AI-Designed Proteins Bind Four Scorpion Toxins in First Claude Science Campaign

Anthropic on October 20, 2026, published results from the first disclosed wet-lab campaign run on Claude Science, its AI-powered biological research platform: an unreleased internal research version of Claude autonomously generated 318,484 de novo protein sequences targeting five deadly scorpion neurotoxins, then selected candidates that went on to bind four of the five toxins in independent laboratory tests. The campaign also surfaced the binding research's central limitation: all 151 attempts to design a protein capable of gripping more than one toxin simultaneously produced zero successes — a result that directly constrains any path from these binders to a practical antivenom.

The campaign ran for approximately 64 hours and consumed roughly 10,000 NVIDIA H100 GPU hours. Of 211 candidate proteins sent for wet-lab testing, 17 met stringent binding criteria across two independent laboratories. The best-performing design, Cn2-B1, bound Centruroides noxius toxin 2 with a dissociation constant of approximately 1.7 nanomoles per liter at one lab and 3.1 nanomoles per liter at the other — a level competitive with engineered antibodies, according to Anthropic's technical report.

Why Scorpion Venom — and Why Conventional Antivenoms Fall Short

Scorpion envenomation kills more than 3,250 people per year, injures more than 1.2 million, and strikes hardest at children in under-resourced regions of Latin America, North Africa, and the Middle East — the same areas where the five targeted toxins originate. Scorpion envenomation remains a neglected tropical disease largely because it generates no pharmaceutical market returns sufficient to justify expensive development pipelines.

Conventional antivenom manufacturing still depends on immunizing horses or other large animals with toxin fractions, drawing blood, and extracting polyclonal antibodies — a process measured in months. The resulting serum is expensive (scorpion venom has been cited at roughly $39 million per gallon by Citrini Research, the investment firm that proposed this project to Anthropic) and geographically restricted: antivenom effective against Mexican Centruroides species provides no protection against North African Androctonus toxins, because each serum is trained on local venom fractions. Cold-chain requirements further constrain distribution in high-burden regions. Mexico, Brazil, India, and Algeria together account for the majority of documented scorpion fatalities; in parts of northern Brazil's Piauí and Minas Gerais states, local Tityus serrulatus populations are so prolific that antivenom demand routinely outstrips supply.

Citrini Research framed the challenge to Anthropic as a test case: could a group with no formal biology training use Claude to make progress on a genuine public-health problem? The answer involved targeting "binders" — short engineered proteins designed to physically block the surface that each toxin uses to attack neuronal sodium channels.

Inside the 64-Hour Campaign: 12 Tools and 553,052 Co-Folds

The design pipeline drew on approximately 12 specialized computational programs, assembled and orchestrated autonomously by an internal research version of Claude.

Claude retrieved three-dimensional toxin structures from the Protein Data Bank and AlphaFold's database, then identified the binding "hotspot" on each toxin — the surface region that normally contacts voltage-gated sodium channels (Nav). Scorpion neurotoxins disable neurons by locking Nav channels open, producing sustained depolarization and, in severe cases, cardiac and pulmonary failure. A designed binder that occupies this hotspot would physically prevent the toxin from reaching the channel.

With hotspots defined, Claude applied diffusion-based backbone generation tools analogous to RFdiffusion to create novel protein scaffolds shaped to complement each hotspot geometry, then used inverse-folding tools similar to ProteinMPNN to find amino acid sequences that would reliably fold into those shapes. This produced 318,484 candidate designs. Each was then evaluated by co-folding prediction — structural simulation of the candidate docked against its target — across 553,052 design–toxin pairs. Candidates lacking compatible geometry were discarded; those too similar to natural proteins or to the toxins themselves were filtered. After roughly 64 hours of this generate-and-evaluate loop, Claude delivered 180 shortlisted proteins. The research team added 31 more with Claude's assistance, yielding 211 total candidates.

What Bound and What Didn't: Results Across Five Toxins

The 211 candidates were split between Adaptyv Bio (cell-free expression) and Twist Bioscience (HEK-cell antibody fusions). Both labs used surface plasmon resonance (SPR) to measure binding affinity in real time. To qualify, a candidate needed a dissociation constant (KD) at or below 1 micromole per liter in all replicates, with no signal against CLEC12A — an unrelated human protein used as a non-specific-binding control. For four toxins, candidates needed independent confirmation from both labs. LqhIII was excepted because the toxin did not reach Twist Bioscience in time; its two qualifying designs were evaluated at Adaptyv Bio only.

Results by toxin: Cn2 yielded 7 confirmed binders; AaH I and AaH II each yielded 4; LqhIII yielded 2 (single-lab). Ts1, the Brazilian toxin, produced zero qualifying binders. Overall hit rate: 8.1 percent of all 211 candidates; 18.3 percent among the 60 single-toxin-targeted designs.

