Anthropic acquired Mithril Medicine, a one-person company, and reported on Claude AI’s work in drug development with six partners

Anthropic acquired Mithril Medicine, a one-person company, and reported on Claude AI’s work in drug development with six partners

On October 27, Anthropic reported on Claude’s work with four biotech companies, the pharmaceutical company Bristol Myers Squibb, and the investment firm The Column Group. Almost all of the projects have yet to reach human trials. Mithril, acquired in July, was founded by Cody Tranbarger.

As Tranbarger explains on X, Mithril was partly an experiment to see how far one person could get with Claude and part-time consultants. He writes about the benefits of AI:

I see today’s easy gains in coordinating work and running it in parallel, rather than in speeding it up and compressing timelines.

He would wager that drug developers make better decisions when they run several programs at once and spend their time exercising scientific judgment.

Mithril is developing an antibody for LAMA2 muscular dystrophy, which is caused by the loss of the protein laminin-α2. Of 42 children followed in London who completely lacked the protein, one could walk independently. Laminin-α2 helps attach the sheath surrounding a muscle fiber to the fiber itself, while the alternative protein, laminin-α4, binds weakly. The antibody is designed to bridge laminin-α4 and the muscle fiber: one “arm” binds to β-dystroglycan on the fiber, and the other binds to laminin-α4. There is a precedent: adding a gene for mini-agrin, a protein that attaches laminin to the fiber, increased median lifespan in affected mice from 15,5 to 50 weeks.

The Mithril report, which has not been peer reviewed, examines a week that the authors selected as their best recent example. Claude itself wrote the account. A contractor sent 341 pages of data on 113 antibody variants, and Claude proposed a next step for each in less than 8 minutes. Claude recommended measuring binding rates for all of them, but one author set aside three laminin-binding “arms” and, with them, 24 antibodies. Claude revised the plan in 4 minutes, leaving 89.

Claude drafted emails to laboratories, which a human edited and sent. Of the six errors examined, at least three reached the laboratories, two went unnoticed for more than four days, and Claude identified five itself. Claude’s median initial response time was 96 seconds, compared with 8,5 hours for the laboratories. The authors concluded that laboratory responses and human decisions set the pace, while Claude handled work that grew with the number of antibodies.

In vitro, an “arm” stays bound to β-dystroglycan for only minutes, so the authors added a second to hold the antibody in place while the first rebinds. They compared antibodies with one or two “arms” for each target. Mice in the pilot experiment were assigned to groups without randomization: the group receiving antibodies with two “arms” was tested last, 126 and 161 days after the control group. Peak triceps force in its 7 mice was almost twice as high as in the 4 control mice. The antibody with one “arm” reached the muscles, but muscle force was no different from that in controls. The authors suggest that the bridge may need a firmer grip or more “arms.” Anthropic expects human trials no earlier than the end of 2028, provided the drug candidates prove viable.

In partner BigHat’s report, Claude and an Anthropic researcher reexamined the raw data: across antibody fragment variants, the temperature at which the protein began to aggregate remained around 40 °C, even though aggregation rates differed by more than tenfold. The authors describe this as “time disguised as temperature” and replaced the metric with aggregation rate. A Claude-based system developed by another partner, Enveda, reviewed 1 061 pages of manufacturing documents and identified all 11 issues flagged by human reviewers across different runs. In one run, it also found seven minor issues the reviewers had missed.