OpenAI agents analyze mouse brain gene atlas in 32 minutes

Accelerating neuroscience research

OpenAI, PerturbAI, and Scripps Research recently demonstrated a significant leap in how scientists can process complex biological data. By deploying nearly 2,000 AI agents, the team analyzed a massive mouse brain dataset in just 32 minutes. This speed represents a major shift from traditional methods, which the team estimates would take human researchers between several months and a full year to complete.

The project utilized a specialized model called GPT-Rosalind, which was specifically built to assist in life sciences research. The goal was to navigate a whole-brain CRISPR atlas created by PerturbAI. CRISPR is a gene-editing tool that allows scientists to switch off specific genes to see how those changes affect different types of cells in the brain.

The results of this study were shared as a bioRxiv preprint, which means the work has not yet undergone formal peer review. Despite this, the findings show that AI can now keep pace with, and in some cases exceed, the depth of human analysis in complex biological systems. The team successfully merged the individual reports from the thousands of agents into a single reasoning atlas.

How the multi-agent system works

The researchers used a 'split' approach to handle the enormous amount of data. Instead of asking one AI to do everything, they assigned each of the nearly 2,000 agents a single gene change to study. Each agent was responsible for looking at how its specific gene change affected brain cells and then searching the existing scientific literature for relevant information.

Once the agents finished their individual tasks, they suggested possible explanations for the biological responses they observed. The team then combined these thousands of individual reports into a comprehensive reasoning atlas. This final product links specific genes to their cell effects, relevant scientific papers, and the likely biological mechanisms involved in those interactions.

This method proved to be more effective than using a single large AI model. In blinded tests, human experts preferred the reports generated by this multi-agent split approach 75% of the time. Furthermore, the many agents working together were able to cite more than 700 references, whereas a single agent only cited 11 for the same workload.

Identifying new medical leads

The AI agents were able to reproduce findings from previous human analyses while also uncovering new leads for rare diseases. One notable discovery involved two genes, Tsc1 and Tsc2, which are both involved in controlling how cells grow. The agents found that switching off Tsc2 changed dozens of genes related to communication between nerve cells in the cerebellum, while switching off Tsc1 resulted in very little change.

This specific gap in behavior between the two genes may be significant for understanding tuberous sclerosis, a rare genetic disorder. By identifying these differences, the AI helps scientists pinpoint which genes are most relevant to the disease's progression. The agents are capable of identifying these nuances across thousands of data points simultaneously.

In another instance, the agents analyzed the Dctn1 gene, which is linked to a rare brain disease known as Perry syndrome. They observed that switching off Dctn1 raised the activity of genes responsible for making cholesterol. Based on this, the agents suggested specific lab experiments to test if this cholesterol activity contributes to nerve cell damage in patients.

Open data and future impact

The researchers have made the results of this project accessible to the broader scientific community. The reasoning atlas is currently available on GitHub, a platform for hosting software and data. Additionally, the underlying dataset used for the study has been placed on Hugging Face, a popular hub for sharing machine learning models and data.

Sarah Teichmann, chair in stem cell medicine at the University of Cambridge, noted that these cell atlases are built so other scientists can make discoveries they haven't imagined yet. By providing the tools and the data, the team is enabling other researchers to use AI to explore new biological questions. This move toward open science helps democratize access to high-powered computational analysis.

Looking forward, this project sets a precedent for how AI agents can be used to navigate 'big data' in biology. As more gene-editing data becomes available, the ability to process it in minutes rather than months could drastically shorten the timeline for identifying new drug targets. The success of the multi-agent approach suggests that scaling AI to handle complex, multi-step reasoning is a viable path for scientific discovery.