Anthropic buys a one-person biotech to start its own work on rare childhood diseases

What Anthropic announced

As itmedia.co.jp reported, Anthropic said on October 27, U.S. time, that it is running drug discovery in-house. The news came with an update on its partnerships with six organizations that develop new medicines.

To get started, it acquired Mithril Medicine, a U.S. company built around treatments for rare muscular dystrophies in children. Founder Cody Tranberger was its only employee, and he continues the work at Anthropic. He plans to launch several new drug programs.

One founder, Claude and eleven labs

Tranberger used Claude, Anthropic's AI model, to handle contracts, logistics and experiment scheduling with eleven outside research institutions. He also handed much of his other work to the model.

According to the report, this setup took one of the two target diseases from a treatment idea to promising mouse results in under a year and a half.

The science, in plain terms

One target is LAMA2-related congenital muscular dystrophy. In this childhood disease, a missing protein normally anchors muscle fibers to the membrane around them, and there is no treatment that fixes the root cause.

Mithril's approach uses a bispecific antibody, a protein engineered to grab two different targets at once. Here it links a protein still present in the muscle to the muscle fiber, re-anchoring the membrane. The same idea might also work for a second disease, alpha-dystroglycanopathy.

In a small preliminary mouse test, one of two prototype antibodies was given for 6 weeks. Those mice showed about twice the forelimb and grip strength of mice given a liquid with no drug.

Where Claude slipped

Claude also helped narrow the candidates. It read a very long stack of manufacturing data and sorted more than a hundred candidate antibodies into three groups: test as is, purify first, or skip. That took under 8 minutes. The plan involved evaluation at five contract research organizations, outside firms that run experiments for hire.

Claude drafted emails to those firms, and staff edited them before sending. Still, at least 3 flawed requests got through, including one that asked for less antibody than the recipient wanted. The technical report concludes that human reviewers tend to catch mistakes in judgment, such as which candidate to test, but miss errors in quantities and underlying assumptions.

What comes next

If the candidates stay promising in later tests, human clinical trials could begin in a few years. That is a hope, not a result.

Anthropic had already signaled this direction when it announced Claude Science, an AI workspace for researchers. It said it would target diseases that traditional drug companies have found unattractive, and stick to preclinical work, the research done before any human testing.