Categorical and Dimensional Alterations Along Two Principal Cortical Gradient Axes Across the Schizophrenia-Bipolar Spectrum

Technical Summary: Categorical and Dimensional Alterations Along Two Principal Cortical Gradient Axes Across the Schizophrenia-Bipolar Spectrum

Problem Statement
Schizophrenia (SZ) and bipolar disorder (BD) are severe psychiatric conditions characterized by significant clinical overlap, shared genetic risk, and convergent neurobiological alterations, particularly in dopaminergic and motivational circuitry. Despite these overlaps, traditional categorical diagnoses often fail to capture the heterogeneity within and between these disorders. While large-scale meta-analyses have identified widespread structural and functional brain disruptions, it remains unclear how macroscale functional connectivity gradients—the continuous spatial axes that organize cortical regions into brain-wide hierarchies—are altered across the SZ-BD spectrum. Specifically, it is unknown whether gradient organization reflects shared transdiagnostic alterations, disorder-specific signatures, or how individual-level gradient variations map onto dimensional clinical phenotypes.

Methodology
The study utilized resting-state functional MRI (rs-fMRI) data from 187 adults (110 healthy controls [HC], 37 SZ, 40 BD) from the UCLA Consortium for Neuropsychiatric Phenomics. Diagnoses were confirmed via the Structured Clinical Interview for DSM-IV. The analysis followed a three-pronged approach:

1. Gradient Estimation and Categorical Analysis:

   • Subject-specific functional connectomes were constructed using 400 cortical parcels (Schaefer atlas).
   • Macroscale gradients were estimated using diffusion map embedding (BrainSpace toolbox), retaining the first two principal gradients (G1 and G2).
   • G1 was identified as a visual-to-somatomotor axis.
   • G2 was identified as a unimodal-to-transmodal axis (sensorimotor to default mode network).
   • Group differences were assessed via Procrustes disparity (global deviation from normative architecture) and linear modeling at the parcel and network levels (adjusting for age and sex).
   • Subcortical gradient projections were computed for seven bilateral structures (thalamus, caudate, putamen, pallidum, amygdala, nucleus accumbens, hippocampus) to assess subcortical-cortical coupling.

2. Multivariate Brain-Behavior Analysis (PLS):

   • Partial Least Squares (PLS) regression was applied to the combined patient sample (N=77) to identify latent dimensions maximizing covariance between parcel-wise gradient scores and clinical/cognitive phenotypes (symptom severity, diagnosis, cognitive composites).
   • The model was validated using a half-splitting procedure (100 random train-test splits) to ensure generalizability, retaining the first latent variable (LV1) for each gradient.
   • Spatial correspondence between PLS brain salience maps and normative gradient maps was tested using spin permutation to account for spatial autocorrelation.

3. Individual-Level Similarity Indices (SI-PLS):

   • To position individuals within the identified gradient-behavior space, PLS-based similarity indices (SI-PLS) were computed. This metric quantifies the alignment between an individual's cortical gradient topography and the group-derived PLS brain salience maps.
   • Healthy controls were withheld from the PLS fitting, serving as an independent normative reference to validate the dimensional separation of patient groups.

Key Results

• Global Architecture: The global triangular manifold of cortical organization (anchored by visual, somatomotor, and default mode networks) was preserved across all groups. However, both patient groups showed significantly elevated Procrustes disparity along both G1 and G2 compared to HC, indicating greater deviation from the normative gradient configuration.
• Network-Level Alterations:
  • Transdiagnostic: The frontoparietal network showed significant compression along the G2 (unimodal-transmodal) axis in both SZ and BD.
  • SZ-Specific: SZ exhibited a specific contraction of the visual pole along G1 and a lower amygdala gradient projection score along G1, indicating reduced effective coupling along the sensory hierarchy.
  • BD-Specific: BD showed a relative displacement of the somatomotor network along G2.
• Brain-Behavior Dimensions (PLS):
  • G1-LV1: This dimension was primarily associated with a BD profile, characterized by lower symptom burden (positive, negative, and disorganization symptoms) and a preserved/expanded gradient architecture relative to HC.
  • G2-LV1: This dimension was primarily associated with an SZ profile, characterized by greater symptom severity and broad cognitive impairment. The brain salience map for G2 ran counter to the normative G2 axis, indicating a compression of the unimodal-transmodal hierarchy driven by frontoparietal regions.
• Individual Positioning (SI-PLS):
  • When projected into the joint SI-PLS space (G1 × G2), SZ and BD occupied largely distinct regions, with HC positioned between them.
  • BD participants showed high alignment with the G1 pattern (positive SI-PLS G1), while SZ participants showed high alignment with the G2 pattern (positive SI-PLS G2).
  • The separation was graded rather than categorical, with significant within-group spread and partial overlap, suggesting that individual position within this two-dimensional neural space is not fully determined by diagnostic category.

Significance and Claims
The paper claims that differences across the SZ-BD spectrum are best understood through a multi-axis dimensional framework rather than a strictly categorical or single-axis model. The study demonstrates that:

1. Transdiagnostic vs. Specific: While higher-order association systems (frontoparietal) show transdiagnostic alterations, primary sensory systems (visual and somatomotor) carry disorder-specific signatures (visual contraction in SZ, somatomotor displacement in BD).
2. Dimensional Heterogeneity: Clinical variance does not collapse onto a single dimension; instead, it distributes across two principal gradient axes, with G1 capturing BD-associated variation and G2 capturing SZ-associated severity and cognitive burden.
3. Individual-Level Utility: The introduction of SI-PLS allows for the continuous positioning of individuals within a shared neural space. This approach reveals that healthy controls occupy an intermediate position between SZ and BD, supporting dimensional accounts of psychosis where disorders share common neural axes but express them in different combinations and degrees.

The authors conclude that these findings support a framework where clinical heterogeneity in psychosis is captured by individual positions along two principal cortical gradient axes, offering a potential pathway for developing biomarkers that reflect dimensional brain-behavior phenotypes.