Claude designs proteins against scorpion venom in 64 hours: 17 bind to their target in the lab

Every year, an estimated one and a half million people worldwide are stung by a scorpion. More than 3,250 of them die, mostly children in areas where good medical care is far away. An antidote exists, but it is made in a laborious way and usually works against only a handful of toxins.

AI company Anthropic announced on 20 October 2026 that an internal research version of Claude had designed over 318,000 small proteins in 64 hours that are meant to clamp onto scorpion venom. Two laboratories made 211 of them and tested them. Seventeen gripped their toxin firmly, and the two strongest appear to do so at the spot where the venom becomes dangerous.

Why scorpion venom is so dangerous

Nerve cells pass on their signals through gates in their wall, called sodium channels. The most dangerous substances in scorpion venom, neurotoxins, clamp onto those gates. As a result, nerves fire at the wrong moment and the body floods with stress hormones such as adrenaline. In severe cases the heart fails or fluid builds up in the lungs.

An antidote does exist, but producing it takes a great deal of work. Venom is extracted from thousands of scorpions, horses are injected with it, and researchers then extract the antibodies from those horses' blood. That takes months. Such antidotes often work against only a few toxins, are expensive to make and are scarce in many countries. Because so little money is involved, researchers argue for officially recognising scorpion stings as a neglected tropical disease. It is not yet on the World Health Organization's list.

Proteins drawn from scratch

An alternative is a small protein that fits the venom exactly, like a key in a lock. If the protein grabs the venom at the spot with which it attacks the nerve gates, the venom can no longer do anything. Such small, custom-made proteins are called minibinders, and they are designed de novo.

Proteins have been designed from scratch against snake venom before, which protected mice. For scorpion venom, this had not yet been done.

The assignment came about at the request of Citrini Research, an investment research firm that asked Anthropic whether motivated non-specialists could tackle a real health problem with AI. That team chose the subject itself, and so five toxins from four regions ended up on the list: Mexico, Brazil, North Africa and the Middle East. Those five broadly resemble one another, but their building blocks overlap by only a quarter to just over sixty per cent. That made them a tough test.

318,484 designs in 64 hours

Claude was given an extensive assignment and then worked independently for 24 hours. The researchers steered the model a total of four times, mainly towards proteins that had to catch multiple toxins at once. The model itself looked up the blueprints of the toxins in public databases, chose five spots per toxin where its designs had to attach, and drew up the requirements a design had to meet itself.

Using twelve existing design programmes, Claude produced 318,484 designs and worked through more than 553,000 combinations of design and toxin. After 64 hours there was a list of 180 designs; the team later added 31 more with Claude's help. All that computing took roughly 10,000 hours on powerful graphics cards. Claude then also examined the measurement data, made the graphs and wrote the draft report.

Seventeen hits

The 211 designs went to Adaptyv Bio and Twist Bioscience, two companies that make and test proteins to order. Both produced the proteins in their own way and checked how firmly they attached to the venom. A design only counted as a hit if it bound firmly enough in every measurement and did not keep sticking to a human protein included as a control. As a rule, this had to succeed at both labs.

Seventeen designs cleared that bar, each against one toxin: seven against the Mexican venom Cn2, four each against the two North African toxins, and two against the one from the Middle East. That last venom was only tested at Adaptyv Bio, so those two hits lack confirmation from a second lab. Against the Brazilian venom, no hit emerged from the test. Of the sixty designs aimed at a single toxin, eleven were hits: nearly one in five, a typical score for this kind of design work. Proteins that have to catch multiple toxins at once did not succeed yet. That is one of the questions for follow-up research, because with such broad-acting proteins, one remedy could help on several continents.

The two strongest proteins both target the Mexican venom Cn2 and attached to it exceptionally firmly. In an additional test, they proved able to compete with an existing antibody that neutralises Cn2 in mice: all three want to sit on or very near the same spot. According to Anthropic, this makes it likely that these new proteins can also switch off the venom. It is not certain yet; for that, the trials on cells and animals that are now to follow are needed.

From binder to medicine

Anthropic explicitly does not call the result an antidote yet. The tests were done with toxins recreated in the lab, not with venom from a real scorpion, and the proteins have not yet been tested on cells or animals. The next steps are ready: first confirm that the proteins really do grab the venom, then measure whether the nerve gates are thereby left alone, and then trials in mice. A remedy must ultimately also be safe, stable and affordable, and reach the places where the stings happen.

At Anthropic's request, two experts reviewed the work: Susana Vázquez Torres of the Spanish cancer research institute CNIO and Timothy P. Jenkins of the Technical University of Denmark, both authors of the earlier research into snake venom. They call it a promising first step. Vázquez Torres states that the binding itself is no longer the major bottleneck; according to Jenkins, the gains are now to be made in the rest of the road towards a remedy that is safe, affordable and available. Precisely for diseases where there is little profit to be made, that can make the difference. All 211 designs and their measurement data have been made public, so that other researchers can build on them.

What does this mean for the road to the singularity?

The most striking thing about this project is not the number of designs, but how much of the selection work lay with the model. Five people from Citrini chose the subject and wrote an assignment of some thirty-five pages; after that, Claude looked up the targets, determined where the proteins had to attach, set the requirements, selected the candidates, ordered them by e-mail from the lab and wrote the report. The researchers steered four times. So the direction came from people, the work in between from the model. A week earlier, another search with AI agents had already yielded a first candidate for a new antibiotic, there not by designing but by combing through databases.

On top of that, a language model is not confined to a single field. The fact that an approach that worked against snake venom is now being applied to scorpions is exactly the kind of cross-pollination such models are good at; a physicist showed how that moves computational methods between fields. And although this campaign took a lot of computing power, Anthropic reports that a simpler set-up with a single agent on a single graphics card achieves comparable scores on the computer in 24 hours. If that also holds up in the lab, this kind of design work will come within reach of labs that do not have a computing centre at their disposal.

The challenge now lies in the steps that come after: neutralising, testing in animals, and making a remedy that is affordable and arrives where it is needed. The pace there is set by the lab: the design took two days, and the team then waited about two more months for the results of the bench tests. The fact that the design phase goes from months to days does make that road considerably shorter.

What is the technological singularity? Read the explanation →