How Claude is accelerating protein design and analytical chemistry

An unreviewed summary of a Science blog post reporting two experiments, a largely autonomous protein binder design campaign whose designs were tested by two outside laboratories and a routine chemical characterization task on raw instrument files, alongside the post's own statement that such capabilities are dual-use.

Anthropic (Science team) · 2026-08-18 · original

Not reviewed. This note reports what the post itself says. Unlike the other notes in this collection, its claims have not been checked against the underlying papers, reports or data. The work concerns biological design that could be misused, and in-depth analysis of it triggers the safety safeguards of the AI system used to prepare these notes, so it was left at the level of the post's own summary.

What the post reports

Two experiments, both offered as evidence that AI is beginning to shorten work in fields where verification is slow and expensive. The post sets this against mathematics, where results can be checked quickly and where it says agents have lately been making visible progress.

The first is a protein binder design campaign run largely autonomously. Designing a binder from scratch, the post says, has historically taken a specialist weeks or months per target, and even with machine-learning design tools still takes computational experts days or weeks of orchestration. Claude models (Mythos Preview and Opus 4.8) designed against 15 targets and succeeded on 14, producing 354 binders out of 1,320 designs. Hit rates (the share of individual designs that actually bound) were 26.7% and 22.6% when the models worked on all targets at once in a 48-hour session, and 35.1% when one model worked on each target separately; the post gives 10 to 15% as typical for campaigns today. It claims high-affinity binders against at least six targets and results matching or exceeding the best reported affinity against at least four. For comparison, it notes that the two largest public collections of such designs together hold roughly 770 binders out of 5,700 designs. On one benchmark target where a public competition had been run, the post reports a 40% hit rate against 3.7% among entrants. It also gives counterexamples: two targets remained hard, and against one of them none of the 90 designs was confirmed to bind.

The second experiment is analytical chemistry rather than design: interpreting measurements of a molecule already made. Given only a contract lab's raw instrument files for a routine quality-control sample and a short plain-language prompt, with no vendor software and no operator, a generally available model (Opus 5) returned processed results in 23 and 19 minutes working in parallel. The post says these matched the lab's own processing (hydrogen counts within 0.08 of the lab's, purity 96.4% against the lab's 96.33%), where a chemist typically spends half an hour to an hour per sample and the lab's finished report arrived four days after the first measurement. One instrument's file format was undocumented, and the post says Claude worked out the encoding and then checked its reading against the instrument's own recorded totals before analyzing anything. It notes Claude returned its own list of caveats about how far the results could be trusted, and that it caught and corrected an overstatement in its first reading of one measurement.

Who did the work and who checked it

Anthropic's Science team publishes the post, with two technical reports linked and prompts, computational models and experimental data released publicly. The laboratory work was done by two outside contract organizations, Adaptyv Bio and Twist Bioscience, which the post says independently produced and tested the designs; the competition comparisons use Adaptyv Bio's published results. No peer review is claimed. The post discloses that its own models are the subject, and that life science research tasks are blocked in its most capable model pending an access program for scientists.

Caveats the post itself gives

The post says it intends more extensive characterization to confirm its own hit rates and affinity measurements. Results for one of 16 targets were dropped as inconclusive. It reports being unsure why one model succeeded on a hard target where another failed, and says model capability should be assessed holistically rather than from single targets. Human involvement was not zero, and the post lists approvals and infrastructure handling in a footnote; it expects expert designers giving active guidance would do better still. The chemistry result is a single routine sample. Most pointedly, a section headed "Agentic biological discovery is dual-use" states that these capabilities could, without robust safety measures, help bad actors perform dangerous research, and that protein design and related capabilities remain unavailable in general access.

How it connects

References