Mapping Neurochemical Signatures onto Brain Structure for Neurotransmitter-Informed Discrimination of Schizophrenia Patients from Healthy Controls

BackgroundSchizophrenia (SCZ) is associated with widespread gray matter volume (GMV) reductions, yet the underlying mechanisms driving these alterations remain unclear. This study investigates whether SCZ-related GMV changes co-localize with normative neurotransmitter (NT) density maps derived from healthy individuals. We further examine the diagnostic utility and transdiagnostic relevance of these NT-informed structural patterns within a machine learning (ML) framework. MethodsStructural imaging from 445 SCZ patients and 414 matched healthy controls (HC) were used to train two ML models: (1) an NT-informed model based on spatial cortical and subcortical correlations between GMV and 25 normative NT maps, and (2) an NT-unaware control model using 50 principal components of GMV. Both models were tested for transdiagnostic generalizability in major depression (MDD), bipolar disorder (BD), borderline personality disorder (BPD), and attention-deficit/hyperactivity disorder (ADHD) to determine whether NT-informed and NT-uninformed morphometric patterns reflect psychosis-specific mechanisms. Associations with accelerated brain aging, clinical symptoms, and psychotropic medication exposure were evaluated. Individual feature importance profiles from the NT model were used for clustering to identify potential neurochemical subtypes of SCZ. ResultsThe NT model had a higher sensitivity, but a lower specificity compared to the control model in distinguishing SCZ from HC (sensitivityNT = 61.8% vs. sensitivityPCA = 58.4%; specificityNT = 62.3% vs. specificityPCA = 81.9%). The NT models SCZ-predictive features localized primarily to subcortical regions and mapped onto serotonergic (5-HT4), endocannabinoid (CB1), and cholinergic (VAChT) systems, as well as synaptic (SV2A) and gene expression regulation (HDAC) markers. In contrast, the control model relied on medial temporal and subcortical structures, and its decision scores were associated with accelerated brain aging (R2 = 0.11). The NT model showed modest transfer to BPD (BAC = 61.8%, p = 0.003) and MDD (BAC = 56.6%, p = 0.036), while the control model performed better for MDD (BAC = 61.2%, p < 0.001) and BD (BAC = 61.0%, p = 0.002). Finally, clustering of individual NT feature contributions identified two transdiagnostic NT subtypes: a subcortical dopaminergic/COX-1/mGluR5-driven subtype and a subtype with diffuse cortical and subcortical serotonergic-GABAergic-glutamatergic-cholinergic involvement. ConclusionsSpatial correlations with normative NT distributions provide biologically interpretable insights into SCZ-related GMV alterations, particularly within subcortical systems. The NT-informed model identifies mechanistically distinct subtypes potentially relevant for future NT-based stratification of SCZ. These findings support the future integration of molecular and structural imaging to uncover system-specific vulnerabilities in psychosis and guide more personalized approaches to diagnosis and treatment.