The technique at the heart of the study departs from conventional neuroimaging analyses that focus on the size or thickness of brain regions. Instead, it examines texture, the fine-grained spatial patterns of intensity within brain tissue as captured on magnetic resonance images. Texture features are sensitive to subtle microscopic changes in tissue organization, such as alterations in cell density, myelination, or vascularity, that conventional structural measures can miss. The researchers computed a rich set of radiomic texture descriptors for each brain region and then, crucially, quantified how similar each individual’s texture profile in one region was to their own profiles in other regions. This produced, for every participant, an individualized network in which brain regions were connected by the similarity of their tissue texture rather than by anatomical wiring or synchronized activity.

The logic behind this texture similarity network is that healthy brains tend to show characteristic, coordinated patterns of tissue properties across regions, and that disease may disrupt this coordination in region-specific ways. By comparing each patient’s network against a reference derived from healthy controls, the team could generate a map of network-level texture disruption for every individual. Because these maps are computed per person rather than averaged across groups, they preserve the individual variability that most case-control studies deliberately wash out, and they provide a high-dimensional feature space in which biologically distinct patient groups might separate from one another.

To test the approach, the investigators assembled a substantial multi-site cohort: 437 individuals with schizophrenia, 263 men and 174 women, and 431 healthy controls, 246 men and 185 women, recruited across four sites and scanned on five different MRI scanners. Multi-site data of this kind present a formidable technical challenge, because scanner differences can introduce artifacts that masquerade as biological signals. The team’s ability to find stable subtypes across this heterogeneity of hardware and population is itself an important demonstration that the texture similarity signal is robust enough to survive real-world clinical conditions rather than the pristine single-scanner settings of many imaging studies.

Applying k-means clustering, a widely used machine learning algorithm that partitions data into groups based on feature similarity, to the individual disruption maps, the researchers found that patients fell cleanly into two subtypes. The clustering was not a one-off result: the two subtypes were validated across different samples, different datasets, and different stages of the illness. This kind of reproducibility is the exception rather than the rule in psychiatric neuroimaging, where subtype claims have often failed to replicate. The stability of the division across disease stages is particularly noteworthy, because it suggests the subtypes may reflect underlying biological differences rather than transient states such as acute psychosis or medication effects.

The two subtypes were distinguished by sharply different patterns of texture disruption. One pattern primarily involved connections between prefrontal cortex and sensorimotor regions together with prefrontal-limbic links, while the other was dominated by disruptions between subcortical structures and sensorimotor and limbic areas, along with altered texture relationships within the subcortical structures themselves. These are not arbitrary partitions of the brain. The prefrontal cortex is central to the cognitive and negative symptoms of schizophrenia, the limbic system to emotion and motivation, the sensorimotor network to the motor abnormalities increasingly recognized in the disorder, and subcortical structures to reward processing and psychosis. Distinct disruption of these circuits in different patients points toward divergent pathophysiological routes to what clinicians currently label with one word.

What elevates the study beyond a clustering exercise is the molecular grounding of the two subtypes. The team examined whether the regional pattern of texture disruption in each subtype was associated with the spatial expression of schizophrenia risk genes, drawing on the well-established principle that genes with elevated expression in particular brain regions shape those regions’ vulnerability to disease. The disruption patterns of the two subtypes were indeed differentially associated with the expression of schizophrenia risk genes, and those genes were enriched in biological processes including protein binding, nervous system development, and neuron-specific structures. In other words, the two imaging-defined patient groups appear to carry distinct transcriptional signatures, hinting that different developmental and molecular mechanisms may underlie each form of the illness.

The molecular analysis went a step further, into the realm of neurotransmitter systems. Using publicly available maps of receptor density distributions across the brain, the researchers tested how the density of several neurotransmitter receptors related to the degree of texture disruption in each subtype. Receptors including the muscarinic M1 receptor, the 5HT6 serotonin receptor, and the metabotropic glutamate receptor 5, mGluR5, showed broadly diverse associations with disruption between the two subtypes. This finding has obvious pharmacological resonance. M1 receptors are targets of interest for cognitive symptoms, serotonin receptors are the focus of next-generation antipsychotics, and glutamatergic targets such as mGluR5 have long been pursued as alternatives to dopamine-centered treatments. If the two subtypes genuinely differ in how their pathology relates to these receptor systems, they may also differ in which drugs are most likely to help them.

The implications for clinical practice are considerable, though the authors and the field at large would caution that translation will take time. A texture similarity network can in principle be computed from a standard structural MRI scan, an examination that is inexpensive, widely available, and already routinely performed in psychiatric care. If the two subtypes, or a finer subdivision revealed by future work, prove predictive of treatment response or prognosis, radiomic stratification could become an objective biomarker that guides initial treatment selection, replacing the current sequential trial-and-error approach. The study’s multi-site validation and the planned public release of the data on GitHub will make it easier for other groups to test, refine, and extend the method, which is exactly what a candidate biomarker needs before it can move toward the clinic.

For the broader field of psychiatry, the study adds to a growing conviction that the path forward lies in biologically grounded stratification rather than in ever-broader diagnostic categories. Schizophrenia’s heterogeneity has been described across symptoms, cognition, genetics, and neuroimaging for decades, but few approaches have connected individual-level imaging phenotypes to gene expression and receptor architecture in a single reproducible framework. By showing that the texture of brain tissue, a property most radiologists never consciously assess, can carry enough information to divide patients into stable, molecularly distinct groups, the work suggests that the answers to psychiatry’s hardest classification problems may have been hiding in plain sight, embedded in the images clinicians already collect. The two subtypes identified here are unlikely to be the last word, but they offer a concrete, testable template for how precision psychiatry might actually be built.

Subject of Research: Neuroimaging-based stratification of schizophrenia into subtypes using texture similarity networks and molecular signatures

Article Title: Texture similarity network uncover schizophrenia subtypes with unique molecular signatures

Article References: Luo, Y., Lin, L., Zhang, X., Li, X., Zhang, Y., Xie, Y., Chang, Z., Du, X., Wei, X., Ji, Y., Wei, L., Zhao, Z., Liang, M., Liu, H., Ding, H., Yu, C., & Qin, W. (2026). Texture similarity network uncover schizophrenia subtypes with unique molecular signatures. BMC Psychiatry. https://doi.org/10.1186/s12888-026-08606-9

Image Credits: AI Generated

DOI: 10.1186/s12888-026-08606-9

Keywords: schizophrenia, texture similarity network, neuroimaging, brain subtypes, radiomics, gene expression, neurotransmitter receptors, precision psychiatry, machine learning, biomarkers, BMC Psychiatry, Texture