Anthropic had Claude design proteins itself, hitting 14 out of 15 targets—a result that amazed researchers in the field.
On August 18, Anthropic released an experimental result: Claude (Mythos Preview and Opus 4.8) autonomously completed an entire protein design workflow.
Given 15 protein targets, it was asked to independently design new proteins to bind to them, succeeding for 14 targets.
Out of 1320 designs, 354 were validated as effective by two independent labs, with an overall success rate of 26.8%.
The industry average in this field is 10%–15%.
Protein binders are the foundation for how many drugs work: first, design a molecule that can grab the target protein, then it might become a drug.
This design process typically requires protein engineers weeks of computation, optimization, and screening.
From AlphaFold to Claude: Predicting Proteins and Designing Proteins Are Two Different Things
What AlphaFold did in 2020: given a protein sequence, predict the shape it folds into.
The input is known, the output is a prediction.
It is a specially trained protein model, doing one thing extremely well.
What Claude did this time is different.
It was only given the name of a target protein, with the task of designing 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 from a prompt.
The tools Claude used were all off-the-shelf—RFdiffusion, ProteinMPNN, ESMFold2. These open-source models for protein design and structure prediction already exist and can be downloaded by any lab.
Claude didn't invent new tools; what it did was orchestration: the research team wrote a prompt of approximately 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 did not specify which face of the protein to target, which generation method to use, or preset any sequences.
It was given to Claude, along with a cloud server account, and left to run.
In multi-target mode, a single 48-hour Session handled 14 targets simultaneously; in single-target mode, each target took 24 hours.
Human operators only did three things: approved network access, monitored infrastructure, and sent the designs prioritized by Claude to two independent labs (Adaptyv Bio and Twist Bioscience) for synthesis and testing—no one intervened in any design decisions.
Claude autonomously selected targets, selected epitopes, installed tools, ran models, screened and optimized, prioritized deliverables, and ultimately utilized 10 different structure generation methods, combining them into 24 tool combinations.
Results
The current average success rate in the protein design field is 10%–15%.
Claude, in different modes, achieved 22%–35%, two to three times the industry average.
https://x.com/AnthropicAI/status/2089842389682954621
If we look only at Claude's own top-ranked design, the hit rate was 49%: for every two targets, the design ranked first was directly usable for one.
Results for several targets are worth mentioning separately.
Adaptyv Bio previously hosted an open design competition for a protein called RBX1; global participants submitted 245 designs, only 9 of which were successful.
Claude submitted 90 designs for the same target, with 28 successes; its best design bound the target ten times more tightly than the competition champion.
TNFα is an even harder target—Humira, one of the world's best-selling drugs, works by binding to this protein, but multiple expert teams had previously attempted *de novo* binder design, all failing.
Opus 4.8 produced 12 effective designs, some of which could even cross-species binding to human, monkey, and mouse TNFα simultaneously.
There were also failures: for a protein called MBP, all 90 designs failed.
Similar signals exist in analytical chemistry.
Given raw data from an NMR spectrometer and a one-sentence instruction, Claude Opus 5 produced results in 23 minutes, consistent with conclusions from a lab chemist's manual analysis taking half an hour to an hour.
The two experimental directions are different, but the signal is consistent: this generalist model, Claude, autonomously ran and produced lab-validated, expert-level results in 24 to 48 hours (48 hours in multi-target mode, 24 hours in single-target mode) for research tasks that would take experts weeks.
Just Six Years, A World Transformed
Binders are still far from being drugs, with steps like toxicology and clinical trials in between, each taking years.
All structures are computational predictions, not yet verified by experimental structural biology.
Claude used only open-source tools; Anthropic has open-sourced the prompt, data, and all 1440 design models on HuggingFace, allowing any lab to reproduce the work.
https://huggingface.co/datasets/Anthropic/claude-protein-binder-design/tree/main
Such autonomous research capabilities carry dual-use risks; Anthropic has blocked protein design and other biological capabilities in the public version of Claude.
Only six years have passed from AlphaFold predicting protein structures to Claude autonomously designing proteins.
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" (New Wisdom), author: ASI启示录
