Claude speeds up protein design with AI

Anthropic says Claude designed proteins capable of binding to 14 of 15 targets in laboratory tests and analyzed chemical data in minutes, with results almost identical to those from a specialized lab. The company presented both cases as examples of how AI can reduce weeks of scientific work to hours.

Designing proteins without starting from scratch

One of the tests involved creating minibinders: small proteins designed to attach strongly to another protein. This type of binding matters in drug development because it can block a signal, activate a process, or transport a substance within the body.

Designing these molecules usually takes weeks or months of simulations, adjustments, and candidate selection. Anthropic had Claude Mythos Preview and Claude Opus 4.8 coordinate specialized models to generate structures, design sequences, check how they folded, and filter the most promising candidates.

The designs were produced and tested by the external evaluators Adaptyv Bio and Twist Bioscience. In total, Claude generated 1,320 designs, and the experiments confirmed 354 proteins capable of binding to targets across 14 of the 15 cases studied.

The success rate, meaning the proportion of designs that actually bound to their target, ranged from 22.6% to 35.1%, depending on the model and how the work was organized. Anthropic puts the typical rate for this type of campaign at between 10% and 15%.

In one test, Mythos Preview achieved a 40% success rate against RBX1, compared with 3.7% for participants in an earlier competition. The result was not uniform, however: the model did not achieve confirmed binding to MBP and struggled with other complex targets.

Not all the designs were better than existing ones, either. Anthropic still plans to carry out more extensive characterization to confirm the success rates and affinity measurements, which indicate how strongly a protein binds to its target.

Analyzing compounds without specialized software

The second experiment focused on a more routine task: checking which compound a chemist has produced and how pure it is.

Anthropic gave Claude Opus 5 raw files from two commonly used techniques:

• NMR, which helps identify a molecule's structure by observing signals from its hydrogen atoms.
• LC-MS, which separates the components of a sample and measures their mass to estimate its composition and purity.

The model received only the files from a contracted lab and a two-sentence instruction. Working in parallel, it returned the NMR analysis in 23 minutes and the LC-MS analysis in 19 minutes.

Its results matched those from the lab on the main points. Hydrogen counts per signal were less than 0.08 units away from the lab's values, and the calculated purity was 96.4%, compared with 96.33% obtained by the specialists.

Claude also detected how a manufacturer's undocumented file format was encoded. Before analyzing it, the model checked that it could reproduce the totals recorded by the instrument across its 2,664 measurements. It also generated a report, charts, tables, and reusable code for processing similar files.

This does not mean the chemist can stop reviewing the results. The model itself pointed out limitations, such as the fact that the instrument measures mass with an approximate precision of one unit. It also initially made an error when interpreting a test with heavy water, but detected it through a follow-up check and corrected its conclusion.

What changes for you

In practice, these advances will affect laboratories and research teams first. AI can handle some repetitive work, run specialized tools, and deliver an initial report without requiring an expert to intervene at every step.

That can speed up the selection of candidate molecules and free scientists to focus on tasks that require judgment, such as deciding which experiment to run next or checking whether a result is reliable. It does not automatically turn a designed protein into a drug: more testing is still needed for safety, efficacy, production, and use in living organisms.

Anthropic also acknowledges that protein design has dual-use implications. The same capabilities that can help develop therapies could enable dangerous research if offered without safeguards. That is why the advanced biological design features are not available to the general public, and the company has discussed creating an access program for scientists.

The important point is not that Claude has replaced a laboratory, but that it can already connect several stages of scientific work: reading data, operating specialized models, proposing checks, and preparing results. The next thing to watch is whether these demonstrations hold up across more targets, more samples, and independent validations beyond the experiments selected by the company itself.