From Finding Literature to Writing Experimental Protocols: Biomni Lets AI Take Over Multi-Step Biomedical Research Work

The most time-consuming part of biomedical research is often not formulating a question, but connecting scattered papers, databases, analytical software, and experimental procedures into an executable path. A research team published Biomni, a general-purpose AI agent, in Science, seeking to let researchers assign tasks in natural language and have the system plan and complete these interconnected steps.

Biomni does more than generate text answers. It combines large language model reasoning, retrieval-augmented planning, and code execution. It first breaks down a question, then selects data sources and tools, writes code, and adjusts subsequent steps based on computational results. Its workflows do not apply prewritten fixed templates, so the same architecture can be assigned to different tasks, including causal gene prioritization, drug repurposing, rare disease diagnosis, microbiome analysis, and molecular cloning.

To build a working environment that the agent could access, the team first extracted commonly used methods from 2,500 recent bioRxiv papers across 25 biomedical fields. People then organized them into 105 software tools, 150 specialized biology tools, and 59 databases. This manual work is crucial: whether the agent can propose a reasonable plan depends not only on the language model, but is also constrained by whether the tools are reliable, the data are complete, and different resources can be connected correctly.

The research report shows that Biomni can handle eight previously unseen biomedical scenarios without being fine-tuned for individual tasks, outperforming multiple benchmark agent architectures overall. In a molecular cloning test, its accuracy was reported to be close to that of experts, while its completion time was significantly shorter. One gene-editing vector cloning protocol generated by the system was also confirmed through colony growth and Sanger sequencing to have the correct inserted sequence. However, this was a successful validation of a specific experiment and cannot be directly extrapolated to conclude that all AI-designed wet-lab procedures are equally reliable.

The system also organized 458 spreadsheets containing wearable-device and physiological data, and analyzed approximately 336,000 droplets from single-nucleus RNA and chromatin accessibility sequencing. These cases show that an AI agent can link data cleaning, statistics, visualization, and hypothesis generation. The same study also demonstrated protein stability optimization, coordination of laboratory instruments, and the design of protocols suitable for empirical testing, extending the scope of applications from computational analysis toward laboratory operations.

The real test will be whether the system can maintain traceability after leaving carefully selected tests and operating amid uneven data quality, tool updates, restricted permissions, or high-risk experiments. The authors also acknowledge that the current evaluation covers only part of the biomedical field. Before the system can be used for clinical judgment or automated instrument control, mechanisms for human review, error interception, biosafety, and accountability must still be established. This research developed from an early preprint that had not undergone peer review into a formal paper. The conflict-of-interest disclosure states that some authors are employed by Genentech and that one holds equity in Roche.