Seventeen binders cleared dual-lab SPR, and Anthropic still withholds any neutralization claim.
On 20 October 2026, Anthropic published a research post and technical report on a Claude Science campaign against five scorpion neurotoxins that injure or kill people across several regions. An internal research model ran mostly on its own for 64 hours after a Citrini Research-shaped prompt and four follow-ups. Adaptyv Bio and Twist Bioscience then measured binding on 211 designs. Seventeen cleared Anthropic's bar. The company is explicit about what that bar covers and where it stops.
Claude ran a 64-hour design campaign against five toxins
Anthropic frames the novelty as a closed loop: a language-model research agent picks structures and hotspots, runs a multi-tool design campaign, emails sequences to a CRO, drafts figures and a report, and then survives a dual-lab binding filter with negatives published in full.
The toxin set spans Mexico (Cn2), Brazil (Ts1), North Africa (AaH I and AaH II), and the Middle East (LqhIII). Claude searched the Protein Data Bank and AlphaFold database for structures, chose five hotspots per toxin on the channel-binding face, and was asked for four binder classes: single-toxin, β-family (Ts1 and Cn2), α-family (AaH II, AaH I, LqhIII), and all five at once. Traditional horse-serum antivenoms are slow and narrow; a 2024-2025 snake-toxin de novo programme showed neutralization in mice. Anthropic says it knows of no prior published de novo scorpion-toxin binders.
The campaign used 12 design tools, produced 318,484 protein designs, co-folded and screened 553,052 design-toxin pairs, and consumed on the order of 10,000 GPU hours. Claude selected 180 designs for wet lab; Anthropic later added 31 with Claude's help, for 211 total: 60 single-toxin, 30 β-family, 30 α-family, and 91 five-toxin designs.
Of the 17 binders, all were single-toxin hits: seven against Cn2, four against AaH II, four against AaH I, two against LqhIII, and none against Ts1. Among the 60 single-toxin designs, 11 bound (18%), which Anthropic calls consistent with typical protein-design hit rates, including its August minibinder work.
Autonomy had limits worth naming. Claude worked alone for the first 24 hours. Anthropic's follow-ups then pushed multitoxin design harder. Claude selected most of the final 180 designs, including the strongest binder Cn2-B1, only after the third follow-up. Twist received sequence emails that Anthropic says Claude wrote. Under supervision, Claude also analysed the experimental data, generated figures, and drafted the technical report. The human role was steering and adding designs, with the wet-lab bar held by the CROs.
The public-health framing is load-bearing. Anthropic cites more than 1.2 million scorpion stings a year and more than 3,250 deaths, concentrated among children where care is scarce, and quotes the neglected-tropical-disease label for envenomation. Horse-serum antivenoms take months, cover few toxins, and remain expensive or scarce in the regions that need them. De novo minibinders are pitched as a path to potency, stability, and supply properties that animal antibodies struggle to deliver. That pitch is the stake of the campaign; it is also why the missing neutralization data matter.
Adaptyv and Twist measured binding under two expression formats
Both CROs measured binding by surface plasmon resonance. Adaptyv expressed designs in a cell-free system as monovalent proteins. Twist expressed them in mammalian HEK cells as Fc fusions. Anthropic required a binder label only when both labs saw binding in every replicate, every KD was 1 µM or tighter, and the design showed no signal against CLEC12A, an unrelated human protein used as a stickiness control. LqhIII was assessed only at Adaptyv.
Cn2-B1 and Cn2-B2, the strongest designs, bound Mexican Centruroides noxius toxin Cn2 with single-digit nanomolar KD at both labs. Competition experiments suggest both attach on or near the epitope of antibody LR, which has been reported to neutralize Cn2 in mice and covers the sodium-channel contact. Anthropic treats that competition as a reason to expect neutralization, and states that neither binder has been tested in cells or animals.
