OpenAI uses thousands of AI agents to study the brain

OpenAI, PerturbAI, and Scripps Research used thousands of AI agents to analyze data from millions of mouse brain cells in 32 minutes. The system found disease-related connections involving conditions such as Alzheimer’s and Parkinson’s that could open up new lines of research.

This is not a treatment or a confirmed clinical discovery. The AI examined existing experimental data and proposed hypotheses for scientists to test in the lab.

Thousands of analyses in parallel

The team divided the work among thousands of agents powered by GPT-Rosalind, a model specialized in biological reasoning. Each agent studied a genetic alteration, observed how cells responded, reviewed scientific papers, and proposed an explanation.

The results were assembled into a reasoning atlas linking genes, cellular effects, scientific papers, and possible biological mechanisms. The goal is to turn a vast number of observations into specific questions that can be investigated.

The data came from a brain atlas created using CRISPR. This type of experiment makes it possible to alter thousands of genes and measure what happens in different cell types. The problem is that a human team can take months or even a year to review each alteration and connect it with decades of scientific literature.

According to the researchers, the agents reproduced findings previously obtained by humans and uncovered additional clues. Experts who evaluated the results without knowing whether they came from people or AI also considered their assessments of novelty and scientific value comparable.

A difference between two related genes

One of the most specific clues emerged from comparing Tsc1 and Tsc2, two genes that work together to regulate mTOR, a pathway linked to cell growth, metabolism, and aging.

In several groups of neurons, altering either gene produced similar effects. But in the cerebellum, the difference was notable: modifying Tsc2 changed the activity of dozens of genes involved in neuronal communication, while modifying Tsc1 produced barely any changes.

That result could be relevant to tuberous sclerosis, a disease in which TSC2 mutations are associated with cerebellar abnormalities. Medications that reduce mTOR activity can help with some brain tumors and seizures, but learning and behavioral problems remain difficult to treat.

The AI did not solve that problem. It did identify a specific group of neurons and cellular changes that researchers could study to determine which alterations might be corrected with existing treatments and which require new strategies.

Another clue: cholesterol and neurodegeneration

The system also focused on DCTN1, a gene whose mutations cause Perry syndrome, a neurodegenerative disease with symptoms similar to Parkinson’s.

When the researchers altered the Dctn1 gene in the brain atlas, the activity of genes related to cholesterol production increased. The agents connected that observation with studies suggesting that problems transporting cholesterol within cells can cause them to produce more of it.

That led to a testable hypothesis: altering DCTN1 could change how cholesterol is distributed inside neurons and contribute to their deterioration. Further experiments will have to determine whether restoring that distribution can protect them.

Human experts considered the idea novel and worthy of follow-up. But it remains a possibility, not evidence that cholesterol causes the disease or a promise of treatment.

Why one agent is not enough

The team compared this distributed system with a single agent that received the same dataset and the same resources. As the workload increased, the individual agent consulted less literature: in the most demanding test, it cited 11 references, compared with more than 700 from the agents working in parallel.

The difference matters because scientists tend to focus on the strongest signals. A distributed system can also review subtle or unexpected responses and connect them with research from other fields.

For you, the change is not that AI will diagnose Alzheimer’s or Parkinson’s tomorrow. It is that AI can help determine sooner which experiments make the most sense, especially when the data is too extensive to review by hand in full.

The reasoning atlas is already available for other researchers to compare genes, examine unusual responses, and select new hypotheses. The next step does not come from the model, but from the lab: testing which of these connections hold up in experiments and which were only coincidences.