Claude binders hit 14 of 15 protein targets in the lab
Anthropic just put wet-lab numbers on a claim it has been making for months: a general agent can run a specialist's early-stage biology workflow, not just talk about it.
Anthropic said Tuesday that Claude designed de novo protein binders against 15 targets and produced confirmed binders for 14 of them, with independent labs measuring a hit rate well above the field's usual 10 to 15 percent. The company ran Opus 4.8 and an unreleased Mythos Preview inside Claude Science, then sent the sequences to Adaptyv Bio and Twist Bioscience for synthesis and binding tests. Of 1,320 designs with interpretable data, 354 bound. Mythos Preview hit 26.7 percent when it worked all targets at once in a 48-hour session, and 35.1 percent when it focused on one target at a time. Opus 4.8 landed at 22.6 percent in the multi-target setup. Anthropic says some of the strongest designs bound several times more tightly than the best previously published result on the same target, including a 3.9 nanomolar binder to RBX1 that beat a recent open-competition winner measured on the same plate.
This is not a new protein-generation model. Claude installed and drove the same open-source structure, sequence, and folding tools a computational biologist already uses, then cycled candidates until it had 30 designs per target. Humans picked the target list, approved network and compute access, and placed the synthesis orders. After the first prompt, Anthropic says it gave no scientific guidance. That is the actual claim: a general agent as lab operator, not a secret in-house AlphaProteo. RuntimeWire notes the comparison to human campaigns is messy — results from several of the benchmark competitions were available to Claude during design, and Anthropic did not run a matched expert control with the same tools and budget.
The caveats are load-bearing. A binder is not a drug. None of the proteins was tested for biological function, and no protein-target complex was structurally solved. Hit rates count sequence variants of the same backbone as separate designs; collapsing to the best sequence per backbone drops the rate to 24.7 percent. Claude went 0-for-90 against maltose-binding protein, a slippery, water-loving surface that gives a binder almost nothing to grab. Anthropic is explicit that protein-design skills stay gated in its most capable generally available lineup because the same stack is dual-use, and it says a scientist access program is a top priority. We covered Claude's earlier science work when it lifted the Riemann zeta zero bound to 67.2 percent. Math proofs check themselves. Proteins do not.
A second experiment is less flashy and closer to daily lab work. Given only a contract lab's raw NMR and LC-MS files and a two-sentence prompt, generally available Opus 5 processed the spectra in 23 and 19 minutes, matched the lab's hydrogen counts, and reported 96.4 percent purity against the lab's 96.33 percent. It even proposed the same heavy-water follow-up the lab had already run, then corrected its own first pass when a self-check showed only two of four flagged peaks had vanished. That is the part most chemists will actually use this week. The binder campaign is the part that will get cited in every drug-discovery slide deck, including the ones that skip the sentence about function, structure, and safety.
What to watch: whether the scientist access program ships before competitors publish a matched human-versus-agent campaign on the same targets.
If a general agent can run the first step of binder design, should the wet-lab failure modes be public before the access program opens? Tell us in the comments.
Sources: Anthropic · Anthropic technical report · RuntimeWire · Techmeme
