The Biomni AI agent selects scientific databases and software, writes code, and drafts laboratory protocols: Science published the paper describing the system
The Biomni AI agent selects scientific databases and software, writes code, and drafts laboratory protocols: Science published the paper describing the system
The system identifies the scientific tools needed for a task, develops a plan, and executes it step by step. It was tested on identifying genes that may cause disease, finding new uses for drugs, diagnosing rare diseases, and designing DNA experiments.
In a paper published on July 9, the Biomni team described a system in which a language model opens the appropriate database, prepares the data, runs a program, detects errors, and saves intermediate results. The open version of the environment contains 150 specialized tools, 105 software packages, and 59 databases. The agent selects the resources it needs, writes code in Python, R, or Bash, and revises its plan as the work proceeds.
The team built this environment by reviewing recent papers from 25 areas of biomedicine. A separate AI system extracted tools, databases, and methods from the publications, after which people reviewed the selected components. Biomni uses this library to assemble a sequence of actions for each task instead of following a predefined workflow for every type of analysis. In June, the integration of Boltz models with Biomni showed how this works: the agent selected a drug target, ran a molecular design workflow, and returned a list of candidates for laboratory testing.
On a held-out portion of a test covering biological databases, Biomni answered 74.4% of the questions correctly, compared with 74.7% for experts. On questions about protein and DNA sequences, it scored 81.9%, compared with 78.8% for humans. A more difficult assessment of broad biomedical knowledge produced a score of 17.3%. The authors used 12.5% of one test and designed some of the applied tasks themselves. Access to tools improves the model substantially, but the system still makes frequent errors.
In a laboratory test, Biomni selected a guide RNA that directs CRISPR to the required region of DNA and designed its insertion into a plasmid, a small circular DNA molecule used to deliver genetic instructions. The agent prepared both an assembly protocol and a validation method. A researcher followed the instructions without changes, and two cultured bacterial colonies contained the correct sequence. The agent carried an initial design through to a physically verified result, although the experiment covered only one familiar type of DNA work.
In aging research, a single dataset often contains thousands of cells and millions of measurements of gene activity. Biomni can analyze these data, match genes against databases, and develop a testable hypothesis while preserving the code and a record of the computational steps. This allows a researcher to test more possibilities in the same amount of time and discard weak ideas sooner.
The authors acknowledge that Biomni performs poorly in clinical judgment and in generating new experimental ideas. Its library favors recent papers and may miss older methods that remain useful. People must be able to inspect the code, intermediate data, and independently validated results. Without this oversight, a fast agent can propagate an unnoticed error just as quickly.
Access to validated procedures allows a scientific agent to complete a sequence of steps and produce a measurable result. If code and data validation become faster alongside automation, an individual researcher will be able to test more hypotheses about aging and move the strongest candidates into the laboratory sooner.