Anthropic Describes How Research Labs Are Building Claude-Powered Science Agents

Case study details Claude use in research labs

Anthropic published a case study on Jan. 15, 2026 describing how several research labs have built custom systems on top of Claude to run parts of the scientific process, from choosing experiments to interpreting large datasets. The company says the report grew out of its Claude for Life Sciences suite, launched in October 2025, and its AI for Science program, which provides free API credits to researchers on high-impact projects. Anthropic says Claude Opus 4.5 showed improvements on benchmarks covering figure interpretation, computational biology and protein understanding.

Biomni: a general-purpose biomedical agent

Anthropic highlights Biomni, an agentic AI platform built at Stanford University that lets a Claude-powered agent draw on hundreds of biological tools, software packages and databases from a single interface. Researchers describe their requests in plain English, and the system selects the relevant resources, forms hypotheses, designs experimental protocols and runs analyses across more than 25 biological subfields, according to Anthropic.

Anthropic cites several examples from the Biomni team. In an early trial, a genome-wide association study — a process that normally takes months because of data cleaning and interpretation — was completed in 20 minutes. In a blind evaluation, a molecular cloning protocol Biomni designed matched one from a postdoctoral researcher with more than five years of experience. The system analyzed more than 450 wearable-device data files from 30 people in 35 minutes, a task Anthropic says was estimated to take a human expert three weeks, and it processed gene activity data from more than 336,000 individual cells from human embryonic tissue, confirming known regulatory relationships and identifying new transcription factors not previously linked to human embryonic development. Anthropic notes the system includes guardrails meant to detect when Claude goes off track, and that when working with the Undiagnosed Diseases Network on rare disease diagnosis, the team documented an expert clinician's diagnostic process and taught it to Claude after finding the model's default approach differed from a clinician's.

Cheeseman Lab: interpreting gene-knockout clusters

At the Whitehead Institute and MIT, Iain Cheeseman's lab uses CRISPR to knock out thousands of genes across tens of millions of cells and photographs the results, then uses a pipeline called Brieflow to group genes with similar effects. According to Anthropic, PhD student Matteo Di Bernardo built a Claude-powered system called MozzareLLM, developed by working closely with Cheeseman to encode his interpretation process, that identifies shared biological processes among gene clusters, flags understudied genes, and assigns confidence levels to its findings. Cheeseman is quoted saying: "Every time I go through I'm like, I didn't notice that one! And in each case, these are discoveries that we can understand and verify." Anthropic says Di Bernardo tested multiple AI models and found Claude outperformed the alternatives, including correctly identifying an RNA modification pathway that other models dismissed as noise.

Lundberg Lab: testing AI-generated gene hypotheses

Anthropic also describes the Lundberg Lab at Stanford, which runs smaller, targeted gene screens that can cost upwards of $20,000 and typically rely on researchers manually compiling candidate gene lists. The lab built a map of molecular relationships among proteins, RNA and DNA, and is testing whether Claude, navigating that map, can identify candidate genes more effectively than the conventional approach. According to Anthropic, the lab is running this comparison using primary cilia, a understudied cell structure, as the test case, with a genome-wide screen planned to establish which approach performs better.

Source: Anthropic, "How scientists are using Claude to accelerate research and discovery," published Jan. 15, 2026.