Anthropic reveals Claude drug-research results and one-person biotech acquisition

Anthropic disclosed results from six drug-development partnerships on October 27, alongside its acquisition of one-person biotech Mithril Medicine. Researchers used Claude for experimental planning, data analysis and manufacturing-record review as well as molecule design. The company describes most of the work as preclinical.

The Mithril technical report, not peer reviewed, dates the acquisition to July 2026. Founder Cody Tranbarger’s rare-muscle-disease program moved into Anthropic. During one week examined in the report, Claude reviewed data on 113 antibody constructs and proposed follow-up tests. A researcher decided on plans, edited emails and sent them to outside laboratories.

The report also examines six errors; at least three reached laboratories in human-edited emails. Its mouse work was a pilot without randomized treatment allocation, and safety was outside the study’s scope. Further work is needed to establish whether the approach could be effective and safe in people.

Testing designs—and the measurements used to judge them

Experimental feedback helped narrow designs for the receptors that redirect immune cells toward cancer. Generate Biomedicines worked with AI agents over three rounds, producing an intracellular signaling component of 45 amino acids against a 154-amino-acid reference. Scientists set objectives, reviewed proposed sequences and selected 186 candidates for the initial screen.

Generate’s compact design increased receptor expression on cell surfaces and signaling triggered by its target. With transient mRNA delivery, expression and target-cell killing lasted longer. The report presents cell and mouse experiments, with proposed designs repeatedly checked against physical measurements to guide the next round of work.

At BigHat Biosciences, the measurements themselves became a research question. Two antibody-fragment campaigns ran from June to September, covering about 1,010 design variants over seven rounds. Researchers and Claude revisited raw heating signals and found that established metrics could misrepresent the molecular properties they were meant to capture.

The BigHat team used a physical model to select different metrics, then changed protein concentration and heating rate to test its predictions. Both leads improved in binding, thermal stability and aggregation while maintaining or improving yield and purity. Candidate design and the validity of the tests were investigated together.

For manufacturing review, Enveda tested its system on four records from one clinical production campaign: 1,061 scanned pages. Each of 11 actionable issues found by people appeared in at least one system run. A Claude Opus 4.7 pass added seven actionable findings that human reviewers had not raised.

Enveda recorded 1,553 minutes for unaided review versus 166 for AI processing plus human checks of its comments. This was an exploratory comparison on one campaign. Source records stayed unchanged and quality staff retained batch-release decisions. The advisory tool is not validated as a GMP computerized system.

Access for biological research has separate conditions. Anthropic’s Life Sciences Verification Program, announced September 17, vets organizations’ credentials, security and ethical oversight. Standard grants cover team research; work requiring additional scrutiny receives project-specific grants. The program launched in beta for teams and institutions, with more permissive biology safeguards than general-access models.

Anthropic expects a Mithril candidate could be ready to begin human trials as early as late 2028 if further results hold up. It gives no fixed start date. The company also plans to expand its clinical-trial-planning collaboration with Bristol Myers Squibb in 2027.