Skip to content

Anthropic

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

4 min read

Written by an AI model from the linked sources and checked against them by a second model. How we work

Video: x.com

Claude designed protein binders from scratch against 14 of 15 targets, and outside labs built and measured them. Anthropic said on X on August 18 that between 22% and 35% of Claude’s designs bound successfully, depending on the setup, against the 10% to 15% typical in protein design campaigns today. Adaptyv Bio and Twist Bioscience independently built and tested the designs.

A binder is a protein meant to stick to one target and nothing else. Most antibody cancer drugs are binders, and so are the capture reagents in a pregnancy test. Designing one has historically taken a specialist weeks or months per target. Here a human expert wrote the design prompt, and Claude (Mythos Preview and Opus 4.8) designed the binders autonomously.

What the lab measured

Anthropic chose 16 targets from Adaptyv’s previous public competitions, hackathons and BenchBB. Each Claude version was prompted similarly and had access to the same publicly available tools. Adaptyv then ran the designs anonymized: in its write-up the company says it had no information on which Claude model designed which protein. Every sequence went through the same Affinity Characterization assay, which measures binding strength by SPR at five target concentrations with duplicate measurements.

95% of the designs expressed, meaning the proteins were actually produced in the lab. Adaptyv says that three years ago this would have been an impressive headline on its own. It matched the best expression rates from the company’s EGFR competition, which had hundreds of expert protein designers, and surpassed other challenges such as the RBX1 one.

Of 1,320 designs, 354 bound their target, an overall hit rate of 26.8%, with per-target rates varying widely. One target, MBP, yielded no binders. A sixteenth, GDF-8 mature, was excluded because the measurements came out low quality in Adaptyv’s setup, likely because the target was aggregating and binding to other molecules of itself.

Against the competition winners

Adaptyv also set Claude next to its own past competitions, subsetting each one down to de novo minibinders for a fair comparison. Claude surpassed all of their hit rates. On TREM2 it reached 80%, against the 38.3% Adaptyv reported in its competition. On 15-PGDH, a trickier target, the success rate came in more than three-fold higher than what is observed on Proteinbase.

Affinity is reported as K_D, and a lower K_D means a tighter binder. The best Claude binders beat 5 of the 6 competition winners: on 15-PGDH the best result moved from 1.7 µM to 33.4 nM, on RBX1 from 25.7 nM to 3.9 nM. Some of the strongest designs bound several times more tightly than the best previously published result.

Nipah was the exception. Claude did not beat the best binder from that competition, which Adaptyv calls surprising. The company also looked at where on each target the designs aimed: for GDF-8, RBX1 and Nipah, Claude appears to have explored a wide range of possible binding sites, while on TREM2 it converged on single well-defined ones.

Adaptyv’s conclusion is that Claude is at least expert-level at orchestrating protein design tools, which it calls good news because those tools are hard to use. Before AI, the company writes, even setting up a protein structure prediction tool “could take hours of debugging opaque conda errors.”

Not a drug

Anthropic states the limit plainly: protein binders are not drugs. Designing a high-affinity binder is only the first step toward a drug-like molecule, and designing the molecule itself is one phase out of the many required to establish that a drug is safe and effective before it reaches people. Adaptyv’s version of the caveat is shorter. Can AI now solve all diseases? “Well no, not yet.”

What Anthropic says it is building on that foundation is teaching Claude to run the entire development process end-to-end for every major type of drug molecule, from antibodies to small molecules. The prompts and data are open-sourced on Hugging Face alongside a technical report, and Adaptyv is releasing the protein sequences Claude designed together with the experimental data on Proteinbase.

Which model scientists can use

The binder work ran on Mythos Preview and Opus 4.8. For outside researchers, Anthropic says Opus 5 remains its most capable model available for life science research, and that launching an access program for scientists to use its most capable models is one of its highest priorities, with more to come soon; no date has been given.

The same write-up carries a second result on that generally available model. Opus 5 was given a contract lab’s raw NMR and LC-MS files, the data chemists use to check what a compound is and how pure it is, along with a two-sentence prompt. It returned finished results in 23 and 19 minutes, matching the lab’s own analysis on hydrogen counts and on purity: 96.4% against the lab’s 96.33%.

No comments yet. Be the first.

Join the conversation

Sign in to leave a comment. It appears after moderation.