In October 2025 Anthropic launched Claude for Life Sciences, a set of connectors and skills intended to make Claude a better scientific collaborator. Since then the company has invested heavily in making Claude suited for scientific work: Opus 4.5 showed improvements in figure interpretation, computational biology, and protein-understanding benchmarks. Those advances reflect collaborations with academic and industry researchers and a focus on how scientists actually use AI to accelerate progress.
Anthropic’s AI for Science program also supplies free API credits to leading researchers working on high-impact projects worldwide. Participating groups have built Claude-powered systems that go beyond literature reviews or coding help: in the labs described, Claude functions across the research lifecycle—helping decide which experiments to run, compressing projects that normally take months into hours, and finding patterns in massive datasets that humans might miss.
Biomni: hundreds of tools and datasets under one agent
One bottleneck in biological research is fragmented tooling: hundreds of databases, software packages, and protocols exist, and researchers spend a lot of time selecting and learning platforms. Biomni, an agentic AI platform from Stanford University, aggregates hundreds of tools, packages, and datasets into a single system navigable by a Claude-powered agent. Researchers give plain-English requests; Biomni automatically selects resources. It can form hypotheses, design experimental protocols, and run analyses across more than 25 biological subfields.
A genome-wide association study (GWAS) illustrates the difficulty. Scanning genomes is relatively straightforward, but cleaning messy genomic formats, controlling for confounders, handling missing data and interpreting hits (which point to genomic locations rather than specific genes) is time-consuming. A single GWAS can take months. In an early Biomni trial the same work completed in 20 minutes.
Is this accurate? The Biomni team validated the system through several case studies. In one, Biomni designed a molecular cloning experiment; in a blind evaluation the protocol matched that of a postdoc with more than five years’ experience. In another, Biomni analyzed data from over 450 wearable-data files (30 people; mix of continuous glucose monitoring, temperature, and activity) in 35 minutes—work an expert estimated would take three weeks. In a third example, Biomni analyzed 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.
Biomni is not perfect, so it includes guardrails to detect when Claude goes off track. Where it falls short, experts can encode their methodology as a skill—teaching the agent an expert’s stepwise approach rather than leaving it to improvise. For example, when working with the Undiagnosed Diseases Network on rare-disease diagnosis, the team found Claude’s default approach differed from clinicians’; after interviewing an expert and documenting their diagnostic process, they taught that workflow to Claude and improved performance.
Biomni is one strategy—a general-purpose system that unifies many tools. Other labs are building more specialized Claude-based systems that target their specific workflow bottlenecks.
MozzareLLM and the Cheeseman lab: automating large-scale knockout interpretation
To learn what a gene does, researchers often remove it and see what breaks. CRISPR made large-scale, precise knockouts possible around 2012, but labs can generate more data than they can analyze. Iain Cheeseman’s lab at the Whitehead Institute and the Department of Biology at MIT faces this exact challenge. Using CRISPR, they knock out thousands of genes across tens of millions of human cells and photograph each cell to observe changes. Image patterns show that genes with similar functions produce similar phenotypes. Software groups those patterns automatically; Cheeseman’s lab built a pipeline called Brieflow to do so.
Interpreting why genes cluster—what they share, which are well studied, and which deserve follow-up—still required manual literature review, gene by gene. A single screen can yield hundreds of clusters, most of which go uninvestigated because of time and expertise limits. Cheeseman used to do much of this himself; he estimates he can recall functions for about 5,000 genes, but analyzing a screen still required hundreds of hours.
PhD student Matteo Di Bernardo worked with Cheeseman to automate that interpretive workflow. They documented the expert’s approach—what data sources to consult, which patterns indicate relevance, and what makes findings interesting—and built a Claude-powered system called MozzareLLM. Given a gene cluster, MozzareLLM identifies likely shared biological processes, flags well-characterized versus understudied genes, and highlights candidates worth follow-up. The system accelerates work and also surfaces discoveries the team had missed. Cheeseman reports Claude repeatedly catching items he overlooked: "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."
MozzareLLM’s usefulness stems from combining diverse information sources and reasoning like a scientist; it also reports confidence levels for findings, helping the lab decide where to invest limited resources. Di Bernardo tested multiple models and found Claude outperformed alternatives—once correctly identifying an RNA modification pathway other models dismissed as noise.
Cheeseman and Di Bernardo plan to publish Claude-annotated datasets so other experts can pursue clusters the lab lacks time to follow up, enabling domain specialists (for example mitochondrial biologists) to explore relevant clusters.
Lundberg Lab: AI-driven hypothesis generation for which genes to study
Cheeseman focuses on interpretation after pooled optical screens, but some labs face a different bottleneck: choosing which genes to target. The Lundberg Lab at Stanford runs smaller, focused screens where costs—typically starting above $20,000—limit how many genes can be tested. Teams conventionally assemble candidate lists in spreadsheets based on literature, expertise, and intuition.
The Lundberg team uses Claude to invert that process: rather than guessing based on what’s already been studied, they ask which genes should be studied based on molecular properties. They built a map of every known cellular molecule—proteins, RNA, DNA—and their relationships: binding partners, gene-product links, and structural similarities. Claude navigates that map to recommend candidate genes for a given cellular structure or process.
To test the approach, the lab chose a poorly studied topic so the model wouldn’t simply reproduce known findings: primary cilia, antenna-like appendages implicated in developmental and neurological disorders. The plan is to run a whole-genome screen to establish ground truth and then compare human-selected candidate lists (the spreadsheet method) to Claude’s molecular-property-based lists. If Claude finds a substantially higher fraction of true hits, or performs similarly but much faster, the method could become a standard pre-screening step for focused perturbation experiments, enabling better-targeted screens without whole-genome infrastructure.
Looking ahead
None of these systems are flawless, but they illustrate how AI agents have become research partners capable of far more than basic tasks. Each model release brings visible improvements: models that two years ago were mostly useful for code and summaries are now beginning to replicate the workflows those papers describe. As models and tools advance, labs continue to refine domain-specific guardrails and skills. Anthropic continues to expand Claude for Life Sciences and to accept applications to its AI for Science program; the company provides documentation and tutorials for teams that want to build or adopt these systems.
