← Back to Home

Revisiting a Mouse Brain Cell Atlas in 32 Minutes: AI Agents Divide the Work to Find Clues for Neurodegeneration Research

Nearly two thousand AI agents analyzed gene perturbation data in parallel, connecting cellular responses with the literature to form traceable research hypotheses. Preliminary assessments support the value of this division of labor, but experiments are still needed to confirm disease mechanisms and develop drugs.

By SURL BioNews

When an experiment can record changes in millions of cells, scientists’ challenge extends from obtaining data to understanding it. OpenAI, PerturbAI, and Scripps Research collaborated to have nearly two thousand AI agents divide the work of examining a whole-brain gene perturbation atlas in mice. According to the team, the agents completed the analysis and literature integration in 32 minutes, generating leads for neurological disease research that can be tested in subsequent work.

The existing dataset covers approximately 7.7 million mouse brain cells, using CRISPR to perturb genes and single-cell RNA measurements to track responses. It allows researchers to compare which genes change their expression in different types of brain cells when the same gene is disrupted. The 32 minutes refers to the AI analysis stage and does not include the animal experiments and sequencing needed to build the atlas.

The study used a method called “distributed biological reasoning,” with agents powered by GPT-Rosalind each responsible for one gene perturbation. They combined cell-type-specific expression data with scientific literature, then consolidated their reports into a queryable reasoning atlas. A public GitHub repository also provides code, prompts, saved predictions, execution logs, and analysis workbooks, allowing other researchers to trace how the conclusions were reached.

Testing of the method went beyond comparing speed. The study abstract states that the agents reproduced conclusions from the original research paper without access to it and predicted perturbation responses that had been held out in advance and were not provided to the model. According to blind evaluation results released by OpenAI, experts preferred the multi-agent reports in 75% of comparisons. However, a preference for a report and experimental confirmation of a newly proposed biological mechanism remain two different levels of evidence.

One lead involves DCTN1. Mutations in this gene, which is involved in transporting materials within nerve cells, can cause Perry syndrome, which features Parkinson’s-like symptoms. The team said the AI noticed increased expression of genes related to cholesterol synthesis after mouse Dctn1 was disrupted, leading it to propose a hypothesis: could an abnormal distribution of cholesterol within cells contribute to neurodegeneration? This provides a direction for subsequent experiments, but it does not yet establish that modifying cholesterol transport can protect neurons.

The findings are currently available as a bioRxiv preprint and have not yet been peer reviewed. RNA changes caused by gene perturbations in mice also cannot be directly equated with human disease or drug effects. The near-term use of this work is to help researchers identify questions within a vast atlas that warrant experimental investigation; whether it can actually improve drug development still depends on how many of these hypotheses hold up after independent validation.

References

  1. Tech Times
  2. Jin Lab Neurogenomics / GitHub