The five toxins share a fold while pairwise sequence identity runs only 26% to 62%. That spread is why Anthropic treated multitoxin design as a hard test rather than a marketing checkbox. Scorpion sodium-channel toxins trip nerves into dumping stress chemicals; severe cases bring heart failure and pulmonary oedema. A binder aimed at the channel-contact face is meant to block that latch. Binding to recombinant toxin is the first measurement in that chain. It is still only the first.
Outside reviewers Susana Vázquez Torres (CNIO) and Timothy P. Jenkins (Technical University of Denmark) called the work a promising first step and stressed that binding has ceased to be the main bottleneck for venoms. Their quotes sit on Anthropic's post as expert colour. They do not constitute an independent replication of the SPR table.
Dual-lab SPR is a hard filter; multitoxin failure stays on the record
The dual-lab, dual-format rule is stricter than a single CRO hit list. It also means 194 of 211 designs failed the published bar, including every multitoxin candidate against more than one toxin. Anthropic lists prompting, tool limits, and biological feasibility as open explanations. That honesty matters more than the headline binder count: the campaign tested whether Claude could design cross-toxin molecules and reported that it could not under these criteria.
Comparison to antibodies remains incomplete by Anthropic's own framing. Three antibody or nanobody molecules already report sub-nanomolar KD against Cn2 or AaH II with mouse protection. Claude's best designs enter that conversation on affinity and epitope overlap for Cn2; they do not yet enter it on function.
Neutralization, venom, safety, and access were left for later
Anthropic lists the two largest gaps up front: no in-vitro neutralization (for example patch-clamp on sodium channels) and no in-vivo efficacy. Binding used recombinant toxins rather than venom-purified protein. Safety, stability, manufacturability, regional pricing, and clinical pathways are named as remaining work. Multitoxin cocktails, the economically attractive path to a more universal antivenom, produced zero multi-hit binders in this round.
The GPU budget is another limit readers should keep. Anthropic notes that 10,000 GPU hours exceeds what most academic groups can spend, then points to separate work making related models about 4× faster and a simplified single-GPU, 24-hour prompt that reaches comparable in-silico scores. Those in-silico scores still leave a second dual-lab SPR campaign undone.
All 211 designs and binding data sit on Hugging Face
Anthropic released the full design set and binding data as dataset Anthropic/scorpion-toxin-binder-design on Hugging Face, with a technical PDF hosted on Anthropic's CDN. The research post links Adaptyv and Twist as the experimental partners and credits Citrini Research on the problem framing. Outside groups can re-express sequences, re-run SPR, or attempt the neutralization steps Anthropic deferred.
Treat the release as a binder dataset and a process receipt
Drug-discovery and neglected-disease teams can pull the 17 binders and the 194 negatives as a public training and methods resource, especially the Cn2 nanomolar pair and the LR competition note. Procurement and science-policy desks should file the piece as evidence that agentic design plus CRO SPR can close a neglected target in weeks, while keeping neutralization, venom, and access off the claim list until someone measures them.
The "17 binders" figure belongs inside Anthropic's own refusal of an antivenom claim. The multitoxin failure should stay visible: a universal cocktail still needs chemistry that this campaign did not deliver. If a team advances Cn2-B1, Anthropic's own next-step list is the checklist: venom-purified binding, patch-clamp neutralization, mouse protection, then safety and access work with clinicians and funders.
This sits after Anthropic's August minibinder demo and snake-toxin de novo work
The August 2026 Claude minibinder study supplied the base prompt pattern. Snake-toxin de novo binders with mouse protection are the therapeutic precedent Anthropic cites for ambition. This October campaign adds a neglected human-health target, multitoxin design pressure, dual-CRO confirmation, and a full public dump of hits and misses. Citrini's non-expert framing question (can motivated outsiders plus Claude tackle a real health target) is answered only up to binding. The next paper in the line that would change Monday reading is a neutralization or mouse study on Cn2-B1 or a true multi-toxin hit that clears the same SPR bar.