How Claude designed protein binders against scorpion neurotoxins
A · Strong evidence Anthropic's science team reports an agent-run protein design campaign with wet-lab binding tests at two independent contract labs, strict hit criteria and all 211 designs released; a vendor showcase with no neutralisation data, so A rather than S.
Summary based on the full text·Original: about 10 minutes to read
In short
An Anthropic research post (20 October 2026) on an internal research model designing minibinders, small proteins that attach tightly to a target, against five scorpion neurotoxins. The project started from a question by Citrini Research, an investment research firm: can AI tools let motivated non-experts work on a neglected health problem? The model ran the design campaign mostly on its own for 64 hours. Two contract labs tested 211 designs and 17 bound their target toxin. None of the designs meant to bind several toxins at once worked. No design has been tested for neutralising venom, and Anthropic does not claim an antivenom.
Key numbers
- 318,484protein designs generated by the model over the campaign, using 12 design tools
- 17 of 211lab-tested designs that met the binder criteria at both labs, all against a single toxin
- 11 of 60 (18%)hit rate among single-toxin designs, which Anthropic calls typical for protein design
- 64 hourstime from first prompt to the final list of 180 designs, with four follow-up prompts
- ~10,000GPU hours used by the campaign
As reported by the source.
English·Web page·opens anthropic.com
What it reports
- Targets: Cn2 (Mexico), Ts1 (Brazil), AaH I and AaH II (North Africa), LqhIII (Middle East). Any two share only 26% to 62% sequence identity.
- The model chose toxin structures and binding hotspots, set quality thresholds, picked among 12 design tools, generated 318,484 designs and co-folded 553,052 design-toxin pairs. It ran 24 hours unprompted, then continued after four follow-ups; most of the 180 chosen designs came after the third.
- Staff added 31 designs with the model's help, for 211 in total: 60 single-toxin, 30 two-toxin, 30 three-toxin, 91 five-toxin.
- 17 of 211 met the binder criteria (seven Cn2, four AaH II, four AaH I, two LqhIII). Single-toxin hit rate 11 of 60 (18%). Zero multi-toxin designs bound more than one toxin.
- The two strongest, Cn2-B1 and Cn2-B2, bound with single-digit nanomolar affinity at both labs and compete with LR, an antibody reported to neutralise Cn2 in mice.
- Context: scorpions sting 1.2 million people or more a year, with more than 3,250 deaths.
How they got it
Adaptyv Bio (cell-free expression) and Twist Bioscience (mammalian cells, Fc fusions) each produced and tested the designs by surface plasmon resonance. A binder had to hit at both labs in every replicate with KD of 1 µM or tighter, and show no signal against an unrelated human protein. LqhIII was tested at one lab only. Two outside researchers reviewed the work; the model also drafted the report and figures under supervision.
Limits
- Binding was measured only against lab-grown recombinant toxins, not venom-purified toxin.
- No cell-based neutralisation or animal tests were run; the link to neutralisation is inferred from antibody competition.
- About 10,000 GPU hours is far beyond most labs; Anthropic says a simplified single-GPU setup reaches comparable in-silico scores, without lab data for it.
- The campaign was directed by follow-up prompts from Anthropic staff, so "autonomous" is partial.
Why it matters for companies
The result shifts where the cost sits in early discovery. Generating credible binders in days is now plausible; testing, formulation and getting a drug to patients remain the slow, expensive part. Firms in life sciences can read this as a template: agent-run computational campaigns paired with outsourced lab validation and strict, pre-set hit criteria. The release of all designs and data, including failures, is the standard to ask of any vendor making similar claims.
Summary based on the full text. Read the original for the full argument and data.Read the original on anthropic.com
- Tags
- AI for science