Can brain scans define schizophrenia subgroups?
A peer-reviewed multisite study used resting-state fMRI from 746 healthy controls and 387 people with schizophrenia to derive two overlapping groups linked mainly to negative symptoms. The result is exploratory and cannot diagnose a patient or select treatment.
The short version
- What happened
- The discovery analysis retained 746 healthy controls and 387 patients across four datasets, then assigned 386 patients to two groups after additional feature-level quality control.
- Who it affects
- The work may help researchers focus on negative symptoms that strongly affect daily life but remain difficult to treat.
- What remains unproven
- The symptom-guided feature selection and symptom comparison used the same discovery cohort, so the clinical separation is partly built into the analytic design.
Keep in mind
The symptom-guided feature selection and symptom comparison used the same discovery cohort, so the clinical separation is partly built into the analytic design.
The two groups overlapped in the reduced imaging space and should not be treated as discrete biological diseases.
Leave-one-site-out analyses showed only moderate feature overlap and inconsistent direction of negative-symptom differences across sites.
The transcriptomic analysis links group-level spatial maps with postmortem gene-expression maps; it does not measure genes in the scanned participants or prove causation.
The 41-person rTMS analysis had no sham group, no correction across outcomes and no lasting four-week difference.
Source analysis · 1 linked source. Primary analysed source date: 7 October 2026.
Read the full report · 7 min The evidence, methods, implications and limitations
The scans produced a research grouping, not a diagnostic test
Researchers report two patterns of altered communication between the pulvinar, a thalamic hub involved in coordinating cortical information, and the rest of the cortex in people with schizophrenia. The groups were linked mainly to differences in negative symptoms such as reduced motivation, expression and social engagement. That may help researchers study why people with the same diagnosis can experience very different illness profiles.
The study does not show that a clinician can place a new patient into a biologically fixed subtype, predict an individual's future or choose a treatment from a scan. The two distributions were only partially separable and were created after selecting imaging features associated with symptom scores in the same discovery data. The paper itself calls for independent replication with a prespecified model before clinical or biological validity can be established.[1]
Four datasets supplied 1,133 people after imaging quality control
The discovery cohort combined resting-state functional MRI from four datasets: COBRE in the United States, the Japanese SRPBS multi-disorder dataset, BrainGluSchi and an in-house schizophrenia cohort. After exclusions for incomplete imaging or average framewise displacement above 0.5, 746 healthy controls and 387 people with schizophrenia remained. The sites differed substantially in age, sex and head motion, so the researchers modelled those variables and site effects rather than treating the pooled sample as uniform.
Each participant contributed correlations between left and right pulvinar seeds and 400 cortical parcels, yielding 800 connectivity edges. Models fitted to the healthy participants estimated expected connectivity for a person with given demographic, motion and site characteristics. Each patient's observed value was then compared with that reference. The resulting deviations were winsorised, sparsely missing values were imputed and the schizophrenia matrix was harmonised using neuroCombat. Every one of those choices can affect the patterns available to clustering.[1]
Symptoms helped choose the features before the algorithm formed groups
The team ranked connectivity deviations by their association with positive, negative and general scores on the Positive and Negative Syndrome Scale. It retained a top set of 200 symptom-related edges, reduced them using principal component analysis and compared different feature counts, clustering methods and numbers of groups. Five principal components explained 48.5% of variance in the selected feature space, and a two-cluster solution had the strongest silhouette score among the candidates.
This design is clinically motivated but it also limits interpretation. Because symptom scores helped select the imaging variables, a later difference in those same symptom domains is not a fully independent confirmation. The final allocation placed 181 patients in Biotype 1 and 205 in Biotype 2, one fewer than the 387 retained after overall imaging quality control. The paper's table and results support 387 as the available patient denominator and 386 as the clustered denominator, a distinction that should remain visible rather than being rounded away.[1]
Negative symptoms separated the groups, but not into clean categories
Biotype 1 had a mean negative-symptom score of 19.47, compared with 16.75 in Biotype 2. The standardised difference was moderate: Cohen's d 0.419, with a reported 95% confidence interval from 0.216 to 0.626. The difference remained after adjustment for site, age, sex and motion. Positive and general psychopathology scores did not differ significantly after correction, and a nominal total-score difference did not survive false-discovery-rate adjustment.
The imaging features were concentrated in somatomotor, visual and ventral-attention networks, with many of the strongest associations involving the right pulvinar. Yet the plotted groups overlap rather than forming two distinct islands. Leave-one-site-out analyses also exposed sensitivity: selected-edge overlap with the full model was moderate, and negative-symptom differences were nominally significant in three of four folds but did not keep the same direction across sites. That variability weakens any claim of a universal subtype boundary.[1]
Gene-expression links and treatment results are exploratory layers
The authors compared the cortical imaging maps with postmortem gene-expression patterns from the Allen Human Brain Atlas. One group showed spatial enrichment involving translation, protein targeting and synaptic organisation; the other involved neurodevelopment, synaptic remodelling and metal-ion homeostasis. Spatial permutations and bootstrap procedures were used, but these are correlations between two maps. They do not show that a gene programme caused an individual's connectivity pattern or establish two distinct molecular diseases.
The frozen discovery model was also projected into an independent cohort of 41 people receiving a 20-day course of low-frequency right orbitofrontal rTMS alongside antipsychotic medication. At two weeks, the predicted groups differed nominally in negative-symptom improvement, with uncorrected p values of 0.048 for percentage improvement and 0.018 for absolute change. There was no sham group, no multiple-comparison correction and no significant four-week difference. The study therefore provides no basis for selecting rTMS by scan.[1]
What this means for people—and what would change the assessment
For people living with schizophrenia, better characterisation of negative symptoms could matter because these symptoms often impair relationships, work and daily functioning and can be less responsive to existing treatment. A research framework that respects individual variation is preferable to assuming every brain differs in the same way. But prematurely labelling someone with a scan-derived biotype could add false certainty, stigma or pressure toward a treatment that has not been validated for that group.
Confidence would rise if an external team froze the full preprocessing, feature-selection and clustering pipeline before applying it to a large, prospectively recruited cohort from different scanners, health systems and populations. The groups would need stable assignment over time, consistent symptom differences and incremental value beyond ordinary clinical assessment. A preregistered, sham-controlled treatment study would then need to show a replicated interaction between subgroup and outcome at a clinically meaningful endpoint. Until then, the result is a useful hypothesis about heterogeneity, not precision psychiatry ready for patients.[1]