Anthropic Says Claude Designed Protein Binders Largely on Its Own — Hitting 14 of 15 Targets

Anthropic has published research claiming its model Claude designed novel protein binders largely on its own, then passed independent wet-lab validation. A protein binder is a small protein engineered to latch onto a specific target protein — a starting point for drugs, diagnostics, and research tools. The notable part is that Claude took one detailed human-written prompt and then handled most of the pipeline itself, from researching targets to selecting binding sites and optimizing designs.

What Claude did

Claude researched the targets, chose binding sites (epitopes), and orchestrated open-source structure- and sequence-design tools to generate candidates, optimizing them in silico. It then screened for solubility, expressibility, and novelty before producing ranked designs. The actual protein synthesis and binding assays were handled by external partners Adaptyv Bio and Twist Bioscience, which independently produced and tested the model-generated sequences.

What the numbers really say

Beating the typical industry range by roughly two-to-one is striking, but the individual cases are more telling. Anthropic says Mythos Preview, in single-target mode, achieved a 40% hit rate against RBX1 — far above the 3.7% average among participants in Adaptyv's public competition — and its top-ranked design was a high-affinity binder that outperformed the contest's winning entry. Some of Claude's strongest designs bound several times more tightly than the best previously published de novo binder.

Validation and pushback

The weight of the result rests on independent verification. Claude did the designing, but synthesis and binding assays were run by two outside labs — not Anthropic — and binding was cross-confirmed by both. Skeptics remain, though. Martin Shkreli, the "Pharma Bro," dismissed the work as "not impressive." Deep-learning tools such as RFdiffusion have already shown strong performance in binder design, so opinions split on whether Claude's contribution is genuinely new science or a capable orchestration of existing tools.

Guardrails and open questions

Anthropic is treating this as a double-edged capability. The company says it still blocks life-science tasks in its most capable model and is preparing a separate access program open only to vetted scientists. Protein design can accelerate drug discovery while also touching on biological misuse concerns, leaving "who gets access, and how far" as a question as important as the technology itself.