Anthropic announced on October 20, 2026 that an internal Claude research model designed 17 protein binders against scorpion neurotoxins, confirmed by two independent laboratories. The binders are not an antivenom and were not tested in cells or animals. That sentence carries the whole story: a genuine acceleration of early-stage drug discovery, and a result that most headlines will oversell. This is the honesty-first breakdown of what happened in those 64 hours, what it means, and where it stops.
What Claude actually did
Working inside Anthropic's Claude Science environment, an internal research model ran a protein-design campaign with what Anthropic calls "considerable autonomy." The initial prompt defined the neurotoxin targets, the design and folding tools, and the rules for novelty, scoring, and filtering. Beyond choosing the toxin structures and amino-acid hotspots, Claude set its own quality thresholds, decided how heavily to lean on each of 12 design tools, and developed strategies for designing binders against multiple toxins at once. It filtered out anything resembling a known toxin or protein, then co-folded and screened 553,052 design-toxin pairs, predicting each pair's structure as one complex. Sixty-four hours after the first prompt, with about 10,000 GPU hours spent, it produced a final list of 180 designs. The Anthropic team added 31 more with Claude's help, bringing the total to 211.
The campaign targeted five scorpion neurotoxins across four regions: Cn2 from Mexico (from the scorpion Centruroides noxius), Ts1 from Brazil, AaH I and AaH II from North Africa, and LqhIII from the Middle East. The designs aimed at each toxin's channel-binding site, the part of the neurotoxin that latches onto sodium channels in nerve cells. Claude picked the structures from the Protein Data Bank and the AlphaFold database, and selected five "hotspot" amino-acid positions per toxin for its designs to latch onto. The toxins share one overall protein fold but any two are only 26 to 62 percent identical in sequence, which made them a deliberately hard test for multi-toxin design.
Then came the part that separates this from a simulation demo. Two contract research organizations, Adaptyv Bio and Twist Bioscience, lab-tested all 211 designs for binding. They used complementary methods: Adaptyv expressed the designs in a cell-free system built from bacterial protein-making machinery, while Twist expressed them in mammalian cells (human kidney cells) as fusions to Fc, the stem of an antibody. Both measured binding with surface plasmon resonance, an instrument that watches a toxin bind to and detach from a design in real time and produces a dissociation constant, or KD. Lower KD means tighter binding. The mechanism matters here because a language model driving the whole pipeline is only as interesting as the part it cannot fake, and the part it cannot fake is the wet lab. Our visual explainer of how transformers actually work walks through the architecture that makes this kind of tool-orchestrating agent possible in the first place.
The strict bar, and what cleared it
Anthropic set a strict definition of a binder. A design counted only if it showed binding in every replicate at both labs, with every measured affinity at 1 micromolar or tighter, and with no signal against CLEC12A, an unrelated human protein used to detect off-target "stickiness." Seventeen designs met all of that, each against a single toxin: seven against Cn2, four against AaH II, four against AaH I, and two against LqhIII. None hit Ts1. None of the multi-toxin designs met the criteria for more than one toxin, despite Claude designing 91 of them against all five toxins simultaneously.
Among the 60 designs aimed at a single toxin (12 per toxin), 11 were binders, an 18 percent hit rate that Anthropic says is consistent with typical protein design campaigns, including its own earlier work. The two strongest binders, named Cn2-B1 and Cn2-B2, bound Cn2 with single-digit nanomolar affinities at both labs. Competition experiments suggest they attach on or near the same site as an antibody called LR, which has been reported to neutralize Cn2 in mice. Presenting that carefully: the competition result is evidence about where the binder sits, not proof that it neutralizes anything. Anthropic has not tested its binders against toxin purified from actual venom, in cells, or in animals.
