Claude designs working protein binders for 14 of 15 targets at up to a 35% hit rate
Claude (Mythos Preview and Opus 4.8) designed binders against 15 protein targets and succeeded on 14, with 354 of 1,320 designs binding in wet-lab tests run by outside labs Adaptyv Bio and Twist Bioscience — hit rates of 22.6% to 35.1% against the 10–15% typical of design campaigns. The runs were autonomous, 24 to 48 hours of wall time on up to 12,500 H100 hours, compressing a task that has taken specialists weeks or months per target. Anthropic published the prompts, computational models, and experimental data on Hugging Face, which makes this an externally validated wet-lab result rather than a benchmark claim.
Source: anthropic.com ↗
Between 22% and 35% of its individual designs bound successfully, depending on the setup, compared to the 10-15% that is typical in protein design campaigns today.
Why this matters
- → AI compressed weeks-to-months protein design work into 24–48 hours with 22–35% hit rates (vs. 10–15% baseline)
- → Wet-lab validation from independent labs (Adaptyv Bio, Twist Bioscience) proves autonomous design actually bin
- → Published prompts and data on Hugging Face enable reproducibility and immediate use by life scientists.