species: Cn2 in Mexico, Ts1 in Brazil, AaH I and AaH II in North Africa, and LqhIII in the Middle East. The researchers chose these targets for their medical importance and their molecular differences, which made designing a protein effective against several toxins at once particularly difficult. A research agent that organizes its own campaign

The team did not simply ask Claude to produce a few protein sequences. It provided it with a scientific objective, the toxins to target, a list of molecular design tools, and selection rules. The agent then had to choose the structures to use, identify the regions of the toxins where its proteins could bind, organize the calculations, and determine which candidates merited experimental validation.

Starting from an initial prompt and four human follow-ups, Claude mobilized twelve design tools, including RFdiffusion, RFdiffusion3, BoltzGen, and Genie 3, as well as several structural prediction systems. In total, it generated 318,484 proposals and evaluated 553,052 associations between designed proteins and toxins, before progressively narrowing down the selection. The calculations represented approximately 10,000 GPU hours, a significant volume that nuances the idea of research achievable without substantial technical resources.

The agent worked for 24 hours after the initial prompt, then continued its campaign through the follow-ups. The researchers notably prompted it to search for proteins capable of interacting with multiple toxins. At the end of 64 hours, Claude had selected 180 candidates. The team then added 31 more with its assistance, to form a set of 211 proteins to be tested. Claude also contributed to analyzing the results, producing figures, and drafting the scientific report, under the supervision of the researchers.

This autonomy thus remains framed: humans set the objectives, selected the broad categories of toxins, and guided certain decisions. The design, filters, and a large portion of the computational work were then conducted by the system. Seventeen proteins capable of recognizing their target

The 211 candidates were evaluated by Adaptyv Bio and Twist Bioscience, two laboratories tasked with measuring their ability to bind to the toxins. The proteins were produced using different processes, then analyzed notably by surface plasmon resonance, a technique that measures the strength and stability of the interaction between two molecules.

To be considered a positive result, a candidate had to satisfy several requirements: produce reproducible binding, reach a minimum affinity, and show no unwanted interaction with a human control protein. Seventeen proteins met these criteria, including seven against Cn2, four against AaH II, four against AaH I, and two against LqhIII. No candidate met the criteria for Ts1. The two proteins targeting LqhIII were, however, only measured by Adaptyv Bio, unlike the other targets evaluated in both laboratories.

The most striking results concern Cn2-B1 and Cn2-B2, two proteins designed to bind to Cn2, a neurotoxin from the Mexican scorpion Centruroides noxius. Their affinity reaches the nanomolar scale, an indicator of tight binding to the target. For Cn2-B1, measurements provide upper limits of approximately 3.1 nanomolar at Adaptyv Bio and 1.7 nanomolar at Twist Bioscience.

The researchers also performed competition experiments with LR, an antibody fragment already known to neutralize Cn2 in mouse experiments. The results suggest that Claude's two proteins bind to a region identical or adjacent to the one recognized by this antibody. The observation is interesting, as this area plays a role in the interaction between the toxin and the sodium channels of nerve cells. It is not sufficient, however, to demonstrate that the new proteins actually prevent toxic action. The search for a versatile antivenom fails for now

On