Claude Designed Proteins That Bind Scorpion Toxins: What the Lab Tests Show

On October 20, 2026, Anthropic said that an internal Claude research model had designed small proteins that attached to scorpion neurotoxins in laboratory tests. Out of 211 designs sent for testing, 17 met the team’s criteria for binding. Anthropic released all 211 designs, their predicted structures, and the experimental results.

It is an eye-catching result, and it is easy to overstate. “Binds a toxin” is not the same as “treats a sting.” This article explains what the team did, what the tests showed, and what is still missing before anyone could call it an antivenom.

Disclosure: this article was drafted with Claude, which is made by Anthropic. The facts below come from third-party coverage of Anthropic’s technical report, and the report itself has not been through journal peer review. Check the primary source before relying on any detail.

What Happened

According to Interesting Engineering, Anthropic assigned an internal Claude research model to design “minibinders,” which are small engineered proteins built to attach to a specific target. The lab testing was done by two outside companies, Adaptyv Bio and Twist Bioscience, with input from Citrini Research.

Coverage describes the work as largely autonomous. After one initial prompt and four follow-ups, Claude selected 180 designs for lab testing, and researchers then added 31 more with the model’s help, for 211 in total. Along the way, the model used 12 protein design tools to generate 318,484 candidates and screened more than 553,000 protein-toxin pairs on a computer. The campaign took about 64 hours and roughly 10,000 GPU hours.

The targets were five neurotoxins from scorpions found in Mexico, Brazil, North Africa, and the Middle East.

What the Lab Tests Measured

Both labs measured binding with a technique called surface plasmon resonance, which tracks how strongly two molecules attach and how quickly they come apart. For a design to count as a success, it had to:

1. Bind consistently at both labs.
2. Meet an affinity threshold.
3. Show no binding to an unrelated human protein used as a control.

Seventeen designs passed. By toxin, that was seven against Cn2, four against AaH I, four against AaH II, and two against LqhIII. The two best candidates against Cn2 bound with affinity in the single-digit nanomolar range, which is considered strong for an early result. Competition experiments suggested they attach near the same site as an antibody fragment that neutralized Cn2 in mice, though that is a clue, not proof.

The Numbers, Explained

The different hit rates are worth understanding. Kingy, an analysis blog, points out that the 8% overall figure reflects early binding, not drug-development success. The higher 18% figure applies only to designs aimed at one toxin. Designs aimed at two or more toxins at once did not succeed: none of the 151 multitoxin designs bound more than one target.

What the Tests Did Not Show

This is the most important section. According to the coverage, none of the following has been shown yet:

• Neutralization. Binding a toxin is not the same as stopping it from working. The team has not shown that any design blocks toxin activity in cells.
• Protection in animals. No design has been tested for protecting an animal from venom.
• Real venom. The experiments used laboratory-produced toxins, not toxins purified from scorpion venom.
• Safety and stability. Immune response, how long a design lasts in the body, stability, and manufacturing cost were not measured.
• Solid affinity numbers. Kingy notes that some affinity values changed when the experimental setup was reversed, and two binders showed a signal against an unrelated protein in an additional test.
• Multitoxin goal. One toxin, Ts1, had no binder at all.

Anthropic itself, as reported, calls this “an early step toward potential antivenom development.” The strongest honest summary is that AI-designed proteins can attach to some of these toxins in a test tube.

Why This Still Matters

Scorpion stings are a neglected problem. Treatment today depends mostly on antivenom made from antibodies of immunized animals. These products work and are the established therapy, though production and regional supply can be difficult. Because the market for these treatments is small, commercial incentives for new research are weak.

In that context, a faster way to find candidate molecules is useful. Kingy highlights three reasons the result is notable:

• It shows an AI agent coordinating many specialized design tools and running a large computational search.
• It produced molecules that contract labs could actually make and measure.
• The released data, including failures, lets other researchers judge the approach for themselves.

Designed proteins might one day offer more consistent production, better stability, or lower cost than animal-derived antivenom. The study does not establish any of those advantages.

Caveats About Claude’s Role

A fair reading also includes the limits of the claim:

• No baseline. There is no matched comparison with an expert team or a conventional pipeline, so Claude’s specific contribution and any time or cost savings cannot be separated from the tools and computing power used.
• Cost. The campaign used about 10,000 high-end GPU hours, and the full project cost was not reported.
• Funding. Anthropic and Citrini funded the contract labs.
• Internal model. The work used an internal research model, not a product available to ordinary users.
• Review. The findings appear in a technical report. Outside expert review described by Anthropic is not the same as formal journal peer review.
• Ranking. The authors say the computational ranking did not reliably pick out the binders that worked.

