Anthropic says any lab can now let a language model agent run the whole protein design stack

Details

Anthropic published results from two AI-in-drug-discovery experiments where Claude models (Mythos Preview and Opus 4.8) operated as autonomous agents in early-stage drug development. The primary experiment targeted 15 proteins for de novo minibinder design — small proteins engineered to bind and block specific molecular targets, a foundational step in drug development. Of 1,320 candidate binders tested in the lab, 354 bound to their target (26.8%). Claude's top-ranked designs succeeded at 49%. In multi-target mode, handling all 15 targets simultaneously within 48 hours, hit rates reached 26.7% (Mythos Preview) and 22.6% (Opus 4.8). The highest rate — 35.1% — was achieved with single-target mode and 2.8× more compute per target. Crucially, Claude did not build new biology tools: it installed and orchestrated existing open-source software including RFdiffusion, ProteinMPNN, ESMFold2, PXDesign, and others, guided by a 16,000-word protocol prompt. About two-thirds of that prompt covered scheduling, delegation, verification, and compute budgeting — not scientific content. AlphaFold-3 and Rosetta were excluded for licensing reasons. The practical implication is significant: a general-purpose language model orchestrating existing tools can now match or exceed specialist performance on at least some drug-discovery tasks. Anthropic explicitly states the workflow is reproducible by any lab with access to the same open-source tools. Independent review is still pending.