Claude designs novel protein binders autonomously, hitting 14 of 15 targets at 2-3x industry success rate
新智元 · wechat · 2026-08-19
Anthropic reported on Aug 18 that Claude (MythosPreview and Opus4.8) autonomously designed novel protein binders given only target names, succeeding on 14 of 15 targets. Of 1,320 designs, 354 were validated by two independent labs — a 26.8% success rate versus an industry average of 10–15%; counting only Claude's top-ranked designs, the hit rate reaches 49%.
How it worked: Claude invented no new tools but orchestrated existing open-source models (RFdiffusion, ProteinMPNN, ESMFold2) guided by a 16,000-word prompt encoding protein-engineer expertise. Multi-target mode ran 48 hours, single-target 24 hours; humans only approved network access, monitored infra, and shipped designs for synthesis — Claude chose epitopes, ran models, filtered and ranked, combining 10 generation methods into 24 tool combos.
Standout targets: On RBX1, a public contest drew 245 designs with only 9 successes; Claude got 28 of 90 validated, with best binding ten times tighter than the contest winner. On TNFα (Humira's target, where expert teams had failed), it produced 12 valid designs, some cross-species; on MBP all 90 failed. Separately, Opus5 matched a chemist's NMR analysis in 23 minutes.
Anthropic open-sourced the prompts, data and all 1,440 designs on HuggingFace; such bio capabilities are blocked in public Claude. Binders remain far from drugs, and all structures are computationally predicted.
Related event: Anthropic Says Claude Designed Protein Binders for 14 of 15 Targets(45 posts)→
More from Research
- Interp researcher pushes back: if you can't simulate a neuron, you haven't explained it — nabla_theta · 2026-10-16
- MLP-Splatting (ECCV 2026) code released: MLPs as light-field primitives instead of Gaussians — AjdDavison · 2026-10-16
- Four-page solution to subcubic APSP builds intuition for Anthropic-AI-aided breakthrough — gautamcgoel · 2026-10-16
- ATAS: Zhejiang-led team builds world-state memory framework for consistent agentic video generation — CSProfKGD · 2026-10-16
- Hermes: a 4B model with 8K context got worse with more test-time compute, pointing to contextual reasoning — mangahomanga · 2026-10-16
- Microsoft Research unveils Magnetar, the fastest zkVM at 10-39x speedups and post-quantum security — jedisct1 · 2026-10-16