Accelerating Scientific Research

Anthropic's report on three research groups using Claude describes a recurring method: a lab writes down how its own expert does a task and has a Claude-powered system follow that method instead of the model's default approach. Stanford's Biomni gathers hundreds of biological tools and databases behind one agent that takes requests in plain English. When the Biomni team worked with the Undiagnosed Diseases Network on rare-disease diagnosis, they found Claude's default approach differed substantially from a clinician's, so they interviewed an expert, recorded the diagnostic process step by step and gave it to the agent as a skill, a packaged set of instructions Claude follows for a specific task. Anthropic's Claude for Life Sciences launch brought these skills to scientific work. Anthropic says the agent then performed well, without giving figures.

The clearest case is at Iain Cheeseman's lab at the Whitehead Institute and MIT. The lab uses CRISPR gene editing to knock out thousands of genes across tens of millions of human cells and photographs each cell; genes with similar jobs tend to cause similar-looking damage, so software groups them into clusters. A single screen can produce hundreds of clusters, and most are never examined because explaining each one means an expert reading the literature gene by gene. Cheeseman and PhD student Matteo Di Bernardo documented how he does that interpretation, including which sources he consults and what he treats as interesting, and built MozzareLLM on Claude. For each cluster it proposes a shared biological process, separates well-studied genes from poorly studied ones, flags candidates for follow-up and attaches a confidence level, which Cheeseman says helps him decide whether a lead is worth more lab time. He says it consistently catches things he missed, and that he can verify those findings.

The evidence is mostly the labs' own accounts. A genome-wide association study that can take months finished in 20 minutes in what Anthropic calls an early trial of Biomni, and the article points to the Biomni team's own case studies for validation. A planned head-to-head test at Stanford's Lundberg Lab has not reported results: it will compare which genes Claude and human experts predict affect primary cilia, a little-studied cell structure, against a whole-genome screen, and the article's figures of 150 versus 80 out of 200 are explicitly hypothetical. The report shows where labs are placing Claude in their work, not yet a systematic measure of how often its conclusions are right.