OpenAI’s agents build a mouse-brain reasoning atlas in 32 minutes

The team calls its method “distributed biological reasoning.”[1] Each agent examined a gene disruption, analyzed cellular responses, searched scientific literature and proposed explanations for the results.[1] The agents’ reports were combined into a reasoning atlas linking genes, cell effects, publications and possible mechanisms.[1] OpenAI says the underlying dataset shows how gene disruptions affect different cell types across the mouse brain.[1] The atlas is available, giving researchers a way to explore these connections without treating a proposed explanation as a confirmed mechanism.[1]

Thousands of GPT-Rosalind agents analyzed existing experimental data from millions of mouse brain cells, and the work took 32 minutes.[1] The engine launched with an eight-million-cell, brain-wide CRISPR atlas.[3] OpenAI says comparable human analysis and literature review, done one perturbation at a time, would take months to a year.[1] That comparison highlights the potential value of parallel review for exploring large datasets; it does not mean the agents performed new experiments or that their proposed explanations are established findings.[1]

OpenAI says the agents reproduced findings from human analyses and uncovered additional leads; their assessments of novelty and scientific value aligned with blinded experts’ judgments.[1] In blinded comparisons, experts preferred reports from distributed agents over single-agent reports 75% of the time.[1] In the largest-workload comparison, distributed agents cited more than 700 literature references, compared with 11 from a single agent.[1] The paper says the agents also recovered the source study’s conclusions without access to its manuscript, predicted held-out perturbation effects and identified relationships between perturbations.[2] The paper is a preprint that has not been certified by peer review.[2]

One example involved Tsc2 disruption, which altered dozens of genes involved in neuronal communication, while Tsc1 disruption produced little change.[1] OpenAI notes that, in people with tuberous sclerosis, cerebellar abnormalities are strongly linked to TSC2 mutations.[1] In another example, Dctn1 disruption increased expression of genes involved in cholesterol production; the proposed explanation was that problems with cholesterol transport may cause cells to make more cholesterol.[1] Human experts considered that DCTN1 hypothesis novel, consequential and worth following up.[1] PerturbAI and OpenAI say they will apply their AI reasoning frameworks to discover new drug targets.[3]