Image: OpenAI

OpenAI says thousands of AI agents running its GPT-Rosalind model sifted through data from millions of mouse brain cells in 32 minutes and turned up new leads on brain disease. The company described the work on its blog with partners PerturbAI and Scripps Research. The results are in a preprint that has not yet been peer reviewed.

GPT-Rosalind surfaced a new lead for drug discovery: could restoring cholesterol transport protect neurons from degeneration?

DCTN1 is a gene linked to Perry syndrome, a rare neurodegenerative disorder. In mouse-brain data, disrupting it increased expression of genes involved in…

— OpenAI (@OpenAI) October 21, 2026

The dataset is a whole-brain CRISPR atlas built by PerturbAI. Scientists switched off nearly 2,000 genes one at a time in mice and recorded how each change affected different brain cells. That produces a mountain of numbers. Making sense of it means comparing each result against decades of papers spread across many fields. The team estimates a human doing that one gene at a time would need months to a year.

This is why biology is hard for AI. A chatbot can read text well, but lab data is noisy, huge and full of weak signals. A real finding often only makes sense next to an old paper from a different field. OpenAI's answer was to split the job. Each agent took one gene, studied the cell data, searched the literature and wrote an explanation. The reports were then merged into a searchable "reasoning atlas."

OpenAI says splitting the work mattered. In its tests, a single agent given the whole dataset cited only 11 references at the largest workload. The distributed agents cited more than 700. Blinded human experts preferred the split approach 75% of the time, according to OpenAI. The preprint says the agents recovered 94.5% of the findings in the original study that could be checked directly.

The headline lead involves DCTN1, a gene linked to Perry syndrome, a rare disease with Parkinson-like symptoms. Switching it off in mice raised activity in cholesterol-making genes. The agents tied that to research suggesting poor cholesterol transport can push cells to make more. That raises a question for the lab. Could fixing cholesterol transport protect nerve cells? Nobody has tested it yet.

The work fits a fast-moving push. A day earlier, OpenAI bought Orchestra Bio, a startup that manages drug development work, and folded it into Rosalind. PerturbAI says its year-long work with OpenAI is now expanding into joint drug target discovery. OpenAI is building tools for the whole chain, from reading data to running programs.

Rivals are taking different routes. Google DeepMind's AlphaFold predicts the shapes of proteins, and its AlphaGenome predicts how DNA changes affect genes. Those are specialised models trained on biology itself. Anthropic recently showed Claude designing proteins that bind scorpion toxins, which is making new molecules. OpenAI's study is about something else. It uses a general reasoning model to read existing results and suggest what to test next.

For ordinary people, the promise is that scientists waste less of the data they already collect. Big experiments often get mined only for their strongest signals. Cheap, fast analysis could surface leads that would otherwise sit unread.

The limits are real. The data comes from mice, and mouse findings often fail in humans. The leads are ideas, not results, and agents can sound convincing while being wrong. The study is not peer reviewed, and the expert ratings came from a setup the partners designed. PerturbAI also has a commercial stake in the outcome. The true test is whether any of these leads holds up in the lab.