OpenAI's GPT-Rosalind Mapped a Brain CRISPR Atlas in 32 Minutes

OpenAI, PerturbAI, and Scripps Research used thousands of reasoning agents to analyze gene-disruption data from millions of mouse brain cells. The system, called GPT-Rosalind, completed its computational pass in 32 minutes and produced hypotheses for neurodegenerative disease research that researchers can test in the lab.

The project report describes a distributed reasoning workflow that divides a large dataset into focused assignments. Each agent examined one gene disruption, measured the cellular response, searched relevant literature, and proposed mechanisms that could explain the result. An aggregation step combined those reports into a reasoning atlas linking genes, cell types, effects, citations, and candidate mechanisms.

One atlas, thousands of assignments

PerturbAI’s whole-brain CRISPR atlas records how disrupting thousands of genes changes gene expression across mouse brain cell types. CRISPR provides the perturbation, while gene-expression measurements reveal how cells respond when a particular gene stops functioning normally.

The team estimated that researchers analyzing each perturbation and reviewing its literature sequentially would need several months to a year. GPT-Rosalind completed the agent analysis in 32 minutes. That timing covers interpretation of an existing dataset; generating the atlas and validating the resulting hypotheses require separate experimental work.

Blinded reviewers also compared the system’s judgments with assessments from human experts. According to the researchers, its ratings of scientific novelty and value tracked the expert evaluations.

Cholesterol transport and a cerebellar split

One hypothesis centers on DCTN1, mutations in which can cause Perry syndrome, a rare Parkinson-like neurodegenerative disorder. Disrupting the mouse gene Dctn1 increased expression of genes involved in cholesterol production. The agents connected that response to studies suggesting that impaired cholesterol transport can prompt cells to synthesize more cholesterol.

That connection yields a concrete experiment: determine whether the increase in cholesterol production compensates for faulty distribution inside neurons, then test whether restoring normal transport reduces degeneration. The atlas establishes the expression pattern, while the proposed mechanism remains unvalidated.

A second hypothesis concerns Tsc1 and Tsc2, genes that jointly regulate mTOR, a signaling pathway involved in cell growth, metabolism, and aging. Disrupting either gene produced similar effects in several neuronal populations. In the cerebellum, however, Tsc2 disruption changed the activity of dozens of genes involved in neuronal communication, while Tsc1 disruption produced little change.

The cell-type-specific difference aligns with clinical observations that TSC2 mutations tend to cause more severe tuberous sclerosis. Establishing a causal link will require targeted experiments in relevant cells and animal models.

Parallelism keeps more evidence in view

OpenAI compared the distributed workflow with a single agent given the entire dataset. Both configurations used the same model and total compute budget, allowing the benchmark to focus on task decomposition and context management.

The single agent’s performance declined as the workload grew because one context had to hold the dataset, literature, intermediate reasoning, and final synthesis. Narrower agents could investigate weaker signals without displacing evidence related to other perturbations. The aggregation stage then reconciled their reports.

Reference volume alone does not establish citation quality, and reviewer preference does not validate a biological mechanism. The benchmark shows that decomposition preserved more literature and produced reports experts preferred within this workload and compute budget.

Code, data, and boundaries

The release includes a bioRxiv preprint, the CRISPR atlas on Hugging Face, and the reasoning atlas with source code on GitHub. Researchers can inspect agent-generated reports, trace claims to citations, compare related genes, and identify hypotheses for follow-up.

The manuscript is a preprint and has not completed peer review. Its findings come from mouse cells, generated mechanisms still need experimental confirmation, and the architecture was tested on a workload that divides cleanly by gene perturbation. OpenAI provides academic access to GPT-Rosalind through its Rosalind Academic Plan.

Where the architecture fits

Developers can apply the same design to research tasks containing many independent records and a large external literature. A practical implementation needs four stages:

1. Partition the workload into narrow units with explicit inputs and output schemas.
2. Run retrieval and reasoning separately for each unit so evidence remains within a manageable context.
3. Preserve provenance by attaching citations, dataset identifiers, and intermediate results to every claim.
4. Reconcile the outputs by deduplicating findings, resolving conflicts, and ranking hypotheses for review.

Tasks with tightly coupled records may require shared state or additional coordination between agents. For partitionable datasets, GPT-Rosalind provides a concrete example of how parallel reasoning can expand evidence coverage while keeping the model and overall compute budget fixed.