Bottom line: In tests, Claude models achieved higher hit rates in de novo protein binder design than typical campaigns and evaluated NMR/LC-MS analytical data in under 25 minutes with lab-comparable accuracy.
Anthropic demonstrates in two experiments that Claude models can both design de novo protein binders and independently evaluate NMR and LC-MS data. For practitioners, this points to a potential shortening of early drug discovery phases that have so far required weeks to months of specialist work.
In the first experiment, Anthropic had the models Claude Mythos Preview and Claude Opus 4.8 design protein binders against 15 different target proteins. At least one successful hit was achieved for 14 of the 15 targets. Depending on the experimental setup, between 22 and 35 percent of individual designs successfully bound to their target — current protein design campaigns typically achieve hit rates of 10 to 15 percent. Some of Claude’s strongest designs achieved binding affinities several times higher than the best previously published results for the respective targets. De novo design of minibinders — small proteins engineered to dock onto a target protein — is regarded as a key early step in drug discovery and has historically cost protein engineers months of computation, optimization, and screening per target.
In the second experiment, the generally available model Claude Opus 5 was provided with raw NMR and LC-MS data from a contract lab along with a two-sentence prompt — with no further preprocessing. Claude delivered finished analysis results within 23 and 19 minutes, respectively, that matched the lab analysis: for purity determination, Claude arrived at 96.4 percent versus 96.33 percent in the lab report, and the hydrogen counts matched.
For practitioners in life-science organizations, this points to a possible shift of computationally intensive, specialized tasks toward generally accessible models. Anthropic frames both examples as part of a broader trend in which AI models are increasingly making progress in experimental fields with more elaborate verification requirements — in contrast to, say, mathematics, where verification is fast and cheap. As reference points, Anthropic cites unsolved Erdős problems, which are currently being resolved at a rate of several per month, as well as a recently improved lower bound for the Riemann zeta function.
Important context: the protein design results were achieved using a combination of Mythos and Opus models, with life-science research tasks currently still locked in the most capable model. Anthropic announced that it is preparing a prioritized access program for scientists and will publish further details on this soon. Until then, Opus 5 remains the most capable generally available model. Anthropic also emphasizes that accelerating early development phases addresses only one aspect of drug discovery — many bottlenecks in the overall process are more regulatory and operational in nature than solvable through model capabilities alone.
Source: www.anthropic.com · Published August 18, 2026
Lumi AI News — AI-assisted curation pursuant to Art. 50 EU AI Act. Paraphrasing and classification by Lumi News Pipeline v1.8.3.