Anthropic tests Claude in de novo protein design and confirms 354 successful laboratory designs

Many medicines and medical treatments work by binding to a specific target in the human body to halt its activity or alter its biological function. The critical first step in developing any new drug begins with engineering a biomolecule capable of binding tightly to its intended target. Traditionally, this foundational phase required weeks or months of intensive, specialized work for each individual target, with researchers screening vast numbers of candidate molecules in hopes of finding the rare few that prove effective. To test whether generative AI could streamline this process, Anthropic evaluated whether its Claude model could design entirely new protein binders from scratch, through de novo design without relying on preexisting protein structures.

The experiment relied on a precise protein-design prompt crafted by a human expert, leaving Claude to generate the protein binders fully autonomously, andthe model successfully designed binders targeting 14 out of 15tested targets, generating a total of 1,320 novel protein designs in the process. The trial tasked generative AI with independently proposing complete molecular structures, attempting to bypass the lengthy and laborious manual screening researchers traditionally face in early-stage laboratory drug discovery.

The evaluation extended beyond computational models into physical biology through a partnership with Adaptive Bio and Twist Bioscience. The two companies synthesized Claude's designs in the laboratory and subjected them to independent assays to verify physical performance. The wet-lab results confirmed that 354 of the model-generated designs were functional and capable of binding in a laboratory environment, with several AI-designed proteins matching or exceeding the binding performance of leading biological binders currently available in the medical field.

Assays showed particularly strong results across at least four targets, pointing to the possibility of developing more potent therapeutics that could be administered at lower concentrations, reducing required patient dosages while maintaining clinical efficacy. Although the published findings did not disclose the specific identities of the 14 biological targets, the trial demonstrates how AI models can compress early-stage biomolecule design by combining autonomous generation with independent laboratory validation.