Claude designed proteins that grab scorpion toxins, and 17 passed lab tests

Scorpions are not a topic you usually see in AI news. But they are a real public health problem. According to Anthropic, scorpions sting at least 1.2 million people every year, and more than 3,250 of them die, most of them children in places with limited access to medical care. Treatments are expensive to make and rarely earn back their development costs, which is why scorpion envenomation is described as a neglected tropical disease.

On Tuesday, 20 October 2026, Anthropic published a new research report in which it asked Claude to attack exactly this problem. An internal research model designed small proteins, called minibinders, that are meant to latch onto scorpion neurotoxins. Two outside labs then tested the best designs. The result is not a cure, and Anthropic is very clear about that. But it is one of the most concrete examples so far of an AI model running a real piece of drug discovery from the computer to the lab bench.

Here is what happened, what the numbers mean, and where the limits are.

The three things you need to know

• Claude did most of the work on its own: Over 64 hours, the model generated 318,484 protein designs with 12 design tools, screened 553,052 design and toxin pairs, and picked 180 designs for the lab. Anthropic's researchers guided it with one prompt and four follow-ups.
• 17 designs passed strict lab tests: Two contract labs, Adaptyv Bio and Twist Bioscience, tested 211 designs. 17 met Anthropic's criteria for a binder. The two strongest bind a Mexican scorpion toxin very tightly and likely attach to the same spot as an antibody that is known to neutralize it.
• It is not an antivenom yet: No design has been tested against cells or in animals. Anthropic says it does not claim to have produced an antivenom, and it is releasing all 211 designs and their data so other researchers can build on them.

How the project started

The project grew out of an earlier study. In August, Anthropic showed that Claude could design protein minibinders from a single prompt. A month before those results were published, the team at Citrini Research, a financial investment research firm, approached Anthropic with a question: could Claude, and AI drug discovery in general, help a group of motivated non experts tackle a real health problem?

Anthropic's life science team brainstormed possible problems together with Claude and settled on scorpion neurotoxins. The longer term hope is that such binders could one day become part of a new kind of scorpion antivenom. The prompt from the earlier binder study was adapted together with Citrini and then handed to an internal research model running inside Claude Science.

Why scorpion venom is so hard to treat

Scorpion venom contains several protein neurotoxins. The most dangerous ones attach to sodium channels, large proteins in the membranes of nerve cells that let those cells fire. When the toxin disturbs these channels, nerves fire at the wrong time and the body releases large amounts of stress chemicals such as adrenaline. In severe cases this can lead to heart failure and fluid in the lungs, and sometimes death.

Antivenoms for scorpion stings exist, but they are made the old fashioned way. Venom is milked from thousands of scorpions and used to immunize horses, and scientists then purify antibodies from the horses' blood. That takes months. The resulting products work best when given early, usually cover only a small number of toxins, and are expensive, so they are often scarce in the regions where stings happen most.

De novo protein design offers a different route. Instead of harvesting antibodies from animals, researchers design a new protein from scratch on the computer, aimed at the exact part of the toxin that grabs the sodium channel. If the binder sits on that spot, the toxin should not be able to reach its target. Such designed proteins can also be tuned for stability and potency, which could make them easier to supply in hot, remote regions. Anthropic says the approach was inspired by an earlier project in which de novo proteins neutralized snake neurotoxins and protected mice, and that, to its knowledge, nobody had described de novo binders against scorpion neurotoxins before.

Five toxins from four regions

The team picked five neurotoxins from scorpions in four parts of the world: Cn2 from Mexico, Ts1 from Brazil, AaH I and AaH II from North Africa, and LqhIII from the Middle East. All five share the same overall fold, but any two of them are only 26 to 62 percent identical in their amino acid sequences. That makes them a tough test.

The main question was whether Claude could design a binder for each toxin. But the team also wanted to know whether it could design binders that catch several toxins at once. That matters in practice: a universal antivenom cocktail would be cheaper and would work regardless of region if fewer molecules were needed. So Claude was asked for four kinds of designs: binders for each single toxin, binders for the two so called beta family toxins (Ts1 and Cn2), binders for the three alpha family toxins (AaH I, AaH II and LqhIII), and binders for all five at the same time.

What Claude actually did

This is the part that makes the study interesting from an AI point of view. The initial prompt defined the targets, the available protein design and folding tools, and rules for novelty, scoring and filtering. Within that frame, Claude made a lot of decisions itself. It searched the Protein Data Bank and the AlphaFold database to choose which toxin structures to use, and it picked five hotspots on each toxin, specific amino acid positions its designs should latch onto. It set the quality thresholds a design had to pass, decided how heavily to rely on each of the 12 design tools, and came up with the strategies for multi toxin binders.

In total, Claude generated 318,484 designs and threw out anything that looked like a toxin or a known protein. It then co-folded and screened 553,052 design and toxin pairs, predicting the structure of each pair as one complex. After 64 hours it delivered a final list of 180 designs. The campaign used on the order of 10,000 GPU hours. Anthropic admits that this is far more compute than most researchers have, but says it has since made these models about four times faster on average, and that a simplified single agent setup on a single GPU over 24 hours reached comparable scores in silico.

Claude worked autonomously for the first 24 hours. The early results encouraged the team to ask it to keep going and look for better designs, and the follow-up prompts mostly pushed it toward multi toxin binders. Most of the 180 final designs, including the strongest binder, called Cn2-B1, were selected only after the third follow-up. At the team's request, Claude even sent the protein sequences to Twist Bioscience itself, in emails that said they were written by Claude. Later, the researchers worked with Claude to add 31 more designs, bringing the total to 211: 60 designs against a single toxin (12 per toxin), 30 against the two beta family toxins, 30 against the three alpha family toxins and 91 against all five.

