Anthropic: Claude designs protein binders for 14 of 15 targets

What did Anthropic show? Claude autonomously designed protein binders — small proteins that dock onto target proteins — and hit 14 of 15 targets. The designs were not just simulated: partners Adaptyv Bio and Twist Bioscience produced and tested them independently in wet labs, as Anthropic documents in its research report.

How good are the results? Hit rates reached 22.6 to 26.7 percent in multi-target mode and up to 35.1 percent in single-target mode — against an industry baseline of 10 to 15 percent. In total, 354 working binders emerged from 1,320 designs. Against the RBX1 target the model achieved a 40 percent hit rate, versus 3.7 percent among competition participants.

How did Claude work? The model orchestrated open-source protein design and folding models, ran optimisation cycles and screened candidates — largely autonomously, using up to 12,500 Nvidia H100 hours across 48 hours of wall time. In a second test, the generally available Claude Opus 5 processed raw NMR and LC-MS instrument data in 23 and 19 minutes respectively, determining purity at 96.4 versus the lab's 96.33 percent.

Where are the limits? Anthropic remains cautious: "protein design and other dual-use research biology capabilities remain unavailable for general access in Claude Fable 5", the report states — biology capabilities with misuse potential stay gated. A controlled access programme for scientists is announced. Critics such as Martin Shkreli argue the results are less impressive than presented.

Why does it matter? Protein binders are building blocks for diagnostics and drug development — from cancer therapies to longevity research. If AI models double hit rates and compress analysis times from days to minutes, biotech's bottleneck shifts from design to the lab. For diagnostics and supplement companies, personalised protein research moves closer to the mass market.