Claude Designed Working Protein Binders for 14 of 15 Targets — at Double the Industry Hit Rate
Anthropic ran Claude as an autonomous protein-design agent, and Adaptyv Bio's wet lab confirmed the results without modification. Claude hit binders on 14 of 15 targets at a 26.8% average rate versus 10–15% for human de novo campaigns.
Anthropic put Claude in charge of a protein-design campaign — planning, reasoning, and iterating on candidate molecules — and then had the designs physically built and tested in a lab. Claude succeeded on 14 of 15 targets, with an average hit rate of 26.8%. The industry baseline for de novo protein binder campaigns is 10–15%. Claude roughly doubled it.
The scale matters. Claude designed 1,320 protein binders and pushed them through an automated pipeline; 354 were confirmed to physically bind their intended targets. Adaptyv Bio and Twist Bioscience synthesized and tested the designs in the wet lab without modification — no human editing of Claude’s output before it hit the bench. Depending on how the sessions were run, hit rates ranged from 22.6% to 35.1%.
Some individual results are the real headline. Against TREM2, Claude reached an 80% hit rate, up from 38.3% in a prior Adaptyv Bio competition. On RBX1, Anthropic’s Mythos Preview model recorded a 40% hit rate versus 3.7% among participants in an earlier competition. Binding affinity improved by orders of magnitude in places: the best Claude binder for 15-PGDH measured 33.4 nanomolar against a previous 1.7 micromolar — roughly a 50x tighter bind. For RBX1, 3.9 nanomolar versus 25.7 nanomolar.
What’s new isn’t that AI can design proteins — specialized models like AlphaFold and RFdiffusion have done structure and design work for years. It’s that a general-purpose language model ran the workflow: reading the problem, selecting tools, reasoning about candidates, and iterating, with humans and lab robots reduced to the physical steps. Claude wasn’t a component in a pipeline. It was the scientist directing the pipeline.
That framing is exactly why the result carries a safety edge, and Anthropic said so directly. A model that can autonomously design functional proteins that bind human targets is dual-use by definition. The same capability that accelerates drug discovery, aging research, and disease-resistance work is the capability that biosecurity teams worry about. Anthropic publishing this alongside its safety framing — rather than burying the lab numbers — is the point: the capability is here, and pretending otherwise helps no one.
For biotech, the practical implication is speed and cost. De novo binder discovery is slow, expensive, and failure-heavy; a model that triples hit rates on hard targets compresses the discovery loop and shifts where human expertise is most valuable — toward problem selection and validation, away from brute-force candidate generation.
The honest caveat: this is one benchmark, run in partnership with the labs that validated it, on 15 targets. It’s a genuine result, not a product. But “a chatbot ran a protein-design campaign and beat the human baseline in a real wet lab” is the kind of sentence that would have read as science fiction two years ago.