Anthropic has published results showing that its Claude AI model designed functional protein binders against 14 of 15 targeted proteins, achieving a hit rate of 22.6 to 35.1 percent that significantly exceeds the typical 10 to 15 percent industry benchmark.
The results, released on August 18, describe a campaign in which Claude, using both its Mythos Preview and Opus 4.8 models, autonomously designed 1,320 protein binder candidates. Of those, 354 were confirmed as working binders across 14 distinct protein targets. Two independent laboratories, Adaptyv Bio and Twist Bioscience, produced and tested the designs in wet-lab conditions without human intervention in the design process.
How Claude Ran the Design Pipeline
The work was conducted inside Claude Science, Anthropic’s research platform integrating standard scientific tools and computing resources. Claude operated in two modes: a multi-target approach where it designed against all 15 proteins in a single session, and an autonomous agent mode in which it orchestrated the full computational design pipeline end to end, selecting tools, running simulations, and iterating on designs without prompts.
Protein binders are small engineered proteins that latch onto specific target proteins. They are not drugs themselves, but represent an early and critical step in therapeutic development. Designing a high-affinity binder traditionally requires weeks or months of expert biochemistry work per target, making Claude’s speed and success rate notable.
What the Results Mean and Where They Fall Short
Anthropic frames the campaign as foundational rather than a breakthrough in drug discovery. Protein minibinders are not a standard therapeutic modality, and even for conventional drug classes such as monoclonal antibodies, a high-affinity binder is only the first stage of a lengthy development pipeline. The company said it plans further characterization to confirm hit rates and measure binding affinities more precisely.
The results are self-reported by Anthropic, though independently validated at the lab level. The company acknowledged that the work carries dual-use implications and said it has taken steps to review potential risks associated with making protein design capabilities broadly accessible.
We are quite hopeful that this release will enable new kinds of research and the creation of new kinds of products.
The protein design work follows Anthropic’s launch of Claude Science earlier this year and its broader push into life sciences applications. The company said it wants Claude to eventually run drug development pipelines end to end across multiple modalities, including antibodies and small molecules.
The release comes amid growing competition between Anthropic, OpenAI, and Google in applying frontier AI models to scientific research, with each company publishing results demonstrating capabilities in biology, chemistry, and materials science.
Sources: Anthropic research blog; CNBCTV18; The Next Web; Adaptyv Bio case study
discussion