How the designs were tested

Both labs produced the proteins, but in different ways. Adaptyv Bio used a cell free system, which uses protein making machinery extracted from bacteria. Twist Bioscience made the designs in human kidney cells, fused to the stem part of an antibody. Anthropic combined both methods to raise confidence in the results.

Both labs measured binding with surface plasmon resonance. In simple terms, the toxin flows over the designed protein while an instrument records in real time how fast it sticks and how fast it lets go. That produces a number called the dissociation constant, or KD. The lower the KD, the tighter the binding.

The bar for calling a design a binder was strict. It had to be a hit at both labs, show binding in every single replicate, and reach a KD of 1 micromolar or tighter every time. It also had to show no signal against CLEC12A, an unrelated human protein used to catch designs that are simply sticky. One exception: LqhIII was only tested at Adaptyv Bio.

The results

Of the 211 designs, 17 met these criteria, each against a single toxin. Seven bind Cn2, four bind AaH II, four bind AaH I and two bind LqhIII. Among the 60 designs aimed at one toxin, 11 were binders, a hit rate of 18 percent. Anthropic says this is in line with typical protein design campaigns, including its own earlier work.

The multi toxin goal did not work. None of the designs made against several toxins met the criteria for more than one of them. Anthropic lists several possible reasons, from Claude's current capabilities and the prompting to the limits of today's design tools, and even the possibility that a binder for several of these toxins is not biologically feasible.

The highlight is the pair Cn2-B1 and Cn2-B2. Both target Cn2, a neurotoxin of the Mexican scorpion Centruroides noxius, and both bound it at both labs with affinities in the single digit nanomolar range. For comparison, Anthropic notes that three antibodies and nanobodies with affinities below 1 nanomolar against Cn2 or AaH II are known and have protected mice, so the designed proteins are not yet better than the best existing molecules.

What makes the Cn2 binders promising is a competition experiment. It tests whether two molecules can attach to the same target at the same time. If one blocks the other, they probably sit on the same spot. The experiments suggest that the Cn2 binders attach on or near the same site as an antibody called LR, which has been reported to neutralize Cn2 in mice. That site covers the part of the toxin that grabs the sodium channel. Anthropic says this suggests the binders will likely neutralize Cn2, but stresses that this has not been tested.

What is still missing

Anthropic is unusually direct about the limits. All binding results come from recombinant toxins, grown in the lab from each toxin's gene rather than purified from real venom. No design was tested for neutralization in cells or for protection in living animals. Those are the two biggest gaps.

If the team were to push one design toward a therapy, Cn2-B1 would be a top candidate. The next steps would be to confirm it binds toxin purified from venom, to check with patch clamp recordings whether it stops the toxin from acting on sodium channels, and then to test whether it protects mice. Beyond that, a real treatment must be safe, stable, affordable and available where stings happen, which needs clinicians, funders and policymakers, not just a model.

Two outside experts, Susana Vázquez Torres of the Spanish National Cancer Research Centre in Madrid and Timothy P. Jenkins of the Technical University of Denmark, reviewed the work and called it a promising first step. Vázquez Torres made a point that sums up the state of the field well: in her view, binding itself is no longer the main bottleneck. The hard part is turning binders into real drugs.

Why it matters

There are two stories here. The first is about medicine. Neglected diseases stay neglected because they do not pay. If an AI model can produce credible binders against several hard toxins within days, and labs can confirm them within weeks, the early discovery phase gets much cheaper. That will not solve the expensive later stages, but it lowers the barrier to starting at all, and the open data lets academic groups pick up where Anthropic stopped.

The second story is about AI agents. Claude did not just answer a question. It chose structures, set thresholds, balanced 12 tools, ran a multi day campaign and even emailed a lab. Only a handful of prompts came from humans. It fits a pattern we have seen this month, from Claude designing quantum circuits that beat published designs to other labs pushing models into real research. At the same time, the honest negative result on multi toxin binders and the clear list of missing tests are a good reminder that "AI designed a drug" headlines usually skip the hardest parts.

Dany's take

I really like this one, and not only because it is a nice break from all the agent security news this week. A model working mostly on its own for almost three days, generating hundreds of thousands of designs and ending up with 17 real lab hits, is impressive. What I like even more is how Anthropic reports it: they say clearly that this is not an antivenom, they publish the failures together with the successes, and they release all 211 designs. That is how AI science announcements should look. My caution is the compute. Around 10,000 GPU hours is not something a small lab in a country with many scorpion stings can just run, even if Anthropic says a much smaller setup gets comparable in silico scores. If this approach is supposed to help neglected diseases, the tools have to become cheap and accessible, not just powerful.

What this means for you

If you work in biotech or academic research, the released dataset with all 211 designs and their binding data is the most useful part. It shows what a fully agentic design campaign produces, including the misses, and it is a good benchmark for your own pipelines. If you are a developer building agents, this is a strong example of a long running workflow with clear human checkpoints: the humans set the goal and the rules, the model does the search, and independent labs do the verification. And if you are simply following AI, keep an eye on the next step. Neutralization tests in cells and protection tests in mice will tell us whether Cn2-B1 is a real lead or just a tight binder. Could AI help fix neglected diseases? Tell me in the comments under the video.

Sources

• How Claude designed protein binders against scorpion neurotoxins (Anthropic, 20 October 2026)