Brain Myelin Covariance Networks: Gradients, Cognition, and Higher-Order Landscape

Abstract

Myelin is essential for efficient neural signaling and can be quantitatively evaluated using the T1-weighted/T2-weighted (T1w/T2w) ratio as a proxy for regional myelin content. Myelin covariance networks (MCNs) reflect correlated myelin patterns across brain regions, enabling the investigation of topological organization. However, a vertex-level map of myelin covariance gradients and their cognitive associations remains underexplored. The objective of this study was to construct and characterize vertex-level MCNs, identify their principal gradients, map their higher-order topological landscape, and determine their associations with cognitive functions and other multimodal cortical features. We conducted a cross-sectional, secondary analysis of publicly available data from the Human Connectome Project (HCP). The dataset included T1w/T2w MRI data from 1096 healthy adult participants (age 22-37). All original data collection and sharing procedures were approved by the Washington University institutional review board. Our procedures involved (1) constructing a vertex-wise MCN from T1w/T2w ratio data; (2) applying gradient analysis to identify principal organizational axes; (3) calculating network connectivity strength; (4) performing cognitive meta-analysis using Neurosynth; and (5) using graphlet analysis to assess higher-order topology. Our results show that the primary myelin gradient (Gradient 1) spans from sensory-motor to association cortices, strongly associates with connectivity strength (r = 0.66), and shows a functional dissociation between affective processing and sensorimotor domains. Furthermore, Gradient 2, as well as the positive and full connectivity strength, showed robust correlations with fractional anisotropy (FA), a DTI metric reflecting white matter microstructure. Our higher-order analysis also revealed that negative and positive myelin covariance connections exhibited distinct topologies. Negative connections were dominated by star-like graphlet structures, while positive connections were dominated by path-like and triangular structures. This systematic vertex-level investigation offers novel insights into the organizational principles of cortical myelin, linking gray matter myelin patterns to white matter integrity, and providing a valuable reference for neuropsychological research and the potential identification of biomarkers for neurological disorders.

Keywords: cognitive functions; gradient analysis; myelin covariance networks; neurological disorders; vertex-level.

Figure 1

Methodological Pipeline for Myelin Covariance Network Analysis. This flowchart illustrates the comprehensive analytical pipeline for investigating cortical MCN using data from the Human Connectome Project. The analysis begins with acquisition of T1w MPRAGE and T2w SPACE neuroimaging data from 1096 healthy participants (age 22–37). Following HCP minimal preprocessing pipelines, myelin content was estimated using the T1w/T2w ratio method, which enhances cortical myelin contrast by approximating the squared myelin content. Vertex-wise MCNs were constructed after regressing out confounding variables (age, age², gender, and age × gender interactions), yielding a 59,412 × 59,412 correlation matrix. Six parallel analytical approaches were then implemented: (1) gradient analysis using Fast Connectivity Gradient Approximation (FCGA) with 3000 random landmarks to extract principal components of connectivity; (2) connectivity strength analysis examining negative, positive, and full network connections; (3) mapping of gradients and connectivity strength to five Economo–Koskinas cytoarchitectonic types; (4) cognitive meta-analysis using the BrainStat toolbox and Neurosynth; (5) correlation analysis with 10 multimodal cortical features including anatomical hierarchy, functional hierarchy, and gene expression, etc.; and (6) higher-order topological analysis using graphlets at various correlation thresholds (−0.4, −0.2, 0.2, 0.4).

Figure 2

Representations of gradient components, connectivity strength patterns, and their correlations in the MCN. This figure presents the spatial patterns and relationships between principal gradients and connectivity strength measures in the vertex-wise MCN. (a) Brain surface renderings of the first three principal gradients (Gradient 1, Gradient 2, and Gradient 3) extracted using Fast Connectivity Gradient Approximation (FCGA). Color scales represent gradient values, with blue indicating lower values and yellow indicating higher values. (b) Brain surface renderings of three connectivity strength measures: full network strength (Strength_Full), negative correlation strength (Strength_Negative), and positive correlation strength (Strength_Positive). (c) Correlation matrix showing Pearson correlation coefficients between gradients and connectivity strength measures across all 59,412 vertices. The color scale ranges from dark blue (strong positive correlation) to white (no correlation) to yellow (negative correlation). Crossed-out cells (X) indicate non-significant correlations at p < 0.05 after Holm multiple comparison adjustment. Strong positive correlations exist between Gradient 1 and all strength measures, particularly with Strength_Full (r = 0.66), while Gradient 3 shows weak or non-significant correlations with strength measures.

Figure 3

Distribution of Myelin Covariance Network Properties Across Economo–Koskinas Cytoarchitectural Classes. (a) Radar plots showing the distribution of three principal gradients of the MCN across five Economo–Koskinas cytoarchitectural classes (Agranular, Frontal, Parietal, Polar, and Granular). Gradient 1 shows highest values in Polar regions (0.1651) and lowest in Granular regions (−0.1975). Gradient 2 shows highest values in Agranular regions (0.0654) and lowest in Parietal regions (−0.0774). Gradient 3 shows highest values in Agranular regions (0.1558) and lowest in Frontal regions (−0.0443). (b) Radar plots illustrating connectivity strength distributions across the same cytoarchitectural classes for three different network types. Strength_Negative shows strongest negative connections in Granular regions (−826.32). Strength_Positive shows strongest positive connections in Agranular regions (854.33). Strength_Full (combined positive and negative connections) shows highest values in Frontal regions (49.07) and lowest (negative) values in Granular regions (−17.97). (c) Left panel: Brain surface renderings highlighting the anatomical locations of the five Economo–Koskinas cytoarchitectural classes. Middle panel: Line graph comparing the distribution patterns of the three principal gradients across cytoarchitectural classes. Right panel: Line graphs showing the distribution patterns of the three connectivity strength measures across cytoarchitectural classes.

