Claude designed protein binders in the wet lab while models learned to brainstorm

Claude goes to the wet lab

Anthropic researchers set an internal research model a concrete wet-lab problem: design minibinders — small proteins that attach tightly to a target — against scorpion neurotoxins. Over 64 hours, working mainly autonomously inside the Claude Science environment, Claude generated 318,484 protein designs, a campaign that consumed on the order of 10,000 GPU hours. Claude picked the toxin structures itself by searching the Protein Data Bank and the AlphaFold database. m:…3199

The designs left the computer. Adaptyv Bio and Twist Bioscience experimentally tested 211 of them for binding — Adaptyv in a cell-free system, Twist in mammalian cells — and 17 met the criteria: evidence of binding in every replicate, with every dissociation constant (KD) 1 µM or tighter. The two strongest binders, which the researchers call Cn2-B1 and Cn2-B2, target Cn2, a neurotoxin from the Mexican scorpion Centruroides noxius. The stated motivation is real: scorpions sting 1.2 million people or more each year, causing more than 3,250 deaths. m:…3199

The caveat, stated plainly

Anthropic states the limit plainly, and it is worth carrying in full: the team has not confirmed whether these binders neutralize scorpion venom in a cellular environment, so it does not claim to have produced an antivenom. That is the honest shape of the result — a binding result, not a therapeutic one. The researchers are releasing all 211 designs and their data as a resource for the field, which is the other half of the story: machine-generated protein design campaigns can now be published as open datasets. m:…3199

Models learn to brainstorm

Meanwhile, models are getting better at the part of science that comes before the experiment: coming up with ideas. In a preprint, researchers describe a set-level reinforcement learning objective that trains language models to generate diverse, high-quality, document-grounded ideas — a single generation writes K structured idea cards grounded in source papers, and the set is scored jointly, with each idea weighted by how different it is from the ideas before it. Their 8B model, the paper reports, achieves higher creative utility — quality plus diversity — than Claude Opus 4.8 and GPT-6 Astra on a contamination-free benchmark, and in a pilot human study its idea sets were judged as promising as Claude Opus 4.8's. m:…4770

Sakana AI is pushing a related idea into decision-making with QD-Rec, a recommendation method that applies open-ended exploration — quality-diversity evolution — to decision support. Instead of recommending candidates alone, it recommends candidate-hypothesis pairs: the candidate plus the reasoning for choosing it, with new hypotheses generated by mutation and crossover, and unique perspectives kept as stepping stones even when they aren't currently the best. The company reports it outperformed baselines in average ratings and recommendation set size in most settings across its experiments — company-run experiments, not independent replication. m:…0825

Discovery as a workflow

Reasoning, not reporting: the thread across all three is the loop closing. Claude ran a design campaign end to end and handed it to wet labs; the 8B paper suggests idea generation itself can be trained as a set-level objective rather than left to luck; QD-Rec treats hypotheses as the unit of recommendation. None of this is a breakthrough claim — it is infrastructure for a world where the model does the brainstorming, the experiment, and the triage, and humans set the criteria. The binder campaign is the proof that the loop can touch reality, and the caveat is the proof the authors know the difference. m:…3199 m:…4770 m:…0825