Anthropic had Claude design proteins itself—14 out of 15 targets hit—a result that surprised expert researchers.
On August 18, Anthropic announced experimental results showing that Claude (Mythos Preview and Opus 4.8) autonomously completed an entire protein design workflow.
Give it 15 protein targets, and it designed new proteins to bind to them on its own—14 targets succeeded.
Out of 1,320 designs, 354 were validated as effective by two independent laboratories, resulting in an overall success rate of 26.8%.
The industry average in this field is 10%–15%.
Protein binders form the foundation of many drugs: first, design a molecule that can bind to the target protein, then it becomes possible to develop it into a medication.
This design process typically requires protein engineers several 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.
Input is known; output is predicted.
It is a specially trained protein model that excels at one specific task.
But what Claude did this time is different.
It only receives the name of the target protein and is tasked with designing a completely new protein from scratch to bind to it.
Input is a name, output is a new protein.
One is describing what you see in a picture; the other is writing an essay on a given topic.
Claude uses readily available tools—RFdiffusion, ProteinMPNN, and ESMFold2—open-source models for protein design and structure prediction that have long existed and can be downloaded by any lab.
Claude did not invent new tools; instead, it orchestrated: the research team wrote a prompt of approximately 16,000 words, embedding the expertise of protein engineers, including the stages of experiments, the tools available at each stage, and the screening criteria.
This prompt does not specify which face of the protein to target, does not prescribe any generation method, and makes no assumptions about the sequence.
Give it to Claude, provide a cloud server account, and let it run.
In multi-target mode, one session lasts 48 hours and processes 14 targets simultaneously; in single-target mode, each target takes 24 hours.
Human operators performed only three tasks: approving network access, monitoring infrastructure, and sending Claude’s well-organized designs to two independent laboratories (Adaptyv Bio and Twist Bioscience) for synthesis and testing—no one intervened in any design decisions.
Claude independently selected targets and epitopes, assembled tools, ran models, screened and optimized results, and ranked deliverables, ultimately invoking 10 structural generation methods to create 24 tool combinations.
Result
The current average success rate in the field of protein design is 10%–15%.
Claude achieves 22%–35% across different modes, two to three times the industry average.
https://x.com/AnthropicAI/status/2089842389682954621
If you look only at Claude’s highest-ranked design, the hit rate is 49%: one out of every two targets has a top-ranked design that is directly usable.
The results for several targets are worth mentioning individually.
Adaptyv Bio previously held an open design competition for a protein called RBX1, receiving 245 submissions from participants worldwide, of which only 9 were successful.
Claude submitted 90 designs on the same target, 28 of which succeeded; the best design bound to the target ten times more tightly than the competition winner.
TNFα is a more challenging target—Humira, one of the best-selling drugs worldwide, works by binding to this protein, but multiple expert teams previously failed in attempts to de novo design binders.
Opus 4.8 generated 12 effective designs, some of which can simultaneously target TNFα in humans, monkeys, and mice.
There was also a failure: all 90 designs on a protein called MBP fell through.
There are similar signals in the field of analytical chemistry.
Given raw data from an MRI machine and a one-sentence instruction, Claude Opus 5 delivered results in 23 minutes, matching the manual analysis conclusions of laboratory chemists, which typically take 30 minutes to an hour.
Two experimental approaches differed, yet yielded consistent signals: The general-purpose Claude model autonomously generated expert-level results, validated by laboratory testing, within 24 to 48 hours—48 hours in multi-target mode and 24 hours in single-target mode—tasks that typically require experts weeks to complete.
In just six years, everything has changed dramatically.
The binder is still far from the drug, with steps such as toxicology and clinical trials in between, each taking several years.
All structures are computational predictions and have not been experimentally validated.
Claude uses only open-source tools, and Anthropic has open-sourced the prompts, data, and all 1,440 design models on Hugging Face, allowing any lab to reproduce them.
https://huggingface.co/datasets/Anthropic/claude-protein-binder-design/tree/main
This dual-use risk associated with autonomous research capabilities led Anthropic to block biological functions such as protein design in the public version of Claude.
It has only been six years since AlphaFold predicted protein structures to Claude autonomously designing proteins.
Reference materials:
https://www.anthropic.com/research/Claude-accelerates-protein-design
https://www-cdn.anthropic.com/30bf50e22a01388bb29bf077ee3f244531594b7a.pdf
This article is from the WeChat public account "New Intelligence Yuan," authored by ASI Revelation.