How Scientists Are Using Claude to Accelerate Research and Discovery
Published: January 15, 2026
Introduction
Anthropic launched Claude for Life Sciences last October, introducing connectors and skills that enhanced Claude's capabilities as a scientific collaborator. Since then, the company has invested substantially in strengthening Claude's performance across scientific domains. Opus 4.5 demonstrated notable gains in figure interpretation, computational biology, and protein understanding benchmarks—improvements informed by partnerships with academic and industry researchers.
The company operates the AI for Science program, which distributes free API credits to prominent researchers pursuing high-impact scientific initiatives globally.
How Scientists Use Claude
Researchers have built custom systems leveraging Claude in ways extending far beyond literature reviews or coding help. In participating labs, Claude functions as a collaborator across all research stages: determining which experiments to conduct, employing diverse tools to condense months-long projects into hours, and detecting patterns in vast datasets that humans might miss. The system frequently eliminates bottlenecks and handles tasks requiring deep knowledge previously impossible to scale, sometimes enabling entirely novel research approaches.
Biomni: General-Purpose Biomedical Agent
Stanford University's Biomni addresses a critical bottleneck: fragmentation of scientific tools. Hundreds of databases, software packages, and protocols exist, yet researchers invest considerable time mastering various platforms rather than conducting experiments or analyzing data.
Biomni consolidates hundreds of tools, packages, and datasets into a single system. Researchers submit requests in plain language; Biomni automatically selects appropriate resources. The system can form hypotheses, design experimental protocols, and conduct analyses across 25+ biological subfields.
Genome-Wide Association Studies (GWAS) Example
Consider a GWAS searching for genetic variants linked to specific traits or diseases. Researchers scan large populations' genomes for variants appearing more frequently in one group versus another. While genome scanning itself is relatively straightforward, analyzing and interpreting results proves time-consuming.
Genomic data requires extensive cleaning and formatting; researchers must control for confounding factors and handle missing data. After identifying "hits," they must determine nearby genes, cell type expression patterns, and affected biological pathways. Each step potentially involves different tools, file formats, and manual decisions—a tedious process typically requiring months. Biomni completed the same analysis in 20 minutes.
Validation and Accuracy
The Biomni team validated their system through multiple case studies:
- Molecular cloning: Biomni designed an experiment matching a postdoc's protocol in blind evaluation
- Wearable data analysis: Processed 450+ files from 30 people in 35 minutes (estimated three-week task for human experts)
- Gene activity analysis: Examined 336,000 individual cells from embryonic tissue, confirming known regulatory relationships while identifying previously unknown transcription factors connected to embryonic development
The system includes guardrails detecting when Claude goes off-track. When falling short, experts can encode methodology as skills—teaching Claude how specialists approach problems. When working with the Undiagnosed Diseases Network on rare disease diagnosis, the team found Claude's default approach differed from clinical practice, so they documented expert processes step-by-step and trained Claude accordingly.
Cheeseman Lab: Automating Gene Knockout Interpretation
MIT's Cheeseman Lab faces a different bottleneck. Using CRISPR, they knock out thousands of genes across tens of millions of human cells, photographing each cell to observe changes. Genes performing similar functions produce similar-looking damage when removed. Software detects these patterns and groups genes automatically using Brieflow (a pipeline named after the cheese).
The Interpretation Challenge
Interpreting what gene groupings mean—why genes cluster together and whether relationships are known or novel—requires human experts to review scientific literature gene-by-gene. This process is slow; a single screen generates hundreds of clusters that mostly remain uninvestigated due to time and bandwidth constraints.
MozzareLLM Solution
Cheeseman estimates he can recall ~5,000 genes' functions from memory, yet analysis still requires hundreds of hours. PhD student Matteo Di Bernardo developed MozzareLLM, a Claude-powered system trained on Cheeseman's interpretive approach.
The system identifies shared biological processes among gene clusters, distinguishes well-understood from poorly-studied genes, and highlights candidates for follow-up. Cheeseman reports Claude consistently identifies overlooked connections that withstand verification. The system provides confidence levels in findings, helping determine whether to invest resources in pursuing conclusions.
During testing, Claude outperformed alternative models, correctly identifying an RNA modification pathway others dismissed as noise.
Lundberg Lab: AI-Led Hypothesis Generation
Stanford's Lundberg Lab runs smaller, focused screens where the bottleneck emerges earlier: determining which genes to target initially.
Cost and Conventional Approaches
Focused screens cost $20,000+, with expenses rising with scale. Labs typically target a few hundred genes believed most likely involved in specific conditions. The conventional process involves teams sitting around spreadsheets, adding candidate genes with brief justifications or paper links—educated guessing informed by literature reviews and intuition but constrained by human bandwidth and fallible recall.
Claude-Powered Approach
Lundberg Lab reversed the approach: instead of asking "what guesses can we make from studied research?" they ask "what should be studied based on molecular properties?"
The team mapped every known cellular molecule—proteins, RNA, DNA—and their relationships: protein binding patterns, gene-product coding, structural similarities. Claude navigates this map to identify candidate genes based on biological properties and relationships.
Primary Cilia Study
The lab is testing this approach using primary cilia—antenna-like cell appendages with limited research but connections to developmental and neurological disorders. They're comparing human experts (using spreadsheet methods) to Claude (using the molecular relationship map). If Claude identifies more genes affecting cilia formation than experts, it validates the approach. Even similar discovery rates would likely show Claude's efficiency advantage, streamlining the research process.
If successful, this could become standard practice in focused perturbation screening, enabling informed gene selection without requiring whole-genome screening infrastructure.
Looking Forward
These systems aren't perfect, yet they demonstrate how scientists increasingly incorporate AI as research partners capable of far more than basic tasks—increasingly able to accelerate, and sometimes replace, various research aspects.
A common theme emerged: tool usefulness grows alongside AI capabilities. Each model release brings noticeable improvements. Where earlier models two years ago were limited to code writing or paper summarization, more powerful agents now increasingly replicate the work those papers describe.
As tools advance and AI models grow more intelligent, understanding how scientific discovery develops alongside these technologies remains crucial.
For more information: Expanded Claude for Life Sciences capabilities are detailed here, with tutorials available. Researchers can submit applications to the AI for Science program, reviewed by subject matter experts.