Anthropic’s Claude Science team designed proteins to stick to scorpion neurotoxins, had 211 of them tested in outside laboratories, and published the whole set. The campaign generated 318,484 protein designs and screened them as 553,052 design-toxin pairs. Seventeen designs met the company’s own criteria for binding. None of the seventeen was tested for whether it stops the toxin working. Anthropic states that it has not confirmed whether these binders neutralise scorpion venom in a cellular environment, and that it does not claim to have produced an antivenom.

That gap between a binding measurement and a medicine is where the campaign stopped, and much of the surrounding coverage stepped over it.

Citrini asked the question and Anthropic ran the campaign

Citrini Research, an investment research firm, approached Anthropic with a question about whether an AI model could design a new drug. Anthropic ran the work with an internal research model operating inside Claude Science. The starting prompt came from Anthropic’s earlier protein-binder study, modified with Citrini, and four follow-up prompts directed the work. Anthropic says the model worked mainly autonomously for 24 hours after the first prompt, and that it selected most of the 180 designs produced there, including the strongest binder, only after the third follow-up.

The design stage produced 318,484 protein candidates using 12 design tools, after filtering out anything resembling a toxin or a known protein, and screened them as 553,052 design-toxin pairs. About 64 hours after the first prompt there were 180 candidate designs. Anthropic’s own team added 31 more with Claude’s help, which brought the set sent for testing to 211. All 211 designs and their per-design measurements are published on Hugging Face under a CC BY 4.0 licence.

The five targets were the scorpion neurotoxins Cn2, AaH I, AaH II, LqhIII and Ts1. Cn2 is a neurotoxin of the Mexican scorpion Centruroides noxius. The design set was split by ambition: 60 designs aimed at a single toxin, twelve against each of the five; 30 aimed at the two beta-family toxins; 30 at the three alpha-family toxins; and 91 at all five at once.

What the wet lab measured

Two contract research organisations did the bench work, Adaptyv Bio and Twist Bioscience. Both measured binding by surface plasmon resonance, or SPR, an assay that detects binding in real time from the mass that accumulates on a sensor surface. Affinity is reported as KD, the dissociation constant. Lower values mean tighter binding; a nanomolar dissociation constant is a billionth of a mole per litre.

Anthropic’s criteria for calling a design a binder appear in four parts in the released dataset, which reproduces the report’s own wording. A design needed a hit against the toxin at both labs in every replicate, a KD from the main screen of 1 µM or tighter, a fitted maximum response at Twist Bioscience within the theoretical maximum, and no signal above background against CLEC12A, an unrelated human protein used to check for off-target stickiness. The thresholds are Anthropic’s own, and the published material cites no external standard and no preregistration. One carve-out shows they were not applied uniformly: LqhIII was assessed only by Adaptyv Bio, so for the two LqhIII binders the requirement of a hit at both labs could not have applied.

Seventeen designs met the criteria: seven against Cn2, four against AaH II, four against AaH I and two against LqhIII. None was a binder against Ts1. All seventeen bound one toxin, and no design aimed at several toxins met the criteria against more than one of them. Two hit rates in the write-up measure different things. Eleven of the 60 single-toxin designs were binders, a rate of 18 per cent. Across all 211 tested designs the rate is around 8 per cent, and the six-binder difference between the two counts came from the multi-toxin design groups, even though each of those binders still bound only one toxin.

The strongest molecules were Cn2-B1 and Cn2-B2. Cn2-B1 measured 3.1 nM at Adaptyv Bio and 1.7 nM at Twist Bioscience, both flagged in the dataset as upper limits, and Cn2-B2 measured 5.57 nM and 3.43 nM. That is tight binding in the single-digit nanomolar range.

Binding measures attachment, not effect

A binder is a protein that attaches tightly to a specific target. Anthropic calls them “minibinders”, its shorthand for small proteins that attach tightly to a target. A binding result says the protein sticks to the toxin, and nothing about whether the toxin has stopped working.

Anthropic did probe the next question with a competition experiment. Its Cn2 binders attach on or near the same epitope, or binding site, as an antibody called LR, the anti-Cn2 neutralising antibody from a 2016 paper in the Journal of Biological Chemistry. There LR was measured at a KD of 0.05 nM and protected mice against Cn2 and whole Centruroides noxius venom.

Anthropic then drew an inference beyond its own binding data, and the sentence is best quoted exactly: “We have not tested our strongest binders, Cn2-B1 and Cn2-B2, in a cellular environment or in live animals, but the fact that they compete with LR suggests that they will likely neutralize Cn2.” The dataset’s caption for the strongest molecule is more careful: “That Cn2-B1 should neutralize Cn2 as LR does is an expected outcome, not yet tested by patch-clamp recordings in cells.” The expectation is Anthropic’s, and we do not adopt it as a result.

