# Anthropic’s Claude Designed 17 Working Protein Binders Against Scorpion Venom Toxins > Source: > Published: 2026-10-20 17:19:40+00:00 There has been some criticism over AI having not yet made progress on [curing diseases](https://officechai.com/stories/ai-will-lead-to-100-years-of-scientific-progress-in-5-10-years-anthropic-ceo-dario-amodei/), and Anthropic appears to have made some progress in that regard. Anthropic has shared results from a new experiment in which its AI model was tasked with a real problem in drug discovery: designing proteins that could one day form the basis of a scorpion antivenom. The project, carried out through Anthropic’s Claude Science research environment, saw an internal research model work largely on its own for 64 hours to produce hundreds of candidate protein designs, 17 of which were confirmed in lab tests to bind tightly to dangerous scorpion neurotoxins. The idea for the project came from Citrini Research, a financial investment research firm, which approached Anthropic with a question: could Claude, and AI-driven drug discovery more broadly, let a small, motivated team make real progress on a genuine health problem? Anthropic’s life sciences researchers landed on scorpion antivenom as the test case. ## Why Scorpion Stings Are a Neglected Problem Scorpion stings affect more than 1.2 million people a year and are linked to over 3,250 deaths annually, with children in regions with poor access to healthcare bearing the brunt of it. Existing antivenoms are made the old-fashioned way — venom is extracted from scorpions and used to immunize horses, after which antibodies are purified from the horses’ blood, a process that takes months and is costly to scale. Because the market for these treatments is small and unprofitable, scorpion envenomation is often classified as a neglected tropical disease, meaning it attracts little commercial R&D investment despite the human toll. This is exactly the kind of gap Anthropic has said it wants Claude’s life sciences push to address: problems that are medically significant but not commercially attractive enough for traditional pharma to pursue. ## How the Experiment Worked Anthropic gave Claude a prompt describing the goal, the protein design and folding tools available to it, and the criteria a design would need to meet to count as a success — largely reusing a setup from an earlier protein-binder study, adapted with input from Citrini. From there, the model worked mostly autonomously inside Claude Science, with Anthropic’s researchers stepping in four times with follow-up guidance, mostly to push Claude toward attempting binders that could work against more than one toxin at once. Within that window, Claude searched protein structure databases (the Protein Data Bank and the AlphaFold database) to pick which toxin structures to target, then identified specific binding “hotspots” on five different neurotoxins from scorpions found in Mexico, Brazil, North Africa, and the Middle East. It used a dozen different protein design tools to generate 318,484 candidate protein designs, screened over half a million design-toxin pairings by predicting their combined 3D structure, and narrowed the field down to a final set of 180 designs, later expanded to 211 with Anthropic’s help. Claude also set its own quality thresholds for what counted as a promising design and decided how heavily to lean on each of the available modeling tools — decisions that would typically be made by a human protein engineer with years of specialized training. ## The Lab Results Anthropic sent the 211 designs to two outside contract research organizations, Adaptyv Bio and Twist Bioscience, which independently produced the proteins and tested them for binding strength using surface plasmon resonance, a standard lab technique for measuring how tightly two molecules stick together. To count as a genuine “hit,” a design had to bind consistently across every test at both labs, with a binding strength at or better than 1 micromolar, and show no attraction to an unrelated protein used as a control for stickiness. Under those strict criteria, 17 of the 211 designs qualified as binders — all of them aimed at a single toxin rather than multiple toxins at once. The breakdown: seven against the Mexican scorpion toxin Cn2, four against AaH II and four against AaH I from North African scorpions, and two against LqhIII from a Middle Eastern species. None of the designs built to target several toxins simultaneously met the bar for more than one, something Anthropic says could come down to its own prompting, limitations in the current design tools, or simply that multitoxin binding is a harder biological problem. The standout results were two binders against Cn2, named Cn2-B1 and Cn2-B2, which bound their target with single-digit nanomolar affinity at both testing labs. Follow-up competition experiments suggest these two binders attach at or near the same site on the toxin as LR, an antibody already shown to neutralize Cn2 in mice — a strong hint, though not proof, that Claude’s designs could work similarly. ## Still Early-Stage Work Anthropic is careful to frame this as a first step rather than a finished antivenom. None of the 211 designs have been tested for whether they actually block toxin activity in living cells or protect animals in vivo — the next stages that would be needed before anything could move toward a real treatment. Anthropic says that if it were to advance one candidate further, Cn2-B1 would be among the strongest contenders, but it would still need to clear tests against toxin purified from real venom, functional studies measuring nerve-channel activity, and eventually trials in mice. Two independent experts who reviewed the work, Susana Vázquez Torres of the Spanish National Cancer Research Centre and Timothy P. Jenkins of the Technical University of Denmark, both called it a promising proof of concept. Anthropic is publicly releasing all 211 designs along with their binding data, hoping other labs will build on the results. The release adds to a string of life-sciences announcements from Anthropic this year, including the launch of its Claude Science research workbench and an earlier protein-binder campaign against more conventional drug targets. It also comes at a time when Anthropic’s broader initiatives — including a recently announced “[public benefit](https://officechai.com/ai/anthropic-criticized-for-opaque-nature-of-its-public-benefit-fund/)” fund — are drawing scrutiny over how exactly the company plans to deploy its resources toward causes like this one. For now, the scorpion antivenom work stands as another data point in Anthropic’s running argument: that an AI model, pointed at a real scientific problem and given the right tools, can meaningfully shrink the time it takes to get from “target identified” to “candidate molecule in hand” — even if turning that molecule into an actual medicine remains a long road ahead.