On October 21, 2026, OpenAI, PerturbAI, and Scripps Research announced that about 2,000 AI agents powered by GPT-Rosalind had analyzed and interpreted a large set of mouse brain experiments in 32 minutes. The system examines how cells respond when gene activity is altered, then cross-checks those results against the existing scientific literature to generate new biological hypotheses. Advances in experimental technology have made it possible to collect huge volumes of data in the life sciences, but interpreting that data one piece at a time has become a major burden for researchers. How far can dividing the analysis among many AI agents solve this problem? And how useful will the hypotheses the AI produces be in real research?
AI agents match each genetic perturbation to data and literature
The analysis targeted a "whole-brain CRISPR atlas," which records the effects of genetic perturbations across the entire mouse brain. An atlas, in this context, is a dataset that systematically compiles the activity and state of genes across many cells. According to PerturbAI, the underlying atlas covers about 8 million cells and was published in March 2026. The 32 minutes reported in the announcement is the time it took to analyze data that had already been collected and compile the findings. It does not mean the entire experimental process, including collecting and measuring cells, was completed in 32 minutes.
The research team assigned an independent AI agent to each genetic perturbation. Each agent analyzes the cellular response to its assigned perturbation, surveys the relevant scientific literature, and proposes a hypothesis that could explain the observed results. The individual reports are then combined into what the team calls a "reasoning atlas," which links genetic perturbations to cellular changes and the biological mechanisms that may underlie them. Researchers can use it to look up which cells change, and how, when a particular gene is perturbed, and which literature offers clues to explain the response.
One could imagine simply having a single AI agent process the entire dataset. But as the number of genes under analysis grows, so does the range of cell types, related papers, and research fields that must be consulted. In this approach, the detailed investigation of each perturbation is divided among many agents, and the results are integrated at the end to address that problem. The team aims to find new research targets by looking not only at genes with large, conspicuous effects but also at responses that have traditionally been overlooked.
The work is a collaboration led by Sun and colleagues, and a preprint has been posted on bioRxiv. The analysis code and reasoning results have also been made public. However, it has not yet been peer-reviewed and should not be considered a definitive research result. Its significance lies not in the AI having demonstrated new biological facts, but in its having extracted, at scale, hypotheses worth testing by linking experimental data with existing scientific knowledge.
What do "32 minutes" and "75%" actually mean?
OpenAI says it compared analysis by a single agent with distributed analysis by many agents, using the same AI model and the same computing resources. Under the heaviest workload condition, the single agent cited 11 papers, while the distributed approach cited more than 700. Investigating many genetic perturbations in parallel allowed it to draw on a much wider range of literature. However, citing more papers does not necessarily mean the analysis is more accurate or the hypotheses more correct.
The quality of the reports was also assessed through a blinded comparison by biology experts. According to PerturbAI, the experts examined the novelty and scientific grounding of the reports, and also rated the priority of follow-up experiments and the level of confidence in the conclusions. As a result, reports produced by the distributed agents were preferred 75% of the time. This figure shows which report the experts judged to be better; it does not mean that 75% of the hypotheses proposed by the AI were scientifically correct.
The speed claim also requires careful reading of the comparison conditions. The research team estimates that if humans were to investigate each genetic perturbation one by one, check the relevant literature, and write up the reports, it would take several months to about a year. But the team did not actually run a team of researchers and the AI side by side under identical conditions to compare the time required. It is therefore not possible to directly calculate from the 32-minute figure how many times faster drug discovery research as a whole would become.
The announcement also does not disclose details such as the number of experts who took part in the blinded comparison, the number of comparisons, the computing resources used, or the cost of the analysis. More detailed information on the conditions is needed to judge whether the 75% preference rate and the difference in citation counts would be reproduced for other research subjects or in other computing environments. At this stage, all that can be said is that the research team reports that, in this data analysis, the distributed approach broadened the range of literature explored and was also favored in expert comparisons of the reports.
Hypotheses emerging from cerebellar gene responses and cholesterol metabolism
One concrete lead found by the AI agents is a difference between two genes, Tsc1 and Tsc2. Both regulate the activity of mTOR, which controls cell growth and metabolism. In several groups of neurons, disrupting either Tsc1 or Tsc2 produced similar responses. In the cerebellum, however, the results differed. According to OpenAI, disrupting Tsc2 changed the expression of dozens of genes involved in communication between neurons, whereas disrupting Tsc1 caused relatively small changes.
