The Limits of the Cartesian Machine

An MRI asks little of the person inside it. You lie on a narrow table with your head in a plastic cradle. For eight minutes, a superconducting magnet tips the protons in your brain and listens as they relax. Software uses that timing to reconstruct the folded sheet of your cortex and estimate its thickness, usually two to four millimeters, at more than three hundred thousand points. You do not have to think about anything. The magnet measures tissue on the theory that years of feeling have left a mark. Sadness itself remains beyond its field.

The familiar result is an inflated cortical surface, smoothed into a walnut without its wrinkles. Blue marks the places where patients’ gray matter runs thinner than controls’. Red marks the places where it runs thicker. Every such picture is a group average, often built from a few dozen patients and a few dozen controls. It carries a promise: collect enough scans in enough places and the colors should settle. Depression should acquire a face.

Two papers in Nature Neuroscience, published two weeks apart this summer, put that promise under pressure from opposite directions. One asks whether psychiatric brain maps reproduce across sites. The other asks whether human cortical neurons can be treated as scaled-up mouse neurons. Both come from researchers invested in the methods they audit. The criticism is coming from inside the work.

Why expect a diagnosis to settle into a map? Because localization has worked before. In 1861, Paul Broca autopsied a patient known as Tan. Louis Victor Leborgne had spent twenty-one years able to produce almost no syllable but that one. Broca found a lesion in his left frontal lobe. A specific loss of function met a specific hole in the tissue. For the next century, neurology filled in its map one tragedy at a time.

Psychiatry inherited that hope. Around the turn of the twentieth century, Emil Kraepelin divided severe mental illness into dementia praecox and manic-depressive illness according to symptoms and course. He assumed distinct pathology would eventually vindicate those categories. Schizophrenia would become a thing in the tissue, much as tabes dorsalis had. When MRI arrived in the 1980s, it looked like an autopsy without death. Scan living patients, compare them with controls, and let the structural signature emerge.

Instead, the findings scattered. Researchers reported both more and less gray matter in the striatum of people with schizophrenia. The anterior cingulate was larger in some bipolar studies and smaller in others. Depression appeared in the orbitofrontal cortex, the insula, the thalamus, or nowhere, depending on the paper.

At first, the field blamed sample size. That was a hopeful diagnosis: signal buried under noise. Forty patients per group could not pin down a subtle effect, so the obvious answer was to pool. Paul Thompson and colleagues founded the ENIGMA consortium in 2009 on that logic. By 2018, an ENIGMA analysis could compare cortical measures in 4,474 people with schizophrenia against 5,098 controls. In 2022, Scott Marek and colleagues argued in Nature that brain-wide association studies need thousands of participants before their findings replicate. More people, cleaner map.

Trang Cao, Alex Fornito, and colleagues at Monash University tested that explanation directly. They assembled scans from fifty-nine sites: 2,437 patients with schizophrenia, schizoaffective disorder, autism, major depression, or bipolar disorder, plus 2,065 controls. They kept the sites separate, making a case-control difference map for each one, processing them all through the same pipeline, and asking how well any two agreed. The test was generous. Published studies usually differ in their processing choices, which gives them one more way to diverge.

The Alzheimer’s disease benchmark showed that the machinery can work. Across seven sites, the median correlation between one Alzheimer’s map and another was 0.54, with some pairs above 0.7. Different scanners in different cities, looking at different patients, still recovered a recognizable structural footprint.

Nothing comparable happened in psychiatry. Schizophrenia produced the strongest result, a median correlation of 0.16. Schizoaffective disorder followed at 0.15. Bipolar disorder reached 0.06, autism 0.04, and major depression 0.01.

The last point deserves a slower reading. For autism and depression, the observed agreement was statistically indistinguishable from the agreement between randomized maps with the same spatial smoothness. Two depression maps made from real patients on real scanners resembled each other about as much as two structured noise fields did.

The researchers pushed on the result. They changed the amount of smoothing, added and removed statistical thresholds, swapped vertices for coarse anatomical parcels, and turned harmonization on and off. Agreement barely moved. They also tested nineteen demographic, clinical, and scanner variables. Sites with similar ages, sex ratios, medication patterns, illness duration, or hardware did not reliably agree more. The inconsistency could not be pinned on one obvious technical choice.

That leaves two broad explanations. One keeps the diagnoses intact. The effects may be real, small, and highly individual. Each patient’s brain differs from the norm somewhere, just not in the same place as the next patient’s. A group average then erases the very signal it was meant to reveal. The schizophrenia data give this view some support: cross-site agreement rises sharply once samples pass roughly two hundred people per group. A comparison between two enormous depression datasets, UK Biobank and ENIGMA, reached 0.64. Cao and colleagues estimate that depression studies may need thousands of participants.

The second explanation questions the categories. A diagnosis assembled from symptoms and committee rules, and satisfiable by many combinations of those symptoms, may not correspond to one structural pattern. “Depression” could name a family of conditions that happen to share a waiting room.

The study cannot choose between these accounts. The first calls for larger samples. The second calls for better categories. For an ordinary case-control study, the practical result is the same: the group average does not yield a stable portrait.

