Generalist AI Agent Matches Experts on Biomedical Research Tasks While Slashing Analysis Time

Biomedical research is often bogged down by fragmented, time-consuming workflows that require stitching together specialized tools, vast datasets, and deep domain knowledge. A study published today in Science describes an artificial intelligence agent that can autonomously tackle a broad range of these tasks—from analyzing millions of wearable sensor readings to designing cloning experiments that were successfully carried out in the lab—often matching the accuracy of seasoned human experts in a fraction of the time.

Led by senior author Jure Leskovec at Stanford University, with first author Kexin Huang, the team built a system called Biomni that can reason across genetics, genomics, microbiology, pharmacology, and other fields without being pre-programmed for each specific job. The agent draws on a curated environment of 150 specialized tools, 105 software packages, and 59 databases, then writes and executes its own analysis code on the fly.

When tested on a benchmark of 443 queries spanning tasks such as rare disease diagnosis, causal gene detection, and CRISPR screen design, Biomni achieved an average accuracy of 57%, outperforming a leading large language model alone (30%) and a bioinformatics agent equipped with the same tool set (44%).

Across several head-to-head comparisons with human experts, Biomni reached comparable accuracy while dramatically cutting the time required: rare disease diagnosis took roughly 3 minutes versus over 110 minutes for experts, and GWAS causal gene detection took about 4 minutes instead of 90.

To boost performance on especially tricky tasks, the researchers applied reinforcement learning, fine-tuning an open-source model by rewarding it for successful task completion. This lifted the smaller 8-billion-parameter version from an average score of 0.32 to 0.59—surpassing the much larger closed-source Claude 4 Sonnet—while the 32-billion-parameter model climbed to 0.67.

The team demonstrated Biomni’s versatility through several real-world case studies. In one, the agent independently reproduced a complex analysis of wearable heart-rate and step data from over 1,000 participants during the COVID-19 pandemic, identifying the same six physiological biomarkers and their correlations that had originally required expert human analysis. In another, it took raw single-cell RNA and ATAC sequencing data from developing human embryos and constructed gene regulatory networks, recovering known relationships such as RUNX2’s role in bone formation and flagging several underexplored transcription factors that may control skeletal cell fate.

Perhaps most strikingly, Biomni designed a complete molecular cloning protocol for inserting a guide RNA into a CRISPR plasmid. When researchers followed the AI-generated instructions exactly, they obtained bacterial colonies with perfect sequence alignment on Sanger sequencing—performance that blinded reviewers judged equivalent to that of a senior expert. The agent also orchestrated protein engineering tools to suggest three thermostability-improving mutations and turned natural-language instructions into executable code for robotic liquid handlers.

The authors caution that Biomni does not perform at expert level across the board. It sometimes stumbles on tasks demanding nuanced clinical judgment or deep biological synthesis, and highly complex workflows can benefit from more structured prompts. The vastness of biomedical research means many domains remain untested, and the agent’s reliance on recent literature may overlook older yet valuable methods. The team also highlights the importance of biosecurity guardrails as AI systems become more capable in the life sciences.

“Biomni and its successors could become foundational infrastructure in an AI-powered biomedical ecosystem, working seamlessly with human experts to unlock insights into health and disease,” the researchers write. For now, the system stands as a powerful proof of concept that generalist AI agents can meaningfully augment—not replace—the human scientist.

Reference: Kexin Huang et al. Autonomous biomedical research with an artificial intelligence agent. Science (2026). DOI: 10.1126/science.adz4351