report · 2026-10-20
De novo design of protein binders against scorpion toxins with Claude
On this page
Why it matters here
Anthropic reports a protein-design campaign run by an unnamed internal Claude model within Claude Science. Working from a human-written prompt and four follow-ups, it selected 180 designs in 64 hours; humans added 31 later with Claude. Of the 211 tested designs, the authors labeled 17 as binders against four of five scorpion toxins under their criteria. The versioned released table contains the same 211 design identifiers and 17 labels. The report separates computational selection from paid laboratory testing and discloses post-data binding criteria, a one-lab exception and assay limitations. This supplies context about an AI-assisted scientific workflow; it does not demonstrate toxin neutralization, causal research acceleration or recursive self-improvement.
What to keep in mind
- One campaign using an unnamed internal Claude research model. The exact model version and reasoning configuration are not public, and this does not establish capabilities available in ordinary Claude subscriptions.
- Humans selected the problem, targets, tools, ranking rules and a roughly 19,000-word prompt; four follow-ups extended and redirected the campaign. The 64-hour figure covers the initial 180 selected designs; 31 designs were added later with Claude. It is not the duration of the complete laboratory study.
- The campaign used about 10,000 H100 GPU hours. No matched expert or fixed-pipeline design baseline using the same tools and resources was run, so the model's contribution and time/cost savings cannot be isolated.
- Binding criteria were set after seeing the data. The 17 are the authors' qualified binder labels, corroborated as labels in the released table, not independently re-fitted assay results. LqhIII was tested at only one lab, so its two binders are not directly comparable to the two-lab cases.
- No design met the criteria against Ts1 or against multiple toxins. Only Cn2 and AaH I had positive controls working at both labs; the Ts1 results may depend on recombinant-toxin quality.
- Affinities are preliminary. AaH I classification depends on the tagged-toxin mass assumption; reverse BLI reproduced hits for 12 of 15 binders tested at Twist, and two binders bound the unrelated ovalbumin control. Missing, ambiguous and unassessable measurements are not zeros.
- No neutralization, efficacy in cells or animals, or clinical benefit was demonstrated. Binding to recombinant toxins is not an antivenom result.
- Anthropic and Citrini paid the contract laboratories. Financial-sector coauthors helped conceive and direct the project; external advice/review and paid laboratory measurements are not independent experimental replication or formal journal peer review.
- The released-data check covered counts/identities, binder labels, selected controls, reverse BLI, ovalbumin and one kinetics example. Full per-design traces, all supplementary panels and unavailable capture-level/quality-control data were not independently audited.
- This is a bounded example of AI-assisted scientific research, not AI-system improvement, a retained recursive research loop, a cross-system capability ranking or a tracker autonomy assessment. No numeric observation or public figure is included.