Safety and Responsible Release

Work on toxins naturally raises a question: is it safe to publish? The coverage I could read does not describe the biosecurity safeguards Anthropic used or what, if anything, it withheld. That is a gap, and the technical report is the place to check. Anthropic released designs of proteins that bind the toxins, not information about how to make the toxins themselves. Readers who want to judge the release should read the primary report rather than rely on summaries, including this one.

A Plain-Language Guide to the Terms

Minibinder. A small engineered protein designed to stick to one specific target, like a key cut for a single lock.

Neurotoxin. A toxin that interferes with nerve signaling. Scorpion neurotoxins act on the proteins nerves use to send signals, which is why stings can be dangerous.

Affinity. How tightly two molecules hold on to each other. “Single-digit nanomolar” means a very tight grip by laboratory standards.

Surface plasmon resonance. A measurement method that detects molecules attaching to a surface in real time, including how fast they bind and how fast they let go.

Neutralization. Stopping a toxin from doing its job. This is the property an antivenom needs, and it is different from simply binding.

Recombinant toxin. A toxin produced in a lab from its genetic recipe, instead of being extracted from venom. It is safer to handle but may not behave exactly like the natural version.

Binding Versus Treating: An Analogy

Imagine a thief who uses a particular crowbar to open doors. A tool that grabs the crowbar proves it can hold it. It does not prove the thief stops opening doors, because the tool might grab the wrong end, be too weak, or fall off. To prove it works, you must watch the doors stay shut.

That is the gap between a binding result and a treatment. The 17 designs showed they can hold on to a toxin. What remains is to show that holding on actually stops the harm, in cells and then in living animals, without causing new problems.

Why Hit Rates Matter in Drug Discovery

An 8% hit rate sounds low, but early discovery is a numbers game. Researchers typically test many candidates to find the few worth pursuing, and each stage that follows removes more. A design that binds in a test tube may fail because it does not neutralize the toxin, is unstable, triggers an immune reaction, or is too costly to make.

The useful question is not whether 17 successes is a lot. It is whether the method finds good starting points faster or more cheaply than traditional approaches. The coverage says that comparison has not yet been made, which is why independent replication and a conventional baseline would be valuable.

What This Says About AI in Science

The headline detail is not just the protein result. It is that an AI model coordinated a multi-step scientific workflow with limited human guidance: choosing tools, generating candidates, screening them, and selecting what to send to a lab. That points to a broader shift, in which AI moves from answering questions to running parts of the research process.

That shift brings benefits and risks. The benefit is speed, particularly in neglected areas where funding is scarce. The risk is that errors can compound across automated steps, and that results need careful human review. The researchers’ own note that the computer ranking did not reliably find the working designs is a good reminder: lab testing, not model confidence, decided which ones counted.

What to Watch Next

1. Neutralization tests. The next milestone is showing that a binder stops a toxin from working in cells.
2. Animal studies. Only these can show real protection.
3. Independent replication. Other labs testing the released designs would strengthen or weaken the result.
4. Peer review. Publication in a journal would add outside scrutiny.
5. Biosecurity discussion. Expect debate about how AI-designed biological molecules should be released.

For other recent AI news, see our coverage of why OpenAI cancelled GPT-6.1 Astra and our October 2026 updates roundup.

Frequently Asked Questions

Did Claude invent an antivenom? No. It designed proteins that bound scorpion toxins in lab tests. No design has been shown to neutralize venom or protect an animal.

How many designs worked? Seventeen of 211 tested designs met the binding criteria, about 8% overall.

Was this done by a public Claude model? Coverage says it was an internal Claude research model, not a product available to ordinary users.

Is the work peer reviewed? Not yet. It appears in a technical report.

Could this lead to a treatment? Possibly, but it would need successful tests for neutralization, safety, stability, and effectiveness, plus a way to produce it affordably.

Can I see the data? Coverage says Anthropic released all 211 designs, predicted structures, and experimental results. Runtime Wire reports they were posted on Hugging Face on October 20, 2026.

The Bottom Line

The result is a real early step: an AI model coordinated a large design campaign and produced proteins that attached to scorpion neurotoxins in two independent labs. But it is a binding result in a test tube, not a treatment. The honest headline is “promising start, long road ahead.”

Sources

• Interesting Engineering: Claude’s 64-hour experiment finds 17 binders for scorpion toxins
• Runtime Wire: Anthropic releases 211 scorpion-toxin designs
• Kingy: Why this early antivenom research matters
• AI Weekly: Claude autonomously designs 17 scorpion-toxin binders