Anthropic reports Claude designed working protein binders

Anthropic reported that its models, prompted to design proteins that bind to a chosen target, produced working “minibinders” for 14 of 15 lab-validated targets, with the wet-lab confirmation carried out by outside partners Adaptyv Bio and Twist Bioscience. The company said this exceeded the roughly 10–15% hit rate typical of industry design campaigns, though the figures varied considerably by model and setting: Claude Mythos Preview, an experimental model, reached a 35.1% hit rate when designing against one target at a time, and lower rates — 26.7% for Mythos Preview and 22.6% for Claude Opus 4.8 — when handling several targets at once.

In a separate demonstration, Anthropic said Claude Opus 5 parsed raw output files from laboratory instruments — nuclear magnetic resonance and liquid chromatography–mass spectrometry data — converting undocumented vendor formats into calibrated spectra without being told how the files were encoded, completing both analyses within 25 minutes.

Anthropic was explicit about the limits of the work. The models struggled against some targets, including BBF-14 and maltose-binding protein, and the company noted that a minibinder that sticks to its target is only an early step toward a usable drug or reagent — affinity, specificity and manufacturability all still need separate work. The company also flagged the dual-use risk inherent in a model that can design binding proteins to order, saying the underlying capability is restricted rather than exposed in its general-access products.

The results extend a run of AI-for-science claims from the major labs, following Google DeepMind’s Co-Scientist reaching Nature publication earlier in the year, and add a biology capability to the cyber-offense concerns Anthropic had already raised about Mythos when it withheld the model from public release in April.