Claude designs binders against scorpion neurotoxins
Event confirmedAwaiting review
The takeaway
On Oct 20, 2026 Anthropic’s Science Blog reported that an internal Claude research model, working mostly autonomously for 64 hours, designed protein binders against five scorpion neurotoxins.
Status
- Claim
Event confirmedAwaiting review
- Our reporting
- High confidence
- Verification
- Unverified claim
- Importance
- 3 of 5
- Last verified
- 21 October 2026
Your AI and this story
- GPT-6 Astra173 days after its cutoff
- Claude Opus 5.5112 days after its cutoff
- Gemini 4 Argon203 days after its cutoff
- Grok 4.7142 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 112 days before it.
Key facts
- Partners: Anthropic’s Science team with Citrini Research; a test of whether Claude can help motivated non-experts tackle a health problem. Model: ‘an internal research model, operating within Claude Science’ (not named)
- Autonomy: 64 hours, four follow-up prompts, about 24 hours unattended after the first prompt. Claude chose toxin structures (PDB, AlphaFold DB), five hotspot residues per toxin, thresholds and multi-toxin strategies, and emailed sequences to Twist Bioscience itself
- Scale: 12 design tools, 318,484 designs, 553,052 design–toxin pairs co-folded and scored, ~10,000 GPU hours; 180 designs chosen by Claude plus 31 added later with its help = 211
- Targets: Cn2 (Mexico), Ts1 (Brazil), AaH I and AaH II (North Africa), LqhIII (Middle East), aimed at each toxin’s sodium-channel binding site
- Validation: Adaptyv Bio (cell-free) and Twist Bioscience (human cells, Fc fusions), surface plasmon resonance at both. Criteria: hit in both labs, KD ≤ 1 µM, no binding to off-target CLEC12A. 17/211 passed (Cn2 7, AaH II 4, AaH I 4, LqhIII 2); single-toxin designs 11/60 (~18%). No multi-toxin design passed for more than one toxin
- Leads: Cn2-B1 and Cn2-B2, single-digit nM at both labs; competition experiments suggest they bind at or near the site of the LR antibody, reported to neutralize Cn2 in mice
- Caveats (Anthropic): tested only against recombinant toxins, not venom-purified ones; no in vitro or in vivo neutralization; ‘we do not claim to have produced an antivenom’. All 211 designs and data released on Hugging Face
- Reviewers quoted: Susana Vázquez Torres (CNIO) called it a promising first step; Timothy P. Jenkins (DTU) said binding ‘is no longer the main bottleneck’
Show 1 more
- Context: >1.2M scorpion stings and >3,250 deaths a year, mostly children; Anthropic’s X post had ~1.38M views (fxtwitter, Oct 21)
What happened
Anthropic and Citrini Research let an internal Claude model run a protein-design campaign against five scorpion toxins with little human input, then had the chosen designs made and tested by two commercial labs. About 8% of all designs, and about 18% of single-toxin designs, bound their targets, a hit rate the authors call typical for protein design. The attempt to make one binder that catches several toxins failed.
Why it matters
It is one of the first published cases of a frontier lab’s model running a wet-lab-validated design campaign mostly by itself, with data released for checking. It also shows the limits: binding is not neutralization, and the hardest goal (a broad binder) was not reached. Anthropic chose a project with clear medical value, while its CEO also points to bio misuse risk as a reason for strict biology safeguards.
What is disputed or not yet verified
| Verification | Lab-validated binding (SPR at two labs); no neutralization or peer review |
|---|
Sources
2 sources from 2 sites. Numbers match the chips in the text.
2 sources: 2 primary
Primary
- Anthropic: How Claude designed protein binders against scorpion neurotoxinsanthropic.com, official
- Anthropic on Xx.com, official
Changes
- Filed