Skip to content
NewsResearch

Claude designed protein binders against 14 of 15 drug targets in independent lab tests, at hit rates roughly double the industry average

· by Pondero Newsdesk

The short version

Anthropic published research on August 18, 2026, showing Claude autonomously ran a protein design campaign that produced 354 confirmed binders from 1,320 designs, a 26.8 percent hit rate versus the 10 to 15 percent typical in professional campaigns today.

Claude designed protein binders against 14 of 15 drug targets in independent lab tests, at hit rates roughly double the industry average

Two independent labs physically tested Claude's protein designs last week and found it outperformed the winners of professional design competitions on the same targets. The result, published by Anthropic on August 18, marks the first time a commercial AI model has completed a multi-target protein binder campaign end-to-end with wet-lab verification by an outside party.

What happened

Claude Mythos Preview and Opus 4.8 designed protein binders against 15 targets inside Anthropic's Claude Science environment over a 48-hour session, with no scientific guidance after the initial prompt, per Anthropic's research post. Adaptyv Bio and Twist Bioscience then independently manufactured the proteins and ran binding experiments in automated wet labs. Of 1,320 submitted designs, 354 bound their target. That works out to a 26.8 percent hit rate, per Adaptyv Bio's case study, compared to 10 to 15 percent typical in professional protein design campaigns today.

Mythos Preview running in a separate single-target mode, with 24 hours per protein rather than all at once, reached 35.1 percent overall. On RBX1, a protein involved in targeted protein degradation, it achieved a 40 percent hit rate against a 3.7 percent rate recorded by participants in an earlier open competition on the same target, per Anthropic. Adaptyv Bio reported that Claude's designs would have won 5 of the 6 competitions the company has run, producing tighter binders than the human winners on those targets.

Two targets proved difficult. Against BBF-14, a computationally designed beta-barrel protein with no natural equivalent, Claude still produced three confirmed binders. Against maltose-binding protein, none of the 90 designs bound.

A second experiment, run with the publicly available Claude Opus 5 model, put Claude on a routine analytical chemistry task: processing raw NMR and LC-MS instrument files for a quality-control sample with a two-sentence plain-language prompt, per the same research post. Claude returned finished results in 23 minutes for NMR and 19 minutes for LC-MS, and its purity measurement of 96.4 percent matched the contract lab's own reading of 96.33 percent. The lab's report for the same sample arrived four days after the first spectrum was acquired.

Anthropic published all 1,320 design sequences, the full binding dataset, and the prompts used to the Hugging Face dataset "claude-protein-binder-design." The sequences also appear on Adaptyv Bio's Proteinbase open protein data platform.

Why it matters

Protein binder design is a foundational step in drug discovery, and the timeline matters: a task that previously occupied specialist teams for weeks ran in 24 to 48 hours with Claude as the sole operator. The chemistry result adds a second data point in a different domain, with a routine analytical step that chemists spend 30 minutes to an hour per sample on completed in under 25 minutes.

For operators tracking how far current frontier models reach into lab science, the wet-lab verification is the critical detail. Earlier AI protein design claims relied on computational scoring alone. Here, an independent facility manufactured and tested the physical proteins, so the hit-rate numbers reflect binding in actual biological conditions.

Protein design and related biology capabilities remain unavailable for general access in Claude Fable 5, per Anthropic's safety policies for dual-use biological research. Opus 5, which ran the chemistry task, is publicly available. Mythos Preview and Opus 4.8 were used under controlled access for this research.

What to watch next

Anthropic said it expects to share details soon on a scientist access program for life researchers who need Mythos 5 for biology tasks. A second round of characterization is also planned to confirm hit rates and affinity measurements from this campaign. A full technical report with methodology details is linked from Anthropic's research post.

Sources