Figure 4

Correlations Between Principal Myelin Covariance Gradients and Neurosynth Meta-Analysis Maps. The figure displays correlations between three principal myelin covariance gradients and Neurosynth meta-analysis maps, with both word cloud visualizations (left) and correlation plots (right) showing the top 30 correlations for each gradient. (a) Gradient 1 (top panel) shows a functional dissociation between systems. Positive correlations (orange) appear with limbic and affective processing regions, with strongest associations to medial prefrontal, ventromedial prefrontal, and orbitofrontal cortices (r ≈ 0.22–0.25). Terms related to subgenual cortex, ventral anterior regions, hypothalamus, and amygdala also show robust positive correlations. In contrast, negative correlations (blue, r ≈ −0.2) are observed with sensorimotor systems, including primary motor, sensorimotor cortex, motor function, and rehabilitation terms. (b) Gradient 2 (middle panel) primarily relates to white matter microstructural properties, with strongest positive correlations (orange-yellow, r ≈ 0.28–0.29) with fractional anisotropy measures, diffusion tensor imaging metrics, and white matter properties. The correlation strength gradually decreases across the 30 top terms (r = 0.29 to 0.17), with no strong negative correlations observed, suggesting this gradient primarily captures variations in white matter organization patterns. (c) Gradient 3 (bottom panel) reveals another functional dissociation. Positive correlations (orange-yellow, r ≈ 0.13–0.18) are observed with motor system terms including cortex m1, primary motor, sensorimotor cortex, and motor premotor areas. Negative correlations (blue, r ≈ −0.12 to −0.14) are seen with terms related to higher cognitive functions and pathological states, including cognitive impairment, memory retrieval, novelty, dementia, and polymorphism, suggesting this gradient may distinguish between basic sensorimotor functions and higher-order cognitive processes.

Figure 5

Correlations Between Myelin Covariance Network Connectivity Strength and Neurosynth Meta-Analysis Maps. This figure illustrates the correlations between three different MCN connectivity strength measures (Negative, Positive, and Full) and Neurosynth meta-analysis terms, presented as word clouds (left) and correlation plots (right) showing the top 30 correlations for each strength measure. (a) Strength_Negative (top panel) demonstrates predominantly negative correlations (blue, r ≈ −0.21 to −0.32) with motor and sensorimotor systems. The strongest negative correlations appear with cortex m1, primary motor, sensorimotor cortex, motor premotor, and motor cortex terms. Interestingly, a few terms show positive correlations (orange, r ≈ 0.2), including dementia and cognitive impairment, suggesting this network may capture an inverse relationship between motor function and neurodegenerative conditions. (b) Strength_Positive (middle panel) shows robust positive correlations (orange-yellow, r ≈ 0.28–0.30) with white matter microstructural properties. The most strongly correlated terms include fractional anisotropy, anisotropy fa, diffusion tensor, t1 weighted, white matter, corpus callosum, and gray matter. This pattern suggests that positive connectivity strength in the MCN relates primarily to white matter integrity and structural connectivity metrics derived from diffusion tensor imaging. (c) Strength_Full (bottom panel), representing the combined positive and negative connectivity, reveals a pattern similar to Strength_Positive but with some additional terms. The strongest correlations (orange-yellow, r ≈ 0.27–0.29) are with fractional anisotropy, anisotropy fa, and white matter structural terms. Additionally, terms related to affective and cognitive processes such as subgenual, addiction, anxiety, and dementia show moderate positive correlations, indicating that the full connectivity strength measure captures both structural and functional dimensions of brain organization.

Figure 6

Cortical feature correlations with myelin covariance network gradients and strength measures. (a) Brain surface renderings of ten cortical features: Anatomical Hierarchy, Functional Hierarchy, Evolutionary Hierarchy, Allometric Scaling, Aerobic Glycolysis (top row); Cerebral Blood Flow, Gene Expression, NeuroSynth, Externopyramidization, and Cortical Thickness (bottom row). Each pair shows lateral and medial views with corresponding value ranges below. (b) Correlation coefficient heatmap between three MCN gradients (Gradient 1, Gradient 2, Gradient 3) and the ten cortical features. Gradient 1 exhibited the strongest correlations, notably with Anatomical Hierarchy (r = −0.50), Gene Expression (r = −0.28), and Externopyramidization (r = −0.26). (c) Correlation coefficient heatmap between three MCN strength networks (Strength-Negative, Strength-Positive, Strength-Full) and the same cortical features. In both heatmaps, orange indicates positive correlations, blue indicates negative correlations, and “X” marks non-significant correlations after Bonferroni correction (p > 0.05/30). These findings suggest that MCNs capture distinct aspects of cortical organization.

Figure 7

Higher-order topological organization of myelin covariance networks through graphlet analysis. (a) Visualization of 30 distinct graphlets (G0–G29) representing all possible connected subgraphs with 2–5 nodes. Each graphlet captures a specific local connectivity pattern. (b) Distribution of graphlets across four correlation thresholds (r > 0.4, r > 0.2, r < −0.2, r < −0.4) shown as a flow diagram. Line thickness represents the normalized frequency of each graphlet, with color indicating the corresponding threshold. The circular heatmap on the left displays normalized graphlet frequencies (darker green indicating higher frequency). Notable patterns emerge where star-like structures (particularly G10, G11) dominate negative correlations, while path-like and triangular structures (G09, G13) predominate in positive correlations. This analysis reveals fundamentally different topological organizations between positively and negatively correlated components of the MCN, suggesting distinct underlying biological mechanisms governing these relationships.