| Toxin | Region | Confirmed binders |
|---|---|---|
| Cn2 | Mexico | 7 (incl. Cn2-B1, Cn2-B2) |
| AaH II | North Africa | 4 |
| AaH I | North Africa | 4 |
| LqhIII | Middle East | 2 |
| Ts1 | Brazil | 0 |
The origin story: a Wall Street question
The project did not start in a pharma lab. It started with a question from Citrini Research, a financial investment research firm, which approached Anthropic about a month before the campaign: could Claude, and AI drug discovery more broadly, enable a group of motivated nonexperts to tackle a genuine problem in human health? Anthropic's life science team brainstormed problems with Claude and landed on scorpion antivenom. The motivation is real enough. Scorpions sting 1.2 million people or more every year, killing more than 3,250, mostly children in parts of the world with limited access to medical care. Scorpion envenomation is classed as a neglected tropical disease, starved of investment because new treatments are unlikely to earn back their development cost. Existing antivenoms are made by milking venom from thousands of scorpions, immunizing horses, and purifying antibodies from horse blood, a process that takes months, works against only a small number of neurotoxins, and is expensive or unavailable in many of the places where stings happen.
Two outside experts reviewed the work. Susana Vázquez Torres, a postdoctoral researcher at the Spanish National Cancer Research Centre, called binding "no longer the main bottleneck" and said the real impact will come from tackling everything downstream of binding. Timothy P. Jenkins, an associate professor at the Technical University of Denmark, called it a great showcase of the AI-driven paradigm shift in therapeutic discovery, especially for neglected diseases starved of innovation by low commercial incentives. Anthropic is releasing all 211 designs and their binding data for other researchers to build on.
Evergreen: how de novo binder design works, and why binding is not a cure
A binder is a molecule that attaches tightly to a specific target; antibodies are the natural example. In de novo design, protein binders are generated computationally from scratch against a chosen target, then optimized for druglike properties like potency and stability. The modern pipeline runs roughly like this: pick a target structure, choose binding hotspots on it, generate new protein backbones aimed at those hotspots with diffusion-based design tools, predict how each design folds against the target, filter by predicted quality, then make the best candidates in a lab and measure whether they actually bind. That last step is the honesty gate. Computational scores predict; surface plasmon resonance measures.
Binding is the first and easiest milestone on a long road, and Anthropic's own report lays out the rest. Even the strongest binder, Cn2-B1, would need to be confirmed against toxin purified from venom, then tested for function with patch-clamp recordings that measure whether it stops the toxin from acting on sodium channels in cells, then tested in mice to see whether it actually protects them. Beyond neutralizing the toxin, a treatment must be safe, stable, affordable, and available where stings occur, which takes clinicians, funders, and policymakers. None of that has happened. So "17 binders" means 17 molecules that stuck to their target in two labs under strict criteria. It does not mean an antivenom, a drug, or a treatment. Anyone telling you otherwise is selling the headline, not the science.
Why this still matters
The honest reading is not that nothing happened. In January 2025, David Baker's team published de novo designed proteins that neutralized lethal snake-venom toxins and protected mice, in Nature. Anthropic's campaign, a year and a half later, shows a general-purpose AI agent running a comparable early-stage campaign mostly on its own, from a handful of prompts, in under three days. The bottleneck in therapeutic discovery is moving upstream: from "can we make a molecule that binds?" to "can we turn binders into drugs?" Vázquez Torres's verdict is the quotable one: designing a binder against almost any target no longer seems unrealistic, and binding itself is no longer the main bottleneck.
There is a harder question underneath, and it is about release. Anthropic published all 211 designs openly, and this same dual-use tension, models that accelerate both medicine and harm, is why we mapped every AI shutdown proposal on the table in The Kill-Switch Map: the capability to design binders against toxins is the same capability class that biosecurity frameworks exist to govern. Anthropic's bet is that open data accelerates neglected-disease research faster than it enables misuse. That bet deserves scrutiny alongside the celebration, because the campaign's most important precedent may not be the 17 binders. It may be that a motivated group with no lab experience and access to a frontier model can now run a credible early-stage therapeutic design campaign at all.
By ByteHerald Staff. All campaign figures from Anthropic, "How Claude designed protein binders against scorpion neurotoxins," October 2026. Outside expert quotes as reported by Anthropic. Milestone dates verified: AlphaFold 2 at CASP14 (2020); RoseTTAFold in Science (2021); RFdiffusion preprint (2022), published in Nature (Aug 2023); snake-venom neutralizing proteins in Nature 639 (Jan 2025); Anthropic single-prompt minibinder study (Aug 2026).
Comments
Post a Comment