Claude Found 17 Proteins That Bind Scorpion Toxins in 64 Hours

If you have ever built a complex agentic workflow, the pattern will be familiar: the model receives a goal, finds data, calls a dozen tools, checks intermediate results, and gradually narrows the search. Except here, the output was neither a report nor a pull request, but 211 new protein designs that were subsequently manufactured and tested by two laboratories.

Anthropic's research model spent 64 hours running a campaign to find proteins capable of binding to five scorpion neurotoxins. Claude generated 318,484 variants, computationally evaluated 553,052 protein-toxin pairs, and selected 180 candidates. Humans working with the model added another 31 designs. Of the final 211 proteins, 17 met the laboratory criteria.

This is not a finished antidote. But the experiment clearly shows where agentic systems are already useful in practice—and where the presentation still runs ahead of reality.

What Claude Actually Did

Think of the task as searching for a key to an exceptionally complicated lock. The toxin is the lock, while a small artificial protein is the key that must firmly attach to the right surface. If it binds the region the toxin uses to attach to a nerve cell's sodium channel, it could theoretically stop the toxin from doing its job.

Humans defined the objectives, available tools, novelty rules, and basic constraints. After the initial prompt, four more clarifications were required. Claude then searched public databases for toxin structures, selected binding sites, set quality thresholds, combined 12 design tools, and assembled a list for physical synthesis. The model worked autonomously for the first 24 hours, while the full computational campaign consumed about 10,000 accelerator-hours.

The proteins were independently tested by Adaptyv Bio and Twist Bioscience. Seven designs bound to the Cn2 toxin, four each to AaH II and AaH I, and another two to LqhIII. No protein met the established criteria for the fifth toxin, Ts1. The best candidate, Cn2-B1, demonstrated binding in the nanomolar range—a strong early-stage laboratory result.

Where the Agent Succeeded—and Where It Broke Down

The most useful figure is not the 318,000 generated variants, but the 11 successful proteins among 60 designs targeting one specific toxin. That is about 18%, which the authors consider a normal result for campaigns of this kind. Generation produced not a pretty collection of molecules, but candidates that survived laboratory testing.

Universal proteins performed worse. Claude designed separate groups of candidates against two, three, and all five toxins at once. Under the specified criteria, not a single design could bind more than one. The complete failure of all 91 proteins designed to target all five is particularly revealing.

The model may not be the only reason. The toxins have similar overall shapes, but their amino acid sequences overlap by only 26–62%. The selected tools may be poorly suited to this task, the prompts may have directed the search the wrong way, or a universal protein may simply be difficult to obtain in the proposed form. The experiment honestly leaves this question open.

There is also a more important limitation: the researchers tested binding only. They did not test toxin neutralization in cells or animals. The proteins' safety, stability, dosage, manufacturing cost, and behavior inside the body remain unknown. The result therefore cannot be called an antidote.

What Developers and Product Teams Should Take From This

The main lesson is that an agent benefits from receiving not one magical command, but a controllable research loop: a goal, a set of tools, measurable thresholds, a decision log, and an expensive external validation step at the end. Claude did not replace the laboratory. It dramatically expanded the space of possibilities, while the laboratory separated genuine hits from convincing computational predictions.

For a product team, this is a familiar architecture: the model generates and ranks hypotheses, but the final criterion sits outside the model—in tests, user data, revenue, or a physical experiment. The more costly the error, the more important independent validation becomes.

Anthropic released all 211 designs and the measurement results. That means the work can be assessed from the data rather than from a promotional description. For now, the sober conclusion is that an autonomous agent can already complete a meaningful portion of a complex scientific search in a matter of days. But the path from a discovered candidate to a medicine remains longer and harder than the generation process itself.