Anthropic Claude designs proteins for 14 of 15 targets

Anthropic has delivered protein design work generated entirely by its AI model Claude to two external laboratories for physical manufacturing and testing. The results show that out of 15 target proteins, Claude successfully designed functional binders for 14 targets. The success rate for individual designs ranged between 22% and 35%, significantly outperforming the industry standard of roughly 10% to 15%.

This disclosure is part of two recent experiments conducted by Anthropic to measure how effectively AI can accelerate life sciences research. In a separate experiment, raw data files from a chemical analysis instrument were given to Claude Opus 5 with just a two-sentence prompt, prompting the model to return analysis results matching professional lab standards within 23 minutes and 19 minutes.

The task assigned to Claude was to design Minibinders, which are small proteins engineered to bind tightly to specific target proteins. This binding mechanism forms the foundation of numerous modern therapeutics, enabling drug molecules to inhibit, activate, or deliver payloads to specific targets within the body.

Stock photo for illustration only, not from the actual event

Historically, performing De Novo Design—creating entirely new binders without copying nature—required protein engineers months of effort per target, involving extensive calculation, refinement, and screening. Although machine learning models have recently assisted in generating and ranking candidate designs, executing those tools still demanded human computational experts to coordinate workflows over days or weeks.

In this project, Anthropic placed Claude in control of the entire pipeline through Claude Science, a dedicated research workspace. The model autonomously selected binding sites on target proteins, generated structural and amino acid sequence candidates, directed specialized models for structure and sequence design alongside co-folding prediction tools, and iteratively filtered options to yield novel, water-soluble, and effective binders. Human involvement was strictly limited to approving network access requests and ordering final designs for synthesis.

The experiment was split into two modes. The first mode designed all targets concurrently within a single 48-hour session using up to 12,500 H100 GPU hours, where Mythos Preview achieved a 26.7% success rate and Opus 4.8 reached 22.6%. The second mode focused on targets individually across 24-hour sessions with 2,500 GPU hours per target, raising Mythos Preview's overall success rate to 35.1%.

"Mythos Preview in the single-target focused mode achieved a 40% success rate, while participants in the Adaptyv Bio challenge averaged just 3.7%."

Anthropic

Stock photo for illustration only, not from the actual event

Source: Techsauce