The studied limits are set out in the same release. Anthropic says the two biggest limitations are that it did not test any design for neutralisation in vitro, against cells grown in the lab, or for efficacy in vivo, in living animals such as mice, and that binding was measured against recombinant toxins, grown in the lab from each toxin’s gene, rather than venom-purified toxin. No patch-clamp recording and no animal work were done on these binders.

Why the treatment problem sits downstream of the binder

Scorpion antivenom is made from the plasma of immunised horses, so it is a mixture of polyclonal antibodies. Scorpion venoms contain many neurotoxins, so a product has to cover the ones that matter, and existing antivenoms are effective against only a small number of them; supplies are limited and prices are high in the regions where stings are common. Anthropic cites BMJ Global Health for the scale: more than 1.2 million stings a year and more than 3,250 deaths, mostly among children with limited access to medical care. That paper was not accessible to us, so the figures remain Anthropic’s cited context rather than a count we verified.

A binder that sticks to one toxin is therefore a starting point. Susana Vázquez Torres, of the Spanish National Cancer Research Centre in Madrid, said in a comment Anthropic included in its release: “We can now obtain very tight binders remarkably quickly, and binding itself is no longer the main bottleneck… the real impact will come from tackling the parts of the therapeutic pipeline that go beyond binding.” She is the first author of the 2025 Nature paper on designed snake-neurotoxin binders that Anthropic cites as the precedent for this campaign.

That precedent shows how far the pipeline runs. Those snake binders were shown to neutralise their targets in patch-clamp assays, electrical recordings of the ion channels the toxins attack, and to protect mice. For scorpion toxins, Anthropic’s own comparison is three antibodies or nanobodies with KD values below 1 nM against Cn2 or AaH II that have been shown to protect mice. The two strongest binders here are in the single-digit nanomolar range, which is weaker, and Cn2-B1’s 1.7 to 3.1 nM is roughly 34 to 60 times weaker than LR’s 0.05 nM.

What the claim audit shows about the antivenom framing

The coverage that followed the release compressed it. Parts of the cycle described an AI that had produced a scorpion antivenom, and The Verge answered that compressed version with the headline “Claude did not invent scorpion antivenom”. Anthropic’s own position sits between the two, and it is narrower than either.

The Verge’s technical objection is the distinction this article has been drawing: “Binding to a toxin does not mean a protein can actually neutralize it, let alone function as a safe and effective treatment.” Its critics include Andreas Hougaard Laustsen-Kiel of the Technical University of Denmark, who said that discovering a good binder against snake and scorpion toxins is easy-ish and that “The hard part is to translate it into a medicine”, and Andreas Bender, who argued that existing screening methods “are often able to find a multitude of hits at manageable cost” and concluded that “It’s exactly this translation that lacks”.

One detail complicates how those quotes are usually read. Vázquez Torres and Timothy P. Jenkins, the two experts Anthropic quotes approvingly, are co-authors of the 2025 Nature paper that Anthropic cites as its precedent. Melisa Benard Valle and Laustsen-Kiel, two of the Verge’s critical experts, are co-authors of the same paper. Neither Anthropic nor the experts disclosed a paid relationship, and the disagreement between them is about how much therapeutic promise to read into binding data rather than about the data themselves.

What the release leaves open

The binding criteria were set by the lab that ran the campaign, and a different threshold would change the count. The published rows include qualifying binders recorded between 350 and 580 nM, well inside the 1 µM cut-off, and near-misses recorded from 1.29 µM to 3.44 µM as weak or ambiguous. A tighter cut-off would remove the weakest of the seventeen; a looser one would admit the near-misses. Citrini reports 35 design-toxin pairs that hit without passing the criteria; that figure counts pairs, and the 17 is a count of designs. Counts at other thresholds have not been computed.

No group outside the collaboration has replicated any of it, and the model used was an internal research model that outside researchers cannot currently run. The campaign used on the order of 10,000 GPU hours, and Anthropic’s own footnote concedes that this “is much more than most researchers have access to”. The same footnote reports a mitigation: the models were made about 4x faster, and a simplified single-agent, single-GPU, 24-hour prompt reached comparable in silico scores. That mitigation is a prediction from the model, not a lab-confirmed binding result. The Verge reports that Anthropic covered a compute and token bill of nearly $50,000, while Citrini’s own account lists the same figure inside its project budget; the release does not say who paid.

Two things can be checked without the unreleased model. The 211 designs and their per-design measurements are public under CC BY 4.0, so anyone can re-analyse the binding claims. Whether a protein that binds Cn2 would help a person who has been stung is a different experiment, and so far it has not been run.