Claude's protein binders hit 14 of 15 targets in wet lab

Claude designed novel protein binders from scratch against 14 of 15 targets, with 22% to 35% of individual designs binding successfully depending on the setup, against the 10-15% typical of protein design campaigns today. A human expert wrote the design prompt, and Claude ran the rest on its own, Anthropic said on X.

At a glance

• The designs were built and tested independently by Adaptyv Bio and Twist Bioscience, with Adaptyv Bio using cell-free protein synthesis on automated workcells and measuring binding strength by SPR.
• Of 1,320 submitted designs, 95% expressed and 354 bound their target, an overall hit rate of 26.8%, according to Adaptyv Bio; one target, MBP, produced no binders at all.
• Anthropic says protein binders are not drugs and that a high-affinity binder is only the first step, and it is now teaching Claude to run development end-to-end from antibodies to small molecules.

Binder design is the narrow front end of drug discovery, but it is the part that has historically consumed weeks or months of specialist time per target, and it is now the part an agent appears able to run at expert level. The result reads less as a therapeutic breakthrough than as evidence that orchestrating existing design tools, long the bottleneck for anyone outside the field, is becoming automatable. The wet-lab loop remains the slow step.

Claude's best binders beat five of six competition winners on affinity

Adaptyv Bio compared the designs against the winners of its own public protein design competitions, subsetting each competition to de novo minibinders for a fair comparison. Claude's strongest 15-PGDH binder improved on the competition's best from 1.7 uM to 33.4 nM, and its best RBX1 binder went from 25.7 nM to 3.9 nM.

The exception was the Nipah virus competition, where Claude did not beat the winning binder. On hit rates, Adaptyv Bio reported 80% for TREM2 against the 38.3% recorded in its own competition, and more than a three-fold improvement over Proteinbase on the harder 15-PGDH target. According to Adaptyv Bio, Claude explored a wide range of epitopes on GDF-8, RBX1 and Nipah, while converging on a single well-defined epitope for TREM2.

One of the 16 targets was dropped after it aggregated in the assay

Anthropic picked 16 targets from Adaptyv Bio's earlier public competitions, hackathons and its BenchBB set, and benchmarked Claude Mythos Preview and Opus 4.8 through Claude Science, with similar prompts and the same publicly available tools. GDF-8 mature was excluded because of low-quality measurements, likely caused by the target binding to copies of itself.

Adaptyv Bio received the designs anonymized, without knowing which model produced which sequence, and ran its affinity characterization assay at five target concentrations in duplicate, which it says keeps the results within its quality control standards. The sequences and the experimental data are being released on Proteinbase, the open protein data platform.

Opus 5 matched a contract lab's purity figure, 96.4% versus 96.33%

In a second experiment, Anthropic gave Claude Opus 5, a generally available model, raw NMR and LC-MS files from a contract lab and a two-sentence prompt. Those are the measurements chemists use to check the identity and purity of a compound. The model returned finished analyses in 23 and 19 minutes, matching the lab's hydrogen counts and its 96.33% purity figure with 96.4%.

Anthropic has also published a technical report on the work and open-sourced its prompts and data on Hugging Face. Adaptyv Bio says the 95% expression rate matched the best rates from its EGFR competition, which drew hundreds of expert protein designers, and surpassed other challenges such as the RBX1 one.

When the scientist access program opens

Anthropic calls launching an access program for scientists to use its most capable models one of its highest priorities and says it expects to share more soon. No timing has been given, and Opus 5 remains the most capable model it offers for life science research; the designs in the binder experiment came from Claude Mythos Preview and Opus 4.8.