Anthropic: Claude designed 1,320 miniproteins for 15 targets, and two laboratories tested their binding
Anthropic: Claude designed 1,320 miniproteins for 15 targets, and two laboratories tested their binding
On August 18, Anthropic published the campaign results, the full technical report, and the open dataset. Of the 1,320 sequences, 354 were classified as binders based on the combined assessment of two contract laboratories. The researchers defined the overall protocol, computational budget, and target list in advance. Claude independently selected protein sites and the software used to design candidates.
In this campaign, each miniprotein was intended to bind to a selected target protein. This required choosing a site on the target surface, determining the candidate's structure and amino acid sequence, producing the candidate, and measuring whether binding occurred.
Anthropic selected the targets and established the overall protocol. Claude examined each target, chose a site on its surface, ran open-source software to generate structures and select amino acids, evaluated the candidates computationally, and prepared the final lists for laboratory testing. This process converted decisions about what to design into sequences that could be synthesized and measured.
In June, the Boltz model and the Biomni environment had already provided an agent with a workflow: select a target, run protein-design software, and use computational binding predictions to choose 50 small antibodies. In that study, the agent followed a predefined workflow and worked with a separate set of laboratory data. Anthropic tested a longer process in which the agent independently chose which tools and methods to use for design, while the laboratory measured every final candidate across a series of targets.
In May, Adaptyv Bio tested 100 protein candidates that ten human teams and six AI agents had designed in one day for TREM2, a protein found on immune cells in the brain. In the current campaign, the same laboratory helped test candidates against 15 targets, and Anthropic disclosed the result for every submitted sequence.
Anthropic sent the final sequences to two contract laboratories, Adaptyv Bio and Twist Bioscience. The laboratories did not know which model, campaign, or list position corresponded to each sequence, and they measured binding using different methods. The report's authors published the raw measurement curves and the rule used to combine the results into a binary classification of binding or no binding.
The candidate rankings were also evaluated experimentally. Across 41 lists from three campaigns covering at least 13 targets, the first-ranked candidate bound in 49% of cases, a candidate among the top five bound in 44%, and a candidate among all thirty bound in 28%. For every sequence, the public dataset reports its position in the list, the laboratory measurements, and the final classification.