Claude Designed Working Protein Binders for 14 of 15 Targets

Anthropic said on Tuesday that its Claude models designed working protein binders for 14 of 15 biological targets in a controlled experiment, a result the company offers as early evidence that a general AI system can take on work once reserved for specialist tools in a drug-discovery lab.

A protein binder is a molecule shaped to latch onto a specific target in the body and change what it does. Designing one that binds tightly is usually the opening move in developing a drug, and it can take a trained scientist weeks or months for a single target. In the work Anthropic published, Claude was handed a prompt written by a human expert and asked to produce 30 candidate binders for each of the 15 targets, driving publicly available protein-design and co-folding models on its own.

The designs were not graded on paper. Two outside labs, Adaptyv Bio and Twist Bioscience, synthesised the candidates and tested them at the bench. Between 22 and 35 percent of Claude's designs bound to their targets, depending on the setup, against a typical industry hit rate of 10 to 15 percent. The company reported high-affinity binders for at least six targets, and binders that matched or beat the best previously reported affinity for at least four.

Why the framing matters

Specialist models for protein design already exist and are very good. Anthropic's claim is a different one: that a general-purpose model, from the same family that writes code and drafts email, can orchestrate those specialist tools end to end and produce results a working scientist would take seriously. The runs used Claude Opus 4.8 and an internal Mythos preview.

Not everyone is convinced. Martin Shkreli, the former pharmaceutical executive, publicly dismissed the results as "not impressive work," arguing that generating a binder is the easy part of drug development and that the real difficulty comes later. He has a point worth holding onto. A molecule that sticks to its target in a dish is a long way from a medicine, and we take that reality check apart separately in our Myths and Fears section.

The result also lands in a delicate spot for Anthropic, which has spent much of the year tightening and then re-tuning the biology safeguards on its models. Only last week the company loosened Fable 5's biology filters to cut false alarms while keeping the hard limits in place. Protein design is exactly the sort of dual-use capability those guardrails exist to police, because the same skill that speeds a cancer therapy could, in the wrong hands, speed something far worse.

For now the claim is narrow and, unusually for this beat, checked by independent wet-lab work rather than a self-reported benchmark. That is the part I find most persuasive. Whether it holds up beyond 15 targets is the question the next round of experiments will have to settle.