When whole-brain averages disappoint, moving down a level is tempting. Cells seem firmer than diagnoses. They can be stained, traced, and perturbed. Much of that work happens in mice, with an assumption so familiar that it often goes unstated: a human cortical neuron is basically a larger mouse cortical neuron.

Zhixi Yun, Wen Ye, and colleagues in Hanchuan Peng’s group carried out the most systematic test of that assumption yet. They reconstructed the dendrites of 2,363 human cortical neurons, dye-filled one by one in tissue removed during surgery from twenty-three patients, and compared them with 16,011 mouse neurons traced from whole-brain image volumes.

The matching was unusually careful. Human and mouse regions were aligned by anatomy, function, and cross-species transcriptomic cell composition. Those criteria determined the closest mouse counterpart for a neuron from the human superior parietal lobule; anatomical proximity alone was insufficient. The researchers also normalized for size. Human neuronal cell bodies were about 1.68 times larger, and their dendrites were larger too. The question was whether the same architecture remained after scale was divided out.

The shapes still did not match. Human dendrites branched more often and at shorter intervals, with more first-order branches leaving the cell body. Uniformly enlarging a mouse neuron could not produce that compact, densely branching local architecture.

Other groups had seen pieces of the same difference. Ruth Benavides-Piccione and colleagues reported in 2020 that human hippocampal pyramidal cells are not stretched mouse cells. In 2025, Lida Kanari and colleagues used topological methods to identify a dense collar of branching near the cell body in human layer 2/3 pyramidal neurons, another feature that scaling laws missed. The new study adds breadth. It also found that human neurons vary more across cortical lobes than mouse neurons do, suggesting that regional specialization reaches down into the shape of the cell.

Mouse neuroscience remains indispensable. Dendritic integration, plasticity, and the molecular machinery of the synapse are conserved well enough that mouse work helped build modern neuroscience. The caution concerns the distance between a model and its target. Shorter branch intervals change how synaptic currents weaken on their way to the cell body. Denser proximal branching changes what a neuron can compute. A mouse neuron can answer a nearby question beautifully while missing a human cortical computation.

Descartes belongs here for a reason more interesting than the usual complaint about dualism. He wanted an address where the account resolved. For him it was the pineal gland, a single coordinate where mind met mechanism. Psychiatric imaging has often made a similar wager. Somewhere on the cortical sheet, the average would converge and the disorder would be located.

That wager inspired a remarkable amount of work. A PubMed search for structural MRI studies of schizophrenia returns 2,819 papers from 1986 through 2024. ENIGMA was the field’s honest response to the limitations of small samples: combine them. After four decades, the strongest cross-site audit finds a median agreement of 0.16 for schizophrenia.

The Alzheimer’s benchmark shows that MRI can recover a shared structural footprint when one exists. The psychiatric results change the question we can reasonably ask of a group average. Depression can be wholly biological without producing one common map. The biology may be distributed differently in different people, or the diagnosis may join several biological routes under one name.

Researchers closest to this problem are already changing the unit of analysis. Normative modeling charts the expected range of brain measures across a population, much like a pediatric growth curve, then asks where one person departs from it. Several of the Monash authors helped develop the approach. Precision imaging makes another trade: hours of scanner time for one person instead of a few minutes each for forty people.

These methods are slower, more expensive, and less likely to produce one iconic brain map. What they produce instead is an individual trajectory: one cortex measured well, followed over time, and compared with itself.

Eight minutes later, the table slides out. The scan belongs to one person. We should probably start there.

Sources

1. Cao, T., Pang, J. C., et al. “The cross-site reproducibility of MRI morphometric phenotypes in psychiatric disorders.” Nature Neuroscience (2026). doi:10.1038/s41593-026-02359-0.
2. Yun, Z., Ye, W., Ji, N., et al. “A framework for comparative analysis of human and mouse cortical neuron dendrites in corresponding brain regions.” Nature Neuroscience (2026). doi:10.1038/s41593-026-02376-z.
3. Marek, S., et al. “Reproducible brain-wide association studies require thousands of individuals.” Nature 603 (2022): 654–660. doi:10.1038/s41586-022-04492-9.
4. Benavides-Piccione, R., et al. “Differential structure of hippocampal CA1 pyramidal neurons in the human and mouse.” Cerebral Cortex 30 (2020): 730–752. doi:10.1093/cercor/bhz122.
5. Kanari, L., Shi, Y., et al. “Of mice and men: Dendritic architecture differentiates human from mouse neuronal networks.” iScience 28 (2025): 112928. doi:10.1016/j.isci.2025.112928.
6. van Erp, T. G. M., et al. “Cortical brain abnormalities in 4474 individuals with schizophrenia and 5098 control subjects via the ENIGMA Consortium.” Biological Psychiatry 84 (2018): 644–654.
7. Thompson, P. M., et al. “The ENIGMA Consortium: large-scale collaborative analyses of neuroimaging and genetic data.” Brain Imaging and Behavior 8 (2014): 153–182.
8. Kochunov, P., et al. “Functional vs structural cortical deficit pattern biomarkers for major depressive disorder.” JAMA Psychiatry 82 (2025): 582–590.