Written by Ralph Sun
Thousands of AI agents worked through data from millions of mouse brain cells in thirty-two minutes — work the team estimates would take a human months to a year. The interesting part is not the speed. It is that the clues were already there.
OpenAI, PerturbAI and Scripps Research deployed thousands of AI agents to analyse existing experimental data from millions of mouse brain cells. The analysis took 32 minutes. The team estimates that comparable human analysis and literature review, conducted one perturbation at a time, would take months to a year.
What the agents actually did matters more than the headline number. They worked through a whole-brain CRISPR atlas — a dataset showing how gene disruptions affect different cell types across the mouse brain. Each agent took one gene disruption, analysed the cellular response, searched the scientific literature, and proposed an explanation. Their reports were then combined into a “reasoning atlas” linking genes, cellular effects, relevant publications and possible mechanisms. The team calls the method distributed biological reasoning.
Two of the leads point at brain disease. The first concerns Tsc1 and Tsc2, which work together to control mTOR, a major regulator of cell growth, metabolism and aging. Disrupting either gene had similar effects in several groups of neurons. But in the cerebellum, which helps coordinate movement, Tsc2 disruption altered the activity of dozens of genes involved in neuronal communication while Tsc1 disruption produced little change. That asymmetry may matter clinically: in people with tuberous sclerosis, cerebellar abnormalities are strongly linked to TSC2 mutations, and while medicines that reduce mTOR activity treat certain brain tumours and seizures, the learning and behavioural difficulties associated with the disease remain hard to address.
The second concerns DCTN1, a gene whose mutations cause Perry syndrome, a neurodegenerative disease with Parkinson-like symptoms. DCTN1 is involved in moving cargo inside neurons. Disrupting it increased the expression of genes involved in making cholesterol, which connects to research suggesting that problems with cholesterol transport cause cells to produce more cholesterol. The question the agents raised is whether DCTN1 disruption changes how cholesterol is distributed within neurons, and whether restoring normal distribution could protect them.
The measurement I find most telling came from a control experiment. The team tested whether a single agent given the entire dataset could match thousands working in parallel, holding model and resources fixed. As the workload grew, the single agent searched less of the literature — citing only 11 references at the largest workload, against more than 700 from the distributed agents. In blinded comparisons, human experts preferred the distributed reports 75% of the time.
Read that carefully, because it is a specific failure mode rather than a general point about scale. Under load, a single agent narrows. It stops reading, follows the strongest signal, and produces a confident answer from less evidence. That is recognisably how a human researcher behaves under deadline pressure, which makes the result less a story about machine intelligence than about attention as a scarce resource.
Which brings me to the framing the researchers use, and the reason this belongs in a neuro AI column rather than a computer science one. Modern experiments generate far more data than scientists can thoroughly investigate. Researchers can now alter thousands of genes and measure their effects across millions of cells. Understanding what that means for human disease requires connecting complex experimental results with decades of scattered knowledge. The agents did not discover new biology. They read the biology that already existed, at a scale no team could.
The honest limits are worth stating plainly. These are hypotheses, not findings. Nothing was validated at the bench. The agents reproduced findings from human analyses and their novelty assessments aligned with blinded experts, but agreement with expert judgement is not evidence that a proposed mechanism is real.
There is also a second-order risk. A reasoning atlas that is cheap and persuasive can flood a field with plausible, untested leads. If the downstream bottleneck is experiments, then generating more hypotheses does not accelerate science so much as lengthen the queue. The team’s own framing — to help prioritise which experiments to run — is the right one, and the atlas is public, which at least invites scrutiny.
The best line in their write-up is about what was already there: clues scattered across different fields, never connected. That is a claim about attention, not intelligence. And attention has been the scarce resource in biology for a very long time.
