Anthropic let Claude design proteins itself, hitting 14 out of 15 targets—a result that surprised experienced researchers.
On August 18, Anthropic published the results of an experiment: Claude (Mythos Preview and Opus 4.8) autonomously completed an entire protein design workflow.
Given 15 protein targets, it was asked to design new proteins that could bind to them, and it succeeded on 14 targets.
Of 1,320 designs, 354 were validated by two independent labs, for an overall success rate of 26.8%.
The industry average in this field is 10%–15%.
Protein binders are the basis for many drugs: first design a molecule that can grab the target protein, and only then can it potentially become a drug.
This design process typically requires weeks of computation, optimization, and screening by protein engineers.
From AlphaFold to Claude: predicting proteins and designing proteins are two different things
What AlphaFold did in 2020: given a protein sequence, predict what shape it folds into.
The input is known, the output is a prediction.
It is a specially trained protein model that does one thing to perfection.
What Claude did this time is different.
It only received the name of the target protein, and the task was to design a brand-new protein from scratch to bind to it.
The input is a name, the output is a new protein.
One is describing a picture, the other is writing an essay on a given topic.
The tools Claude used were all off-the-shelf—RFdiffusion, ProteinMPNN, ESMFold2—these open-source models for protein design and structure prediction have long existed, and any lab can download them.
Claude did not invent new tools; it did orchestration: the research team wrote a prompt of about 16,000 words, encoding the working knowledge of protein engineers, including the stages of the experiment, the tools available at each stage, and screening criteria.
This prompt does not specify which face of the protein to target, does not specify which generation method to use, and does not presuppose any sequence.
It was handed to Claude, along with a cloud server account, and let it run.
In multi-target mode, one session processed 14 targets in 48 hours; in single-target mode, each target took 24 hours.
Human operators did only three things: approve network access, monitor infrastructure, and send Claude's ranked designs to two independent labs (Adaptyv Bio and Twist Bioscience) for synthesis and testing—no human intervened in any design decision.
Claude itself selected targets, chose epitopes, installed tools, ran models, screened and optimized, ranked and delivered, ultimately invoking 10 structure generation methods and combining 24 tool combinations.
Results
The current average success rate in the protein design field is 10%–15%.
Claude achieved 22%–35% in different modes, two to three times the industry average.
https://x.com/AnthropicAI/status/2089842389682954621
If you only look at Claude's own top-ranked design, the hit rate is 49%: for every two targets, the top-ranked design is directly usable.
The results for several targets deserve special mention.
Adaptyv Bio previously held a public design competition for a protein called RBX1, where global participants submitted 245 designs, and only 9 succeeded.
Claude submitted 90 designs on the same target, 28 succeeded, and the best design bound the target ten times more tightly than the competition champion.
TNFα is a harder target—one of the world's best-selling drugs, Humira, works by binding this protein, but multiple expert teams had previously attempted de novo binder design and all failed.
Opus 4.8 produced 12 effective designs, some of which could cross-species bind human, monkey, and mouse TNFα simultaneously.
There were also failures: on a protein called MBP, all 90 designs failed.
Similar signals appeared in analytical chemistry.
Given Claude Opus 5 raw data from an NMR instrument and a one-sentence instruction, it produced results in 23 minutes, consistent with the manual analysis conclusions of a lab chemist taking half an hour to an hour.
The two experimental directions differ, but the signal is consistent: Claude, a general-purpose model, autonomously produced expert-level results validated by labs in 24 to 48 hours (48 hours in multi-target mode, 24 hours in single-target mode) on research tasks that would take experts weeks.
In just six years, earth-shaking changes
Binders are still far from drugs, with toxicology, clinical trials, and other steps in between, each taking years.
All structures are computational predictions, not experimentally validated.
Claude uses all open-source tools. Anthropic has open-sourced the prompts, data, and all 1,440 design models on HuggingFace, and any lab can reproduce them.
https://huggingface.co/datasets/Anthropic/claude-protein-binder-design/tree/main
This kind of autonomous research capability also carries dual-use risks. Anthropic has blocked biological capabilities such as protein design in the public version of Claude.
From AlphaFold predicting protein structures to Claude autonomously designing proteins, only six years have passed.
References:
https://www.anthropic.com/research/Claude-accelerates-protein-design
https://www-cdn.anthropic.com/30bf50e22a01388bb29bf077ee3f244531594b7a.pdf
This article is from the WeChat public account "新智元", author: ASI启示录