The standout result was Cn2-B1, with KD values well below 10 nM at both labs. Cn2-B2 reached 5.6 nM at Adaptyv Bio and 3.4 nM at Twist. Anthropic noted that SPR can underestimate affinity for some binding geometries, suggesting these figures may be conservative.

A further competition assay confirmed that Cn2-B1 competes with LR — an antibody fragment already known to neutralize Cn2 — for the same surface on the toxin. That surface is the functionally critical region the toxin normally uses to contact Nav channels. The competition result does not prove neutralization, but it indicates that Cn2-B1 targets the therapeutically relevant epitope rather than an inert patch.

The Multitoxin Wall: 151 Designs, Zero Successes

The most consequential result in the dataset goes largely unaddressed in initial reporting: all 151 candidates specifically designed to bind two or more toxins simultaneously failed to pass binding criteria for more than one.

This is not a statistical anomaly. The five targeted toxins share the inhibitory cystine knot (ICK) scaffold — a fold stabilized by conserved disulfide bonds common across venom toxins from scorpions, spiders, and cone snails. Sequence identity across toxin pairs ranges from roughly 26 to 62 percent. A conserved fold might, in principle, allow a single binder to occupy the equivalent surface on multiple toxins. In practice, the functional binding surfaces — the Nav-contact regions a therapeutic binder would need to block — differ enough between toxins to defeat every design attempt.

The practical consequence is significant. A scorpion antivenom covering multiple geographic regions would need to neutralize multiple distinct toxins. Single-toxin binders, regardless of their affinity, address only part of that requirement. Solving the breadth problem will require either a breakthrough in multitoxin design strategy, a cocktail of single-toxin binders formulated together, or a different therapeutic approach entirely. A cocktail strategy is technically feasible but complicates manufacturing, dosing, and regulatory approval — three obstacles that have historically slowed even well-funded antivenom programs.

From Binding to Neutralization: The Gap That Remains

Anthropic is explicit about what this study does not demonstrate. The company stated in its technical report that it "has not confirmed whether these binders neutralize scorpion venom in a cellular environment" and does not claim to have produced an antivenom.

The tests used recombinant toxins — laboratory-produced proteins matching each natural toxin's sequence — rather than venom drawn from actual scorpions. Recombinant toxins allow precise experimental control but do not capture the full complexity of natural venom. More importantly, no cellular assay testing whether the binders protect neurons or cardiac cells from toxin activity has been published, and no animal protection study has been announced. A standard cellular assay would involve exposing cultured neurons or cardiomyocytes to active toxin in the presence and absence of a candidate binder, then measuring whether Nav-channel activity is suppressed — a step that requires BSL-2 biosafety facilities and access to functional recombinant or native toxin at relevant concentrations.

The Baker Lab comparison defines what a successful endpoint would look like. In January 2025, University of Washington researchers published in Nature that de novo proteins designed using RFdiffusion neutralized three-finger toxins from elapid snakes in mice, achieving 80 to 100 percent survival from otherwise lethal doses. That study took the field from binding proof-of-concept to in-vivo functional protection. The Cn2-B1 competition assay result is the most encouraging signal that the scorpion work could eventually follow the same trajectory, but the cellular and animal studies required to demonstrate this have not begun.

Read more: Anthropic's autonomous biology claims under scrutiny

Open Data, GPU Scale, and the Dual-Use Question

All 211 protein sequences and SPR measurements were released on HuggingFace under a CC BY 4.0 license, including designs that failed — a dataset that may be as useful to the research community as the successes. A 69-page technical report documents the full pipeline and raw measurements. Anthropic also described a simplified single-GPU version capable of completing a comparable design run in roughly 24 hours, acknowledging that the campaign's ~10,000 H100 GPU hours substantially exceed typical academic resources.

The campaign used an internal research version of Claude unavailable via the public API. That distinction matters because Claude Fable 5, the current public model, actively restricts biology-relevant queries carrying dual-use risk. Anthropic is simultaneously building a closed-access biological research capability and restricting analogous queries on its open platform — a bifurcated approach that will require ongoing justification as Claude Science's capabilities grow.

The immediate next steps are technically clear: cellular neutralization assays, animal protection studies, and renewed multitoxin design attempts using strategies that target conserved structural rather than functional surfaces. Whether Anthropic's platform or academic groups using the open dataset pursue those milestones was not indicated. What this campaign has established is that autonomous AI-directed protein design can now operate at the scale and selection quality needed to populate early-stage antivenom pipelines — a credible starting point, even if the distance to a deployable therapeutic remains substantial.