OpenAI says thousands of AI agents found brain-research leads in 32 minutes

Thirty-two minutes of AI analysis produced new questions about brain disease—not proven answers. OpenAI, PerturbAI and Scripps Research said on October 21 that thousands of GPT-Rosalind-powered agents examined existing data from millions of mouse brain cells. Their output connects gene disruptions with possible biological mechanisms, giving scientists leads to test rather than validated treatments.

The researchers estimate that comparable human analysis and literature review, tackling one gene disruption at a time, would take months to a year. That is the team’s estimate, not a measured head-to-head race. The project’s more concrete comparison asked whether dividing the work among agents produced better interpretations than assigning the entire workload to one.

One gene disruption per investigator

The starting material was a whole-brain CRISPR atlas: a dataset recording how disrupting genes affects different cell types across the mouse brain. The challenge was to interpret those cellular responses alongside findings scattered through scientific literature. The agents analyzed an existing experiment; they did not generate the underlying mouse-cell data.

Each agent investigated a gene disruption, examined the cellular changes, searched publications and proposed explanations. Thousands worked in parallel. The team combined their reports into a reasoning atlas linking genes, cell effects, relevant papers and possible mechanisms. OpenAI calls this division of labor distributed biological reasoning.

According to OpenAI, the agents reproduced findings from human analyses and identified additional leads. Their assessments of novelty and scientific value also aligned with judgments from blinded human experts. Those evaluations concern how useful the interpretations appear for further research, rather than whether the proposed mechanisms have been demonstrated experimentally.

75%Expert preference for distributed-agent reports

OpenAI says blinded human experts preferred distributed-agent reports over single-agent reports 75% of the time.

The single agent searched less literature

To test the division of labor, the researchers held the model and resources fixed while comparing a single agent with distributed agents. As the workload increased, the single agent searched less of the literature. At the largest workload, it cited 11 references; the distributed agents cited more than 700.

The researchers’ rationale is that scientists have limited time and often concentrate on the strongest signals. Splitting the work lets agents examine a wider range of responses, including subtle or unexpected ones, while connecting each with relevant publications. The comparison therefore tests a way of organizing analysis, not a new model against an older one.

Two clues with specific laboratory questions

• Different effects from related genes. Tsc1 and Tsc2 work together to control mTOR, a regulator of cell growth, metabolism and aging. Disrupting either produced similar effects in several neuron groups. But in the cerebellum, which helps coordinate movement, Tsc2 disruption changed the activity of dozens of genes involved in neuronal communication; Tsc1 disruption produced little change. The agents identified a specific neuron group and cellular changes for scientists to investigate.
• A possible cholesterol connection. Disrupting Dctn1 increased activity in genes involved in making cholesterol. The agents connected that result with literature suggesting that cholesterol-transport problems can prompt cells to make more cholesterol. They proposed testing whether DCTN1 disruption changes cholesterol distribution inside neurons and whether that contributes to neurodegeneration. Human experts considered the hypothesis novel, consequential and worth pursuing.

The reasoning atlas is available for researchers to explore. They can compare related genes, inspect unexpected responses and choose leads to pursue. Its records bring proposed explanations together with the cellular effects and publications behind them, making the output a resource for planning subsequent experiments rather than simply a set of standalone answers.