Claude autonomously designs functional protein binders

Across 15 targets, Claude-run campaigns produced binders for 14, drawing from 1,320 total designs of which 354 were experimentally confirmed as binders - an overall hit rate near 26.8%, well above the 10-15% typical for expert-led campaigns ^[1][2]. Hit rates varied by setup: Mythos Preview reached 26.7% and Opus 4.8 reached 22.6% in multi-target mode, while Mythos Preview climbed to 35.1% in single-target mode with a larger compute budget ^[1][2]. The starkest data point is on the RBX1 target, where Claude's best design bound roughly 10x tighter than the winning entry from a public design competition (3.9 nM versus 45 nM), and Claude's own hit rate on that target (40%) dwarfed the roughly 3.7% rate among competition participants ^[1][2]. Anthropic also reported cross-species binding - 130 of 233 designs also bound the mouse version of their target - and said Claude's top-ranked pick within each campaign succeeded 49% of the time ^[2].

The campaigns ran against a roughly 16,000-word protocol written by a human expert, with Claude then working autonomously - but what it did with that protocol matters for how the result should be read. Rather than inventing a new protein-design method, Claude directed roughly a dozen existing, publicly available open-source specialist models (tools such as PXDesign, RFdiffusion3, SolubleMPNN and ESMFold2) through their computational pipelines ^[1][2]. Anthropic's own framing emphasizes autonomy - a workflow that historically took a specialist weeks to months per target compressed into an unsupervised run ^[1]- while credit for the underlying protein-design science still belongs to the teams that built those specialist tools. Anthropic said it is open-sourcing the prompts, data and design models involved so other labs can reproduce or extend the approach ^[2].

Separate from the binder campaigns, Anthropic handed Claude Opus 5 nothing but a contract lab's raw NMR and LC-MS instrument files and a short plain-language prompt. Working the two analyses in parallel, Claude returned finished results in 23 minutes for the NMR data and 19 minutes for the LC-MS data, matching the lab's own conclusions - hydrogen counts within 0.08 ppm and purity measured at 96.4% versus the lab's 96.33% ^[1]. For the LC-MS file, Claude could not find a suitable reader program, so it decoded the proprietary file format itself and exactly reproduced the instrument's stored summary values across all 2,664 measurement points ^[1]. Unlike the binder-design hit rates, this result needs no independent-lab caveat about generalization - it is a direct, reproducible speed and accuracy comparison against a human analyst's own numbers.

Coverage of the results flagged two structural limits: there was no parallel campaign run by human experts on the same 15 targets to serve as a control, and none of Claude's binding models have been structurally resolved - every prediction shown is computational, not crystallographic ^[2]. The results themselves have not been through independent peer review; they come from Anthropic's own write-up, corroborated by the two contract labs that ran the physical testing ^[1][2]. Anthropic itself is careful to note that a protein binder is not a drug - designing one is only the first step toward a drug-like molecule, with the harder, slower stages of drug development still ahead ^[1]. Pharma industry critic Martin Shkreli went further, calling the results unimpressive and arguing that existing monoclonal antibodies already serve the same extracellular-probe function the binders demonstrated ^[3].