This result suggests that even genes involved in the same regulatory system may have different effects depending on the brain region and cell type. OpenAI also points to existing findings that, in human tuberous sclerosis complex, cerebellar abnormalities are strongly associated with TSC2 mutations. The differences in gene expression observed in mice cannot be taken as the direct cause of patients' symptoms, but they provide clues for narrowing down the cells and molecular mechanisms that merit closer study. According to OpenAI, there are brain tumors and seizures that can be treated with drugs that suppress mTOR activity, but addressing difficulties with learning and behavior remains challenging. The results also point to a new question to test: in which cells do changes occur that existing drugs could improve?
Another notable result concerns the relationship between DCTN1, which is involved in transporting materials inside neurons, and cholesterol metabolism. Mutations in DCTN1 are known to cause Perry syndrome, a neurodegenerative disease with symptoms resembling Parkinson's disease. In the mouse experimental data, disrupting Dctn1 increased the expression of a group of genes involved in cholesterol synthesis. This observation itself is based on experimental data, but why the expression of cholesterol-related genes increased needs to be determined separately.
The AI agent linked this result to existing research showing that when problems arise in cholesterol transport within cells, cells may increase synthesis to make up for the shortfall. From there, it derived the hypothesis that reduced DCTN1 function alters the distribution of cholesterol within neurons and that this contributes to neurodegeneration. Whether restoring cholesterol distribution to a normal state could protect neurons is another question that could be tested.
What matters is to distinguish between facts observed in experiments and hypotheses proposed by the AI. What was confirmed here is the fact that, in mouse cells in which Dctn1 function was lost, the expression of a group of genes involved in cholesterol synthesis increased. The explanation that abnormal cholesterol distribution causes neurodegeneration, and the possibility that correcting that distribution could protect neurons, have not yet been confirmed experimentally. Even so, linking gene expression data with the existing literature has value in making concrete what researchers should measure next and what interventions they should try.
What can be verified with the released code and data?
The research team provides the mechanism and related materials in a public GitHub repository. The repository includes a guide to a hosted service where the generated reports can be viewed, as well as instructions for running the system locally. In the hosted version, users can browse the findings and claims extracted by the AI, along with the supporting evidence, in read-only form. However, because raw data is not included, it is not possible to load the original experimental data and recompute the statistics. Viewing the AI's reasoning results, reproducing the analysis, and verifying the results from raw data are each distinct tasks.
The ways of using the published materials, organized by what each allows you to verify, are as follows.
| How to use | What you can check | Requirements and limitations |
|---|---|---|
| Browse in the hosted version | Findings and claims reported by the AI, and the evidence supporting them | Raw data is not included, and statistics cannot be directly recomputed |
| Reproduce saved analyses | Figures, tables, and analyses for the paper based on published run logs and analysis results | Requires obtaining the repository and necessary data. Reproducing saved analyses requires no new model calls or API authentication |
| Verify using raw data | Access to experimental data and recomputation of statistics | Requires obtaining the raw data and running the necessary processing in a local environment |
According to the repository's documentation, reproducing saved analysis results does not require calling an AI model anew or performing API authentication. However, this does not mean that new AI analyses can be run for free and without limit. Moreover, even if the figures and tables published in the paper can be reproduced, that does not demonstrate the correctness of the biological hypotheses the AI derived. For example, testing the relationship between reduced DCTN1 function and cholesterol distribution would require actually measuring cholesterol inside cells and examining how those changes affect neurons.
PerturbAI has also indicated plans to expand its collaboration with OpenAI, which has run for about a year, into joint exploration of drug targets using its own in vivo experimental data. If AI can shorten the time needed to extract hypotheses from large volumes of experimental data, researchers may be able to devote more time to validating promising candidates. However, even if hypotheses are found faster, the step of confirming whether they hold in living organisms remains. For the proposed link between cholesterol distribution and neurodegeneration in this study, too, the promise of the target as a therapeutic option can be assessed with more confidence only once the existence of the distribution abnormality and the neuroprotective effect of correcting it have been confirmed experimentally.
Rather than demonstrating biological discoveries themselves, the result shows the potential of a new research method that efficiently finds, from vast amounts of data, hypotheses that researchers should test. How much the 32-minute analysis by about 2,000 agents contributes to future research and development will depend on the reproducibility of the proposed hypotheses and on the results of subsequent experiments.