The Experiment
Anthropic announced that it tested whether its Claude models could design brand-new protein binders "from scratch," a process known as de novo design. Many drugs work by binding tightly to a specific target in the body to block or change what that target does, and identifying a molecule that binds well has traditionally required weeks or months of specialist work per target, according to Anthropic. Working from a protein design prompt written by a human expert, Claude autonomously designed binders against 14 of 15 targets. Anthropic then had the designs independently built and tested by two outside firms, Adaptyv Bio and Twist Bioscience, so the results would not rely solely on Anthropic's own assessment.
How Claude Ran the Campaign
According to Anthropic's research post announcing the work, the company ran two models, Claude Opus 4.8 and an experimental Mythos Preview, inside an internal research workbench it calls Claude Science. The models were given a detailed protocol, internet access, connectors to specialized tools, and a large compute budget, then largely left to work on their own. Anthropic's post and subsequent reporting from The Next Web describe Claude choosing where on each target protein to bind, orchestrating existing open-source structure- and sequence-design tools, running optimization rounds, and screening candidates for properties like solubility and novelty before ranking 30 designs per target for synthesis. Two of the 15 targets, 15-PGDH and GDF-8, were drawn from recent design competitions specifically so Claude could not lean on pre-recorded solutions in its training data, Anthropic said.
The Results
Anthropic reported that of 1,320 designs with usable lab measurements across the 15 targets, 354 were confirmed binders, an overall hit rate of roughly 27 percent depending on how the campaign was run, compared with a typical field success rate of 10 to 15 percent that Anthropic cited. The company said the results included high-affinity binders against at least six targets and binders that matched or exceeded the best previously reported affinity for at least four targets. Coverage from The Next Web highlighted one standout case: against a target called RBX1, Claude's Mythos Preview model reportedly reached a 40 percent success rate in a single-target run, compared with a 3.7 percent success rate among human entrants in a related design competition. Not every target went smoothly — reporting from Dataconomy and The Next Web noted that none of 90 designs against maltose-binding protein, described as a notoriously difficult target, were confirmed to bind.
Caveats and Context
Anthropic was careful to frame the work as an early step rather than a finished capability. As stated in the original announcement, protein binders are not drugs, and designing a high-affinity binder is only the first stage in a much longer process of developing a safe, effective drug-like molecule. The Next Web noted that the headline figures come from Anthropic itself rather than an independent peer-reviewed source, even though the physical testing was outsourced to Adaptyv Bio and Twist Bioscience. Dataconomy reported that the claims drew public skepticism from Martin Shkreli, a former pharmaceutical executive, who argued on social media that the binding affinities were weak for this class of molecule and that none of the targets were intracellular proteins. Anthropic said in its announcement that launching broader scientist access to its most capable models for life-science work remains a high priority, with more details expected soon.